WO2023092310A1

WO2023092310A1 - Information processing method, model generation method, and devices

Info

Publication number: WO2023092310A1
Application number: PCT/CN2021/132593
Authority: WO
Inventors: 田文强
Original assignee: Oppo广东移动通信有限公司
Priority date: 2021-11-23
Filing date: 2021-11-23
Publication date: 2023-06-01
Also published as: CN117678257A

Abstract

The present application relates to an information processing method, a model generation method, a terminal device, a network device, an electronic device, a chip, a computer readable storage medium, a computer program product, and a computer program. The method comprises: a terminal device receives first information; and the terminal device sends second information obtained on the basis of the first information, wherein the second information is obtained by processing the first information via a first model, the second information is used for being processed via a second model to obtain channel information, and the first model and the second model are obtained by means of joint training.

Description

Information processing method, model generation method and device

technical field

The present application relates to the communication field, and more specifically, relates to an information processing method, a model generation method, a terminal device, a network device, an electronic device, a chip, a computer-readable storage medium, a computer program product, and a computer program.

Background technique

For channel information acquisition and feedback methods, wireless communication systems mainly rely on basic models and pre-configured feedback parameter sets for channel information determination and feedback. In this process, the error between the feedback channel information and the real channel information is relatively large. Therefore, a wireless communication solution based on artificial intelligence (AI, Artificial Intelligence) is proposed in some studies to make up for the above-mentioned deficiency. However, in AI-based wireless communication solutions, how to ensure the overall performance of the network becomes a problem that needs to be solved.

Contents of the invention

Embodiments of the present application provide an information processing method, a model generation method, a terminal device, a network device, an electronic device, a chip, a computer-readable storage medium, a computer program product, and a computer program, which can at least solve the above problems.

An embodiment of the present application provides an information processing method, including:

The terminal device receives the first information;

The terminal device sends second information obtained based on the first information;

Wherein, the second information is obtained by processing the first information through the first model, and the second information is used for processing through the second model to obtain channel information; the first model and the second model are a joint obtained by training.

The network device sends the first information;

The network device receives second information; wherein, the second information is obtained by processing the first information through a first model;

The network device processes the second information based on a second model to obtain channel information; wherein, the first model and the second model are obtained through joint training.

The embodiment of the present application provides a method for generating a model, including:

performing joint training on the first preset model and the second preset model by using the training samples to obtain the trained first model and the second model;

Wherein, the first model is the first preset model after training, and the second model is the second preset model after training; the first model is used to process the first information to obtain Second information; the second model is used to process the second information to obtain channel information.

An embodiment of the present application provides a terminal device, including:

a first communication unit, configured to receive first information; send second information obtained based on the first information;

An embodiment of the present application provides a network device, including:

The second communication unit is configured to send the first information; receive the second information; wherein, the second information is obtained by processing the first information through the first model;

The second processing unit is configured to process the second information based on a second model to obtain channel information; wherein, the first model and the second model are obtained through joint training.

An embodiment of the present application provides an electronic device, including:

The third processing unit is configured to use training samples to jointly train the first preset model and the second preset model to obtain the trained first model and the second model;

An embodiment of the present application provides a terminal device, including a processor and a memory. The memory is used to store computer programs, and the processor is used to call and run the computer programs stored in the memory, so that the terminal device executes the above information processing method.

An embodiment of the present application provides a network device, including a processor and a memory. The memory is used to store a computer program, and the processor is used to call and run the computer program stored in the memory, so that the network device executes the above-mentioned information processing method.

An embodiment of the present application provides an electronic device, including a processor and a memory. The memory is used to store a computer program, and the processor is used to call and run the computer program stored in the memory, so that the network device executes the above-mentioned model generation method.

An embodiment of the present application provides a chip configured to implement the above information processing method or model generation method.

Specifically, the chip includes: a processor, configured to call and run a computer program from the memory, so that the device installed with the chip executes the above-mentioned information processing method or model generation method.

An embodiment of the present application provides a computer-readable storage medium for storing a computer program, and when the computer program is run by a device, the device is made to execute the above-mentioned information processing method or model generation method.

An embodiment of the present application provides a computer program product, including computer program instructions, which enable a computer to execute the above-mentioned information processing method or model generation method.

An embodiment of the present application provides a computer program that, when run on a computer, causes the computer to execute the above-mentioned information processing method or model generation method.

In this embodiment of the present application, when the terminal device receives the first information, it can process the first information through the first model to obtain the second information and send it, so that the receiving end can use the second model to process the second information The channel information obtained through processing is obtained through joint training of the first model and the second model. Since the processing, transmission, and analysis of the second information are realized by using the first model and the second model obtained through joint training, the performance requirements of the entire information processing, transmission, and analysis can be taken into account, ensuring the overall performance of the network.

Description of drawings

Fig. 1 is a schematic diagram of an application scenario according to an embodiment of the present application.

Fig. 2 is a schematic flowchart of a method for sending and receiving information in a wireless communication system according to an embodiment of the present application.

Fig. 3 is a schematic diagram of transmission and reception scenarios of pilot signals according to an embodiment of the present application.

Fig. 4 is a schematic flowchart of a channel information feedback method according to an embodiment of the present application.

5 is a schematic diagram of the basic structure of a neural network according to an embodiment of the present application;

FIG. 6 is a schematic flowchart 1 of an information processing method according to an embodiment of the present application;

FIG. 7 is a schematic diagram of eigenvector information of channel information according to an embodiment of the present application;

Figures 8a to 8d are schematic diagrams of the composition and structure of the model according to the embodiment of the present application;

FIG. 9 is a schematic diagram of a second training sample according to an embodiment of the present application;

FIG. 10 is another schematic diagram of a second training sample according to an embodiment of the present application;

FIG. 11 is a schematic flowchart II of an information processing method according to an embodiment of the present application;

FIG. 12 is a schematic flowchart three of an information processing method according to an embodiment of the present application;

FIG. 13 is a schematic flowchart 4 of an information processing method according to an embodiment of the present application;

Fig. 14 is a schematic flowchart of a model generation method according to an embodiment of the present application;

15 is a schematic diagram of a first model and a second model according to an embodiment of the present application;

FIG. 16 is a schematic flow of channel estimation for a reference signal according to an embodiment of the present application;

FIG. 17 is a schematic flowchart of channel state information feedback according to an embodiment of the present application;

FIG. 18 is a first schematic block diagram of a terminal device according to an embodiment of the present application;

FIG. 19 is a second schematic block diagram of a terminal device according to an embodiment of the present application;

FIG. 20 is a schematic block diagram of a network device according to an embodiment of the present application;

Fig. 21 is a schematic block diagram of an electronic device according to an embodiment of the present application;

Fig. 22 is a schematic block diagram of a communication device according to an embodiment of the present application;

Fig. 23 is a schematic block diagram of a chip according to an embodiment of the present application;

Fig. 24 is a schematic block diagram of a communication system according to an embodiment of the present application.

Detailed ways

The technical solutions in the embodiments of the present application will be described below with reference to the drawings in the embodiments of the present application.

The technical solution of the embodiment of the present application can be applied to various communication systems, such as: Global System of Mobile communication (Global System of Mobile communication, GSM) system, code division multiple access (Code Division Multiple Access, CDMA) system, broadband code division multiple access (Wideband Code Division Multiple Access, WCDMA) system, General Packet Radio Service (GPRS), Long Term Evolution (LTE) system, Advanced long term evolution (LTE-A) system , New Radio (NR) system, evolution system of NR system, LTE (LTE-based access to unlicensed spectrum, LTE-U) system on unlicensed spectrum, NR (NR-based access to unlicensed spectrum) on unlicensed spectrum unlicensed spectrum (NR-U) system, Non-Terrestrial Networks (NTN) system, Universal Mobile Telecommunications System (UMTS), Wireless Local Area Networks (WLAN), Wireless Fidelity (Wireless Fidelity, WiFi), fifth-generation communication (5th-Generation, 5G) system or other communication systems, etc.

Generally speaking, the number of connections supported by traditional communication systems is limited and easy to implement. However, with the development of communication technology, mobile communication systems will not only support traditional communication, but also support, for example, Device to Device (Device to Device, D2D) communication, Machine to Machine (M2M) communication, Machine Type Communication (MTC), Vehicle to Vehicle (V2V) communication, or Vehicle to everything (V2X) communication, etc. , the embodiments of the present application may also be applied to these communication systems.

Optionally, the communication system in the embodiment of the present application may be applied to a carrier aggregation (Carrier Aggregation, CA) scenario, may also be applied to a dual connectivity (Dual Connectivity, DC) scenario, and may also be applied to an independent (Standalone, SA) deployment Web scene.

Optionally, the communication system in the embodiment of the present application may be applied to an unlicensed spectrum, where the unlicensed spectrum may also be considered as a shared spectrum; or, the communication system in the embodiment of the present application may also be applied to a licensed spectrum, where, Licensed spectrum can also be considered as non-shared spectrum.

The embodiments of the present application describe various embodiments in conjunction with network equipment and terminal equipment, wherein the terminal equipment may also be referred to as user equipment (User Equipment, UE), access terminal, user unit, user station, mobile station, mobile station, remote station, remote terminal, mobile device, user terminal, terminal, wireless communication device, user agent or user device, etc.

The terminal device can be a station (STAION, ST) in the WLAN, a cellular phone, a cordless phone, a Session Initiation Protocol (Session Initiation Protocol, SIP) phone, a wireless local loop (Wireless Local Loop, WLL) station, a personal digital processing (Personal Digital Assistant, PDA) devices, handheld devices with wireless communication functions, computing devices or other processing devices connected to wireless modems, vehicle-mounted devices, wearable devices, next-generation communication systems such as terminal devices in NR networks, or future Terminal equipment in the evolved public land mobile network (Public Land Mobile Network, PLMN) network, etc.

In the embodiment of this application, the terminal device can be deployed on land, including indoor or outdoor, handheld, wearable or vehicle-mounted; it can also be deployed on water (such as ships, etc.); it can also be deployed in the air (such as aircraft, balloons and satellites) superior).

In this embodiment of the application, the terminal device may be a mobile phone (Mobile Phone), a tablet computer (Pad), a computer with a wireless transceiver function, a virtual reality (Virtual Reality, VR) terminal device, an augmented reality (Augmented Reality, AR) terminal Equipment, wireless terminal equipment in industrial control, wireless terminal equipment in self driving, wireless terminal equipment in remote medical, wireless terminal equipment in smart grid , wireless terminal equipment in transportation safety, wireless terminal equipment in smart city, or wireless terminal equipment in smart home.

As an example but not a limitation, in this embodiment of the present application, the terminal device may also be a wearable device. Wearable devices can also be called wearable smart devices, which is a general term for the application of wearable technology to intelligently design daily wear and develop wearable devices, such as glasses, gloves, watches, clothing and shoes. A wearable device is a portable device that is worn directly on the body or integrated into the user's clothing or accessories. Wearable devices are not only a hardware device, but also achieve powerful functions through software support, data interaction, and cloud interaction. Generalized wearable smart devices include full-featured, large-sized, complete or partial functions without relying on smart phones, such as smart watches or smart glasses, etc., and only focus on a certain type of application functions, and need to cooperate with other devices such as smart phones Use, such as various smart bracelets and smart jewelry for physical sign monitoring.

In the embodiment of the present application, the network device may be a device for communicating with the mobile device, and the network device may be an access point (Access Point, AP) in WLAN, a base station (Base Transceiver Station, BTS) in GSM or CDMA , or a base station (NodeB, NB) in WCDMA, or an evolved base station (Evolutional Node B, eNB or eNodeB) in LTE, or a relay station or access point, or a vehicle-mounted device, a wearable device, and an NR network The network equipment (gNB) in the network or the network equipment in the future evolved PLMN network or the network equipment in the NTN network, etc.

As an example but not a limitation, in this embodiment of the present application, the network device may have a mobile feature, for example, the network device may be a mobile device. Optionally, the network equipment may be a satellite or a balloon station. For example, the satellite can be a low earth orbit (low earth orbit, LEO) satellite, a medium earth orbit (medium earth orbit, MEO) satellite, a geosynchronous earth orbit (geosynchronous earth orbit, GEO) satellite, a high elliptical orbit (High Elliptical Orbit, HEO) satellite. ) Satellite etc. Optionally, the network device may also be a base station installed on land, water, and other locations.

In this embodiment of the present application, the network device may provide services for a cell, and the terminal device communicates with the network device through the transmission resources (for example, frequency domain resources, or spectrum resources) used by the cell, and the cell may be a network device ( For example, a cell corresponding to a base station), the cell may belong to a macro base station, or may belong to a base station corresponding to a small cell (Small cell), and the small cell here may include: a metro cell (Metro cell), a micro cell (Micro cell), a pico cell ( Pico cell), Femto cell, etc. These small cells have the characteristics of small coverage and low transmission power, and are suitable for providing high-speed data transmission services.

FIG. 1 exemplarily shows a communication system 100 . The communication system includes a network device 110 and two terminal devices 120 . Optionally, the communication system 100 may include multiple network devices 110, and the coverage of each network device 110 may include other numbers of terminal devices 120, which is not limited in this embodiment of the present application.

Optionally, the communication system 100 may also include other network entities such as a mobility management entity (Mobility Management Entity, MME), an access and mobility management function (Access and Mobility Management Function, AMF), etc. Not limited.

Wherein, the network equipment may further include access network equipment and core network equipment. That is, the wireless communication system also includes multiple core networks for communicating with access network devices. The access network device may be a long-term evolution (long-term evolution, LTE) system, a next-generation (mobile communication system) (next radio, NR) system or an authorized auxiliary access long-term evolution (LAA- Evolved base station (evolutional node B, abbreviated as eNB or e-NodeB) macro base station, micro base station (also called "small base station"), pico base station, access point (access point, AP), Transmission point (transmission point, TP) or new generation base station (new generation Node B, gNodeB), etc.

It should be understood that a device with a communication function in the network/system in the embodiment of the present application may be referred to as a communication device. Taking the communication system shown in Figure 1 as an example, the communication equipment may include network equipment and terminal equipment with communication functions. It may include other devices in the communication system, such as network controllers, mobility management entities and other network entities, which are not limited in this embodiment of the present application.

In order to facilitate the understanding of the embodiments of the present application, the following briefly describes the basic processes and basic concepts involved in the embodiments of the present application. It should be understood that the basic processes and basic concepts described below do not limit the embodiments of the present application.

As shown in Figure 2, the workflow of the wireless communication system may include: the sending end performs operations such as encoding, modulating, and encrypting on the information source to form the sending information to be transmitted; the sending information is transmitted through the wireless space, and at this time the sending information will be received by the channel The impact of the environment and interference noise; at the receiving end, the received information is decoded, decrypted and demodulated, and finally the source information is restored. In the above processing, the sending end may be a network device in the aforementioned communication system shown in FIG. 1, and the receiving end may be a terminal device in the aforementioned communication system shown in FIG. 1; or, the sending end It may be a terminal device in the aforementioned communication system shown in FIG. 1 , and the receiving end may be a network device in the aforementioned communication system shown in FIG. 1 .

In the above-mentioned workflow of the wireless communication system, the quality of the channel environment and whether the current channel environment can be accurately estimated are crucial to the performance of the wireless communication system. When designing a wireless communication system, the sending end (such as a network device) will send some pilot signals, such as sending some channel state information reference signal (CSI-RS, Channel State Information Reference Signal), demodulation reference signal (DMRS, Demodulation Reference Signal), phase tracking reference signal (PT-RS, Phase Tracking Reference Signal), synchronization signal and PBCH block (SSB, Synchronization Signal and PBCH block), etc., used to assist the receiving end (such as terminal equipment) to obtain and estimate the current channel characteristics. Furthermore, the receiving end (such as a terminal device) can feed back corresponding channel information to the sending end (such as a network device) based on the estimated and recovered channel characteristics, and finally the sending end (such as a network device) performs corresponding coding according to the acquired channel information , Modulation, etc. As shown in Figure 3, the sending end (such as a network device) sends out a specific pilot signal, and the pilot signal is received by the receiving end (such as a terminal device) after being transmitted through a channel. The receiving end (such as a terminal device) can The pilot signal and the actual pilot signal estimate the channel condition through which the pilot signal passes and determine the channel information based on the channel condition. After the receiving end (such as a terminal device) obtains the channel information, it can also feed back the channel information to the sending end (such as a network device), and then the sending end (such as a network device) can perform subsequent data Scheduling etc. Wherein, the channel information acquired by the receiving end may be the channel information where the pilot signal is located (such as the time domain and/or frequency domain resource where the pilot signal is located); correspondingly, basic interpolation, etc. The method restores the channel information in each time slot of the complete broadband based on the received channel information where the pilot signal is located, and then performs corresponding data scheduling and other processing.

The processing method for channel information feedback may include: after the terminal device estimates the channel information, it feeds back the channel information to the network device in a channel state information feedback manner in the current communication system. The feedback of channel information is very important in LTE system and NR system, which determines the performance of MIMO transmission. Taking the CSI feedback process as an example in conjunction with FIG. 4 for illustration, it may include the following steps: S410: The network device configures the feedback parameter information indicated by channel state information (CSI, Channel State information), for example, the network device configures the terminal device to feedback channel quality Which information in the indication (CQI, Channel Quality Indicator), precoding matrix indication (PMI, Precoding Matrix Indicator), rank indication (RI, Rank Indication) and other information; at the same time, the network device will configure some reference signals for CSI measurement , such as SSB or CSI-RS. S420: The network device sends the reference signal to the terminal device. S430: The terminal device generates CSI by measuring the above reference signal. S440: The terminal device feeds back the CSI to the network device. Then perform S450: the network device configures a data transmission mode based on the CSI, that is, the network device can configure a reasonable and efficient data transmission mode based on the CSI. The CSI may include indications of information such as CQI, PMI, and RI.

The above-mentioned processing method of channel information feedback can further introduce research and design of artificial intelligence represented by neural network, for example, the estimation of wireless channel can be realized through the design of neural network. As shown in Figure 5, the basic structure of the neural network includes: an input layer, a hidden layer and an output layer; the input layer is responsible for receiving data, the hidden layer processes the data, and the final result is produced at the output layer . Each node in Figure 5 represents a processing unit, and each processing unit can simulate a neuron, and multiple neurons form a layer of neural network, and multi-layer information transmission and processing construct an overall neural network. With the continuous development of neural network research, neural network deep learning algorithms have been proposed in recent years, more hidden layers have been introduced, and feature learning is performed through layer-by-layer training of neural networks with multiple hidden layers, which greatly improves the learning of neural networks. And processing capabilities, and are widely used in pattern recognition, signal processing, optimization combination, anomaly detection, etc.

It should be understood that the terms "system" and "network" are often used interchangeably herein. The term "and/or" in this article is just an association relationship describing associated objects, which means that there can be three relationships, for example, A and/or B can mean: A exists alone, A and B exist simultaneously, and there exists alone B these three situations. In addition, the character "/" in this article generally indicates that the contextual objects are an "or" relationship.

It should be understood that the "indication" mentioned in the embodiments of the present application may be a direct indication, may also be an indirect indication, and may also mean that there is an association relationship. For example, A indicates B, which can mean that A directly indicates B, for example, B can be obtained through A; it can also indicate that A indirectly indicates B, for example, A indicates C, and B can be obtained through C; it can also indicate that there is an association between A and B relation.

In the description of the embodiments of the present application, the term "corresponding" may indicate that there is a direct or indirect correspondence between the two, or that there is an association between the two, or that it indicates and is indicated, configuration and is configuration etc.

In order to facilitate the understanding of the technical solutions of the embodiments of the present application, the related technologies of the embodiments of the present application are described below. The following related technologies can be combined with the technical solutions of the embodiments of the present application as optional solutions, and all of them belong to the embodiments of the present application. protected range.

Fig. 6 is a schematic flowchart of an information processing method 600 according to an embodiment of the present application. The method can optionally be applied to the system shown in Fig. 1, but is not limited thereto. The method includes at least some of the following.

S610. The terminal device receives first information.

S620. The terminal device sends second information obtained based on the first information;

In the above S610, the first information may be a reference signal, specifically, the reference signal may be a reference signal of the current channel, such as a downlink reference signal of the current channel. The downlink reference signal may include at least one of CSI-RS, DMRS, and PT-RS.

The first information may be distributed in the first dimension and/or the second dimension.

Wherein, the first dimension is a time domain dimension; the first information is distributed in at least one time unit in the time domain dimension. Each time unit in the at least one time unit may include one of the following: 1 time slot and 1 Orthogonal Frequency Division Multiplexing (OFDM, Orthogonal Frequency Division Multiplexing) symbol. For example, the first signal is a downlink reference signal, and the downlink reference signal may be distributed in one time slot in the time domain dimension, or the downlink reference signal may be distributed in two or on 4 time slots.

The second dimension is a frequency domain dimension; the first information is distributed on at least one frequency domain resource in the frequency domain dimension; wherein, each frequency domain resource in the at least one frequency domain resource can be one of the following One: one resource block (RB, Resource Block), one subcarrier. For example, the first signal is a downlink reference signal, and the downlink reference signal may be distributed in 1 RB in the frequency domain dimension, or the downlink reference signal may be distributed in 2 or 4 RBs in the time domain dimension. on RBs.

The first dimension and the second dimension above can be used in combination, that is, the first information can be distributed on the first dimension and the second dimension; for example, the first information can be distributed on a RBs in the frequency domain dimension , distributed in b time slots in the time domain dimension; both a and b are positive integers. For example, the first information is a downlink reference signal, and the downlink reference signal may be distributed in 4 RBs in the frequency domain, and may be distributed in 6 time slots in the time domain dimension.

Still further, the first information can also be expressed as a complex number, that is, the first information is also distributed in the third dimension; the third dimension is a complex dimension; the first information includes the first information sample The real part of and the imaginary part of the first information sample. For example, the real part of the first information is distributed on the a RBs of the frequency domain resources and the b time slots of the time domain resources, and the imaginary part of the first information is distributed on the a RBs of the frequency domain resources. b time slots of the time domain resource.

Optionally, before receiving the first information, the terminal device may first receive configuration information sent by the network device, and the configuration information may be configured with the first information for the terminal device to measure. Taking the first information as an example of a downlink reference signal, the configuration information may be configuring the terminal device to measure SSB or CSI-RS and so on.

After completing S610, that is, after the terminal device receives the first information, the terminal device may process the first information based on the first model to obtain the second information.

In an example, the second information is channel compression information; the first model is configured to process the input first information to obtain channel compression information. That is, the input information of the first model is the first information, and the second information output by the first model is the channel compression information.

It should be pointed out that the first model may also be called an encoding model or an encoding network, as long as the input information is the first information and the output information is the channel compression information, the model or neural network is within the protection scope of this embodiment.

Wherein, the first model may specifically include the following sub-models: an estimation sub-model and a compression sub-model;

The estimation sub-model is used to perform channel estimation based on the first information to obtain channel estimation information;

The compression sub-model is used to compress the channel estimation information to obtain channel compression information.

The estimation sub-model can also be called a channel estimation sub-model or a channel estimation sub-neural network, and the estimation sub-model can use one of a fully connected network, a convolutional neural network, a residual network, and a self-attention mechanism network. Or a variety of network structure construction. The compression sub-model may be called a channel compression sub-model or a channel compression sub-neural network, and the compression sub-model may use one of a fully connected network, a convolutional neural network, a residual network, a self-attention mechanism network, or A variety of network structure construction.

The estimation method adopted by the estimation sub-model may include algorithms such as minimum mean square error (MMSE).

The compression sub-model can compress the data volume of the input information. The compression rate between the output information of the compressed sub-model and the input information can be determined during training, for example, the compression rate can be 5/1000, 2/1000, 10%, etc. lift.

The terminal device processes the first information based on the first model to obtain the second information, which may specifically be:

The terminal device inputs the first information into the estimation sub-model, and obtains channel estimation information output by the estimation sub-model;

The terminal device inputs the channel estimation information into the compression sub-model, and obtains channel compression information output by the compression sub-model.

In this example, the first information may be a reference signal, specifically, the first information may be a reference signal of a current channel, for example, the first information may be a downlink reference signal of a current channel. Correspondingly, the channel information may be used to characterize at least one of channel quality, channel state, and channel estimation result obtained based on the first information.

The channel information may be represented by a matrix of T dimensions, where T is an integer greater than or equal to 2. Alternatively, the channel estimation information may also be represented by a matrix of T dimensions. The channel information is used as an example for description below, and the description of the channel estimation information is similar to that and will not be repeated.

The matrix of the T dimensions may specifically be a two-dimensional matrix of M×N; wherein, M represents the number of first granularities in the fourth dimension, and N represents the number of second granularities in the fifth dimension; M and N are all positive integers. That is to say, the channel information may be composed of a two-dimensional matrix with a size of M×N, which has M first granularities in the fourth dimension and N second granularities in the fifth dimension; the above M and N May or may not be equal. The specific numerical indication in the two-dimensional matrix represents the received signal strength at a certain first granularity of the channel quality, where the unit of the numerical value in the two-dimensional matrix may be dBm, or the numerical value in the two-dimensional matrix has no unit It is the value obtained after normalization. In addition, the two-dimensional matrix of M×N can also be synthesized into one-dimensional data of size 1×(M×N) or (M×N)×1. The specific transformation can be the fourth dimension first and then the fifth dimension. It may also be the fifth dimension first and then the fourth dimension, which is not limited in this embodiment.

The fourth dimension may be a frequency domain dimension; the first granularity includes one of the following: L1 resource blocks RB, L2 subcarriers; L1 and L2 are positive integers. Alternatively, the fourth dimension may be a time-domain dimension; the first granularity includes one of the following: K1 microseconds, K2 symbol lengths, and the number of sampling points of K3 symbols; K1, K2, and K3 are positive integers . The symbols are Orthogonal Frequency Division Multiplexing (OFDM, Orthogonal Frequency Division Multiplexing). Here, when the fourth dimension is a time domain dimension, the first granularity may also be called a delay granularity.

For example, when the fourth dimension is the frequency domain dimension, the first granularity may be L1 RBs (L1 is greater than or equal to 1, such as 2RB, 4RB, 8RB), then the channel information in the frequency domain dimension The distribution range is the frequency domain range corresponding to M×L1 RBs; or the first granularity can be L2 subcarriers (L2 is greater than 1, such as 4 subcarriers, 6 subcarriers, and 18 subcarriers), then the channel The distribution of information in the frequency domain dimension is the frequency domain range corresponding to M×L2 subcarriers. When the fourth dimension is a time-domain dimension, the first granularity may be a delay granularity, for example, a first granularity is the number of sampling points of K1 microseconds, or K2 symbol lengths, or K3 symbols, Here, the symbol may be an OFDM symbol; when the fourth dimension is the time domain dimension and the first granularity is K1 microseconds, the distribution range of the channel information on the time domain dimension is M×K1 The time domain range corresponding to microseconds.

The fifth dimension may be a space domain dimension; correspondingly, the second granularity is an interval between a pair of transmitting and receiving antennas or an angle of arrival. That is to say, the fifth dimension is the space domain dimension, specifically, it may be an antenna dimension, and the second granularity may be a pair of transmitting and receiving antennas. Alternatively, the fifth dimension is a space domain dimension, specifically, an angle domain dimension, and the second granularity may be an interval of arrival angles.

Still further, the value of the ijth position in the two-dimensional matrix representing the channel information is used to represent the i-th first granularity in the fourth dimension and the j-th second granularity in the fifth dimension The channel quality of ; i and j are both positive integers. That is to say, a numerical value (or an indicator value) at a certain position in the two-dimensional matrix used to represent the channel information represents the channel quality under the combination of the fourth dimension and the fifth dimension. Wherein, the channel quality may be characterized by a signal strength value; the unit of the signal strength value may be dBm, or the signal strength value has no unit but a value obtained after normalization.

The T dimensions may also include a sixth dimension. The matrix of T dimensions may be a three-dimensional matrix of M×N×W; wherein, M represents the number of first granularities in the fourth dimension, N represents the number of second granularities in the fifth dimension, and W represents The number of third granularities under the sixth dimension; M, N and W are all positive integers.

Exemplarily, the sixth dimension may be a complex dimension, the third granularity is 1, and the quantity W of the third granularity under the complex dimension is 2. For example, when the fourth dimension represents the time domain dimension, the first granularity is the delay granularity; the fifth dimension is the spatial domain dimension, specifically the angle dimension, and the second granularity is the interval of arrival angles ; The sixth dimension is a complex dimension, W is 2, k is 1 to indicate the real part, and k is 2 to indicate the imaginary part. When i=4, j=5, k=1, the value (or indicator value) at the ijkth position of the above-mentioned three-dimensional matrix represents the fourth delay granularity in the fifth spatial granularity (such as the interval of arrival angle) The real part of the channel quality on . If i=4, j=5, k=2, then the value (or indicator value) at the ijkth position of the above-mentioned three-dimensional matrix represents the channel quality on the 4th delay granularity in the 5th spatial granularity imaginary part.

In the subsequent description, for the sake of simplicity, the above channel information is illustrated by using a two-dimensional matrix formed by the fourth dimension and the fifth dimension. However, it should be clarified that the dimension of the above channel information matrix is not limited to two dimensions.

In another example, the second information may be channel compression information; the channel compression information includes eigenvector information of compressed channel estimation information; correspondingly, the first model is used for the input first The information is processed to obtain the eigenvector information of the compressed channel estimation information.

In this example, the first information may be a reference signal, specifically, the first information may be a reference signal of a current channel, for example, the first information may be a downlink reference signal of a current channel. The channel information output by the second model may specifically be feature vector information of the channel information.

It should be pointed out that the first model can also be called an encoding model or an encoding neural network, etc., as long as the input information is the first information and the output information is the eigenvector information of the compressed channel estimation information or the neural network. Within the protection scope of the embodiment.

Wherein, the first model may specifically include the following sub-models: estimation sub-model, channel generation sub-model and compression sub-model;

The estimation sub-model is used to perform channel estimation on the first information to obtain channel estimation information;

The channel generation sub-model is used to perform eigendecomposition on the channel estimation information to obtain eigenvector information of the channel estimation information;

The compression sub-model is used to compress the eigenvector information of the channel estimation information to obtain compressed eigenvector information of the channel estimation information.

Wherein, the eigenvector information of the channel information includes R groups of eigenvector sequence information; R is a positive integer. For example, R may be 1, then the eigenvector information of the channel information includes a set of eigenvector sequence information. R may be 2, then the eigenvector information of the channel information includes 2 sets of eigenvector sequence information. The above value of R may be determined according to the actual situation, or may be specified during the training of the first model. Correspondingly, the eigenvector information of the channel estimation information may also include R groups of eigenvector sequence information, which will not be described in detail.

In the above R sets of feature vector sequence information, each set of feature vector sequence information may include a feature sequence of a preset length. Wherein, the lengths of the feature sequences included in the feature vector sequence information of different groups are the same. The preset length can be set according to the actual situation or can be set during training, for example, it can be any one of 16, 32, 48, 64, 128, 256, and of course it can be longer or shorter. The embodiment does not exhaustively list all possible values of the preset length. In conjunction with FIG. 7, for example, the preset length is 32 (but it can be a bit), and R is 4, that is, the feature vector information of the channel information includes 4 sets of feature vector sequence information, wherein each set of feature vector sequence information contains a feature sequence of length 32.

The function of the estimation sub-model is the same as that of the foregoing embodiment, and no repeated description is given.

The manner of performing eigendecomposition in the channel generation sub-model may specifically be a singular value decomposition (SVD, Singular Value Decomposition) manner. For example, the channel generation sub-model performs SVD eigendecomposition on the input channel estimation information to obtain eigenvector information of the channel estimation information after eigendecomposition. The channel estimation information may be represented by a matrix, and the specific description is the same as that of the foregoing embodiment, and will not be repeated here.

The compression sub-model may be to compress the data volume of the input information. The compression rate between the second information output by the compression sub-model and the input information can be determined during training, for example, the compression rate can be five thousandths, two thousandths, ten percent, etc., not here Exhaustive.

In this example, the channel estimation information output by the estimation sub-model may be different from the channel information output by the second model, and the channel estimation information output by the estimation sub-model may specifically be a matrix of channel information, such as It is represented by a matrix of T dimensions; the channel information output by the second model may be eigenvector information of the channel information, for example, may include R groups of eigenvector sequence information. Certainly, the channel information output by the second model in this example may also be the same as the channel estimation information output by the estimation sub-model.

The terminal device may process the first information based on the first model to obtain the second information, which may specifically be:

The terminal device inputs the channel estimation information into the channel generation sub-model, and obtains eigenvector information of the channel estimation information output by the channel generation sub-model;

The terminal device inputs the eigenvector information of the channel estimation information into the compression sub-model, and obtains the eigenvector information of the compressed channel estimation information output by the compression sub-model.

After the above processing is completed, the terminal device may execute S620, where the terminal device sends second information obtained based on the first information.

Specifically, the terminal device sending the second information obtained based on the first information may be: the terminal device sends the second information obtained based on the first information to a network device. Wherein, the network device may be an access network device (such as a base station, or eNB, or gNB) serving the terminal device, or the network device may refer to an access network device that communicates with the terminal device (such as a base station, or eNB, or gNB).

Wherein, the second information may be carried by one of the following information: information included in the random access process, radio resource control (RRC, Radio Resource Control) message, uplink control information (UCI, Uplink Control Information). The information contained in the random access process includes one of the following: message A in the two-step random access process; Msg1 in the four-step random access process; Msg3 in the four-step random access process.

The above describes how the terminal device uses the first model in detail. There are two ways for the terminal device to obtain the first model: the first way: the terminal device obtains the first model directly; the second One way: obtained by the terminal device through training. The two methods are described below:

In a first manner, the terminal device receives the first model.

Specifically, the terminal device receives the first model sent by the electronic device; for example, the terminal device may receive model parameters of the first model sent by the electronic device.

Here, the electronic device may be an electronic device that obtains the first model and the second model through joint training.

Exemplarily, the electronic device may be a network device, and the network device may be an access network device serving the terminal device, such as a base station, eNB, gNB, and so on. Alternatively, the electronic device may be other devices than the network device serving the terminal device, for example, it may be a server, or a desktop computer, or a notebook or other device with data processing capabilities, which is not described in this embodiment. Exhaustive.

Wherein, when the electronic device is a network device serving the terminal device, receiving the first model sent by the electronic device by the terminal device may specifically be: the terminal device receives the first model sent by the network device. a model. Wherein, the first model (or the model parameters of the first model) may be carried by at least one of the following: downlink control signaling, media access control (MAC, Media Access Control) control element (CE, Control Element) Messages, RRC messages, broadcast messages, downlink data transmission, and downlink data transmission for artificial intelligence business transmission requirements.

In the case that the electronic device is other than the network device serving the terminal device, the first model (or the model parameters of the first model) may be transmitted through a wired connection, or other transmitted over a wireless connection. For example, the electronic device transmits the first model (or the model parameters of the first model) to the terminal device through a wired connection with the terminal device. Alternatively, the electronic device transmits the first model (or the model parameters of the first model) to the terminal device through other wireless connections with the terminal device; wherein, the other wireless connection method may be a Bluetooth connection Ways or wireless fidelity (Wi-Fi, Wireless Fidelity) connection ways, etc., are not exhaustive here.

Alternatively, the terminal device may receive multiple sub-models respectively, and then combine the multiple received sub-models to obtain the first model.

In one case, the first model includes an estimation sub-model and a compression sub-model. Correspondingly, the terminal device receives the estimated sub-model and the compressed sub-model; the terminal device generates the first model based on the estimated sub-model and the compressed sub-model. Specifically, the terminal device receives the model parameters of the estimated sub-model and the model parameters of the compressed sub-model sent by the electronic device; parameters to obtain the first model.

Here, the terminal device may receive the estimated sub-model and the compressed sub-model sent by the electronic device at the same time; or, it may receive the estimated sub-model and the compressed sub-model sent by the electronic device separately, for example, it may be received first The estimated sub-model sent by the electronic device then receives the compressed sub-model sent by the electronic device, or first receives the compressed sub-model sent by the electronic device and then receives the estimated sub-model sent by the electronic device.

In the case that the electronic device is a network device serving the terminal device, the estimated sub-model (or the model parameters of the estimated sub-model) and the compressed sub-model (or the compressed sub-model) may be carried simultaneously or separately by one of the following information Model parameters of the model): downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements.

In the case that the electronic device is other than the network device serving the terminal device, the electronic device may use the above estimated sub-model (or the model of the estimated sub-model) through a wired connection with the terminal device parameters) and the compressed sub-model (or the model parameters of the compressed sub-model) are sent to the terminal device at the same time or separately. Or, the electronic device sends the above-mentioned estimated sub-model (or the model parameters of the estimated sub-model) and the compressed sub-model (or the model parameters of the compressed sub-model) to the terminal at the same time or separately through other wireless connections with the terminal device device; wherein, the other wireless connection methods may be a Bluetooth connection method or a WIFI connection method, etc., which are not exhaustive here.

In yet another case, the first model includes an estimation sub-model, a channel generation sub-model, and a compression sub-model.

In this case, the terminal device receives the estimation sub-model, the compression sub-model and the channel generation sub-model; the terminal device generates the first model. Specifically, the terminal device receives the model parameters of the estimated sub-model, the model parameters of the compression sub-model and the model parameters of the channel generation sub-model sent by the electronic device; , model parameters of the compression sub-model, and model parameters of the channel generation sub-model to obtain the first model.

Here, the terminal device may simultaneously receive the estimation sub-model, compression sub-model and channel generation sub-model sent by the electronic device. Alternatively, the terminal device may respectively receive the estimated sub-model, the compressed sub-model and the channel generation sub-model sent by the electronic device, for example, the estimated sub-model, the compressed sub-model and the channel generation sub-model are respectively received; or, the estimated Any two of the sub-model, compression sub-model and channel generation sub-model are received separately from the remaining one. For example, the terminal device may first receive the estimated submodel sent by the electronic device, then receive the channel generation submodel sent by the electronic device, and finally receive the compressed submodel sent by the electronic device; or, the terminal The device first receives the compressed submodel and the channel generation submodel sent by the electronic device, and then receives the estimated submodel sent by the electronic device. It should be pointed out that the above is only an exemplary description, and does not mean that there are only several combinations of the above-mentioned exemplary sub-models, compression sub-models, and channel generation sub-models that are actually sent or received respectively, but this embodiment is not exhaustive. lift.

When the electronic device is a network device serving the terminal device, when the terminal device can simultaneously receive or separately receive the estimation submodel, compression submodel and channel generation submodel sent by the network device, the The estimation sub-model, compression sub-model and channel generation sub-model are carried by one of the following at the same time or separately: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink for artificial intelligence business transmission requirements data transmission.

In the case that the electronic device is other than the network device serving the terminal device, the electronic device can generate the above-mentioned estimated sub-model, compressed sub-model and channel through a wired connection with the terminal device. The sub-models are sent to the terminal devices simultaneously or separately. Alternatively, the electronic device sends the above-mentioned estimation submodel, compression submodel and channel generation submodel to the terminal device simultaneously or separately through other wireless connections with the terminal device; wherein, the other wireless connection methods may be Bluetooth or WIFI, etc., are not exhaustive here.

Through the above processing, the terminal device may receive the first model, and then may execute the foregoing processing of S610-S620.

In this manner, the method may further include: the terminal device receiving the second model. Specifically, the terminal device may receive model parameters of the second model. Still further, the terminal device may receive the second model sent by the electronic device, specifically, the terminal device may receive model parameters of the second model sent by the electronic device.

When the electronic device is a network device serving the terminal device, the second model is carried by one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, manual Downlink data transmission required for intelligent service transmission.

In the case that the electronic device is other than the network device serving the terminal device, the second model (or the model parameters of the second model) may be transmitted through a wired connection, or other transmitted over a wireless connection. For example, the electronic device transmits the second model (or the model parameters of the second model) to the terminal device through a wired connection with the terminal device. Alternatively, the electronic device transmits the second model (or the model parameters of the second model) to the terminal device through other wireless connections with the terminal device; wherein, the other wireless connection methods may be Bluetooth or WIFI, etc., are not exhaustive here.

It should be noted that the above-mentioned second model and the first model may also be sent at the same time, or may be sent separately.

In this manner, the method may further include: the terminal device receiving the third model. Specifically, the terminal device may receive model parameters of the third model. Still further, the terminal device may receive the third model sent by the electronic device, specifically, the terminal device may receive model parameters of the third model sent by the electronic device.

Wherein, the third model is used to perform data transformation processing on the second information output by the first model and input it into the second model; the first model, the second model and the third model are obtained through joint training .

The data transformation processing includes: convolution processing or Fourier transform processing. For example, the Fourier transform processing may specifically include: converting to the frequency domain through Fourier transform, multiplication, and then converting to the time domain through inverse Fourier transform.

When the electronic device is a network device serving the terminal device, the third model is carried by one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, manual Downlink data transmission required for intelligent business transmission.

In the case that the electronic device is other than the network device serving the terminal device, the third model (or the model parameters of the third model) may be transmitted through a wired connection, or other transmitted over a wireless connection. For example, the electronic device transmits the third model (or the model parameters of the third model) to the terminal device through a wired connection with the terminal device. Alternatively, the electronic device transmits the third model (or the model parameters of the third model) to the terminal device through other wireless connections with the terminal device; wherein, the other wireless connection methods may be Bluetooth or WIFI, etc., are not exhaustive here.

The above-mentioned first model, second model, and third model can be sent separately or simultaneously; or the above-mentioned first model, second model, and third model can also be sent separately; or, it is also possible It is any two combinations of which are sent at the same time, and the remaining one is sent separately, and so on.

It should be noted that the foregoing first model is a model that the terminal device needs to use when receiving the first information. However, in this method, besides the first model, the terminal device can also receive the second model and/or the third model, and the reasons are as follows:

Taking the joint training of the first model and the second model as an example, if the terminal device wants to evaluate the first model and the second model as a whole, it needs to obtain the first model and the second model , and then the terminal device may decide whether to use the first model and the second model received this time after completing the overall evaluation of the first model and the second model. If the overall evaluation result of the terminal device on the first model and the second model is poor (for example, the compression rate is low or the accuracy rate of recovered channel information is low, etc.), the first model may not be used. If the terminal device determines not to use the above-mentioned first model, the terminal device may re-train the first model and the second model jointly to update the model parameters of the first model and the second model, or the terminal device Train yourself to get a new first model and a new second model. In addition, if the above-mentioned terminal device obtains the new first model and the new second model after performing joint training or updating again, it can also synchronize the new first model and the new second model to the network device; correspondingly, After the network device receives the new first model and the new second model, it can also replace the original first model and the second model used by itself, and can also send the new first model and the new second model to other terminal equipment. Regarding other subsequent related processes, this embodiment does not exhaustively enumerate them. Through the above processing, it can be ensured that the first model with the best performance and its corresponding second model are used in the whole system, so that the communication performance of the whole system can be further guaranteed.

When the first model, the second model, and the third model are the whole obtained through joint training, the terminal device can also perform an overall evaluation on the first model, the second model, and the third model after receiving it and subsequent Processing, the specific processing method is the same as the above, and will not be repeated.

In a second manner, the terminal device trains itself to obtain the above-mentioned first model.

The terminal device uses training samples to jointly train the first preset model and the second preset model to obtain the trained first model and the second model;

Wherein, the first model is the first preset model after training, and the second model is the second preset model after training.

In the second manner, the training may use the first loss function or the second loss function. The following describes the training using the above two loss functions:

The loss function used in the training is a first loss function; the first loss function is based on the difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model degree of difference is constructed.

The degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model is determined based on a distance, or determined based on a degree of similarity.

The specific calculation method for determining the degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model based on the distance can use mean square error (MSE, Mean Squared Error ) or normalized mean square error (NMSE), etc., which are not exhaustive in this embodiment.

For example, the output information of the second preset model may be a matrix, and correspondingly, the input information of the compressed preset sub-model may also be a matrix, and here, the output matrix of the second preset model It is called matrix 1, and the matrix of the input of the compressed preset submodel is called matrix 2; the output information of the second preset model and the compressed preset submodel of the first preset model are determined based on the distance The way of the degree of difference between the input information is the MSE way, for example, its calculation may include: calculating the difference between matrix 1 and matrix 2, and taking the square of the difference as the difference degree.

The specific calculation method for determining the degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model based on the degree of similarity may use cosine similarity or cosine similarity squared etc., which are not exhaustive in this embodiment.

For example, the output information of the second preset model may be R sets of feature vector sequence information, and correspondingly, the input information of the compressed preset sub-model may also be R sets of feature vector sequence information. Here, the The R sets of feature vector sequence information output by the second preset model are called feature vector sequence 1, and the R sets of feature vector sequence information input by the compressed preset sub-model are called feature vector sequence 2. The method of determining the degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model based on the degree of similarity may be cosine similarity, for example, its calculation may include: The cosine angle between the eigenvector sequence 1 and the eigenvector sequence 2 is used to determine the degree of similarity, and the degree of similarity is used as the degree of difference.

In the process of using the above-mentioned first loss function for training, due to the difference in the sub-models contained in the first preset model and whether the third preset model for simulating the wireless channel environment is included for joint training, the way of joint training will be different. Therefore, the following four situations are described separately:

Case 1, the first preset model includes an estimation preset sub-model and a compression preset sub-model.

Referring to FIG. 8 a , it illustrates a first preset model 800 , a second preset model 810 , and an estimated preset sub-model 801 and a compressed preset sub-model 802 included in the first preset model 800 . The above-mentioned first preset model 800, the second preset model 810, and the input-output relationship between the estimated preset sub-model 801 and the compressed preset sub-model 802 contained in the first preset model 800 can be: estimated The input information of the preset submodel 801 is the first training sample 920; the output information of the estimated preset submodel 801 is used as the input information of the compressed preset submodel 802; the output information of the compressed preset submodel 802 As the input information of the second preset model 810 .

The terminal device uses training samples to jointly train the first preset model and the second preset model, including:

The terminal device inputs the first training sample into the estimated preset sub-model of the first preset model, and obtains initial information output by the estimated preset sub-model;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into the second preset model to obtain restoration information output by the second preset model;

determining the first loss function based on the degree of difference between the restoration information and the initial information;

performing reverse conduction according to the first loss function to update the first preset model and the second preset model.

Wherein, the first training samples may be reference signal samples. The reference signal samples may be original reference signals or processed reference signals obtained through historical acquisition. More specifically, the reference signal samples may be downlink reference signal samples. It should be understood that this embodiment does not limit that the first training samples must be the downlink reference signal samples, and uplink reference signal samples or other reference signal samples may also be used, which are not exhaustive in this embodiment.

It should be pointed out that, in addition to the first training samples, the information input into the estimated preset sub-model of the first preset model may also be wireless channel or other scene-related information, for example, may include at least one of the following: Channel signal-to-noise ratio, signal-to-interference-noise ratio, channel type, bandwidth information, delay information, etc. Whether one or more of the above information is input during the joint training process may be determined according to actual conditions or actual scenarios, and is not limited here.

A specific function of the estimation preset sub-model of the first preset model may be: perform channel estimation based on the first training samples to obtain initial information. Wherein, the channel estimation may adopt algorithms such as minimum mean square error (MMSE).

The aforementioned initial information may be a matrix, and the dimension of the matrix is not limited here, and may be a matrix of two or more dimensions. The value at each position in the matrix is used to represent the corresponding channel quality at the corresponding granularity corresponding to multiple dimensions. Wherein, the channel quality may be characterized by a signal strength value; the unit of the signal strength value may be dBm, or the signal strength value has no unit but a value obtained after normalization.

The compressed preset sub-model of the first preset model compresses the data volume of the input information. The compression rate between the output information of the compressed sub-model and the input information can be determined during training, for example, the compression rate can be 5/1000, 2/1000, 10%, etc. lift. The compressed information obtained by compressing the preset sub-model contains less data than the input initial information. The form of the above-mentioned compressed information is the same as that of the initial information, for example, the initial information is a matrix, and the corresponding compressed information is also a matrix, and the matrix dimensions of the initial information and the compressed information are the same but the amount of data is different.

The function of the second preset model may be to decompress its input information. In this case, the input information of the second preset model is the compressed information, and the output of the second preset model is the restored information. Ideally, the decompression rate of the second preset model should make the obtained restored information contain the same data content as the original information.

The performing reverse conduction update of the first preset model and the second preset model based on the first loss function may specifically refer to performing reverse conduction update based on the degree of difference determined by the first loss function. The model parameters of the estimated preset sub-model, the model parameters of the compressed preset sub-model, and the model parameters of the second preset model.

Regarding the above training, it should also be pointed out that the manner of the above training convergence may include at least one of the following: judging whether the number of iterative training reaches a preset number, and judging whether the degree of difference is smaller than a preset threshold. Wherein, the preset number of times and the preset threshold value can be set according to actual conditions, and are not exhaustive. That is to say, when it is determined that the training is completed based on the above manner, the first preset model after the training can be used as the first model, and the second preset model after the training can be used as the second model.

Case 2, the first preset model includes an estimation preset submodel, a preset channel generation submodel, and a compression preset submodel.

Referring to Fig. 8b, it shows a first preset model 800, a second preset model 810, and the estimation preset sub-model 801, compression preset sub-model 802 and channel generation preset included in the first preset model 800. Set sub-model 803 . Between the above-mentioned first preset model 800, second preset model 810, and the estimation preset submodel 801, compression preset submodel 802 and channel generation preset submodel 803 contained in the first preset model 800 The input-output relationship of can be as follows: the input information of the estimated preset sub-model 801 is the first training sample 920; the output information of the estimated preset sub-model 801 is used as the input information of the channel generation preset sub-model 803; The output information of the channel generation preset submodel 803 is used as the input information of the compression preset submodel 802 ; the output information of the compression preset submodel 802 is used as the input information of the second preset model 810 .

The terminal device uses training samples to jointly train the first preset model and the second preset model, which may include:

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

inputting the compressed feature vector information into the second preset model to obtain restored feature vector information output by the second preset model;

determining a first loss function based on the degree of difference between the restored eigenvector information and the eigenvector information of the initial information;

Wherein, the specific description about the first training sample is the same as the above-mentioned case 1, so repeated description will not be given. It should be pointed out that, in addition to the first training samples, the information input into the estimated preset sub-model of the first preset model may also be wireless channel or other scene-related information, for example, may include at least one of the following: Channel signal-to-noise ratio, signal-to-interference-noise ratio, channel type, bandwidth information, delay information, etc. Whether one or more of the above information is input during the joint training process may be determined according to actual conditions or actual scenarios, and is not limited here.

A specific function of the estimation preset sub-model of the first preset model may be: perform channel estimation based on the first training samples to obtain the initial information. Wherein, the channel estimation may adopt algorithms such as minimum mean square error (MMSE). The initial information output by the estimated preset sub-model above may be a matrix, and the dimension of the matrix is not limited here, and may be a two-dimensional or more dimensional matrix. The value at each position in the matrix is used to represent the corresponding channel quality at the corresponding granularity corresponding to multiple dimensions. Wherein, the channel quality may be characterized by a signal strength value; the unit of the signal strength value may be dBm, or the signal strength value has no unit but a value obtained after normalization.

A function of the channel generation preset submodel may be to perform eigendecomposition on the initial information to obtain eigenvector information of the initial information. Wherein, the eigenvector information of the initial information may include R groups of eigenvector sequences. The method of performing eigendecomposition on the initial information may adopt a method of Singular Value Decomposition (SVD, Singular Value Decomposition).

The compressed preset sub-model of the first preset model compresses the data volume of the input information. The compression rate between the output information of the compressed sub-model and the input information can be determined during training, for example, the compression rate can be 5/1000, 2/1000, 10%, etc. lift. In the above processing, the compressed eigenvector information obtained by compressing the preset sub-model contains less data than the eigenvector information of the input initial information. The above compressed feature vector information is in the same form as the feature vector information of the initial information. For example, the feature vector information of the initial information is a sequence of R groups of feature vectors, and the compressed feature vector information is also a sequence of feature vectors of groups R but The amount of data contained in the two is different.

The function of the second preset model may be to decompress its input information. In the above processing, the input information of the second preset model is compressed feature vector information, and the output of the second preset model is restored feature vector information. The decompression rate of the second preset model should be such that the obtained restored feature vector information contains the same or substantially the same data as the feature vector information of the initial information. Of course, the function of the above-mentioned second preset model can also include recovering the input information to obtain the restored initial information. At this time, the decompression rate of the second preset model should make the recovered initial information contain the same content as the initial information. or basically the same data.

The performing reverse conduction to update the first preset model and the second preset model based on the degree of difference determined by the first loss function may specifically refer to: performing a reverse conduction based on the degree of difference determined by the first loss function performing reverse conduction to update model parameters of the estimated preset submodel, model parameters of the channel generation preset submodel, model parameters of the compression preset submodel, and model parameters of the second preset model.

The method for determining the convergence of the above training is the same as that of the above-mentioned case 1, and repeated explanations are not repeated.

In case three, the terminal device uses training samples to jointly train the first preset model and the second preset model to obtain the trained first model and the second model, including:

The terminal device uses training samples to jointly train the first preset model, the second preset model and the third preset model to obtain the trained first model, the second model and the third model; wherein , the third model is a trained third preset model.

Referring to FIG. 8c, it shows a first preset model 800, a second preset model 810, a third preset model 830, and the estimated preset sub-model 801 contained in the first preset model 800, the compression preset Submodel 802. The above-mentioned first preset model 800, the second preset model 810, and the input-output relationship between the estimated preset sub-model 801 and the compressed preset sub-model 802 contained in the first preset model 800 can be: estimated The input information of the preset submodel 801 is the first training sample 920; the output information of the estimated preset submodel 801 is used as the input information of the compressed preset submodel 802; the output information of the compressed preset submodel 802 is used as the The input information of the third preset model 830; the output information of the third preset model is used as the output information of the second preset model 810.

The terminal device uses training samples to jointly train the first preset model, the second preset model and the third preset model, including:

inputting the compressed information into a third preset model to obtain transformed information output by the third preset model;

inputting the transformed information into the second preset model to obtain restoration information output by the second preset model;

determining a first loss function based on the degree of difference between the restoration information and the initial information;

performing reverse conduction according to the first loss function to update the first preset model, the second preset model and the third preset model.

The specific description about the first training sample is the same as the above-mentioned case 1 or case 2, so no repeated description is given. It should also be pointed out that, in addition to the first training samples, the information input into the estimated preset sub-model of the first preset model can also be wireless channel or other information related to the scene, for example, it can include at least one of the following: channel Signal-to-noise ratio, signal-to-interference-noise ratio, channel type, bandwidth information, delay information, etc. In the process of training, whether one or more of the above information is input may be relevant according to the actual situation or the actual scene, and it is not limited here.

The specific functions of the estimation preset sub-model of the first preset model and the compression preset sub-model of the first preset model are the same as those in the first case, so the description will not be repeated.

In case three, a third preset model is added relative to case one, and the function of the third preset model is to simulate the channel environment, and the specific processing can be to perform data transformation on input information to obtain data transformed information as output information . Wherein, the specific processing method of the data transformation may include convolution processing or data processing equivalent to convolution; wherein, the data processing equivalent to convolution may be multiple Fourier transform processing, for example, may In order to transform the input information of the third preset model into the frequency domain through Fourier transform, multiply and then transform into the time domain through inverse Fourier transform, the process of convolution in the time domain is equivalent.

The function of the second preset model may be to decompress its input information. In the processing of the third case, the input information of the second preset model is transformed information, and the output of the second preset model is restored information. The decompression rate of the second preset model should make the obtained restored information contain the same data as the original information.

The updating of the first preset model, the second preset model, and the third preset model based on the first loss function may specifically refer to: performing reverse conduction based on the first loss function Conductively updating model parameters of the estimated preset sub-model, model parameters of the compressed preset sub-model, model parameters of the second preset model, and model parameters of the third preset model.

The manner of the above-mentioned training convergence is the same as that of the foregoing case 1 or case 2, and will not be repeated here.

Case 4 is different from the above case 3 in that the first preset model includes an estimation preset sub-model, a preset channel generation sub-model, and a compression preset sub-model.

Referring to Fig. 8d, it shows a first preset model 800, a second preset model 810, a third preset model 830, and the estimated preset sub-model 801 contained in the first preset model 800, the compression preset Submodel 802 and channel generation preset submodel 803 . The above-mentioned first preset model 800, second preset model 810, third preset model 830, and the estimation preset sub-model 801, compression preset sub-model 802 and channel generation included in the first preset model 800 The input-output relationship between the preset sub-models 803 may be: the input information of the estimated preset sub-model 801 is the first training sample 920; the output information of the estimated preset sub-model 801 is used as the channel generation preset sub-model The input information of 803; the output information of the channel generation preset submodel 803 is used as the input information of the compression preset submodel 802; the output information of the compression preset submodel 802 is used as the input of the third preset model 830 information; the output information of the third preset model 830 is used as the input information of the second preset model 810 .

Inputting the compressed feature vector information into a third preset model to obtain transformed feature vector information output by the third preset model;

inputting the transformed eigenvector information into the second preset model to obtain restored eigenvector information output by the second preset model;

The specific description about the first training sample is the same as any one of the foregoing case 1, case 2, and case 3, so no repeated description is given.

The specific function of the estimated preset sub-model of the first preset model is the same as any one of the foregoing case 1, case 2, and case 3.

The compressed preset sub-model of the first preset model compresses the data volume of the input information. The compression rate between the output information of the compressed sub-model and the input information can be determined during training, for example, the compression rate can be 5/1000, 2/1000, 10%, etc. lift.

The function of the third preset model is to simulate the wireless channel environment, and the specific processing may be to perform data transformation on input information to obtain information after data transformation as output information. Wherein, the specific processing method of the data transformation may include convolution processing or data processing equivalent to convolution; wherein, the data processing equivalent to convolution may be multiple Fourier transform processing, for example, may In order to transform the input information of the third preset model into the frequency domain through Fourier transform, multiply and then transform into the time domain through inverse Fourier transform, the process of convolution in the time domain is equivalent.

The function of the second preset model may be to decompress its input information. The input information of the second preset model is transformed feature vector information, and the output of the second preset model is restored feature vector information. The decompression rate of the second preset model should make the obtained restored feature vector information and the feature vector information of the initial information contain close to or the same data.

The performing reverse conduction update based on the degree of difference determined by the first loss function to update the first preset model, the second preset model, and the third preset model may specifically refer to: based on the first The degree of difference determined by a loss function is used to perform reverse conduction to update the model parameters of the estimated preset sub-model, the model parameters of the channel generation preset sub-model, the model parameters of the compressed preset sub-model, the The model parameters of the second preset model and the model parameters of the third preset model.

The manner of the above-mentioned training convergence is the same as that of the foregoing embodiment, and no repeated description is given.

The scenario where the first loss function is used for joint training is described above. In this embodiment, the scenario where the second loss function is used for joint training can also be provided, as follows:

The loss function used in the training is a second loss function; the second loss function is based on the difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model The first difference degree of the first preset model and the second difference degree between the output information of the estimated preset sub-model of the first preset model and the second training sample; wherein, the second training sample and the input of the estimated Corresponds to the first training sample of the preset sub-model.

The first degree of difference is determined based on a distance, or is determined based on a degree of similarity; and/or, the second degree of difference is determined based on a distance, or is determined based on a degree of similarity.

The specific calculation method for determining the first degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model based on the distance can use a mean square error (MSE, Mean Squared Error) or normalized mean square error (NMSE), etc., this embodiment is not exhaustive.

For example, the output information of the second preset model may be a matrix, and correspondingly, the input information of the compressed preset sub-model may also be a matrix, and here, the output matrix of the second preset model It is called matrix 3, and the matrix of the input of the compressed preset submodel is called matrix 4; the output information of the second preset model and the compressed preset submodel of the first preset model are determined based on the distance The way of the degree of difference between the input information is the MSE way, for example: calculate the difference between matrix 3 and matrix 4, and use the square of the difference as the degree of difference.

The specific calculation method for determining the first degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model based on the degree of similarity may use cosine similarity or cosine similarity Degree square and other methods are not exhaustive in this embodiment.

For example, the output information of the second preset model may be R sets of feature vector sequence information, and correspondingly, the input information of the compressed preset sub-model may also be R sets of feature vector sequence information. Here, the The R group of feature vector sequence information output by the second preset model is called feature vector sequence 3, and the R group of feature vector sequence information input by the compressed preset sub-model is called feature vector sequence 4. The method of determining the degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model based on the degree of similarity may be cosine similarity, for example: feature vector sequence 3 and the cosine angle of the eigenvector sequence 4 to determine the degree of similarity, and use the degree of similarity as the degree of difference.

The specific calculation method for determining the second degree of difference between the output information of the estimated preset sub-model of the first preset model based on the distance and the second training sample can use mean square error (MSE, Mean Squared Error) or normalization The methods such as normalized mean square error (NMSE) are not exhaustive in this embodiment.

For example, the output information of the estimated preset sub-model may be a matrix, and correspondingly, the second training sample may also be a matrix. Here, the output matrix of the estimated preset sub-model is called matrix 5 , the matrix of the input of the compressed preset sub-model is called matrix 6; the second degree of difference between the output information of the estimated preset sub-model of the first preset model and the second training sample is determined based on the distance In the MSE mode, for example: calculate the difference between matrix 5 and matrix 6, and use the square of the difference as the degree of difference.

Based on the specific calculation method of the second degree of difference between the output information of the estimated preset sub-model of the first preset model and the second training sample based on the degree of similarity, the specific calculation method can use cosine similarity or cosine similarity squared, etc. The methods are not exhaustive in this embodiment.

When the first degree of difference and the second degree of difference are combined to construct the second loss function, the method of connection may be to add the weights of the first degree of difference and the second degree of difference, for example, the two each account for 50%; or , the joint method can be the addition of unequal weights between the first difference degree and the second difference degree. The difference before and after the compression and recovery between the two can be assigned a greater weight, or the above-mentioned second degree of difference can be assigned a larger weight, that is, the accuracy of the output information of the above-mentioned estimated preset sub-model is assigned a larger weight; or, its combination The method can be in the form of multiplying the first degree of difference and the second degree of difference; or the joint method can be that the first degree of difference and the second degree of difference can be calculated by cross entropy, such as p1*log(first degree of difference)+p2*log (the second degree of difference), where both p1 and p2 can be set according to actual conditions, and are not limited here.

In the process of using the above-mentioned second loss function for training, since the sub-models contained in the first preset model are different and whether the third preset model for simulating the wireless channel environment is included for joint training is different, Therefore, the following four situations are described separately:

Case five, the terminal device uses training samples to jointly train the first preset model and the second preset model to obtain the trained first model and the second model; wherein, the first preset model includes estimated preset sub-models Model and compression preset submodels.

In this case, the composition of each model and the input-output relationship between each model are the same as the previous case 1. For details, please refer to FIG. 8 a , which will not be repeated here.

Determine a second loss function based on a first degree of difference between the restored information and the initial information, and based on a second degree of difference between the initial information and a second training sample; the second training sample and The first training sample corresponds to;

performing reverse conduction according to the second loss function to update the first preset model and the second preset model.

The specific function of the estimation preset sub-model of the first preset model may be: perform channel estimation based on the first training samples to obtain initial information. Wherein, the channel estimation may adopt algorithms such as minimum mean square error (MMSE).

The aforementioned initial information may be a matrix, and the dimension of the matrix is not limited here, and may be a matrix of two or more dimensions. The value at each position in the matrix is used to represent the corresponding channel quality at the corresponding granularity corresponding to multiple dimensions. Wherein, the channel quality can be characterized by a signal strength value; the unit of the signal strength value can be decibel milliwatt (dBm, decibel relative to one milliwatt"), or the signal strength value has no unit but is normalized values obtained afterwards.

The function of the second preset model may be to decompress its input information. The decompression rate of the second preset model should make the obtained restored information contain the same data as the original information.

The performing reverse conduction based on the second loss function to update the first preset model and the second preset model may specifically refer to: performing reverse conduction based on the second loss function to update the estimated preset sub-model , the model parameters of the compressed preset sub-model and the model parameters of the second preset model.

For the above-mentioned training, it should also be pointed out that the way of the above-mentioned training convergence can include at least one of the following: judging whether the number of iterative training reaches the preset number of times, judging whether the first difference degree is less than the first preset threshold value, judging the second difference Whether the degree is smaller than the second preset threshold value. Wherein, the preset times, the first preset threshold value and the second preset threshold value can be set according to actual conditions, and are not exhaustive. That is to say, when it is determined that the training is completed based on the above manner, the first preset model after the training can be used as the first model, and the second preset model after the training can be used as the second model.

Case 6 is different from Case 5 in that the first preset model includes an estimation preset sub-model, a preset channel generation sub-model, and a compression preset sub-model.

The composition of each model in this case and the input-output relationship between each model are the same as those in the second case, for details, please refer to FIG. 8 b , which will not be repeated here.

determining a second loss function based on a first degree of difference between the restored feature vector information and the feature vector information of the initial information, and based on a second degree of difference between the initial information and a second training sample; The second training sample corresponds to the first training sample;

The specific description about the first training sample is the same as that of the fifth case above, so the description will not be repeated. It should be pointed out that, in addition to the first training samples, the information input into the estimated preset sub-model of the first preset model may also include other information related to wireless channels or scenes, for example, may include at least one of the following: Signal-to-noise ratio, signal-to-interference-noise ratio, channel type, bandwidth information, delay information, etc. Whether one or more of the above information is input during the joint training process may be determined according to actual conditions or actual scenarios, and is not limited here.

The specific function of the estimation preset sub-model of the first preset model is the same as the fifth case above.

The function of the second preset model may be to decompress its input information. In the above processing, the input information of the second preset model is compressed feature vector information, and the output of the second preset model is restored feature vector information. The decompression rate of the second preset model should be such that the obtained restored feature vector information contains the same or substantially the same data as the feature vector information of the initial information.

Performing reverse conduction according to the second loss function to update the first preset model and the second preset model may specifically refer to: performing reverse conduction based on the second loss function to update the estimated preset sub-model model parameters of the channel generation preset sub-model, model parameters of the compression preset sub-model and model parameters of the second preset model.

The method for determining the above-mentioned training convergence is the same as that of the above-mentioned case five, and repeated explanations are not repeated.

In case seven, the terminal device uses training samples to jointly train the first preset model and the second preset model to obtain the trained first model and the second model, including:

inputting the compressed information into a third preset model of the preset model to obtain transformed information output by the third preset model;

performing reverse conduction according to the second loss function to update the first preset model, the second preset model, and the third preset model.

The composition of each model in this case and the input-output relationship between each model are the same as those in the third case above, which can be referred to FIG. 8c, and repeated explanations are not repeated here.

The specific description about the first training sample is the same as the foregoing case five or six, so no repeated description is given.

The specific functions of the estimation preset sub-model of the first preset model and the compression preset sub-model of the first preset model are the same as those of the fifth case above, so repeated descriptions will not be made.

In case seven, a third preset model is added relative to case five. The function of the third preset model is to simulate the channel environment, and the specific processing can be to perform data transformation on input information to obtain data transformed information as output information . Wherein, the specific processing method of the data transformation may include convolution processing or data processing equivalent to convolution; wherein, the data processing equivalent to convolution may be multiple Fourier transform processing, for example, may In order to transform the input information of the third preset model into the frequency domain through Fourier transform, multiply and then transform into the time domain through inverse Fourier transform, the process of convolution in the time domain is equivalent.

The function of the second preset model may be to decompress its input information.

The performing reverse conduction update based on the degree of difference determined by the second loss function to update the first preset model, the second preset model, and the third preset model may specifically refer to: based on the first The second loss function performs reverse conduction to update the model parameters of the estimated preset sub-model, the model parameters of the compressed preset sub-model, the model parameters of the second preset model, and the model of the third preset model parameter.

The manner of the above-mentioned training convergence is the same as that of the foregoing case five or six, and no repeated description is made.

Case 8 is different from the above case 7 in that the first preset model includes an estimation preset submodel, a preset channel generation submodel, and a compression preset submodel.

inputting the compressed feature vector information into a third preset model among the preset models, to obtain transformed feature vector information output by the third preset model;

In this case, the composition of each model and the input-output relationship between each model are the same as the foregoing case 4, which can be referred to FIG. 8d , and repeated descriptions are not repeated here.

The specific description about the first training sample is the same as any one of the above-mentioned case 5, case 6, and case 7, so the description will not be repeated.

The specific function of the estimated preset sub-model of the first preset model is the same as any one of the fifth, sixth, and seventh cases.

A function of the channel generation preset submodel may be to perform eigendecomposition on the initial information to obtain eigenvector information of the initial information. Wherein, the eigenvector information of the initial information may include R groups of eigenvector sequences. For example, the method of performing eigendecomposition on the initial information may adopt a method of Singular Value Decomposition (SVD, Singular Value Decomposition).

The function of the second preset model may be to decompress its input information. The input information of the second preset model is transformed feature vector information, and the output of the second preset model is restored feature vector information. The decompression rate of the second preset model should be such that the obtained restored feature vector information and the feature vector information of the initial information contain data that are close to or identical.

The manner of the above-mentioned training convergence is the same as any one of the aforementioned cases 5, 6, and 7, so repeated explanations will not be made.

The terminal device may obtain the first model and the second model after its own joint training, or obtain the first model, the second model and the third model after joint training by adopting the above second method. Furthermore, the above-mentioned processing of S610 to S620 may be performed.

The training samples are used in the processing of the joint training of the terminal device itself to obtain the first model and the second model, and the joint training of the terminal device itself to obtain the first model, the second model and the third model provided by the second method above. , the following is a detailed description of the training samples:

The training samples may include a first training sample. The first training samples may be reference signal samples. The reference signal samples may be original reference signals or processed reference signals obtained through historical acquisition. Wherein, the original reference signal may refer to a reference signal that has not been transmitted through a wireless channel. The method for acquiring and processing the reference signal may include: using the reference signal received after the original reference signal passes through the wireless channel (or the real wireless channel, or the real wireless channel) as the processed reference signal. Alternatively, the method for obtaining a processed reference signal may include: using a reference signal received after the original reference signal passes through a simulated wireless channel as a processed reference signal. Still further, the original reference signal may be a downlink reference signal or an uplink reference signal.

The first training samples are distributed in the first dimension and/or the second dimension.

Wherein, the first dimension is a time domain dimension; the first training samples include first information samples distributed in m time units in the time domain dimension; m is a positive integer. Wherein, n first information samples may be distributed in each of the m time units, where n is a positive integer. Each time unit may include at least one time slot, or at least one symbol (such as an OFDM symbol).

For example, the first information sample is a downlink reference signal sample, the number of time slots contained in each time unit may be c (c is a positive integer), and there are n downlink reference signals in each c time slot A sample, combination of c and n can be e.g. (1,1)(1,2)(1,3)(1,4)(1,6)(2,1)(4,1)(5,1) (8,1)(10,1).

The second dimension is a frequency domain dimension; the first training samples include first information samples distributed on x frequency domain resources in the frequency domain dimension; x is a positive integer. Wherein, y first information samples may be distributed in each of the x frequency domain resources, and y is a positive integer. Each frequency domain resource may include at least one resource block (RB), or at least one subcarrier.

For example, the first information sample is a downlink reference signal sample, and the number of time slots contained in each frequency domain resource may be d (d is a positive integer), and there are y time slots in every d RBs in the frequency domain Downlink reference signal samples, the combination of d and y can be, for example, (1,1)(1,2)(2,1)(1,3)(3,1)(1,4)(4,1)(1 ,6)(6,1).

The above-mentioned first training samples are distributed in the first dimension and/or the second dimension. It can be understood that the subsequent training can be performed only according to the distribution of the first training samples in the frequency domain dimension, or only based on the distribution of the first training samples in the frequency domain. The subsequent training may be performed according to the distribution of the first training sample in the frequency domain and the time domain. For example, a first training sample contains 10 RBs in the frequency domain dimension and 1 time slot in the time domain dimension, each RB has 3 first signal samples, and each time slot has 1 first signal samples, the first training samples include a total of 30 first signal samples.

The sizes of the first dimension and the second dimension, the time domain dimension and the frequency domain dimension may be equal or unequal. In addition, the above-mentioned time-domain dimension and frequency-domain dimension can also be combined into one dimension. Specifically, the combination can be the time-domain dimension first and then the frequency-domain dimension, or the frequency-domain dimension first and then the time-domain dimension, which is not implemented in this embodiment limited.

It should be noted that, because the original reference signal or the processed reference signal can be represented by complex numbers, the solution provided by this embodiment can be based on the above-mentioned first dimension and second dimension. Increase the presentation form of complex numbers (or it can be understood as adding a dimension, which is caused by the independent presentation of the imaginary part and real part data of the original reference signal or the processed reference signal), specifically: the first training The samples are also distributed in the third dimension. The third dimension is a complex dimension; the first training samples include the real part of the first information sample and the imaginary part of the first information sample. For example, assuming that a first training sample contains 1 time unit (such as 1 time slot) in the time domain dimension, and contains 10 frequency domain resources (such as 10 RBs) in the frequency domain dimension, each first The information sample can be expressed as a real part and an imaginary part, and the first training sample can be a 1×10×2 matrix.

The training samples also include a second training sample corresponding to the first training sample; the second training sample is composed of a matrix of T dimensions; T is an integer greater than or equal to 2. Here, the second training samples may be used to characterize the expected channel quality based on the first training samples, or channel response, or channel state, or channel estimation results, or channel information .

The T dimensions include a fourth dimension and a fifth dimension.

The matrix of the T dimensions may specifically be a two-dimensional matrix of M×N; wherein, M represents the number of first granularities in the fourth dimension, and N represents the number of second granularities in the fifth dimension; M and N are all positive integers.

That is to say, a second training sample consists of a two-dimensional matrix with a size of M×N, which has M first granularities in the fourth dimension and N second granularities in the fifth dimension; the above M and N May or may not be equal. The specific numerical indication in the two-dimensional matrix represents the received signal strength at a certain first granularity of the channel quality. The specific numerical value in the two-dimensional matrix here may refer to the signal strength value, and its unit may be dBm, or There is no unit but the value obtained after normalization. In addition, the two-dimensional matrix of M×N can also be synthesized into one-dimensional data of size 1×(M×N) or (M×N)×1. The specific transformation can be the fourth dimension first and then the fifth dimension. It may also be the fifth dimension first and then the fourth dimension, which is not limited in this embodiment.

Optionally, the fourth dimension is a frequency domain dimension; the first granularity includes one of the following: L1 resource blocks RB, L2 subcarriers; L1 and L2 are positive integers. Alternatively, the fourth dimension is a time domain dimension; the first granularity includes one of the following: K1 microseconds, K2 symbol length, and K3 symbol sampling points; K1, K2, and K3 are positive integers. The symbols are Orthogonal Frequency Division Multiplexing (OFDM, Orthogonal Frequency Division Multiplexing). Here, when the fourth dimension is a time domain dimension, the first granularity may also be called a delay granularity.

For example, when the first training sample is a reference signal sample or a downlink reference signal sample, the second training sample may be a channel information sample corresponding to the reference signal sample, or may also be called a channel state sample Wait, I'm not going to exhaust the names here. When the fourth dimension is the frequency domain dimension, the first granularity can be L1 RBs (L1 is greater than or equal to 1, such as 2RB, 4RB, 8RB), and the distribution range of a second training sample in the frequency domain dimension is M×L1 The frequency domain range corresponding to each RB; or the first granularity can be L2 subcarriers (L2 is greater than 1, such as 4 subcarriers, 6 subcarriers, and 18 subcarriers), then the distribution of a second training sample on the frequency domain dimension is the frequency domain range corresponding to M×L2 subcarriers. When the fourth dimension is a time-domain dimension, the first granularity may be a delay granularity, for example, a first granularity is the number of sampling points of K1 microseconds, or K2 symbol lengths, or K3 symbols, where the symbols It can be an OFDM symbol; when the fourth dimension is the time domain dimension and the first granularity is K1 microseconds, the distribution range of a second training sample in the time domain dimension is the time domain range corresponding to M×K1 microseconds .

The fifth dimension is a space domain dimension; correspondingly, the second granularity is an interval between a pair of transmitting and receiving antennas or an angle of arrival.

For example, the fifth dimension is the space domain dimension, specifically, the antenna dimension, for example, the fifth dimension is composed of N antenna pairs, and correspondingly, the second granularity is a pair of transmitting and receiving antennas. Alternatively, the fifth dimension is a space domain dimension, specifically an angle domain dimension, for example, the fifth dimension is composed of N arrival angles, and the second granularity is the interval between the above N arrival angles.

Still further, the value at the ijth position in the two-dimensional matrix is used to represent the channel quality at the i-th first granularity in the fourth dimension and the j-th second granularity in the fifth dimension; Both i and j are positive integers. That is to say, in the case of using a first training sample, the value (or referred to as an indicator value) at a certain position in the two-dimensional matrix used to represent the second training sample represents the The expected channel quality situation under such a combination of five dimensions. Wherein, the channel quality or the channel quality situation can be characterized by signal strength, and the unit of the value (or indicator value) can be dBm, or there is no unit but a value obtained after normalization.

For example, in conjunction with Figure 9, in the M×N two-dimensional matrix, if the fourth dimension represents the frequency domain dimension, the fifth dimension is the space domain dimension, specifically the antenna dimension, the first granularity is 2RB, and the second granularity is 1 For the transceiver antenna; if the ij position in the two-dimensional matrix of M×N is the i=3j=6 position, then it is the position of the black box on the 3rd row and the 6th column shown in Fig. 9, The value (or indicator value) at this position can be used to represent the channel quality (or channel quality situation) on the third 2RB bandwidth (that is, the fifth RB to the sixth RB) on the sixth pair of transceiver antennas ). In addition, in FIG. 9 , S may also be used to represent the number of second training samples, and S may be an integer greater than or equal to 1, that is, the second training samples may include one or more.

For another example, in the M×N two-dimensional matrix shown in FIG. 10 , the fourth dimension represents the time domain dimension, and when the fourth dimension is the time domain dimension, the first granularity is one delay granularity; the fifth The dimension is the spatial domain dimension, specifically the angle dimension, and the second granularity is the basic granularity of 1 angle (for example, it can be the interval of 1 arrival angle); if the ij-th position in the M×N two-dimensional matrix is i=4j = 5 positions, then it is the position of the black box on the 4th row and 5th column shown in Fig. The channel quality (or channel quality situation) at the 4th delay granularity within the interval of . In addition, in FIG. 10 , S may also be used to represent the number of second training samples, and S may be an integer greater than or equal to 1, that is, the second training samples may include one or more.

The T dimensions also include a sixth dimension. Correspondingly, the matrix of T dimensions is a three-dimensional matrix of M×N×W; wherein, M represents the number of first granularities in the fourth dimension, and N represents the number of second granularities in the fifth dimension, W represents the quantity of the third granularity under the sixth dimension; M, N and W are all positive integers.

The value of the ijkth position in the three-dimensional matrix is used to represent the i-th first granularity in the fourth dimension, the j-th second granularity in the fifth dimension, and the k-th in the sixth dimension channel quality corresponding to the first training sample at a third granularity; i, j and k are all positive integers.

Wherein, the explanations about the fourth dimension and its first granularity, the fifth dimension and its second granularity are the same as those in the foregoing embodiments, and will not be repeated here.

In this embodiment, the sixth dimension may be a complex dimension. This is because the second training samples can be used to characterize the expected channel quality based on the first training samples (or called channel response, or called channel state, or called channel estimation result, or called channel information), and the above-mentioned channel quality can also be presented by a complex number, so a sixth dimension, that is, a complex number dimension, can be added on the basis of the above two dimensions of the second training sample, and the complex number dimension is the second training sample. The imaginary and real parts of the channel quality in the samples are presented independently generated.

Specifically, the sixth dimension is a complex dimension; the third granularity is 1, and the quantity W of the third granularity under the complex dimension is 2. Wherein, the third granularity being 1 specifically refers to a real part or an imaginary part, and the number of the third granularity being 2 means that there may be two third granularities in the complex dimension.

When the k is the first value, the value of the ijkth position in the three-dimensional matrix is used to represent the i-th first granularity in the fourth dimension and the j-th granularity in the fifth dimension The real part of the channel quality at the second granularity;

When the k is the second value, the value of the ijkth position in the three-dimensional matrix is used to represent the i-th first granularity in the fourth dimension and the j-th granularity in the fifth dimension The imaginary part of the channel quality at the second granularity.

Wherein, the first value is different from the second value, for example, the first value can be set to 1 and the second value can be 2, or the first value can be 0 and the second value can be 1, or the first value It can be 1 and the second value can be 0, as long as the first value is different from the second value, it is within the protection scope of this embodiment.

For example, in a three-dimensional matrix of M×N×W, the fourth dimension represents the time domain dimension, and when the fourth dimension is the time domain dimension, the first granularity can also be called the delay granularity; the fifth dimension is the space The domain dimension is specifically the angle dimension, and the second granularity is the interval of the arrival angle; the sixth dimension is the complex number dimension, W is 2, k is 1 for the real part and 2 for the imaginary part. If i=4, j=5, k=1, it means that the value (or indicator value) on the 4th row and 5th column is the 4th delay granularity in the 5th spatial granularity (such as the interval of arrival angle) The real part of the channel quality (or channel quality situation) on . If i=4, j=5, k=2, it means that the value (or indicator value) on the 4th row and 5th column is the 4th delay granularity in the 5th spatial granularity (such as the interval of arrival angle) The imaginary part of the channel quality (or channel quality situation) on .

In addition, it should be noted that the above-mentioned second training samples can also be split and combined on the basis of the above-mentioned fourth dimension, fifth dimension, and sixth dimension. For example, when the fifth dimension is an antenna pair dimension, the It can be split into sending antenna sub-dimensions and receiving antenna sub-dimensions, thereby expanding the dimension of the second training sample. This embodiment does not exhaustively enumerate various possible sub-dimensions after splitting.

In the above-mentioned second method, the terminal device itself performs joint training on the first preset model and the second preset model to obtain the trained first model and the second model; or the terminal device itself trains the first preset model The model, the second preset model and the third preset module are jointly trained to obtain the trained first model, the second model and the third model. In this way, the terminal device can at least send the trained second model. The following is an example of how the terminal device sends the model:

Example one,

The terminal device itself performs joint training on the first preset model and the second preset model to obtain the trained first model and the second model. After the foregoing training is completed, the terminal device sends the second model. Specifically, it may be: the terminal device sends the second model to the network device. Still further, it may also be: the terminal device sends the model parameters of the second model to the network device.

Wherein, the network device may be a network device that provides services for the terminal device, such as an access network device, and specifically may be a base station, eNB, gNB, and the like.

The second model (or model parameters of the second model) is carried by at least one of the following: uplink control signaling, RRC message, uplink data transmission, and uplink data transmission for the transmission requirements of artificial intelligence services.

In such an example, the terminal device may retain the first model itself for processing the first information to obtain the second information; correspondingly, since the network device can receive the second model sent by the terminal device, Therefore, the network device may process the second information based on the second model to obtain channel information. Wherein, the channel information may also be feature vector information of the channel information.

It should be noted that, since one network device can serve multiple terminal devices, the network device can store the second models sent by the multiple terminal devices. Taking the network device as the base station and the terminal device as the mobile phone as an example, base station 1 can serve three mobile phones, namely mobile phone 1, mobile phone 2 and mobile phone 3. of the second model. When the base station 1 receives the second information sent by the mobile phone 2, the base station 1 can process the second information of the mobile phone 2 based on the second model sent by the mobile phone 2 to obtain the channel information corresponding to the mobile phone 2.

Example two,

The terminal device itself performs joint training on the first preset model and the second preset model to obtain the trained first model and the second model. After the above training is completed, on the basis that the terminal device sends the second model, the method may further include: the terminal device also sends the first model.

Specifically, it may be: the terminal device sends the first model to the network device. Still further, it may also be: the terminal device sends the model parameters of the first model to the network device.

The first model (or model parameters of the first model) is carried by at least one of the following: uplink control signaling, RRC message, uplink data transmission, and uplink data transmission for the transmission requirements of artificial intelligence services.

The above-mentioned first model and the second model may be sent at the same time, or the above-mentioned first model and the second model may be sent separately, which is not limited in this embodiment.

In this example, the terminal device can process the first information through the first model to obtain the second information; correspondingly, since the network device can receive the second model sent by the terminal device, the The network device may process the second information based on the second model to obtain channel information. Wherein, the channel information may also be feature vector information of the channel information.

Further, after the network device receives the first model and the second model sent by the terminal device, the network device can conduct an overall evaluation of the first model and the second model, and after completing the overall evaluation of the first model and the second model After the evaluation, you can decide whether to use the first model and the second model received this time. If the overall evaluation result is poor (for example, the compression rate is low or the accuracy of the restored channel information is low, etc.), the above first model may not be used. A model and a second model. If the network device decides not to use the above-mentioned first model and the second model, it can also re-train the first model and the second model to update the model parameters of the first model and the second model, or the network device trains itself to obtain New first model as well as second model. It should be pointed out that if the network device jointly trains the first model and the second model again, or updates the first model and the second model, the network device also needs to send the new first model and the second model to the terminal device , or the network device sends the new first model to the terminal device.

It should be noted that, since one network device can serve multiple terminal devices, the network device can also save the first model and the second model sent by the multiple terminal devices. Furthermore, the network device can conduct an overall evaluation of the first model and the second model sent by each terminal device, and can select the target first model with the best overall evaluation result and its corresponding target second model, and then the network device can itself The target second model is reserved, and the target first model is sent to the above-mentioned multiple terminal devices. Whether the overall evaluation result is optimal can be judged by indicators such as compression ratio and recovery accuracy.

It should be understood that after the network device in this example saves the first model and the second model sent by multiple terminal devices, it does not process the first model and the second model sent by each terminal device, but only After receiving the second information sent by any terminal device, the second information is processed based on the second model of the terminal device that sent the second information.

Example three,

The terminal device itself performs joint training on the first preset model and the second preset model to obtain the trained first model and the second model. The difference between this example and Example 2 is that after the above training is completed, on the basis of the terminal device sending the second model, the terminal device can send the estimation sub-model and the compression sub-model in the first model Model. For example, the terminal device may send the estimated sub-model and the compressed sub-model in the first model at the same time; or, the terminal device may send the estimated sub-model and the compressed sub-model in the first model respectively.

More specifically, the terminal device may send the estimated sub-model and the compressed sub-model in the first model to the network device at the same time; or, the terminal device may send the estimated sub-model in the first model to the network device respectively. Submodels and compressed submodels.

Further, the terminal device may send the model parameters of the estimated sub-model and the model parameters of the compressed sub-model in the first model to the network device at the same time; or, the terminal device may send the first model parameters to the network device respectively. Model parameters for the estimated submodel and model parameters for the compressed submodel in the model.

The estimation sub-model and the compression sub-model may be simultaneously carried by at least one of the following: uplink control signaling, RRC message, uplink data transmission, uplink data transmission for artificial intelligence business type transmission requirements;

The estimation sub-model and the compression sub-model may be respectively carried by at least one of the following: uplink control signaling, RRC message, uplink data transmission, and uplink data transmission for the transmission requirements of artificial intelligence services.

In this example, the terminal device can process the first information through the estimation sub-model and the compression sub-model in the first model to obtain the second information; correspondingly, the network device can receive the A second model. Therefore, the network device may process the second information based on the second model to obtain channel information. Wherein, the channel information may also be feature vector information of the channel information.

Further, after the network device receives the estimated sub-model, the compressed sub-model and the second model in the first model sent by the terminal device, the network device may integrate the estimated sub-model, the compressed sub-model and the second model Evaluation, after completing the overall evaluation of the estimated sub-model, compressed sub-model and the second model, you can decide whether to use the estimated sub-model, compressed sub-model and the second model received this time, if the overall evaluation results are poor (such as compressed rate is low or the accuracy of recovering channel information is low, etc.), the estimation sub-model, the compression sub-model and the second model may not be used. If the network device decides not to use the above estimation sub-model, compression sub-model and second model, it can also re-train the estimation sub-model, compression sub-model and second model by itself to update the estimation sub-model, compression sub-model and second The model parameters of the model, or the network device trains itself to obtain a new estimated sub-model, a compressed sub-model and a second model. It should be pointed out that if the network device jointly trains or updates the estimated sub-model, the compressed sub-model and the second model, the network device also needs to send the new estimated sub-model, the new compressed sub-model and the new second model to The terminal device or the network device sends the new estimated sub-model and the new compressed sub-model to the terminal device.

It should be noted that, since one network device may serve multiple terminal devices, the network device may store the estimated sub-model, the compressed sub-model and the second model sent by the multiple terminal devices. Furthermore, the network device can conduct an overall evaluation of the estimation sub-model, compression sub-model and second model sent by each terminal device, and can select the target estimation sub-model, target compression sub-model and their corresponding target sub-models with the best overall evaluation results. The second model, and then the network device can reserve and use the target second model, and send the target estimation sub-model and the target compression sub-model to the above-mentioned multiple terminal devices. Whether the overall evaluation result is optimal can be judged by indicators such as compression ratio and recovery accuracy.

It should be understood that after the network device in this example saves all the models sent by multiple terminal devices, it does not process the models sent by each terminal device, but only receives the first model sent by any terminal device. After receiving the second information, process the second information based on the second model of the terminal device that sent the second information.

Example four,

The terminal device itself performs joint training on the first preset model and the second preset model to obtain the trained first model and the second model; wherein, the first model includes an estimation sub-model, a compression sub-model and a channel generation sub-model . The difference between this example and Example 3 is that after the above training is completed, on the basis of the terminal device sending the second model, the terminal device can send the estimation sub-model, compression sub-model in the first model model and the channel generation submodel. For example, the terminal device may simultaneously send the estimation sub-model, the compression sub-model and the channel generation sub-model in the first model; or, the terminal device may separately send the estimation sub-model, the compression sub-model and the channel generation sub-model submodel.

More specifically, the terminal device may send the estimation sub-model, the compression sub-model and the channel generation sub-model in the first model to the network device at the same time; or, the terminal device may send the first model to the network device respectively. The estimation sub-model, compression sub-model and channel generation sub-model in a model; or, the terminal device may simultaneously send the estimation sub-model, compression sub-model and channel generation sub-model in the first model to the network device Any two of , and then send the remaining sub-model.

The estimation sub-model, compression sub-model and channel generation sub-model may be simultaneously carried by at least one of the following: uplink control signaling, RRC message, uplink data transmission, uplink data transmission for artificial intelligence service class transmission requirements;

The estimation sub-model, compression sub-model and channel generation sub-model may be respectively carried by at least one of the following: uplink control signaling, RRC message, uplink data transmission, and uplink data transmission for the transmission requirements of artificial intelligence services.

Any two of the estimation sub-model, compression sub-model and channel generation sub-model may be simultaneously carried by at least one of the following: uplink control signaling, RRC messages, uplink data transmission, and uplink transmission requirements for artificial intelligence services data transmission. The remaining sub-model can be carried by at least one of the following: uplink control signaling, RRC message, uplink data transmission, and uplink data transmission for the transmission requirements of artificial intelligence services.

In this example, the terminal device can process the first information through the estimation sub-model, compression sub-model and channel generation sub-model in the first model to obtain the second information; correspondingly, the network device can receive The second model sent by the terminal device, therefore, the network device may process the second information based on the second model to obtain channel information. Wherein, the channel information may specifically be: channel information, or may be eigenvector information of the channel information.

Further, after the network device receives the estimated sub-model, the compressed sub-model, the channel generation sub-model and the second model in the first model sent by the terminal device, the network device may perform the estimated sub-model, the compressed sub-model and the second model. The channel generation sub-model and the second model perform an overall evaluation, and the processing after the overall evaluation is similar to the third example above, and will not be repeated here.

Example five,

The terminal device itself performs joint training on the first preset model, the second preset model and the third preset model to obtain the trained first model, the second model and the third model. After the above training is completed, on the basis that the terminal device sends the second model and the first model, the terminal device may send the third model.

More specifically, the terminal device may send the first model, the second model and the third model to the network device at the same time; or, the terminal device may send the first model and the second model to the network device respectively and the third model; or, the terminal device may first send any two of the first model, the second model, and the third model to the network device, and then send the remaining one model to the network device.

The first model, the second model, and the third model may be simultaneously carried by at least one of the following: uplink control signaling, RRC messages, uplink data transmission, and uplink data transmission for the transmission requirements of artificial intelligence services;

Alternatively, the first model, the second model and the third model may be respectively carried by at least one of the following: uplink control signaling, RRC message, uplink data transmission, uplink data transmission for artificial intelligence service class transmission requirements;

Alternatively, any two of the first model, the second model, the third model, and the remaining one model are respectively carried by at least one of the following: uplink control signaling, RRC messages, uplink data transmission, and artificial intelligence services Uplink data transmission for class transmission requirements.

Optionally, in this example, the first model may include an estimation sub-model and a compression sub-model. Correspondingly, sending the first model may refer to simultaneously or separately sending the estimation sub-model and the compression sub-model. Regarding The carrying manner of the estimation sub-model and the compression sub-model is the same as that of the previous example and will not be repeated here.

Optionally, in this example, the first model may include an estimation submodel, a channel generation submodel, and a compression submodel. Correspondingly, sending the first model may refer to sending the estimation submodel, the The channel generation sub-model and the compression sub-model, the carrying manners of the estimation sub-model, the channel generation sub-model and the compression sub-model are the same as the previous examples and will not be repeated here.

Further, after the network device receives the first model, the second model and the third model sent by the terminal device, the network device may perform an overall evaluation on the first model, the second model and the third model , after completing the overall evaluation of the first model, the second model and the third model, it may be decided whether to use the first model, the second model and the third model received this time, if the overall evaluation result is poor (For example, the compression rate is low or the accuracy rate of recovering channel information is low, etc.), the above-mentioned first model, second model and third model may not be used. If the network device decides not to use the above-mentioned first model, second model and third model, it can also re-train the first model, second model and third model by itself to update the estimation sub-model, compression sub-model model and model parameters of the second model, or the network device trains itself to obtain a new first model, a new second model, and a new third model. It should be pointed out that if the network device jointly trains or updates the estimation sub-model, the compression sub-model and the second model, the network device also needs to send the new first model, the new second model and the new third model to The terminal device or the network device sends the new first model to the terminal device.

It should be noted that, since one network device can serve multiple terminal devices, the network device can store the first model, the second model and the third model sent by the multiple terminal devices. Furthermore, the network device can conduct an overall evaluation of the first model, the second model, and the third model sent by each terminal device, and can select the target first model, the target second model, and the target third model with the best overall evaluation results. , and then the network device can reserve and use the target second model, and send the target first model to the above-mentioned multiple terminal devices. Whether the overall evaluation result is optimal can be judged by indicators such as compression ratio and recovery accuracy.

It can be seen that by adopting the above solution, when the terminal device receives the first information, it can process the first information through the first model to obtain the second information and send it, so that the receiving end can use the second model to obtain the second information. The channel information obtained by processing the information is obtained through joint training of the first model and the second model. Since the processing, transmission, and analysis of the second information are realized by using the first model and the second model obtained through joint training, the performance requirements in the entire information processing, transmission, and analysis can be taken into account, and the overall performance of the network is guaranteed. Furthermore, since the above solution uses the first model and the second model obtained through joint training, the functions between the first model and the second model can be made compatible with each other, so that the performance of the first model and the second model can reach In a better state, when the processing, transmission and analysis process of the second information is processed as a whole based on the first model and the second model, the performance of the whole processing can be guaranteed, thereby ensuring the performance of the whole network.

Fig. 11 is a schematic flowchart of an information processing method 1100 according to an embodiment of the present application. The method can optionally be applied to the system shown in Fig. 1, but is not limited thereto. The method includes at least some of the following.

S1110. The network device sends first information.

S1120. The network device receives second information; wherein, the second information is obtained by processing the first information through a first model.

S1130. The network device processes the second information based on a second model to obtain channel information; wherein, the first model and the second model are obtained through joint training.

In the above S1110, the first information may be a reference signal, specifically, the first information may be a reference signal of the current channel, such as a downlink reference signal of the current channel. The downlink reference signal may include at least one of CSI-RS, DMRS, and PT-RS.

The second dimension is a frequency domain dimension; the first information is distributed on at least one frequency domain resource in the frequency domain dimension; wherein, each frequency domain resource can be one of the following: one RB, one subcarrier . For example, the first signal is a downlink reference signal, and the downlink reference signal may be distributed in 1 RB in the frequency domain dimension, or the downlink reference signal may be distributed in 2 or 4 RBs in the time domain dimension. on RBs.

Optionally, before sending the first information, the network device may also send configuration information first, and the configuration information may be configured with first information for terminal device measurement. Taking the first information as an example of a downlink reference signal, the configuration information may be to configure the terminal device to measure SSB or CSI-RS and so on.

After completing S1110, the network device executes S1120. Wherein, the second information may be carried by one of the following information: information included in the random access process, radio resource control (RRC, Radio Resource Control) signaling, and uplink control information (UCI, Uplink Control Information). The information contained in the random access process includes one of the following: message A in the two-step random access process; Msg1 in the four-step random access process; Msg3 in the four-step random access process.

In an example, the second information in S1120 is channel compression information; the second model is used to decompress the channel compression information to obtain channel information.

Correspondingly, in S1130, the network device processes the second information based on the second model to obtain channel information, including: the network device inputs the channel compression information into the second model, and obtains the second model output The channel information of .

The second information is obtained by processing the first information through the first model. That is to say, the first model is used to process the input first information to obtain channel compression information.

The channel information may be used to characterize the channel quality, or channel response, or channel state, or channel estimation result obtained based on the first information.

The channel information may be represented by a matrix of T dimensions, where T is an integer greater than or equal to 2.

For example, when the fourth dimension is the frequency domain dimension, the first granularity may be L1 RBs (L1 is greater than or equal to 1, such as 2RB, 4RB, 8RB), then the channel information in the frequency domain dimension The distribution range is the frequency domain range corresponding to M×L1 RBs; or the first granularity can be L2 subcarriers (L2 is greater than 1, such as 4 subcarriers, 6 subcarriers, and 18 subcarriers), then the channel The distribution of information in the frequency domain dimension is the frequency domain range corresponding to M×L2 subcarriers. When the fourth dimension is a time-domain dimension, the first granularity may be a delay granularity, for example, a first granularity is the number of sampling points of K1 microseconds, or K2 symbol lengths, or K3 symbols, The symbol here may be an OFDM symbol; when the fourth dimension is the time domain dimension and the first granularity is K1 microseconds, the distribution range of the channel information on the time domain dimension is M×K1 The time domain range corresponding to microseconds.

In another example, the second information in S1120 is channel compression information; the channel compression information includes eigenvector information of compressed channel estimation information. The channel information is eigenvector information of the channel information; the second model is used to decompress the compressed eigenvector information of the channel estimation information to obtain the eigenvector information of the channel information. Wherein, the eigenvector information of the channel information includes R groups of eigenvector sequence information; R is a positive integer.

Correspondingly, in S1130, the network device processes the second information based on the second model to obtain channel information, which may include:

The network device inputs the eigenvector information of the compressed channel estimation information into the second model, and obtains the eigenvector information of the channel information output by the second model.

The second information may be channel compression information; the channel compression information includes eigenvector information of compressed channel estimation information. That is to say, the first model is used to process the input first information to obtain eigenvector information of compressed channel estimation information.

Wherein, the eigenvector information of the channel information includes R groups of eigenvector sequence information; R is a positive integer. For example, R may be 1, then the eigenvector information of the channel information includes a set of eigenvector sequence information. R may be 2, then the eigenvector information of the channel information includes 2 sets of eigenvector sequence information. The above value of R may be determined according to the actual situation, or may be specified during the training of the first model.

In this example, the channel estimation information output by the estimation sub-model may be different from the channel information output by the second model, and the channel estimation information output by the estimation sub-model may specifically be a matrix of channel information, such as It is represented by a matrix of T dimensions; the channel information output by the second model may be eigenvector information of the channel information, for example, may include R groups of eigenvector sequence information. Of course, the channel information output by the second model and the channel estimation information output by the estimation sub-model may also be the same, for example, both may be a matrix of channel information.

The above describes in detail how the network device uses the second model. There are two ways for the network device to obtain the second model: the first way: the network device obtains it directly; the second One way: obtained by the network device training. The two methods are described below:

In a first manner, the network device receives the second model.

Specifically, the network device receives the second model sent by the electronic device; for example, the network device may receive model parameters of the second model sent by the electronic device.

Exemplarily, the electronic device may be the terminal device, and in this case, the network device may be an access network device serving the terminal device, such as a base station, eNB, gNB, and so on. Alternatively, the electronic device may be other devices than the terminal device, for example, it may be a server, or a desktop computer, or a notebook, or other devices capable of data processing, which are not exhaustive in this embodiment.

Wherein, when the electronic device is the terminal device, specifically, the network device receives the second model sent by the terminal device. Wherein, the second model is carried by at least one of the following: uplink control signaling, RRC message, uplink data transmission, and uplink data transmission for the transmission requirements of artificial intelligence services.

In the case that the electronic device is other devices, the second model (or the model parameters of the second model) may be transmitted through a wired connection or other wireless connection. For example, the electronic device transmits the second model (or the model parameters of the second model) to the network device through a wired connection with the network device. Alternatively, the electronic device transmits the second model (or the model parameters of the second model) to the network device through other wireless connections with the network device; wherein, the other wireless connection methods may be bluetooth or WIFI, etc., are not exhaustive here.

Based on the above processing, the network device may also receive the first model. For example, the network device may receive the model parameters of the first model sent by the electronic device.

The electronic device may be the terminal device, and in this case, the network device may be an access network device serving the terminal device, such as a base station, eNB, gNB, and so on. Alternatively, the electronic device may be other devices than the terminal device, for example, it may be a server, or a desktop computer, or a notebook, or other devices capable of data processing, which are not exhaustive in this embodiment.

Wherein, when the electronic device is the terminal device, specifically, the network device receives the first model sent by the terminal device. Wherein, the first model is carried by at least one of the following: uplink control signaling, RRC message, uplink data transmission, and uplink data transmission for the transmission requirements of artificial intelligence services.

In the case that the electronic device is other devices, the first model (or the model parameters of the first model) may be transmitted through a wired connection or other wireless connection. For example, the electronic device transmits the first model (or the model parameters of the first model) to the network device through a wired connection with the network device. Alternatively, the electronic device transmits the first model (or the model parameters of the first model) to the network device through other wireless connections with the network device; wherein, the other wireless connection methods may be Bluetooth or WIFI, etc., are not exhaustive here.

The foregoing first model and the foregoing second model may be received at the same time, or the foregoing first model and the foregoing second model may be received separately, which is not limited in this embodiment.

The foregoing first model may include: an estimation submodel and a compression submodel; or, the first model may include: an estimation submodel, a channel generation submodel, and a compression submodel.

It should be pointed out that when the first model includes different sub-models, the functions of different sub-models may be partially different, respectively:

The above-mentioned first model may include: an estimation sub-model and a compression sub-model;

Wherein, the estimation sub-model is used to perform channel estimation on the first information to obtain channel estimation information; the compression sub-model is used to compress the channel estimation information to obtain the second information.

The above-mentioned first model may include: an estimation sub-model, a channel generation sub-model and a compression sub-model;

The method of performing eigendecomposition in the channel generation sub-model may be a singular value decomposition (SVD, Singular Value Decomposition) method. For example, in the processing of the channel generation sub-model, the input channel information may be subjected to SVD eigendecomposition to obtain eigenvector information of channel estimation information after eigendecomposition. The channel information may be represented by a matrix, and the specific description is the same as that of the foregoing embodiment, and will not be repeated here.

The compression sub-model may be to compress the data volume of the input information. The compression rate between the output information of the compressed sub-model and the input information can be determined during training, for example, the compression rate can be 5/1000, 2/1000, 10%, etc. lift.

In this example, the channel estimation information output by the estimation sub-model may be different from the channel information output by the second model, and the channel estimation information output by the estimation sub-model may specifically be a matrix of channel information, such as It is represented by a matrix of T dimensions; the channel information output by the second model may be eigenvector information of the channel information, for example, may include R groups of eigenvector sequence information. Of course, in this example, the channel information output by the second model and the channel estimation information output by the estimation sub-model may also be the same, for example, both may be a matrix of channel information. In the following embodiments, the channel estimation information output by the estimation sub-model and the channel information output by the second model may be the same or different, and the description will not be repeated.

In the above embodiment, the network device can directly receive the first model. In actual processing, since the first model can contain multiple sub-models, the network device can also receive multiple sub-models respectively, and then receive The obtained multiple sub-models are combined to obtain the first model.

In one case, the first model includes an estimation sub-model and a compression sub-model. Correspondingly, the network device receives the estimated sub-model and the compressed sub-model; the network device generates the first model based on the estimated sub-model and the compressed sub-model. Specifically, the network device receives the model parameters of the estimated sub-model and the model parameters of the compressed sub-model sent by the electronic device; the network device based on the model parameters of the estimated sub-model and the model of the compressed sub-model parameters to obtain the first model.

Here, the network device may receive the estimated sub-model and the compressed sub-model sent by the electronic device at the same time; or, it may receive the estimated sub-model and the compressed sub-model sent by the electronic device separately, for example, it may be received first The estimated sub-model sent by the electronic device then receives the compressed sub-model sent by the electronic device, or first receives the compressed sub-model sent by the electronic device and then receives the estimated sub-model sent by the electronic device.

In the case where the electronic device is the terminal device, the estimated sub-model (or the model parameters of the estimated sub-model) and the compressed sub-model (or the model parameters of the compressed sub-model) may be carried simultaneously or separately by one of the following information ): Uplink control signaling, RRC messages, uplink data transmission, and uplink data transmission for artificial intelligence business transmission requirements.

In the case that the electronic device is other devices, the electronic device can use the above-mentioned estimated sub-model (or the model parameters of the estimated sub-model) and the compressed sub-model (or the model parameters of the compressed sub-model) through a wired connection with the network device. ) are sent simultaneously or separately to the network devices. Alternatively, the electronic device sends the above-mentioned estimated sub-model (or the model parameters of the estimated sub-model) and the compressed sub-model (or the model parameters of the compressed sub-model) to the network simultaneously or separately through other wireless connections with the network device device; wherein, the other wireless connection methods may be bluetooth or WIFI, etc., which are not exhaustive here.

In this case, the network device receives the estimation sub-model, the compression sub-model and the channel generation sub-model; the network device generates the first model. Specifically, the network device receives the model parameters of the estimated sub-model, the model parameters of the compression sub-model and the model parameters of the channel generation sub-model sent by the electronic device; the network device based on the model parameters of the estimated sub-model , model parameters of the compression sub-model and model parameters of the channel generation sub-model to obtain the first model.

Here, the network device may simultaneously receive the estimation sub-model, compression sub-model and channel generation sub-model sent by the electronic device. Alternatively, the estimated sub-model, the compressed sub-model and the channel generation sub-model sent by the electronic device may be respectively received, for example, the estimated sub-model, the compressed sub-model and the channel generation sub-model are all received respectively; or, the estimated sub-model, compressed Any two of the sub-models and the channel generation sub-model are received separately from the remaining one. For example, the network device may first receive the estimated submodel sent by the electronic device, then receive the channel generation submodel sent by the electronic device, and finally receive the compressed submodel sent by the electronic device; or, first receive the The compressed sub-model and the channel generation sub-model sent by the electronic device, and then receive the estimated sub-model sent by the electronic device. It should be pointed out that the above is only an exemplary description, and does not mean that there are only several combinations of the above-mentioned exemplary sub-models, compression sub-models, and channel generation sub-models that are actually sent or received respectively, but this embodiment is not exhaustive. lift.

When the electronic device is the terminal device, when the network device receives the estimated sub-model, the compressed sub-model and the channel generation sub-model simultaneously or respectively, the estimated sub-model, the compressed sub-model and the channel generation sub-model The model is carried simultaneously or separately by one of the following: uplink control signaling, RRC message, uplink data transmission, and uplink data transmission for the transmission requirements of artificial intelligence services.

In the case that the electronic device is other devices, the electronic device may send the above-mentioned estimation sub-model, compression sub-model and channel generation sub-model to the network device simultaneously or separately through a wired connection with the network device . Alternatively, the electronic device sends the above estimated sub-model, compression sub-model and channel generation sub-model to the network device simultaneously or separately through other wireless connections with the network device; wherein, the other wireless connection methods may be Bluetooth or WIFI, etc., are not exhaustive here.

In this manner, the method may further include: the network device receiving the third model. Specifically, the network device may receive model parameters of the third model. Still further, the network device may receive the third model sent by the electronic device, for example, the network device may receive model parameters of the third model sent by the electronic device.

In the case that the electronic device is the terminal device, the third model is carried by one of the following: uplink control signaling, RRC message, uplink data transmission, and uplink data transmission for artificial intelligence business transmission requirements.

In the case that the electronic device is other devices, the third model (or the model parameters of the third model) may be transmitted through a wired connection or other wireless connection. For example, the electronic device transmits the third model (or the model parameters of the third model) to the network device through a wired connection with the network device. Alternatively, the electronic device transmits the third model (or the model parameters of the third model) to the network device through other wireless connections with the network device; wherein, the other wireless connection manner may be Bluetooth or WIFI, etc., are not exhaustive here.

It should be noted that the foregoing second model is a model that the network device needs to use when receiving the second information and processing to obtain channel information. However, in this mode, besides the second model, the first model and/or the third model can also be received. This is because when the first model and the second model are a whole obtained through joint training, if the network device wants to The overall evaluation of the first model and the second model needs to obtain the first model and the second model, and then, after the network device completes the overall evaluation of the first model and the second model, it can decide whether to use the received The first model and the second model, if the overall evaluation results of the network device on the first model and the second model are poor (for example, the compression rate is low or the accuracy of recovering channel information is low, etc.), the above-mentioned model may not be used. first model. If it is determined not to use the above-mentioned first model, at this time, the network device may re-train the first model and the second model by itself to update the model parameters of the first model and the second model, or the network device itself A new first model and a new second model are obtained through training. When the first model, the second model, and the third model are the whole obtained through joint training, the network device can also evaluate the first model, the second model, and the third model as a whole and correspondingly Subsequent processing, the specific processing method is the same as the above, and will not be repeated.

In a second manner, the network device trains itself to obtain the above-mentioned first model.

The network device uses training samples to jointly train the first preset model and the second preset model to obtain the trained first model and the second model;

For example, the output information of the second preset model may be a matrix, and correspondingly, the input information of the compressed preset sub-model may also be a matrix, and here, the output matrix of the second preset model It is called matrix 1, and the matrix of the input of the compressed preset submodel is called matrix 2; the output information of the second preset model and the compressed preset submodel of the first preset model are determined based on the distance The way of the degree of difference between the input information is the MSE way, for example: calculate the difference between matrix 1 and matrix 2, and use the square of the difference as the degree of difference.

For example, the output information of the second preset model may be R sets of feature vector sequence information, and correspondingly, the input information of the compressed preset sub-model may also be R sets of feature vector sequence information. Here, the The R sets of feature vector sequence information output by the second preset model are called feature vector sequence 1, and the R sets of feature vector sequence information input by the compressed preset sub-model are called feature vector sequence 2. The method of determining the degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model based on the degree of similarity may be cosine similarity, for example: feature vector sequence 1 and the cosine angle of the eigenvector sequence 2 to determine the degree of similarity, and use the degree of similarity as the degree of difference.

In the process of using the above-mentioned first loss function for training, due to the difference in the sub-models contained in the first preset model and whether the third preset model for simulating the wireless channel environment is included for joint training, the following four cases are respectively Be explained:

Referring to FIG. 8 a , it illustrates a first preset model 800 , a second preset model 810 , and an estimated preset sub-model 801 and a compressed preset sub-model 802 included in the first preset model 800 . The above-mentioned first preset model 800, the second preset model 810, and the input-output relationship between the estimated preset sub-model 801 and the compressed preset sub-model 802 included in the first preset model 300 can be: estimated The input information of the preset submodel 801 is the first training sample 920; the output information of the estimated preset submodel 801 is used as the input information of the compressed preset submodel 802; the output information of the compressed preset submodel 802 As the input information of the second preset model 810 .

In this case, the network device uses training samples to jointly train the first preset model and the second preset model, including:

The network device inputs the first training sample into the estimated preset sub-model of the first preset model, and obtains initial information output by the estimated preset sub-model;

Wherein, the first training samples may be reference signal samples. The reference signal samples may be original reference signals or processed reference signals obtained through historical acquisition. More specifically, the reference signal samples may be downlink reference signal samples. It should be pointed out that this embodiment does not limit the first training samples to be the downlink reference signal samples, and uplink reference signal samples or other reference signal samples may also be used, but this embodiment does not make an exhaustive list.

It should also be pointed out that, in addition to the first training samples, the information input into the estimated preset sub-model of the first preset model can also be wireless channel or other information related to the scene, for example, it can include at least one of the following: channel Signal-to-noise ratio, signal-to-interference-noise ratio, channel type, bandwidth information, delay information, etc. In the process of training, whether one or more of the above information is input may be relevant according to the actual situation or the actual scene, and it is not limited here.

A specific function of the estimation preset sub-model of the first preset model may be: perform channel estimation based on the first training samples to obtain the initial information. Wherein, the channel estimation may adopt algorithms such as minimum mean square error (MMSE).

The aforementioned initial information may be a matrix, and the dimension of the matrix is not limited here, and may be a matrix of two or more dimensions. The value at each position in the matrix is used to represent the corresponding channel quality at the corresponding granularity corresponding to multiple dimensions. Wherein, the channel quality may be in dBm, or may be a value after normalization processing of the channel quality.

The compressed preset sub-model of the first preset model compresses the data volume of the input information. The compression rate between the output information of the compressed sub-model and the input information can be determined during training, for example, the compression rate can be 5/1000, 2/1000, 10%, etc. lift. In the above processing, the compressed information obtained by compressing the preset sub-model contains less content or data volume than the input information. The form of the above-mentioned compressed information is the same as that of the initial information. For example, the initial information is a matrix, and the corresponding compressed information is also a matrix. The dimensions of the matrix of the initial information and the compressed information are the same, but the data The amount is different.

The function of the second preset model may be to decompress its input information. The input information of the second preset model is compressed information, and the output of the second preset model is restored information. Ideally, the decompression rate of the second preset model should make the obtained restored information contain the same data content as the original information.

The network device uses training samples to jointly train the first preset model and the second preset model, which may include:

Wherein, the specific description about the first training sample is the same as the above-mentioned case 1, so repeated description will not be given. It should also be pointed out that, in addition to the first training samples, the information input into the estimated preset sub-model of the first preset model can also be wireless channel or other information related to the scene, for example, it can include at least one of the following: channel Signal-to-noise ratio, signal-to-interference-noise ratio, channel type, bandwidth information, delay information, etc. In the process of training, whether one or more of the above information is input may be relevant according to the actual situation or the actual scene, and it is not limited here.

The compressed preset sub-model of the first preset model compresses the data volume of the input information. The compression rate between the output information of the compressed sub-model and the input information can be determined during training, for example, the compression rate can be 5/1000, 2/1000, 10%, etc. lift. In the above processing, the compressed feature vector information obtained by compressing the preset sub-model contains less content or data volume than the feature vector information of the input initial information. The form of the above-mentioned compressed feature vector information is the same as that of the initial information. For example, the feature vector information of the initial information is an R group of feature vector sequences, and the corresponding compressed feature vector information is also an R group of feature vectors. sequence, but the amount of data contained in the two is different.

The function of the second preset model may be to decompress its input information to obtain restoration information. In the above processing, the input information of the second preset model is compressed feature vector information, and the output of the second preset model is restored feature vector information. The decompression rate of the second preset model should make the obtained restored feature vector information contain the same data content as the feature vector information of the initial information.

In case three, the network device uses training samples to jointly train the first preset model and the second preset model to obtain the trained first model and the second model, including:

The network device uses training samples to jointly train the first preset model, the second preset model and the third preset model to obtain the trained first model, the second model and the third model; wherein , the third model is a trained third preset model.

The network device uses training samples to jointly train the first preset model, the second preset model and the third preset model, including:

The function of the second preset model may be to decompress its input information to obtain restoration information. The input information of the second preset model is transformed information, and the output of the second preset model is restored information. The decompression rate of the second preset model should make the restored information obtained by it contain close to or the same data content as the original information.

The function of the second preset model may be to decompress its input information to obtain restoration information. The input information of the second preset model is transformed feature vector information, and the output of the second preset model is restored feature vector information. The decompression rate of the second preset model should make the obtained restored feature vector information and the feature vector information of the initial information contain close to or the same data content.

In case five, the network device uses training samples to jointly train the first preset model and the second preset model to obtain the trained first model and the second model; wherein, the first preset model includes estimated preset Set submodels and compress preset submodels.

The network device uses training samples to jointly train the first preset model and the second preset model, including:

Wherein, the first training samples may be reference signal samples. The reference signal samples may be original reference signals or processed reference signals obtained through historical acquisition. More specifically, the reference signal samples may be downlink reference signal samples. It should be pointed out that this embodiment does not limit the first training samples to be the downlink reference signal samples, and uplink reference signal samples or other reference signal samples may also be used, but this embodiment does not make an exhaustive list. It should also be pointed out that, in addition to the first training samples, the information input into the estimated preset sub-model of the first preset model can also be wireless channel or other information related to the scene, for example, it can include at least one of the following: channel Signal-to-noise ratio, signal-to-interference-noise ratio, channel type, bandwidth information, delay information, etc. In the process of training, whether one or more of the above information is input may be relevant according to the actual situation or the actual scene, and it is not limited here.

The above initial information may be a matrix. The dimension of the matrix is not limited here, and may be a matrix of two or more dimensions. The value at each position in the matrix is used to represent the corresponding channel quality at the corresponding granularity corresponding to multiple dimensions.

The compressed preset sub-model of the first preset model compresses the data volume of the input information. The compression rate between the output information of the compressed sub-model and the input information can be determined during training, for example, the compression rate can be 5/1000, 2/1000, 10%, etc. lift. In the above processing, the compressed information obtained by compressing the preset sub-model contains less content or data volume than the input initial information. The form of the above-mentioned compressed information is the same as that of the initial information, such as a matrix, and the matrix dimensions of the initial information and the compressed information are the same, but the data content (or data volume) is different.

The function of the second preset model may be to decompress its input information to obtain restoration information. The decompression rate of the second preset model should make the obtained restored information contain the same data content (or data amount) as the original information.

Regarding the above training, it should also be pointed out that the way of the above training convergence may include at least one of the following: judging whether the number of iterative training reaches the preset number, and judging whether the degree of difference determined by the second loss function is less than a preset threshold. That is to say, when it is determined that the training is completed based on the above manner, the first preset model after the training can be used as the first model, and the second preset model after the training can be used as the second model.

The specific description about the first training sample is the same as that of the fifth case above, so the description will not be repeated. It should also be pointed out that, in addition to the first training samples, the information input into the estimated preset sub-model of the first preset model can also be wireless channel or other information related to the scene, for example, it can include at least one of the following: channel Signal-to-noise ratio, signal-to-interference-noise ratio, channel type, bandwidth information, delay information, etc. In the process of training, whether one or more of the above information is input may be relevant according to the actual situation or the actual scene, and it is not limited here.

The compressed preset sub-model of the first preset model compresses the data volume of the input information. The compression rate between the output information of the compressed sub-model and the input information can be determined during training, for example, the compression rate can be 5/1000, 2/1000, 10%, etc. lift. In the above processing, the content or data volume of the compressed feature vector information output by the compressed preset sub-model is smaller than the data volume or data content of the feature vector information of the input initial information. The form of the compressed feature vector information and the feature vector information of the initial information are the same, but the amount of data contained in the two is different.

The function of the second preset model may be to decompress its input information to obtain restoration information. In the above processing, the input information of the second preset model is compressed feature vector information, and the output of the second preset model is restored feature vector information. The decompression rate of the second preset model should make the obtained restored feature vector information contain the same data content (or data amount) as the feature vector information of the initial information.

In case seven, the network device uses training samples to jointly train the first preset model and the second preset model to obtain the trained first model and the second model, including:

The function of the second preset model may be to decompress its input information to obtain restoration information.

The function of the second preset model may be to decompress its input information to obtain restored feature vector information. The input information of the second preset model is transformed feature vector information, and the output of the second preset model is restored feature vector information. The decompression rate of the second preset model should be such that the obtained restored feature vector information and the feature vector information of the initial information contain data content close to or the same.

The network device can obtain the first model and the second model after its own joint training, or obtain the first model, the second model and the third model after joint training by adopting the above second method. Furthermore, the above-mentioned processing of S1110 to S1130 may be performed.

The training samples are used in the process of the joint training of the network device itself to obtain the first model and the second model, and the joint training of the network device itself to obtain the first model, the second model and the third model provided by the second method above. , the following is a detailed description of the training samples:

The above-mentioned first training samples are distributed in the first dimension and/or the second dimension. It can be understood that only the distribution of the first training samples in the frequency domain dimension can be used for subsequent training, or only the first training samples can be used in the frequency domain. The distribution of the first training sample in the frequency domain and the time domain can also be used for subsequent training. For example, a first training sample contains 10 RBs in the frequency domain dimension and 1 time slot in the time domain dimension, each RB has 3 first signal samples, and each time slot has 1 first signal samples, the first training samples include a total of 30 first signal samples.

The T dimensions include a fourth dimension and a fifth dimension.

That is to say, a second training sample consists of a two-dimensional matrix with a size of M×N, which has M first granularities in the fourth dimension and N second granularities in the fifth dimension; the above M and N May or may not be equal. The specific numerical indication in the two-dimensional matrix represents the received signal strength at a certain first granularity of the channel quality, where the unit of the numerical value in the two-dimensional matrix may be dBm, or the numerical value in the two-dimensional matrix has no unit It is the value obtained after normalization. In addition, the two-dimensional matrix of M×N can also be synthesized into one-dimensional data of size 1×(M×N) or (M×N)×1. The specific transformation can be the fourth dimension first and then the fifth dimension. It may also be the fifth dimension first and then the fourth dimension, which is not limited in this embodiment.

For example, when the first training sample is a reference signal sample or a downlink reference signal sample, the second training sample may be a channel information sample corresponding to the reference signal sample, or may also be called a channel state sample Wait, I'm not going to exhaust the names here. When the fourth dimension is the frequency domain dimension, the first granularity can be L1 RBs (L1 is greater than or equal to 1, such as 2RB, 4RB, 8RB), and the frequency domain range indicated by a second training sample is M×L1 or the first granularity can be L2 subcarriers (L2 is greater than 1, such as 4 subcarriers, 6 subcarriers, and 18 subcarriers), then the frequency domain range indicated by a second training sample is the frequency domain of M×L2 scope. When the fourth dimension is a time-domain dimension, the first granularity may be a delay granularity, for example, a first granularity is the number of sampling points of K1 microseconds, or K2 symbol lengths, or K3 symbols, where the symbols It may be an OFDM symbol; when the fourth dimension is the time domain dimension and the first granularity is K1 microseconds, the time domain range indicated by a second training sample is the time domain range of M×K1.

For another example, in the M×N two-dimensional matrix shown in FIG. 10 , the fourth dimension represents the time domain dimension, and when the fourth dimension is the time domain dimension, the first granularity is one delay granularity; the fifth The dimension is the spatial domain dimension, specifically the angle dimension, and the second granularity is the basic granularity of 1 angle (for example, it can be the interval of 1 arrival angle); if the ij-th position in the M×N two-dimensional matrix is i=4j = 5 positions, then it is the position of the black box on the 4th row and 5th column shown in Fig. The channel quality (or channel quality situation) at the 4th delay granularity within the interval of .

In the above-mentioned second method, the network device itself performs joint training on the first preset model and the second preset model to obtain the trained first model and the second model; or the network device itself trains the first preset model The model, the second preset model and the third preset module are jointly trained to obtain the trained first model, the second model and the third model. In this way, the network device can at least send the trained second model. The following is an example of how network devices send models:

Example one,

The network device itself performs joint training on the first preset model and the second preset model to obtain the trained first model and the second model. After the above training is completed, the network device sends the second model. Specifically, it may be: the network device sends the second model to the terminal device. Still further, it may also be: the network device sends the model parameters of the second model to the terminal device.

The second model (or the model parameters of the second model) is carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data for artificial intelligence business transmission requirements transmission.

In this example, the network device may retain the second model by itself to process the second information to obtain channel information; correspondingly, due to the first model that the terminal device can receive, the terminal The device may process the first information based on the first model to obtain second information.

It should be noted that, since one network device may serve multiple terminal devices, the network device may send the first model to all terminal devices (or at least some terminal devices) it serves. Taking the network device as the base station and the terminal device as the mobile phone as an example, base station 1 can serve three mobile phones, namely mobile phone 1, mobile phone 2 and mobile phone 3, and base station 1 can send the first call to mobile phone 1, mobile phone 2 and mobile phone 3 respectively Model.

Example two,

The network device itself performs joint training on the first preset model and the second preset model to obtain the trained first model and the second model. After the above training is completed, after the network device sends the second model, the method may further include: the network device also sends the first model.

Specifically, it may be: the network device sends the first model to the terminal device. Still further, it may also be: the network device sends the model parameters of the first model to the terminal device.

The first model (or the model parameters of the first model) is carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data for artificial intelligence business transmission requirements transmission.

After the terminal device receives the first model and the second model sent by the network device, the terminal device can perform an overall evaluation of the first model and the second model, and after completing the overall evaluation of the first model and the second model, It can be decided whether to use the first model and the second model received this time. If the overall evaluation result is poor (for example, the compression rate is low or the accuracy of recovering channel information is low, etc.), the above-mentioned first model and the second model may not be used. Second model. If the terminal device decides not to use the above-mentioned first model and the second model, it can also re-train the first model and the second model to update the model parameters of the first model and the second model, or the terminal device trains itself to obtain New first model as well as second model. It should be pointed out that if the terminal device jointly trains the first model and the second model again, or updates the first model and the second model, the terminal device also needs to send the new first model and the second model to the network device , or send the new first model to the network device. Correspondingly, if the network device determines that the overall performance of the new first model and the new second model is better after receiving the new first model and the new second model, it can also replace the first model generated by itself and the second model, and synchronize the new first model and the new second model to other terminal devices served by itself. Otherwise, the network device may specify that the terminal device does not use the new first model and the new second model, so as to ensure that information exchanged between the network device and the terminal device can be correctly transmitted and parsed.

Example three,

The network device itself performs joint training on the first preset model and the second preset model to obtain the trained first model and the second model. The difference between this example and Example 2 is that after the above training is completed, on the basis of the network device sending the second model, the network device can send the estimation sub-model and the compression sub-model in the first model Model. For example, the network device may send the estimated sub-model and the compressed sub-model in the first model at the same time; or, the network device may send the estimated sub-model and the compressed sub-model in the first model respectively.

More specifically, the network device may send the estimated sub-model and the compressed sub-model in the first model to the terminal device at the same time; or, the network device may send the estimated sub-model in the first model to the terminal device respectively. Submodels and compressed submodels.

Further, the network device may send the model parameters of the estimation sub-model and the model parameters of the compression sub-model in the first model to the terminal device at the same time; or, the network device may send the first model parameters to the terminal device respectively. Model parameters for the estimated submodel and model parameters for the compressed submodel in the model.

The estimation sub-model and the compression sub-model may be carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements;

Alternatively, the estimation sub-model and the compression sub-model may be respectively carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data for artificial intelligence business transmission requirements transmission.

Further, after the terminal device receives the estimated sub-model, the compressed sub-model and the second model in the first model sent by the network device, the terminal device may integrate the estimated sub-model, the compressed sub-model and the second model Evaluation, after completing the overall evaluation of the estimated sub-model, compressed sub-model, and second model, you can decide whether to use the estimated sub-model, compressed sub-model, and second model received this time. The specific processing method is similar to the previous example three , which will not be repeated here.

Example four,

The network device itself performs joint training on the first preset model and the second preset model to obtain the trained first model and the second model; wherein, the first model includes an estimation sub-model, a compression sub-model and a channel generation sub-model . The difference between this example and Example 3 is that after the above training is completed, on the basis of the network device sending the second model, the network device can send the estimation sub-model, compression sub-model in the first model model and the channel generation submodel. For example, the network device may simultaneously send the estimation sub-model, the compression sub-model and the channel generation sub-model in the first model; or, the network device may separately send the estimation sub-model, the compression sub-model and the channel generation sub-model submodel.

More specifically, the network device may send the estimation sub-model, the compression sub-model and the channel generation sub-model in the first model to the terminal device at the same time; or, the network device may send the first model to the terminal device respectively. The estimation sub-model, the compression sub-model and the channel generation sub-model in a model; or, the network device may simultaneously send the estimation sub-model, the compression sub-model and the channel generation sub-model in the first model to the terminal device Any two of , and then send the remaining sub-model.

The estimation sub-model, compression sub-model and channel generation sub-model may be simultaneously carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink for artificial intelligence business transmission requirements data transmission;

Alternatively, the estimation sub-model, compression sub-model and channel generation sub-model may be respectively carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, transmission requirements for artificial intelligence services downlink data transmission.

Alternatively, any two of the estimation submodel, the compression submodel, and the channel generation submodel may be simultaneously carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, for Downlink data transmission required for artificial intelligence business transmission. The remaining sub-model can be carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements.

After the terminal device receives the estimated sub-model, the compressed sub-model, the channel generation sub-model and the second model in the first model sent by the network device, the terminal device may perform the estimation sub-model, the compressed sub-model and the channel generation sub-model The overall evaluation is performed on the model and the second model, and the processing after the overall evaluation is similar to that of Example 3 above, and will not be repeated here.

Example five,

The network device itself performs joint training on the first preset model, the second preset model and the third preset model to obtain the trained first model, the second model and the third model. After the above training is completed, on the basis that the terminal device sends the second model and the first model, the network device may send the third model.

More specifically, the network device may send the first model, the second model and the third model to the terminal device at the same time; or, the network device may send the first model and the second model to the terminal device respectively and the third model; or, the network device may first send any two of the first model, the second model, and the third model to the terminal device, and then send the remaining one model to the terminal device.

The first model, the second model and the third model may be simultaneously carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data for artificial intelligence business transmission requirements transmission;

Alternatively, the first model, the second model and the third model may be respectively carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, and information for artificial intelligence business transmission requirements downlink data transmission;

Or, any two of the first model, the second model and the third model and the remaining one model are respectively carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data Transmission, downlink data transmission for artificial intelligence business transmission requirements.

After the terminal device receives the first model, the second model and the third model sent by the network device, the terminal device may perform an overall evaluation on the first model, the second model and the third model, specifically The processing is similar to the foregoing example three, and repeated descriptions are not repeated.

Finally, with reference to FIG. 12 , the information processing method performed by the aforementioned terminal device and the information processing method performed by the network device are exemplarily described. The network device is a base station, which may specifically be:

S1201: The base station trains to obtain the first model and the second model.

For example, the base station uses training samples as the main body to jointly train the first preset model and the second preset model to obtain the first model and the second model; wherein, the first model includes an estimation sub-model and a compression sub-model; or, the first The model contains an estimation submodel, a channel generation submodel, and a compression submodel.

Alternatively, the base station uses training samples as the main body to train the first preset model, the second preset model and the third preset model to obtain the first model, the second model and the third model; wherein, the first model includes An estimation submodel and a compression submodel; alternatively, the first model comprises an estimation submodel, a channel generation submodel and a compression submodel.

S1202: The base station sends the first model to the terminal device. Correspondingly, the terminal device may receive the first model sent by the base station.

This step may refer to: the base station transmits at least the first model to the terminal device; or it may be: the base station transmits the estimated sub-model and the compressed sub-model to the terminal device; or it may be: the The base station transmits the estimation sub-model, the channel generation sub-model and the compression sub-model to the terminal device.

Regarding the two sub-models of the base station transmission estimation sub-model and compression sub-model, or the three sub-models of transmission estimation sub-model, channel generation sub-model and compression sub-model, the sub-models to be obtained by training determined during training of the base station relevant. For example, if the base station obtains the two submodels of the estimated submodel and the compressed submodel through training, the base station may transmit to the terminal device the first model including the two submodels of the estimated submodel and the compressed submodel, or, the The base station directly transmits the two sub-models, the estimated sub-model and the compressed sub-model, to the terminal equipment.

It should be understood that this step may also include: the base station may send the second model to the terminal device. Correspondingly, the terminal device may receive the second model sent by the base station.

In addition, if the base station has obtained the first model, the second model and the third model through training, this step may further include: the base station transmitting the third model to the terminal device. Correspondingly, the terminal device may receive the third model sent by the base station.

S1203: The base station sends first information; correspondingly, the terminal device receives the first information.

In this step, the first information may be a downlink reference signal, specifically a downlink reference signal of the current channel, such as SSB or CSI-RS, and this example does not limit its specific content.

S1204: The terminal device processes the first information based on the first model to obtain second information.

Specifically, if the first model received by the terminal device does not include a channel generation sub-model, the second information is compressed channel information output after being processed by an estimation sub-model and a compression sub-model. The channel information can be expressed as a matrix. Regarding the specific description of the matrix, the specific processing of the estimation sub-model and the compression sub-model in the first model is the same as the foregoing embodiment, and will not be repeated here.

If the first model received by the terminal device includes a channel generation sub-model, the second information is the characteristics of the compressed channel information output after being processed by the estimation sub-model, channel generation sub-model and compression sub-model vector information. The eigenvector information of the compressed channel information can be expressed as a matrix. Regarding the specific description of the matrix, the specific processing of the estimation sub-model and the compression sub-model in the first model is the same as that of the foregoing embodiment, and will not be repeated here.

S1205: The terminal device sends the second information; correspondingly, the base station receives the second information.

S1206: The base station processes the second information based on the second model to obtain channel information.

Wherein, if the first model does not include a channel generation sub-model, correspondingly, the channel information may specifically be a matrix representing channel information. The matrix representing channel information has been described in detail in the foregoing embodiments, and here Do not repeat. If the first model includes a channel generation sub-model, correspondingly, the channel information may specifically be the eigenvector information of the channel information, and the eigenvector information of the channel information has been described in detail in the foregoing embodiments, and will not be repeated here. .

In conjunction with FIG. 13, an exemplary description will be given of an embodiment of the information processing method performed by the aforementioned terminal device and an information processing method performed by the network device. The network device is a base station, and may specifically be:

S1301: The terminal device trains to obtain the first model and the second model;

For example, the terminal device is the main body and uses training samples to jointly train the first preset model and the second preset model to obtain the first model and the second model; wherein, the first model includes an estimation sub-model and a compression sub-model; or, the second A model includes an estimation submodel, a channel generation submodel and a compression submodel.

Alternatively, the terminal device uses training samples as the main body to train the first preset model, the second preset model and the third preset model to obtain the first model, the second model and the third model; wherein, the first model An estimation submodel and a compression submodel are included; alternatively, the first model includes an estimation submodel, a channel generation submodel, and a compression submodel.

S1302: The terminal device sends the second model to the base station. Correspondingly, the base station may receive the second model sent by the terminal device.

This step may refer to: the terminal device at least transmits the second model to a base station.

It should be understood that this step may also include: the terminal device sending the first model to the base station. Correspondingly, the base station may receive the first model sent by the terminal device. Here, regarding whether the first information is sent as a whole or each sub-model is sent separately, the processing method may be the same as that provided in the foregoing embodiment, and the description will not be repeated in this example.

In addition, if the base station has obtained the first model, the second model and the third model through training, this step may further include: the terminal device transmitting the third model to the base station. Correspondingly, the base station may receive the third model sent by the terminal device.

S1303: The base station sends first information; correspondingly, the terminal device receives the first information.

In this step, the first information may specifically be a downlink reference signal, such as an SSB or a CSI-RS, and this example does not limit its specific content.

S1304: The terminal device processes the first information based on the first model to obtain second information.

Specifically, if the first model does not include a channel generation sub-model, the second information is compressed channel information output after being processed by an estimation sub-model and a compression sub-model. The channel information can be expressed as a matrix. Regarding the specific description of the matrix, the specific processing of the estimation sub-model and the compression sub-model in the first model is the same as the foregoing embodiment, and will not be repeated here.

If the first model includes a channel generation sub-model, the second information is the eigenvector information of the compressed channel information output after being processed by the estimation sub-model, the channel generation sub-model and the compression sub-model. The eigenvector information of the compressed channel information can be expressed as a matrix. Regarding the specific description of the matrix, the specific processing of the estimation sub-model and the compression sub-model in the first model is the same as that of the foregoing embodiment, and will not be repeated here.

S1305: The terminal device sends the second information; correspondingly, the base station receives the second information.

S1306: The base station processes the second information based on the second model to obtain channel information.

The difference between the example shown in FIG. 13 and the aforementioned example in FIG. 12 also includes: since one base station can communicate with multiple terminal devices, the base station side may have received the second model from multiple terminal devices. In this case, the base station side may use the second model sent by the terminal device to process the second information sent by the terminal device for each terminal device to obtain channel information. Alternatively, the base station may designate one of the multiple second models sent from multiple terminal devices as the target second model, and at least send the target first model corresponding to the target second model to other terminal devices , so that all terminal devices connected or served by the base station use the same target first model and target second model for subsequent processing, which can also reduce the time consumption of searching for different second models of terminal devices on the base station side.

Fig. 14 is a schematic flowchart of a model generation method 1400 according to an embodiment of the present application. The method can optionally be applied to the system shown in Fig. 1, but is not limited thereto. The method includes at least some of the following.

S1410. Using training samples to jointly train the first preset model and the second preset model, to obtain the trained first model and the second model;

The model generation method provided in this embodiment can be applied to electronic equipment, and the electronic equipment can be a network equipment or a terminal equipment; the network equipment can be a server, an access network equipment, etc.; the terminal equipment can be a smart Phones, tablets, laptops, desktops (or desktop computers), and more. That is to say, any electronic device capable of data processing can execute the model generation method provided in this embodiment.

The joint training of the first preset model and the second preset model by using training samples includes:

inputting the first training sample into the estimated preset sub-model of the first preset model, and obtaining initial information output by the estimated preset sub-model;

Using training samples to jointly train the first preset model and the second preset model may include:

The specific function of the estimation preset sub-model of the first preset model may be: perform channel estimation based on the first training samples to obtain initial information. Wherein, the channel estimation may adopt algorithms such as minimum mean square error (MMSE). The initial information output by the estimated preset sub-model above may be a matrix, and the dimension of the matrix is not limited here, and may be a two-dimensional or more dimensional matrix. The value at each position in the matrix is used to represent the corresponding channel quality at the corresponding granularity corresponding to multiple dimensions. Wherein, the channel quality may be characterized by a signal strength value; the unit of the signal strength value may be dBm, or the signal strength value has no unit but a value obtained after normalization.

In the third case, the training sample is used to jointly train the first preset model and the second preset model to obtain the trained first model and the second model, including:

Using training samples to jointly train the first preset model, the second preset model and the third preset model to obtain the trained first model, the second model and the third model; wherein, the first preset model The three models are the third preset models after training.

Using training samples to jointly train the first preset model, the second preset model and the third preset model, including:

In case three, a third preset model is added relative to the first case. The function of the third preset model is to simulate the channel environment. The specific processing can be to perform data conversion on the input information to obtain the information after data conversion as Output information. Wherein, the specific processing method of the data transformation may include convolution processing or data processing equivalent to convolution; wherein, the data processing equivalent to convolution may be multiple Fourier transform processing, for example, may In order to transform the input information of the third preset model into the frequency domain through Fourier transform, multiply and then transform into the time domain through inverse Fourier transform, the process of convolution in the time domain is equivalent.

The function of the third preset model is to simulate the wireless channel environment, and the specific processing may be to perform data transformation on the input information to obtain the information after data transformation as output information. Wherein, the specific processing method of the data transformation may include convolution processing or data processing equivalent to convolution; wherein, the data processing equivalent to convolution may be multiple Fourier transform processing, for example, may In order to transform the input information of the third preset model into the frequency domain through Fourier transform, multiply and then transform into the time domain through inverse Fourier transform, the process of convolution in the time domain is equivalent.

Using training samples to jointly train the first preset model and the second preset model, including:

Wherein, the first training samples may be reference signal samples. The reference signal sample may be an original reference signal obtained through historical acquisition, or a processed reference signal. More specifically, the reference signal samples may be downlink reference signal samples. It should be understood that this embodiment does not limit that the first training samples must be the downlink reference signal samples, and uplink reference signal samples or other reference signal samples may also be used, which are not exhaustive in this embodiment.

Case 7, using training samples to jointly train the first preset model and the second preset model to obtain the trained first model and the second model, including:

By adopting the second method above, the first model and the second model after joint training, or the first model, the second model and the third model after joint training can be obtained.

In the joint training provided by the second method above to obtain the first model and the second model, and the joint training to obtain the first model, the second model and the third model, the training samples are used, and the training samples are described in detail below :

The training samples may include a first training sample. The first training samples may be reference signal samples. The reference signal samples may be original reference signals or processed reference signals obtained through historical acquisition. Wherein, the original reference signal may refer to a reference signal that has not been transmitted through a wireless channel. The method for acquiring and processing the reference signal may include: using the reference signal received after the original reference signal passes through the wireless channel (or the real wireless channel, or the real wireless channel) as the processed reference signal. Or, the method for obtaining and processing the reference signal may include: using the reference signal received after the original reference signal passes through the simulated wireless channel as the processed reference signal. Still further, the original reference signal may be a downlink reference signal or an uplink reference signal.

The T dimensions include a fourth dimension and a fifth dimension.

For example, referring to FIG. 9, in the M×N two-dimensional matrix, if the fourth dimension represents the frequency domain dimension, the fifth dimension is the space domain dimension, specifically the antenna dimension, the first granularity is 2RB, and the second granularity is 1 For the transceiver antenna; if the ij position in the two-dimensional matrix of M×N is the i=3j=6 position, then it is the position of the black box on the 3rd row and the 6th column shown in Fig. 9, The value (or indicator value) at this position can be used to represent the channel quality (or channel quality situation) on the third 2RB bandwidth (that is, the fifth RB to the sixth RB) on the sixth pair of transceiver antennas ). In addition, in FIG. 9 , S may also be used to represent the number of second training samples, and S may be an integer greater than or equal to 1, that is, the second training samples may include one or more.

Finally, an exemplary description of the neural network and its application in this embodiment:

Regarding the neural network as shown in Figure 5, the basic structure of the neural network includes: an input layer, a hidden layer and an output layer. The input layer is responsible for receiving data, the hidden layer processes the data, and the final result is generated in the output layer. Among them, each node represents a processing unit, which can be regarded as simulating a neuron. Multiple neurons form a layer of neural network, and multiple layers of information transmission and processing construct an overall neural network.

Furthermore, combined with the deep learning algorithm of the neural network, more hidden layers are introduced, and the feature learning is performed layer by layer through the multi-hidden layer neural network training, which greatly improves the learning and processing capabilities of the neural network, and in pattern recognition, Signal processing, optimal combination, anomaly detection, etc. are widely used. With the development of deep learning, convolutional neural networks are further studied. In a convolutional neural network, its basic structure includes: an input layer, multiple convolutional layers, multiple pooling layers, a fully connected layer, and an output layer. Exemplarily, in conjunction with FIG. 15, the input and output between the estimation sub-model, the compression sub-model and the second model included in the first model obtained after the joint training in this embodiment are described. The input shown in FIG. 15 is The information of the estimated sub-model of the first model may be a reference signal, specifically, a reference signal sequence with a length of 144. The estimation sub-model inputs the received input information into its own fully connected layer for processing to obtain its output result. For example, the output dimension of a fully connected layer in the estimation sub-model shown in FIG. 15 is 1024 in size, The output dimension of the last fully connected layer is 8192. The output result of the estimation sub-model can be input into the compression sub-model to obtain the output of the compression sub-model. As a result, another part of the fully connected layer is processed to finally obtain an output result with an output dimension of 256. The compressed sub-model inputs the output result obtained by it into the second model to obtain the final result restored by the second model. For example, as shown in Figure 15, in the second model, it can be processed through a part of the fully connected layer to obtain the output dimension For the result of 1024 size, another part of the fully connected layer is processed to finally get the result with the output dimension of 2048. After the last part of the fully connected layer is processed, the final result with the output dimension of 8192 can be obtained. In Figure 15, the final second model can output channel information with a size of 8192, or it can be transformed into a [128,32,2]-dimensional channel information matrix.

In terms of related technologies, in some research and design of wireless communication systems based on artificial intelligence, there are the following two design ideas:

The first is AI-based channel estimation. As shown in Figure 16, the reference signal received by the terminal equipment is used as input, and the reference signal is processed by the AI-based channel estimation module (or AI-based channel estimation model) in the terminal equipment to obtain the channel estimation result . The channel estimation result obtained in the processing shown in FIG. 16 takes the wireless channel to be recovered as the expected output and achieves the best estimation of the wireless channel as the target.

The second is AI-based channel state information feedback. As shown in Figure 17, the channel state information processed by the encoding end is input into the encoding end neural network to obtain an output feedback vector, which may be compressed channel state information; the feedback vector is sent to the decoding end, and the encoding end Inputting the feedback vector into the neural network at the encoding end to obtain output channel state information. In this idea, the channel information to be fed back is taken as the output, and the above-mentioned channel information is compressed to the greatest extent at the sending end and the above-mentioned channel information is restored to the greatest extent at the receiving end as the goal, and a corresponding neural network-based solution is constructed; wherein, the to-be-feedback The channel information may be complete channel information, or partial channel information, or processed channel information; the processed channel information may be a channel feature vector.

The above two design ideas have already reflected the gains and effectiveness of various designs for single-module functions. However, the best performance of a single module for the communication system does not necessarily mean the best performance of the overall communication system solution. For example, when designing a channel estimation module through AI, the design goal is to use information such as reference signals obtained to maximize the effective estimation of the channel and minimize the error. When designing channel information feedback through AI, the design goal is to achieve the best channel information feedback effect with the minimum feedback overhead. When doing the above two modules separately, for the AI training of the channel estimation module, the channel estimation error should be reduced as high as possible, or the accuracy of artificially controlled channel estimation can meet the requirements of lower neural network complexity. However, when the joint channel estimation module and the channel state information feedback module are analyzed together, the following problems will be found: whether the above-mentioned channel estimation error minimization, and the additional channel estimation model and algorithm complexity made to minimize the channel estimation error are meaningful There are doubts; if the accuracy of channel estimation is reduced to reduce the complexity of implementation, this part of the reduced accuracy may be redundant; from the perspective of the channel state information feedback module, under the assumption of the same feedback efficiency, relying on high-precision channels to make solutions , Model design may also bring additional redundancy; the above-mentioned high-precision channel information may also bring redundant overhead to the generation and feedback of channel state information.

In summary, it can be seen that in the processing of channel estimation and channel information feedback in wireless communication systems based on AI, since multiple modules (or models) are divided based on different functions and multiple modules (or models) perform independent task training, it is possible There will be problems that need to be introduced for intermediate design goals that have an impact on the feedback scheme.

However, in the solution provided by this embodiment, since the processing, transmission and analysis of the second information are realized by using the first model and the second model obtained through joint training, it can take into account the performance requirements in the entire information processing, transmission and analysis. , to ensure the overall performance of the network. Furthermore, since the above solution uses the first model and the second model obtained through joint training, the functions between the first model and the second model can be made compatible with each other, so that the performance of the first model and the second model can reach In a better state, when the processing, transmission and analysis process of the second information is processed as a whole based on the first model and the second model, the performance of the whole processing can be guaranteed, thereby ensuring the performance of the whole network. Furthermore, the solution provided in this embodiment can avoid problems such as artificially dividing sub-modules and performing independent task training, such as training objectives and redundant information utilization, through the joint training model, and can integrate the entire channel estimation and channel information feedback. The design process is always aimed at allowing the opposite end to recover channel information with the minimum cost, thereby avoiding unnecessary channel recovery appeals and waste of design and calculation overhead, and ensuring that the overall coordination of multiple models obtained in the end can achieve optimal processing As a result, the overall performance of the system is guaranteed.

Fig. 18 is a schematic block diagram of a terminal device 1800 according to an embodiment of the present application. Can include:

The first communication unit 1801 is configured to receive first information; send second information obtained based on the first information;

The second information is channel compression information;

The first model is used to process the input first information to obtain channel compression information.

The first model includes: an estimation sub-model and a compression sub-model;

Wherein, the estimation sub-model is used to perform channel estimation based on the first information to obtain channel estimation information;

As shown in Figure 19, the terminal device also includes:

The first processing unit 1802 is configured to input the first information into the estimation sub-model to obtain channel estimation information output by the estimation sub-model; input the channel estimation information to the compression sub-model to obtain the compressed Channel compression information for the submodel output.

The first information is a reference signal.

The second information is channel compression information; the channel compression information includes eigenvector information of compressed channel estimation information;

The first model is used to process the input first information to obtain eigenvector information of compressed channel estimation information.

The first model includes: an estimation submodel, a channel generation submodel and a compression submodel;

Wherein, the estimation sub-model is used to perform channel estimation on the first information to obtain channel estimation information;

The eigenvector information of the channel information includes R groups of eigenvector sequence information; R is a positive integer.

The first processing unit 1802 is configured to input the first information into the estimation sub-model to obtain channel estimation information output by the estimation sub-model; input the channel estimation information to the channel generation sub-model to obtain The eigenvector information of the channel estimation information output by the channel generation sub-model; input the eigenvector information of the channel estimation information into the compression sub-model, and obtain the eigenvector information of the compressed channel estimation information output by the compression sub-model .

The first information is a reference signal; the channel information is feature vector information of the channel information.

The first communication unit 1801 is configured to receive the first model.

The first model is carried by at least one of the following: downlink control signaling, media access control MAC control element CE message, radio resource control RRC message, broadcast message, downlink data transmission, downlink data for artificial intelligence service transmission requirements transmission.

The first communication unit 1801 is configured for the terminal device to receive an estimated sub-model and a compressed sub-model;

The first processing unit 1802 is configured to generate the first model based on the estimated sub-model and the compressed sub-model.

The first communication unit 1801 is configured to receive an estimation sub-model, a compression sub-model and a channel generation sub-model;

The first processing unit 1802 is configured to generate the first model based on the estimation sub-model, the compression sub-model and the channel generation sub-model.

Wherein, the estimation sub-model is carried by one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements;

The compressed sub-model is carried by one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements;

The channel generation sub-model is carried by one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements.

The first communication unit 1801 is configured to receive the second model.

The first communication unit 1801 is configured to receive the third model.

The third model is used to perform data conversion processing on the second information output by the first model and then input it into the second model;

The first model, the second model and the third model are obtained through joint training.

The data transformation processing includes: convolution processing or Fourier transform processing.

The first processing unit is configured to use training samples to jointly train the first preset model and the second preset model to obtain the trained first model and the second model;

The loss function used in the training is the first loss function;

The first loss function is constructed based on the degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model.

The first processing unit is configured to input a first training sample into an estimated preset sub-model of the first preset model, and obtain initial information output by the estimated preset sub-model;

The first processing unit is configured to use training samples to jointly train the first preset model, the second preset model and the third preset model to obtain the trained first model, the second model and the third model;

Wherein, the third model is a trained third preset model.

The loss function used in the training is a second loss function;

The second loss function is based on a first degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model, and the first preset The second degree of difference between the output information of the estimated preset sub-model of the model and the second training sample is constructed; wherein the second training sample corresponds to the first training sample input to the estimated preset sub-model.

The first degree of difference is determined based on a distance, or is determined based on a degree of similarity; and/or,

The second degree of difference is determined based on distance, or determined based on similarity.

Wherein, the third model is a trained third preset model.

The training samples include the first training samples.

The first dimension is a time domain dimension; the first training samples include first information samples distributed in m time units in the time domain dimension; m is a positive integer.

The second dimension is a frequency domain dimension; the first training samples include first information samples distributed on x frequency domain resources in the frequency domain dimension; x is a positive integer.

The first training samples are also distributed in the third dimension;

The third dimension is a complex dimension; the first training samples include the real part of the first information sample and the imaginary part of the first information sample.

The training samples also include a second training sample corresponding to the first training sample;

The second training sample is composed of a matrix of T dimensions; T is an integer greater than or equal to 2.

The T dimensions include a fourth dimension and a fifth dimension.

The matrix of the T dimensions is a two-dimensional matrix of M×N; wherein, M represents the quantity of the first granularity under the fourth dimension, and N represents the quantity of the second granularity under the fifth dimension; both M and N are is a positive integer.

The value at the ijth position in the two-dimensional matrix is used to represent the channel quality at the i-th first granularity in the fourth dimension and the j-th second granularity in the fifth dimension; both i and j are is a positive integer.

The T dimensions also include a sixth dimension.

The matrix of T dimensions is a three-dimensional matrix of M×N×W; wherein, M represents the number of first granularities in the fourth dimension, N represents the number of second granularities in the fifth dimension, and W represents the number of granularities in the fifth dimension. The quantity of the third granularity under the sixth dimension; M, N and W are all positive integers.

The value of the ijkth position in the three-dimensional matrix is used to represent the i-th first granularity in the fourth dimension, the j-th second granularity in the fifth dimension, and the k-th in the sixth dimension channel quality at the third granularity; i, j and k are all positive integers.

The fourth dimension is a frequency domain dimension; the first granularity includes one of the following: L1 resource blocks RB, L2 subcarriers; L1 and L2 are positive integers.

The fourth dimension is a time domain dimension; the first granularity includes one of the following: K1 microseconds, K2 symbol length, K3 number of sampling points of symbols; K1, K2 and K3 are positive integers.

The symbols are OFDM symbols.

The fifth dimension is a spatial domain dimension;

The second granularity is an interval between a pair of transmitting and receiving antennas or an angle of arrival.

The sixth dimension is a complex dimension; the third granularity is 1, and the quantity W of the third granularity under the complex dimension is 2.

The first communication unit is configured to send the second model.

The first communication unit is configured to send the first model.

The first communication unit is configured to send the estimated sub-model and the compressed sub-model in the first model.

The first communication unit is configured to send the estimation sub-model, the compression sub-model and the channel generation sub-model in the first model.

The first communication unit is configured to send the third model.

The first model is carried by at least one of the following: uplink control signaling, RRC messages, uplink data transmission, and uplink data transmission for artificial intelligence business transmission requirements;

The second model is carried by at least one of the following: uplink control signaling, RRC messages, uplink data transmission, and uplink data transmission for artificial intelligence business transmission requirements;

The third model is carried by at least one of the following: uplink control signaling, RRC messages, uplink data transmission, and uplink data transmission for artificial intelligence business transmission requirements;

The estimation sub-model is carried by at least one of the following: uplink control signaling, RRC message, uplink data transmission, uplink data transmission for artificial intelligence business type transmission requirements;

The compressed sub-model is carried by at least one of the following: uplink control signaling, RRC messages, uplink data transmission, and uplink data transmission for artificial intelligence business transmission requirements;

The channel generation sub-model is carried by at least one of the following: uplink control signaling, RRC messages, uplink data transmission, and uplink data transmission for the transmission requirements of artificial intelligence services.

The terminal device 1800 in the embodiment of the present application can implement the corresponding functions of the terminal device in the foregoing method embodiments. For the processes, functions, implementations and beneficial effects corresponding to each module (submodule, unit or component, etc.) in the terminal device, refer to the corresponding description in the above method embodiment, and details are not repeated here. It should be noted that the functions described by the modules (submodules, units or components, etc.) in the terminal device 400 of the embodiment of the application can be realized by different modules (submodules, units or components, etc.), or by the same Module (submodule, unit or component, etc.) implementation.

Fig. 20 is a schematic block diagram of a network device 2000 according to an embodiment of the present application. This network equipment can include:

The second communication unit 2001 is configured to send first information; receive second information; wherein, the second information is obtained by processing the first information through the first model;

The second processing unit 2002 is configured to process the second information based on a second model to obtain channel information; wherein, the first model and the second model are obtained through joint training.

The second information is channel compression information;

The second model is used to decompress the channel compressed information to obtain channel information.

The second processing unit 2002 is configured to input the channel compression information into the second model, and obtain the channel information output by the second model.

The second information is channel compression information; the channel compression information includes eigenvector information of compressed channel estimation information; the channel information is eigenvector information of channel information;

The second model is used to decompress the compressed eigenvector information of the channel estimation information to obtain the eigenvector information of the channel information.

The second processing unit is configured to input eigenvector information of the compressed channel estimation information into the second model, and obtain eigenvector information of the channel information output by the second model.

The first information is a reference signal.

The second communication unit 2001 is configured to receive the second model.

The second model is carried by at least one of the following: uplink control signaling, RRC message, uplink data transmission, and uplink data transmission for the transmission requirements of artificial intelligence services.

The second communication unit is configured to receive the first model.

The first model is carried by at least one of the following: uplink control signaling, RRC message, uplink data transmission, and uplink data transmission for the transmission requirements of artificial intelligence services.

The first model includes: an estimation sub-model and a compression sub-model;

The compression sub-model is used to compress the channel estimation information to obtain the second information.

The compression sub-model is used to compress the eigenvector information of the channel estimation information to obtain the second information.

The second communication unit is configured to receive the estimated sub-model and the compressed sub-model;

The second processing unit is configured to generate the first model based on the estimated sub-model and the compressed sub-model.

The second communication unit is configured to receive the estimation sub-model, the compression sub-model, and the channel generation sub-model;

The second processing unit is configured to generate the first model based on the estimation sub-model, the compression sub-model and the channel generation sub-model.

The estimated sub-model is carried by one of the following: uplink control signaling, RRC message, uplink data transmission, uplink data transmission for artificial intelligence business class transmission requirements;

The compressed sub-model is carried by one of the following: uplink control signaling, RRC messages, uplink data transmission, and uplink data transmission for artificial intelligence business transmission requirements;

The channel generation sub-model is carried by one of the following: uplink control signaling, RRC message, uplink data transmission, and uplink data transmission for the transmission requirements of artificial intelligence services.

The second communication unit is configured to receive the third model.

The second processing unit is configured to use training samples to jointly train the first preset model and the second preset model to obtain the trained first model and the second model;

The loss function used in the training is the first loss function;

The second processing unit is configured to input a first training sample into an estimated preset sub-model of the first preset model, and obtain initial information output by the estimated preset sub-model;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining the eigenvector information of the initial information output by the channel generation preset sub-model;

The second processing unit is configured to use training samples to jointly train the first preset model, the second preset model and the third preset model to obtain the trained first model, the second model and the A third model; wherein, the third model is a trained third preset model.

performing reverse conduction according to the first loss function to update the first preset model, the second preset model, and the third preset model.

The loss function used in the training is a second loss function;

The second loss function is based on the first degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model and the first preset model The second difference degree between the output information of the estimated preset sub-model and the second training sample is constructed; wherein, the second training sample corresponds to the first training sample input into the estimated preset sub-model.

inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model, and obtaining the compressed eigenvector information output by the compressed preset sub-model;

The training samples include the first training samples.

The first dimension is a time domain dimension;

The first training samples include first information samples distributed in m time units in the time domain dimension; m is a positive integer.

The second dimension is a frequency domain dimension;

The first training samples include first information samples distributed on x frequency domain resources in the frequency domain dimension; x is a positive integer.

The first training samples are also distributed in the third dimension;

The T dimensions include a fourth dimension and a fifth dimension.

The T dimensions also include a sixth dimension.

The symbols are OFDM symbols.

The fifth dimension is a spatial domain dimension;

The second communication unit is configured to send the second model.

The second communication unit is configured to send the first model.

The second communication unit is used for the estimation sub-model and the compression sub-model.

The second communication unit is configured to send the estimation sub-model, the compression sub-model and the channel generation sub-model.

The second communication unit is configured to send the third model.

The first model is carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements;

The second model is carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements;

The third model is carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements;

The estimated sub-model is carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements;

The compressed sub-model is carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements;

The channel generation sub-model is carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements.

The network device 2000 in the embodiment of the present application can implement the corresponding functions of the network device in the foregoing method embodiments. For the procedures, functions, implementation methods and beneficial effects corresponding to each module (submodule, unit or component, etc.) in the network device, refer to the corresponding description in the above method embodiments, and details are not repeated here. It should be noted that the functions described by each module (submodule, unit or component, etc.) in the network device of the application embodiment can be realized by different modules (submodule, unit or component, etc.), or by the same module (submodule, unit or component, etc.) implementation.

The embodiment of the present application also provides an electronic device 2100, as shown in FIG. 21 , including:

The third processing unit 2101 is configured to use training samples to jointly train the first preset model and the second preset model to obtain the trained first model and the second model;

The loss function used in the training is the first loss function;

The degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model is determined based on a distance, or determined based on a degree of similarity.

The third processing unit is configured to input a first training sample into an estimated preset sub-model of the first preset model, and obtain initial information output by the estimated preset sub-model;

The third processing unit is configured to use training samples to jointly train the first preset model, the second preset model and the third preset model to obtain the trained first model, the second model and the A third model; wherein, the third model is a trained third preset model.

The loss function used in the training is a second loss function;

The training samples include the first training samples.

The first training samples are also distributed in the third dimension;

Wherein, the training samples also include a second training sample corresponding to the first training sample;

The T dimensions include a fourth dimension and a fifth dimension.

The T dimensions also include a sixth dimension.

The symbols are OFDM symbols.

The fifth dimension is a spatial domain dimension;

Fig. 22 is a schematic structural diagram of a communication device 2200 according to an embodiment of the present application. The communication device 2200 includes a processor 2210, and the processor 2210 can invoke and run a computer program from a memory, so that the communication device 2200 implements the method in the embodiment of the present application.

Optionally, the communication device 2200 may further include a memory 2220 . Wherein, the processor 2210 may call and run a computer program from the memory 2220, so that the communication device 2200 implements the method in the embodiment of the present application.

Wherein, the memory 2220 may be an independent device independent of the processor 2210 , or may be integrated in the processor 2210 .

Optionally, the communication device 2200 may further include a transceiver 2230, and the processor 2210 may control the transceiver 2230 to communicate with other devices, specifically, may send information or data to other devices, or receive information or data sent by other devices .

Wherein, the transceiver 2230 may include a transmitter and a receiver. The transceiver 2230 may further include an antenna, and the number of antennas may be one or more.

Optionally, the communication device 2200 may be the network device of the embodiment of the present application, and the communication device 2200 may implement the corresponding processes implemented by the network device in the methods of the embodiment of the present application. For the sake of brevity, details are not repeated here.

Optionally, the communication device 2200 may be the terminal device of the embodiment of the present application, and the communication device 2200 may implement the corresponding processes implemented by the terminal device in each method of the embodiment of the present application. For the sake of brevity, details are not repeated here.

FIG. 23 is a schematic structural diagram of a chip 2300 according to an embodiment of the present application. The chip 2300 includes a processor 2310, and the processor 2310 can call and run a computer program from the memory, so as to implement the method in the embodiment of the present application.

Optionally, the chip 2300 may also include a memory 2320 . Wherein, the processor 2310 may invoke and run a computer program from the memory 2320, so as to implement the method performed by the terminal device or the network device in the embodiment of the present application.

Wherein, the memory 2320 may be an independent device independent of the processor 2310 , or may be integrated in the processor 2310 .

Optionally, the chip 2300 may also include an input interface 2330 . Wherein, the processor 2310 can control the input interface 2330 to communicate with other devices or chips, specifically, can obtain information or data sent by other devices or chips.

Optionally, the chip 2300 may also include an output interface 2340 . Wherein, the processor 2310 can control the output interface 2340 to communicate with other devices or chips, specifically, can output information or data to other devices or chips.

Optionally, the chip can be applied to the network device in the embodiment of the present application, and the chip can implement the corresponding processes implemented by the network device in the methods of the embodiment of the present application. For the sake of brevity, details are not repeated here.

Optionally, the chip can be applied to the terminal device in the embodiments of the present application, and the chip can implement the corresponding processes implemented by the terminal device in the methods of the embodiments of the present application. For the sake of brevity, details are not repeated here.

Chips applied to network devices and terminal devices may be the same chip or different chips.

It should be understood that the chip mentioned in the embodiment of the present application may also be called a system-on-chip, a system-on-chip, a system-on-a-chip, or a system-on-a-chip.

The processor mentioned above can be a general-purpose processor, a digital signal processor (DSP), an off-the-shelf programmable gate array (FPGA), an application specific integrated circuit (ASIC) or Other programmable logic devices, transistor logic devices, discrete hardware components, etc. Wherein, the general-purpose processor mentioned above may be a microprocessor or any conventional processor or the like.

The aforementioned memories may be volatile memories or nonvolatile memories, or may include both volatile and nonvolatile memories. Among them, the non-volatile memory can be read-only memory (read-only memory, ROM), programmable read-only memory (programmable ROM, PROM), erasable programmable read-only memory (erasable PROM, EPROM), electrically programmable Erases programmable read-only memory (electrically EPROM, EEPROM) or flash memory. The volatile memory may be random access memory (RAM).

It should be understood that the above-mentioned memory is illustrative but not restrictive. For example, the memory in the embodiment of the present application may also be a static random access memory (static RAM, SRAM), a dynamic random access memory (dynamic RAM, DRAM), Synchronous dynamic random access memory (synchronous DRAM, SDRAM), double data rate synchronous dynamic random access memory (double data rate SDRAM, DDR SDRAM), enhanced synchronous dynamic random access memory (enhanced SDRAM, ESDRAM), synchronous connection Dynamic random access memory (synch link DRAM, SLDRAM) and direct memory bus random access memory (Direct Rambus RAM, DR RAM), etc. That is, memory in embodiments of the present application is intended to include, but not be limited to, these and any other suitable types of memory.

Fig. 24 is a schematic block diagram of a communication system 2400 according to an embodiment of the present application. The communication system 2400 includes a terminal device 2410 and a network device 2420 .

Wherein, the terminal device 2410 may be used to realize corresponding functions realized by the terminal device in the above method, and the network device 2420 may be used to realize corresponding functions realized by the network device in the above method. For the sake of brevity, details are not repeated here.

In the above embodiments, all or part of them may be implemented by software, hardware, firmware or any combination thereof. When implemented using software, it may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on the computer, the processes or functions according to the embodiments of the present application will be generated in whole or in part. The computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable device. The computer instructions may be stored in or transmitted from one computer-readable storage medium to another computer-readable storage medium, e.g. (such as coaxial cable, optical fiber, digital subscriber line (Digital Subscriber Line, DSL)) or wireless (such as infrared, wireless, microwave, etc.) to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that can be accessed by a computer, or a data storage device such as a server or a data center integrated with one or more available media. The available medium may be a magnetic medium (such as a floppy disk, a hard disk, or a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (Solid State Disk, SSD)), etc.

It should be understood that, in various embodiments of the present application, the sequence numbers of the above-mentioned processes do not mean the order of execution, and the execution order of the processes should be determined by their functions and internal logic, and should not be used in the embodiments of the present application. The implementation process constitutes any limitation.

Those skilled in the art can clearly understand that for the convenience and brevity of the description, the specific working process of the above-described system, device and unit can refer to the corresponding process in the foregoing method embodiment, which will not be repeated here.

The above is only the specific implementation of the application, but the scope of protection of the application is not limited thereto. Anyone familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the application, and should covered within the scope of protection of this application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

An information processing method, comprising:

The terminal device receives the first information;

The terminal device sends second information obtained based on the first information;

Wherein, the second information is obtained by processing the first information through the first model, and the second information is used for processing through the second model to obtain channel information; the first model and the second model are a joint obtained by training.
The method according to claim 1, wherein the second information is channel compression information;

The first model is used to process the input first information to obtain channel compression information.
The method of claim 2, wherein the first model comprises: an estimation sub-model and a compression sub-model;

Wherein, the estimation sub-model is used to perform channel estimation based on the first information to obtain channel estimation information;

The compression sub-model is used to compress the channel estimation information to obtain channel compression information.
The method according to claim 3, wherein the method further comprises:

The terminal device inputs the first information into the estimation sub-model, and obtains channel estimation information output by the estimation sub-model;

The terminal device inputs the channel estimation information into the compression sub-model, and obtains channel compression information output by the compression sub-model.
The method according to any one of claims 1-4, wherein the first information is a reference signal.
The method according to claim 1, wherein the second information is channel compression information; the channel compression information includes eigenvector information of compressed channel estimation information;

The first model is used to process the input first information to obtain eigenvector information of compressed channel estimation information.
The method according to claim 6, wherein the first model comprises: an estimation submodel, a channel generation submodel and a compression submodel;

Wherein, the estimation sub-model is used to perform channel estimation on the first information to obtain channel estimation information;

The channel generation sub-model is used to perform eigendecomposition on the channel estimation information to obtain eigenvector information of the channel estimation information;

The compression sub-model is used to compress the eigenvector information of the channel estimation information to obtain compressed eigenvector information of the channel estimation information.
The method according to claim 7, wherein,

The eigenvector information of the channel estimation information includes R groups of eigenvector sequence information; R is a positive integer.
The method according to claim 7 or 8, wherein the method further comprises:

The terminal device inputs the first information into the estimation sub-model, and obtains channel estimation information output by the estimation sub-model;

The terminal device inputs the channel estimation information into the channel generation sub-model, and obtains eigenvector information of the channel estimation information output by the channel generation sub-model;

The terminal device inputs the eigenvector information of the channel estimation information into the compression sub-model, and obtains the eigenvector information of the compressed channel estimation information output by the compression sub-model.
The method according to any one of claims 6-9, wherein the first information is a reference signal; and the channel information is eigenvector information of the channel information.
The method according to any one of claims 1-10, wherein the method further comprises:

The terminal device receives the first model.
The method according to claim 9, wherein the first model is carried by at least one of the following: downlink control signaling, media access control MAC control element CE message, radio resource control RRC message, broadcast message, downlink data transmission , Downlink data transmission for artificial intelligence business transmission requirements.
The method according to any one of claims 1-10, wherein the method further comprises:

The terminal device receives the estimated sub-model and the compressed sub-model;

The terminal device generates the first model based on the estimated sub-model and the compressed sub-model.
The method according to any one of claims 1-10, wherein the method further comprises:

The terminal device receives an estimation sub-model, a compression sub-model and a channel generation sub-model;

The terminal device generates the first model based on the estimation sub-model, the compression sub-model and the channel generation sub-model.
The method according to claim 13 or 14, wherein the estimated sub-model is carried by one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, transmission requirements for artificial intelligence services downlink data transmission;

The compressed sub-model is carried by one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements;

The channel generation sub-model is carried by one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements.
The method according to any one of claims 11-15, wherein the method further comprises:

The terminal device receives the second model.
The method according to any one of claims 11-16, wherein the method further comprises:

The terminal device receives the third model.
The method according to claim 17, wherein the third model is used to input the second information output by the first model into the second model after data conversion processing;

The first model, the second model and the third model are obtained through joint training.
The method according to claim 18, wherein said data transformation processing comprises: convolution processing or Fourier transform processing.
The method according to any one of claims 1-10, wherein the method further comprises:

The terminal device uses training samples to jointly train the first preset model and the second preset model to obtain the trained first model and the second model;

Wherein, the first model is the first preset model after training, and the second model is the second preset model after training.
The method according to claim 20, wherein the loss function used in the training is a first loss function;

The first loss function is constructed based on the degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model.
The method of claim 21, wherein the degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model is determined based on distance, or determined on the basis of similarity.
The method according to claim 21 or 22, wherein the terminal device uses training samples to jointly train the first preset model and the second preset model, comprising:

The terminal device inputs the first training sample into the estimated preset sub-model of the first preset model, and obtains initial information output by the estimated preset sub-model;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into the second preset model to obtain restoration information output by the second preset model;

determining the first loss function based on the degree of difference between the restoration information and the initial information;

performing reverse conduction according to the first loss function to update the first preset model and the second preset model.
The method according to claim 21 or 22, wherein the terminal device uses training samples to jointly train the first preset model and the second preset model, comprising:

The terminal device inputs the first training sample into the estimated preset sub-model of the first preset model, and obtains initial information output by the estimated preset sub-model;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

inputting the compressed feature vector information into the second preset model to obtain restored feature vector information output by the second preset model;

determining a first loss function based on the degree of difference between the restored eigenvector information and the eigenvector information of the initial information;

performing reverse conduction according to the first loss function to update the first preset model and the second preset model.
The method according to claim 21 or 22, wherein the terminal device uses training samples to jointly train the first preset model and the second preset model to obtain the trained first model and the second preset model. models, including:

The terminal device uses training samples to jointly train the first preset model, the second preset model and the third preset model to obtain the trained first model, the second model and the third model;

Wherein, the third model is a trained third preset model.
The method according to claim 25, wherein the terminal device uses training samples to jointly train the first preset model, the second preset model and the third preset model, comprising:

The terminal device inputs the first training sample into the estimated preset sub-model of the first preset model, and obtains initial information output by the estimated preset sub-model;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into a third preset model to obtain transformed information output by the third preset model;

inputting the transformed information into the second preset model to obtain restoration information output by the second preset model;

determining a first loss function based on the degree of difference between the restoration information and the initial information;

performing reverse conduction according to the first loss function to update the first preset model, the second preset model and the third preset model.
The method according to claim 25, wherein the terminal device uses training samples to jointly train the first preset model, the second preset model and the third preset model, comprising:

The terminal device inputs the first training sample into the estimated preset sub-model of the first preset model, and obtains initial information output by the estimated preset sub-model;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

Inputting the compressed feature vector information into a third preset model to obtain transformed feature vector information output by the third preset model;

inputting the transformed eigenvector information into the second preset model to obtain restored eigenvector information output by the second preset model;

determining a first loss function based on the degree of difference between the restored eigenvector information and the eigenvector information of the initial information;

performing reverse conduction according to the first loss function to update the first preset model, the second preset model and the third preset model.
The method according to claim 20, wherein the loss function used in the training is a second loss function;

The second loss function is based on a first degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model, and the first preset The second degree of difference between the output information of the estimated preset sub-model of the model and the second training sample is constructed; wherein the second training sample corresponds to the first training sample input to the estimated preset sub-model.
The method according to claim 28, wherein the first degree of difference is determined based on a distance, or determined based on a degree of similarity; and/or,

The second degree of difference is determined based on distance, or determined based on similarity.
The method according to claim 28 or 29, wherein the terminal device uses training samples to jointly train the first preset model and the second preset model, comprising:

The terminal device inputs the first training sample into the estimated preset sub-model of the first preset model, and obtains initial information output by the estimated preset sub-model;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into the second preset model to obtain restoration information output by the second preset model;

Determine a second loss function based on a first degree of difference between the restored information and the initial information, and based on a second degree of difference between the initial information and a second training sample; the second training sample and The first training sample corresponds to;

performing reverse conduction according to the second loss function to update the first preset model and the second preset model.
The method according to claim 28 or 29, wherein the terminal device uses training samples to jointly train the first preset model and the second preset model, comprising:

The terminal device inputs the first training sample into the estimated preset sub-model of the first preset model, and obtains initial information output by the estimated preset sub-model;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

inputting the compressed feature vector information into the second preset model to obtain restored feature vector information output by the second preset model;

determining a second loss function based on a first degree of difference between the restored feature vector information and the feature vector information of the initial information, and based on a second degree of difference between the initial information and a second training sample; The second training sample corresponds to the first training sample;

performing reverse conduction according to the second loss function to update the first preset model and the second preset model.
The method according to claim 28 or 29, wherein the terminal device uses training samples to jointly train the first preset model and the second preset model to obtain the trained first model and the second preset model. models, including:

The terminal device uses training samples to jointly train the first preset model, the second preset model and the third preset model to obtain the trained first model, the second model and the third model;

Wherein, the third model is a trained third preset model.
The method according to claim 32, wherein the terminal device uses training samples to jointly train the first preset model, the second preset model and the third preset model, comprising:

The terminal device inputs the first training sample into the estimated preset sub-model of the first preset model, and obtains initial information output by the estimated preset sub-model;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into a third preset model of the preset model to obtain transformed information output by the third preset model;

inputting the transformed information into the second preset model to obtain restoration information output by the second preset model;

Determine a second loss function based on a first degree of difference between the restored information and the initial information, and based on a second degree of difference between the initial information and a second training sample; the second training sample and The first training sample corresponds to;

performing reverse conduction according to the second loss function to update the first preset model, the second preset model, and the third preset model.
The method according to claim 32, wherein the terminal device uses training samples to jointly train the first preset model, the second preset model and the third preset model, comprising:

The terminal device inputs the first training sample into the estimated preset sub-model of the first preset model, and obtains initial information output by the estimated preset sub-model;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

inputting the compressed feature vector information into a third preset model among the preset models, to obtain transformed feature vector information output by the third preset model;

inputting the transformed eigenvector information into the second preset model to obtain restored eigenvector information output by the second preset model;

determining a second loss function based on a first degree of difference between the restored feature vector information and the feature vector information of the initial information, and based on a second degree of difference between the initial information and a second training sample; The second training sample corresponds to the first training sample;

performing reverse conduction according to the second loss function to update the first preset model, the second preset model, and the third preset model.
The method according to any one of claims 20-34, wherein the training samples include a first training sample.
The method of claim 35, wherein,

The first training samples are distributed in the first dimension and/or the second dimension.
The method of claim 35, wherein,

The first dimension is a time domain dimension; the first training samples include first information samples distributed in m time units in the time domain dimension; m is a positive integer.
The method of claim 35, wherein,

The second dimension is a frequency domain dimension; the first training samples include first information samples distributed on x frequency domain resources in the frequency domain dimension; x is a positive integer.
The method according to any one of claims 36-38, wherein the first training samples are also distributed in a third dimension;

The third dimension is a complex dimension; the first training samples include the real part of the first information sample and the imaginary part of the first information sample.
The method according to any one of claims 35-39, wherein the training samples further include a second training sample corresponding to the first training sample;

The second training sample is composed of a matrix of T dimensions; T is an integer greater than or equal to 2.
The method of claim 40, wherein,

The T dimensions include a fourth dimension and a fifth dimension.
The method of claim 41, wherein,

The matrix of the T dimensions is a two-dimensional matrix of M×N; wherein, M represents the quantity of the first granularity under the fourth dimension, and N represents the quantity of the second granularity under the fifth dimension; both M and N are is a positive integer.
The method of claim 42, wherein,

The value at the ijth position in the two-dimensional matrix is used to represent the channel quality at the i-th first granularity in the fourth dimension and the j-th second granularity in the fifth dimension; both i and j are is a positive integer.
The method according to any one of claims 41-43, wherein the T dimensions further include a sixth dimension.
The method of claim 44, wherein,

The matrix of T dimensions is a three-dimensional matrix of M×N×W; wherein, M represents the number of first granularities in the fourth dimension, N represents the number of second granularities in the fifth dimension, and W represents the number of granularities in the fifth dimension. The quantity of the third granularity under the sixth dimension; M, N and W are all positive integers.
The method of claim 45, wherein,

The value of the ijkth position in the three-dimensional matrix is used to represent the i-th first granularity in the fourth dimension, the j-th second granularity in the fifth dimension, and the k-th in the sixth dimension channel quality at the third granularity; i, j and k are all positive integers.
The method according to any one of claims 41-46, wherein,

The fourth dimension is a frequency domain dimension; the first granularity includes one of the following: L1 resource blocks RB, L2 subcarriers; L1 and L2 are positive integers.
The method according to any one of claims 41-46, wherein,

The fourth dimension is a time domain dimension; the first granularity includes one of the following: K1 microseconds, K2 symbol length, K3 number of sampling points of symbols; K1, K2 and K3 are positive integers.
The method of claim 48, wherein the symbols are OFDM symbols.
The method according to any one of claims 41-49, wherein the fifth dimension is a spatial domain dimension;

The second granularity is an interval between a pair of transmitting and receiving antennas or an angle of arrival.
The method according to any one of claims 44-50, wherein,

The sixth dimension is a complex dimension; the third granularity is 1, and the quantity W of the third granularity under the complex dimension is 2.
The method according to claim 51, wherein, when the k is the first value, the value of the ijkth position in the three-dimensional matrix is used to represent the ith first value in the fourth dimension Granularity, the real part of the channel quality at the jth second granularity of the fifth dimension;

When the k is the second value, the value of the ijkth position in the three-dimensional matrix is used to represent the i-th first granularity in the fourth dimension and the j-th granularity in the fifth dimension The imaginary part of the channel quality at the second granularity.
The method according to any one of claims 20-52, wherein the method further comprises:

The terminal device sends the second model.
The method according to any one of claims 20-53, wherein the method further comprises:

The terminal device sends the first model.
The method according to any one of claims 20-53, wherein the method further comprises:

The terminal device sends the estimated sub-model and the compressed sub-model in the first model.
The method according to any one of claims 20-53, wherein the method further comprises:

The terminal device sends the estimation sub-model, the compression sub-model and the channel generation sub-model in the first model.
The method according to any one of claims 20-56, wherein the method further comprises:

The terminal device sends the third model.
The method according to any one of claims 53-57, wherein the first model is carried by at least one of the following: uplink control signaling, RRC messages, uplink data transmission, and uplink transmission requirements for artificial intelligence services data transmission;

The second model is carried by at least one of the following: uplink control signaling, RRC messages, uplink data transmission, and uplink data transmission for artificial intelligence business transmission requirements;

The third model is carried by at least one of the following: uplink control signaling, RRC messages, uplink data transmission, and uplink data transmission for artificial intelligence business transmission requirements;

The estimation sub-model is carried by at least one of the following: uplink control signaling, RRC message, uplink data transmission, uplink data transmission for artificial intelligence business type transmission requirements;

The compressed sub-model is carried by at least one of the following: uplink control signaling, RRC messages, uplink data transmission, and uplink data transmission for artificial intelligence business transmission requirements;

The channel generation sub-model is carried by at least one of the following: uplink control signaling, RRC messages, uplink data transmission, and uplink data transmission for the transmission requirements of artificial intelligence services.
An information processing method, comprising:

The network device sends the first information;

The network device receives second information; wherein, the second information is obtained by processing the first information through a first model;

The network device processes the second information based on a second model to obtain channel information; wherein, the first model and the second model are obtained through joint training.
The method of claim 59, wherein the second information is channel compression information;

The second model is used to decompress the channel compressed information to obtain channel information.
The method according to claim 60, wherein the network device processes the second information based on a second model to obtain channel information, comprising:

The network device inputs the channel compression information into the second model to obtain the channel information output by the second model.
The method according to claim 59, wherein the second information is channel compression information; the channel compression information includes eigenvector information of compressed channel estimation information; the channel information is eigenvector information of channel information;

The second model is used to decompress the compressed eigenvector information of the channel estimation information to obtain the eigenvector information of the channel information.
The method of claim 62, wherein,

The eigenvector information of the channel information includes R groups of eigenvector sequence information; R is a positive integer.
The method according to claim 62 or 63, wherein the network device processes the second information based on a second model to obtain channel information, including:

The network device inputs the eigenvector information of the compressed channel estimation information into the second model, and obtains the eigenvector information of the channel information output by the second model.
The method according to any one of claims 59-64, wherein the first information is a reference signal.
The method according to any one of claims 59-65, wherein the method further comprises:

The network device receives the second model.
The method of claim 66, wherein,

The second model is carried by at least one of the following: uplink control signaling, RRC message, uplink data transmission, and uplink data transmission for the transmission requirements of artificial intelligence services.
The method according to claim 66 or 67, wherein the method further comprises:

The network device receives the first model.
The method of claim 68, wherein,

The first model is carried by at least one of the following: uplink control signaling, RRC message, uplink data transmission, and uplink data transmission for the transmission requirements of artificial intelligence services.
The method of claim 68, wherein the first model comprises: an estimation sub-model and a compression sub-model;

Wherein, the estimation sub-model is used to perform channel estimation on the first information to obtain channel estimation information;

The compression sub-model is used to compress the channel estimation information to obtain the second information.
The method of claim 68, wherein the first model comprises: an estimation submodel, a channel generation submodel, and a compression submodel;

Wherein, the estimation sub-model is used to perform channel estimation on the first information to obtain channel estimation information;

The channel generation sub-model is used to perform eigendecomposition on the channel estimation information to obtain eigenvector information of the channel estimation information;

The compression sub-model is used to compress the eigenvector information of the channel estimation information to obtain the second information.
The method according to claim 66 or 67, wherein the method further comprises:

The network device receives the estimated sub-model and the compressed sub-model;

The network device generates the first model based on the estimated sub-model and the compressed sub-model.
The method according to claim 72, wherein the estimation sub-model is used to perform channel estimation on the first information to obtain channel estimation information;

The compression sub-model is used to compress the channel estimation information to obtain the second information.
The method according to claim 66 or 67, wherein the method further comprises:

The network device receives the estimation sub-model, the compression sub-model, and the channel generation sub-model;

The network device generates the first model based on the estimation sub-model, the compression sub-model and the channel generation sub-model.
The method of claim 74, wherein,

The estimation sub-model is used to perform channel estimation on the first information to obtain channel estimation information;

The channel generation sub-model is used to perform eigendecomposition on the channel estimation information to obtain eigenvector information of the channel estimation information;

The compression sub-model is used to compress the eigenvector information of the channel estimation information to obtain the second information.
The method according to any one of claims 72-75, wherein the estimation sub-model is carried by one of the following: uplink control signaling, RRC message, uplink data transmission, uplink data for artificial intelligence service class transmission requirements transmission;

The compressed sub-model is carried by one of the following: uplink control signaling, RRC messages, uplink data transmission, and uplink data transmission for artificial intelligence business transmission requirements;

The channel generation sub-model is carried by one of the following: uplink control signaling, RRC message, uplink data transmission, and uplink data transmission for the transmission requirements of artificial intelligence services.
The method according to any one of claims 66-76, wherein the method further comprises:

The network device receives a third model.
The method according to claim 77, wherein the third model is used to input the second information output by the first model into the second model after data conversion processing;

The first model, the second model and the third model are obtained through joint training.
The method according to claim 78, wherein said data transformation processing comprises: convolution processing or Fourier transform processing.
The method according to any one of claims 59-65, wherein the method further comprises:

The network device uses training samples to jointly train the first preset model and the second preset model to obtain the trained first model and the second model;

Wherein, the first model is the first preset model after training, and the second model is the second preset model after training.
The method according to claim 80, wherein the loss function used in the training is a first loss function;

The first loss function is constructed based on the degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model.
The method of claim 81, wherein the degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model is determined based on distance, or determined on the basis of similarity.
The method according to claim 81 or 82, wherein the network device uses training samples to jointly train the first preset model and the second preset model, comprising:

The network device inputs the first training sample into the estimated preset sub-model of the first preset model, and obtains initial information output by the estimated preset sub-model;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into the second preset model to obtain restoration information output by the second preset model;

determining the first loss function based on the degree of difference between the restoration information and the initial information;

performing reverse conduction according to the first loss function to update the first preset model and the second preset model.
The method according to claim 81 or 82, wherein the network device uses training samples to jointly train the first preset model and the second preset model, comprising:

The network device inputs the first training sample into the estimated preset sub-model of the first preset model, and obtains initial information output by the estimated preset sub-model;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Input the feature vector information of the initial information into the compressed preset sub-model of the first preset model, and obtain the compressed feature vector information output by the compressed preset sub-model;

inputting the compressed feature vector information into the second preset model to obtain restored feature vector information output by the second preset model;

determining a first loss function based on the degree of difference between the restored eigenvector information and the eigenvector information of the initial information;

performing reverse conduction according to the first loss function to update the first preset model and the second preset model.
The method according to claim 81 or 82, wherein the network device uses training samples to jointly train the first preset model and the second preset model to obtain the trained first model and the second preset model. models, including:

The network device uses training samples to jointly train the first preset model, the second preset model and the third preset model to obtain the trained first model, the second model and the third model; wherein , the third model is a trained third preset model.
The method according to claim 85, wherein the network device uses training samples to jointly train the first preset model, the second preset model and the third preset model, comprising:

The network device inputs the first training sample into the estimated preset sub-model of the first preset model, and obtains initial information output by the estimated preset sub-model;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into a third preset model to obtain transformed information output by the third preset model;

inputting the transformed information into the second preset model to obtain restoration information output by the second preset model;

determining a first loss function based on the degree of difference between the restoration information and the initial information;

performing reverse conduction according to the first loss function to update the first preset model, the second preset model, and the third preset model.
The method according to claim 85, wherein the network device uses training samples to jointly train the first preset model, the second preset model and the third preset model, comprising:

The network device inputs the first training sample into the estimated preset sub-model of the first preset model, and obtains initial information output by the estimated preset sub-model;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

Inputting the compressed feature vector information into a third preset model to obtain transformed feature vector information output by the third preset model;

inputting the transformed eigenvector information into the second preset model to obtain restored eigenvector information output by the second preset model;

determining a first loss function based on the degree of difference between the restored eigenvector information and the eigenvector information of the initial information;

performing reverse conduction according to the first loss function to update the first preset model, the second preset model, and the third preset model.
The method according to claim 80, wherein the loss function used in the training is a second loss function;

The second loss function is based on the first degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model and the first preset model The second difference degree between the output information of the estimated preset sub-model and the second training sample is constructed; wherein, the second training sample corresponds to the first training sample input into the estimated preset sub-model.
The method of claim 88, wherein the first degree of difference is determined based on a distance, or is determined based on a degree of similarity; and/or,

The second degree of difference is determined based on distance, or determined based on similarity.
The method according to claim 88 or 89, wherein the network device uses training samples to jointly train the first preset model and the second preset model, comprising:

The network device inputs the first training sample into the estimated preset sub-model of the first preset model, and obtains initial information output by the estimated preset sub-model;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into the second preset model to obtain restoration information output by the second preset model;

Determine a second loss function based on a first degree of difference between the restored information and the initial information, and based on a second degree of difference between the initial information and a second training sample; the second training sample and The first training sample corresponds to;

performing reverse conduction according to the second loss function to update the first preset model and the second preset model.
The method according to claim 88 or 89, wherein the network device uses training samples to jointly train the first preset model and the second preset model, comprising:

The network device inputs the first training sample into the estimated preset sub-model of the first preset model, and obtains initial information output by the estimated preset sub-model;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

inputting the compressed feature vector information into the second preset model to obtain restored feature vector information output by the second preset model;

determining a second loss function based on a first degree of difference between the restored feature vector information and the feature vector information of the initial information, and based on a second degree of difference between the initial information and a second training sample; The second training sample corresponds to the first training sample;

performing reverse conduction according to the second loss function to update the first preset model and the second preset model.
The method according to claim 88 or 89, wherein the network device uses training samples to jointly train the first preset model and the second preset model to obtain the trained first model and the second preset model. models, including:

The network device uses training samples to jointly train the first preset model, the second preset model and the third preset model to obtain the trained first model, the second model and the third model; wherein , the third model is a trained third preset model.
The method according to claim 92, wherein the network device uses training samples to jointly train the first preset model, the second preset model and the third preset model, comprising:

The network device inputs the first training sample into the estimated preset sub-model of the first preset model, and obtains initial information output by the estimated preset sub-model;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into a third preset model to obtain transformed information output by the third preset model;

inputting the transformed information into the second preset model to obtain restoration information output by the second preset model;

Determine a second loss function based on a first degree of difference between the restored information and the initial information, and based on a second degree of difference between the initial information and a second training sample; the second training sample and The first training sample corresponds to;

performing reverse conduction according to the second loss function to update the first preset model, the second preset model, and the third preset model.
The method according to claim 92, wherein the network device uses training samples to jointly train the first preset model, the second preset model and the third preset model, comprising:

The network device inputs the first training sample into the estimated preset sub-model of the first preset model, and obtains initial information output by the estimated preset sub-model;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

Inputting the compressed feature vector information into a third preset model to obtain transformed feature vector information output by the third preset model;

inputting the transformed eigenvector information into the second preset model to obtain restored eigenvector information output by the second preset model;

determining a second loss function based on a first degree of difference between the restored feature vector information and the feature vector information of the initial information, and based on a second degree of difference between the initial information and a second training sample; The second training sample corresponds to the first training sample;

performing reverse conduction according to the second loss function to update the first preset model, the second preset model, and the third preset model.
The method according to any one of claims 80-94, wherein said training samples include a first training sample.
The method of claim 95, wherein,

The first training samples are distributed in the first dimension and/or the second dimension.
The method of claim 96, wherein,

The first dimension is a time domain dimension;

The first training samples include first information samples distributed in m time units in the time domain dimension; m is a positive integer.
The method of claim 96, wherein,

The second dimension is a frequency domain dimension;

The first training samples include first information samples distributed on x frequency domain resources in the frequency domain dimension; x is a positive integer.
The method according to any one of claims 96-98, wherein said first training samples are also distributed in a third dimension;

The third dimension is a complex dimension; the first training samples include the real part of the first information sample and the imaginary part of the first information sample.
The method according to any one of claims 95-99, wherein the training samples further include a second training sample corresponding to the first training sample;

The second training sample is composed of a matrix of T dimensions; T is an integer greater than or equal to 2.
The method of claim 100, wherein,

The T dimensions include a fourth dimension and a fifth dimension.
The method of claim 101, wherein,

The matrix of the T dimensions is a two-dimensional matrix of M×N; wherein, M represents the quantity of the first granularity under the fourth dimension, and N represents the quantity of the second granularity under the fifth dimension; both M and N are is a positive integer.
The method of claim 102, wherein,

The value at the ijth position in the two-dimensional matrix is used to represent the channel quality at the i-th first granularity in the fourth dimension and the j-th second granularity in the fifth dimension; both i and j are is a positive integer.
The method according to any one of claims 101-103, wherein the T dimensions further include a sixth dimension.
The method of claim 104, wherein,

The matrix of T dimensions is a three-dimensional matrix of M×N×W; wherein, M represents the number of first granularities in the fourth dimension, N represents the number of second granularities in the fifth dimension, and W represents the number of granularities in the fifth dimension. The quantity of the third granularity under the sixth dimension; M, N and W are all positive integers.
The method of claim 105, wherein,

The value of the ijkth position in the three-dimensional matrix is used to represent the i-th first granularity in the fourth dimension, the j-th second granularity in the fifth dimension, and the k-th in the sixth dimension channel quality at the third granularity; i, j and k are all positive integers.
The method according to any one of claims 101-106, wherein,

The fourth dimension is a frequency domain dimension; the first granularity includes one of the following: L1 resource blocks RB, L2 subcarriers; L1 and L2 are positive integers.
The method according to any one of claims 101-106, wherein,

The fourth dimension is a time domain dimension; the first granularity includes one of the following: K1 microseconds, K2 symbol length, K3 number of sampling points of symbols; K1, K2 and K3 are positive integers.
The method of claim 108, wherein the symbols are OFDM symbols.
The method according to any one of claims 101-109, wherein the fifth dimension is a spatial domain dimension;

The second granularity is an interval between a pair of transmitting and receiving antennas or an angle of arrival.
The method according to any one of claims 104-110, wherein,

The sixth dimension is a complex dimension; the third granularity is 1, and the quantity W of the third granularity under the complex dimension is 2.
The method according to claim 111, wherein, when the k is the first value, the value of the ijkth position in the three-dimensional matrix is used to represent the ith first in the fourth dimension Granularity, the real part of the channel quality at the jth second granularity of the fifth dimension;

When the k is the second value, the value of the ijkth position in the three-dimensional matrix is used to represent the i-th first granularity in the fourth dimension and the j-th granularity in the fifth dimension The imaginary part of the channel quality at the second granularity.
The method according to any one of claims 80-112, wherein the method further comprises:

The network device sends the second model.
The method according to any one of claims 80-113, wherein the method further comprises:

The network device sends the first model.
The method according to any one of claims 80-113, wherein the method further comprises:

The network device sends the estimated sub-model and the compressed sub-model.
The method according to any one of claims 80-113, wherein the method further comprises:

The network device sends the estimation sub-model, the compression sub-model and the channel generation sub-model.
The method according to any one of claims 80-116, wherein the method further comprises:

The network device sends the third model.
The method according to any one of claims 113-117, wherein the first model is carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, artificial intelligence Downlink data transmission required for similar business transmission;

The second model is carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements;

The third model is carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements;

The estimated sub-model is carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements;

The compressed sub-model is carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements;

The channel generation sub-model is carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements.
A method of model generation comprising:

performing joint training on the first preset model and the second preset model by using the training samples to obtain the trained first model and the second model;

Wherein, the first model is the first preset model after training, and the second model is the second preset model after training; the first model is used to process the first information to obtain Second information; the second model is used to process the second information to obtain channel information.
The method according to claim 119, wherein the loss function used in the training is a first loss function;

The first loss function is constructed based on the degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model.
The method according to claim 120, wherein the degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model is determined based on a distance, or is determined based on a degree of similarity .
The method according to claim 120 or 121, wherein the joint training of the first preset model and the second preset model using training samples comprises:

inputting the first training sample into the estimated preset sub-model of the first preset model, and obtaining initial information output by the estimated preset sub-model;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into the second preset model to obtain restoration information output by the second preset model;

determining the first loss function based on the degree of difference between the restoration information and the initial information;

performing reverse conduction according to the first loss function to update the first preset model and the second preset model.
The method according to claim 120 or 121, wherein the joint training of the first preset model and the second preset model using training samples comprises:

inputting the first training sample into the estimated preset sub-model of the first preset model, and obtaining initial information output by the estimated preset sub-model;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

inputting the compressed feature vector information into the second preset model to obtain restored feature vector information output by the second preset model;

determining a first loss function based on the degree of difference between the restored eigenvector information and the eigenvector information of the initial information;

performing reverse conduction according to the first loss function to update the first preset model and the second preset model.
The method according to claim 120 or 121, wherein said training samples are used to jointly train the first preset model and the second preset model to obtain the trained first model and the second model, include:

Using training samples to jointly train the first preset model, the second preset model and the third preset model to obtain the trained first model, the second model and the third model; wherein, the first preset model The three models are the third preset models after training.
The method according to claim 124, wherein the joint training of the first preset model, the second preset model and the third preset model using training samples comprises:

inputting the first training sample into the estimated preset sub-model of the first preset model, and obtaining initial information output by the estimated preset sub-model;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into a third preset model to obtain transformed information output by the third preset model;

inputting the transformed information into the second preset model to obtain restoration information output by the second preset model;

determining a first loss function based on the degree of difference between the restoration information and the initial information;

performing reverse conduction according to the first loss function to update the first preset model, the second preset model and the third preset model.
The method according to claim 124, wherein the joint training of the first preset model, the second preset model and the third preset model using training samples comprises:

inputting the first training sample into the estimated preset sub-model of the first preset model, and obtaining initial information output by the estimated preset sub-model;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

Inputting the compressed feature vector information into a third preset model to obtain transformed feature vector information output by the third preset model;

inputting the transformed eigenvector information into the second preset model to obtain restored eigenvector information output by the second preset model;

determining a first loss function based on the degree of difference between the restored eigenvector information and the eigenvector information of the initial information;

performing reverse conduction according to the first loss function to update the first preset model, the second preset model and the third preset model.
The method according to claim 119, wherein the loss function used in the training is a second loss function;

The second loss function is based on the first degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model and the first preset model The second difference degree between the output information of the estimated preset sub-model and the second training sample is constructed; wherein, the second training sample corresponds to the first training sample input into the estimated preset sub-model.
The method of claim 127, wherein the first degree of difference is determined based on a distance, or determined based on a degree of similarity; and/or,

The second degree of difference is determined based on distance, or determined based on similarity.
The method according to claim 127 or 128, wherein the joint training of the first preset model, the second preset model and the third preset model using training samples comprises:

inputting the first training sample into the estimated preset sub-model of the first preset model, and obtaining initial information output by the estimated preset sub-model;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into the second preset model to obtain restoration information output by the second preset model;

Determine a second loss function based on a first degree of difference between the restored information and the initial information, and based on a second degree of difference between the initial information and a second training sample; the second training sample and The first training sample corresponds to;

performing reverse conduction according to the second loss function to update the first preset model and the second preset model.
The method according to claim 127 or 128, wherein the joint training of the first preset model and the second preset model using training samples comprises:

inputting the first training sample into the estimated preset sub-model of the first preset model, and obtaining initial information output by the estimated preset sub-model;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

inputting the compressed feature vector information into the second preset model to obtain restored feature vector information output by the second preset model;

determining a second loss function based on a first degree of difference between the restored feature vector information and the feature vector information of the initial information, and based on a second degree of difference between the initial information and a second training sample; The second training sample corresponds to the first training sample;

performing reverse conduction according to the second loss function to update the first preset model and the second preset model.
The method according to claim 127 or 128, wherein said training samples are used to jointly train the first preset model and the second preset model to obtain the trained first model and the second model, include:

Using training samples to jointly train the first preset model, the second preset model and the third preset model to obtain the trained first model, the second model and the third model; wherein, the first preset model The three models are the third preset models after training.
The method according to claim 131, wherein the joint training of the first preset model, the second preset model and the third preset model using training samples comprises:

inputting the first training sample into the estimated preset sub-model of the first preset model, and obtaining initial information output by the estimated preset sub-model;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into a third preset model to obtain transformed information output by the third preset model;

inputting the transformed information into the second preset model to obtain restoration information output by the second preset model;

Determine a second loss function based on a first degree of difference between the restored information and the initial information, and based on a second degree of difference between the initial information and a second training sample; the second training sample and The first training sample corresponds to;

performing reverse conduction according to the second loss function to update the first preset model, the second preset model, and the third preset model.
The method according to claim 131, wherein the joint training of the first preset model, the second preset model and the third preset model using training samples comprises:

inputting the first training sample into the estimated preset sub-model of the first preset model, and obtaining initial information output by the estimated preset sub-model;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

Inputting the compressed feature vector information into a third preset model to obtain transformed feature vector information output by the third preset model;

inputting the transformed eigenvector information into the second preset model to obtain restored eigenvector information output by the second preset model;

determining a second loss function based on a first degree of difference between the restored feature vector information and the feature vector information of the initial information, and based on a second degree of difference between the initial information and a second training sample; The second training sample corresponds to the first training sample;

performing reverse conduction according to the second loss function to update the first preset model and the second preset model.
The method according to any one of claims 119-133, wherein the training samples include a first training sample.
The method of claim 134, wherein,

The first training samples are distributed in the first dimension and/or the second dimension.
The method of claim 134, wherein,

The first dimension is a time domain dimension; the first training samples include first information samples distributed in m time units in the time domain dimension; m is a positive integer.
The method of claim 134, wherein,

The second dimension is a frequency domain dimension; the first training samples include first information samples distributed on x frequency domain resources in the frequency domain dimension; x is a positive integer.
The method according to any one of claims 135-137, wherein said first training samples are also distributed in a third dimension;

The third dimension is a complex dimension; the first training samples include the real part of the first information sample and the imaginary part of the first information sample.
The method according to any one of claims 134-138, wherein the training samples further include a second training sample corresponding to the first training sample;

The second training sample is composed of a matrix of T dimensions; T is an integer greater than or equal to 2.
The method of claim 139, wherein,

The T dimensions include a fourth dimension and a fifth dimension.
The method of claim 140, wherein,

The matrix of the T dimensions is a two-dimensional matrix of M×N; wherein, M represents the quantity of the first granularity under the fourth dimension, and N represents the quantity of the second granularity under the fifth dimension; both M and N are is a positive integer.
The method of claim 141, wherein,

The value at the ijth position in the two-dimensional matrix is used to represent the channel quality at the i-th first granularity in the fourth dimension and the j-th second granularity in the fifth dimension; both i and j are is a positive integer.
The method according to any one of claims 140-142, wherein the T dimensions further include a sixth dimension.
The method of claim 143, wherein,

The matrix of T dimensions is a three-dimensional matrix of M×N×W; wherein, M represents the number of first granularities in the fourth dimension, N represents the number of second granularities in the fifth dimension, and W represents the number of granularities in the fifth dimension. The quantity of the third granularity under the sixth dimension; M, N and W are all positive integers.
The method of claim 144, wherein,

The value of the ijkth position in the three-dimensional matrix is used to represent the i-th first granularity in the fourth dimension, the j-th second granularity in the fifth dimension, and the k-th in the sixth dimension channel quality at the third granularity; i, j and k are all positive integers.
The method according to any one of claims 140-145, wherein,

The fourth dimension is a frequency domain dimension; the first granularity includes one of the following: L1 resource blocks RB, L2 subcarriers; L1 and L2 are positive integers.
The method according to any one of claims 140-145, wherein,

The fourth dimension is a time domain dimension; the first granularity includes one of the following: K1 microseconds, K2 symbol length, K3 number of sampling points of symbols; K1, K2 and K3 are positive integers.
The method of claim 147, wherein the symbols are OFDM symbols.
The method of any one of claims 140-148, wherein the fifth dimension is a spatial domain dimension;

The second granularity is an interval between a pair of transmitting and receiving antennas or an angle of arrival.
The method according to any one of claims 143-149, wherein,

The sixth dimension is a complex dimension; the third granularity is 1, and the quantity W of the third granularity under the complex dimension is 2.
The method according to claim 150, wherein, when the k is the first value, the value of the ijkth position in the three-dimensional matrix is used to represent the ith first in the fourth dimension Granularity, the real part of the channel quality at the jth second granularity of the fifth dimension;

When the k is the second value, the value of the ijkth position in the three-dimensional matrix is used to represent the i-th first granularity in the fourth dimension and the j-th granularity in the fifth dimension The imaginary part of the channel quality at the second granularity.
A terminal device comprising:

a first communication unit, configured to receive first information; send second information obtained based on the first information;

Wherein, the second information is obtained by processing the first information through the first model, and the second information is used for processing through the second model to obtain channel information; the first model and the second model are a joint obtained by training.
The terminal device according to claim 152, wherein the second information is channel compression information;

The first model is used to process the input first information to obtain channel compression information.
The terminal device according to claim 153, wherein the first model comprises: an estimation sub-model and a compression sub-model;

Wherein, the estimation sub-model is used to perform channel estimation based on the first information to obtain channel estimation information;

The compression sub-model is used to compress the channel estimation information to obtain channel compression information.
The terminal device according to claim 154, wherein the terminal device further comprises:

The first processing unit is configured to input the first information into the estimation sub-model to obtain channel estimation information output by the estimation sub-model; input the channel estimation information to the compression sub-model to obtain the compression sub-model Channel compression information for the model output.
The terminal device according to any one of claims 152-155, wherein the first information is a reference signal.
The terminal device according to claim 152, wherein the second information is channel compression information; the channel compression information includes eigenvector information of compressed channel estimation information;

The first model is used to process the input first information to obtain eigenvector information of compressed channel estimation information.
The terminal device according to claim 157, wherein the first model comprises: an estimation submodel, a channel generation submodel, and a compression submodel;

Wherein, the estimation sub-model is used to perform channel estimation on the first information to obtain channel estimation information;

The channel generation sub-model is used to perform eigendecomposition on the channel estimation information to obtain eigenvector information of the channel estimation information;

The compression sub-model is used to compress the eigenvector information of the channel estimation information to obtain compressed eigenvector information of the channel estimation information.
The terminal device according to claim 158, wherein,

The eigenvector information of the channel information includes R groups of eigenvector sequence information; R is a positive integer.
The terminal device according to claim 158 or 159, wherein the terminal device further comprises:

A first processing unit, configured to input the first information into the estimation sub-model to obtain channel estimation information output by the estimation sub-model; input the channel estimation information to the channel generation sub-model to obtain the channel generating eigenvector information of channel estimation information output by the sub-model; inputting the eigenvector information of the channel estimation information into the compression sub-model to obtain the eigenvector information of the compressed channel estimation information output by the compression sub-model.
The terminal device according to any one of claims 156-160, wherein the first information is a reference signal; and the channel information is eigenvector information of the channel information.
The terminal device according to any one of claims 152-161, wherein the first communication unit is configured to receive the first model.
The terminal device according to claim 162, wherein the first model is carried by at least one of the following: downlink control signaling, media access control MAC control element CE message, radio resource control RRC message, broadcast message, downlink data Transmission, downlink data transmission for artificial intelligence business transmission requirements.
The terminal device according to any one of claims 152-161, wherein the first communication unit is used for the terminal device to receive the estimated sub-model and the compressed sub-model;

The first processing unit is configured to generate the first model based on the estimated sub-model and the compressed sub-model.
The terminal device according to any one of claims 152-161, wherein the first communication unit is configured to receive the estimation sub-model, the compression sub-model and the channel generation sub-model;

The first processing unit is configured to generate the first model based on the estimation sub-model, the compression sub-model and the channel generation sub-model.
The terminal device according to claim 164 or 165, wherein the estimation sub-model is carried by one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, transmission for artificial intelligence services Required downlink data transmission;

The compressed sub-model is carried by one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements;

The channel generation sub-model is carried by one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements.
The terminal device according to any one of claims 162-166, wherein the first communication unit is configured to receive the second model.
The terminal device according to any one of claims 162-167, wherein the first communication unit is configured to receive the third model.
The terminal device according to claim 168, wherein the third model is used to input the second information output by the first model into the second model after data conversion processing;

The first model, the second model and the third model are obtained through joint training.
The terminal device according to claim 169, wherein said data transformation processing comprises convolution processing or Fourier transform processing.
The terminal device according to any one of claims 152-161, wherein the first processing unit is configured to use training samples to jointly train the first preset model and the second preset model to obtain all said first model and said second model;

Wherein, the first model is the first preset model after training, and the second model is the second preset model after training.
The terminal device according to claim 171, wherein the loss function used in the training is a first loss function;

The first loss function is constructed based on the degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model.
The terminal device according to claim 172, wherein the degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model is determined based on distance, Or determined based on the degree of similarity.
The terminal device according to claim 172 or 173, wherein the first processing unit is configured to input a first training sample into the estimated preset sub-model of the first preset model to obtain the estimated preset sub-model The initial information output by the model;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into the second preset model to obtain restoration information output by the second preset model;

determining the first loss function based on the degree of difference between the restoration information and the initial information;

performing reverse conduction according to the first loss function to update the first preset model and the second preset model.
The terminal device according to claim 172 or 173, wherein the first processing unit is configured to input a first training sample into the estimated preset sub-model of the first preset model to obtain the estimated preset sub-model The initial information output by the model;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

inputting the compressed feature vector information into the second preset model to obtain restored feature vector information output by the second preset model;

determining a first loss function based on the degree of difference between the restored eigenvector information and the eigenvector information of the initial information;

performing reverse conduction according to the first loss function to update the first preset model and the second preset model.
The terminal device according to claim 172 or 176, wherein the first processing unit is configured to use training samples to jointly train the first preset model, the second preset model, and the third preset model to obtain the training the first model, the second model and the third model after;

Wherein, the third model is a trained third preset model.
The terminal device according to claim 176, wherein the first processing unit is configured to input a first training sample into an estimated preset sub-model of the first preset model, and obtain an output of the estimated preset sub-model the initial information;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into a third preset model to obtain transformed information output by the third preset model;

inputting the transformed information into the second preset model to obtain restoration information output by the second preset model;

determining a first loss function based on the degree of difference between the restoration information and the initial information;

performing reverse conduction according to the first loss function to update the first preset model, the second preset model and the third preset model.
The terminal device according to claim 176, wherein the first processing unit is configured to input a first training sample into an estimated preset sub-model of the first preset model, and obtain an output of the estimated preset sub-model the initial information;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

Inputting the compressed feature vector information into a third preset model to obtain transformed feature vector information output by the third preset model;

inputting the transformed eigenvector information into the second preset model to obtain restored eigenvector information output by the second preset model;

determining a first loss function based on the degree of difference between the restored eigenvector information and the eigenvector information of the initial information;

performing reverse conduction according to the first loss function to update the first preset model, the second preset model and the third preset model.
The terminal device according to claim 171, wherein the loss function used in the training is a second loss function;

The second loss function is based on a first degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model, and the first preset The second degree of difference between the output information of the estimated preset sub-model of the model and the second training sample is constructed; wherein the second training sample corresponds to the first training sample input to the estimated preset sub-model.
The terminal device according to claim 179, wherein the first degree of difference is determined based on a distance, or determined based on a degree of similarity; and/or,

The second degree of difference is determined based on distance, or determined based on similarity.
The terminal device according to claim 179 or 180, wherein the first processing unit is configured to input a first training sample into the estimated preset sub-model of the first preset model to obtain the estimated preset sub-model The initial information output by the model;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into the second preset model to obtain restoration information output by the second preset model;

Determine a second loss function based on a first degree of difference between the restored information and the initial information, and based on a second degree of difference between the initial information and a second training sample; the second training sample and The first training sample corresponds to;

performing reverse conduction according to the second loss function to update the first preset model and the second preset model.
The terminal device according to claim 179 or 180, wherein the first processing unit is configured to input a first training sample into the estimated preset sub-model of the first preset model to obtain the estimated preset sub-model The initial information output by the model;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

inputting the compressed feature vector information into the second preset model to obtain restored feature vector information output by the second preset model;

determining a second loss function based on a first degree of difference between the restored feature vector information and the feature vector information of the initial information, and based on a second degree of difference between the initial information and a second training sample; The second training sample corresponds to the first training sample;

performing reverse conduction according to the second loss function to update the first preset model and the second preset model.
The terminal device according to claim 179 or 180, wherein the first processing unit is configured to use training samples to jointly train the first preset model, the second preset model, and the third preset model to obtain the training the first model, the second model and the third model after;

Wherein, the third model is a trained third preset model.
The terminal device according to claim 183, wherein the first processing unit is configured to input a first training sample into an estimated preset sub-model of the first preset model, and obtain an output of the estimated preset sub-model the initial information;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into a third preset model of the preset model to obtain transformed information output by the third preset model;

inputting the transformed information into the second preset model to obtain restoration information output by the second preset model;

Determine a second loss function based on a first degree of difference between the restored information and the initial information, and based on a second degree of difference between the initial information and a second training sample; the second training sample and The first training sample corresponds to;

performing reverse conduction according to the second loss function to update the first preset model, the second preset model, and the third preset model.
The terminal device according to claim 183, wherein the first processing unit is configured to input a first training sample into an estimated preset sub-model of the first preset model, and obtain an output of the estimated preset sub-model the initial information;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

inputting the compressed feature vector information into a third preset model among the preset models, to obtain transformed feature vector information output by the third preset model;

inputting the transformed eigenvector information into the second preset model to obtain restored eigenvector information output by the second preset model;

determining a second loss function based on a first degree of difference between the restored feature vector information and the feature vector information of the initial information, and based on a second degree of difference between the initial information and a second training sample; The second training sample corresponds to the first training sample;

performing reverse conduction according to the second loss function to update the first preset model, the second preset model, and the third preset model.
The terminal device according to any one of claims 171-185, wherein the training samples include a first training sample.
The terminal device according to claim 186, wherein,

The first training samples are distributed in the first dimension and/or the second dimension.
The terminal device according to claim 187, wherein,

The first dimension is a time domain dimension; the first training samples include first information samples distributed in m time units in the time domain dimension; m is a positive integer.
The terminal device according to claim 187, wherein,

The second dimension is a frequency domain dimension; the first training samples include first information samples distributed on x frequency domain resources in the frequency domain dimension; x is a positive integer.
The terminal device according to any one of claims 187-189, wherein the first training samples are also distributed in a third dimension;

The third dimension is a complex dimension; the first training samples include the real part of the first information sample and the imaginary part of the first information sample.
The terminal device according to any one of claims 186-190, wherein the training samples further include a second training sample corresponding to the first training sample;

The second training sample is composed of a matrix of T dimensions; T is an integer greater than or equal to 2.
The terminal device according to claim 191, wherein,

The T dimensions include a fourth dimension and a fifth dimension.
The terminal device according to claim 192, wherein,

The matrix of the T dimensions is a two-dimensional matrix of M×N; wherein, M represents the quantity of the first granularity under the fourth dimension, and N represents the quantity of the second granularity under the fifth dimension; both M and N are is a positive integer.
The terminal device according to claim 193, wherein,

The value at the ijth position in the two-dimensional matrix is used to represent the channel quality at the i-th first granularity in the fourth dimension and the j-th second granularity in the fifth dimension; both i and j are is a positive integer.
The terminal device according to any one of claims 192-194, wherein the T dimensions further include a sixth dimension.
The terminal device according to claim 195, wherein,

The matrix of T dimensions is a three-dimensional matrix of M×N×W; wherein, M represents the number of first granularities in the fourth dimension, N represents the number of second granularities in the fifth dimension, and W represents the number of granularities in the fifth dimension. The quantity of the third granularity under the sixth dimension; M, N and W are all positive integers.
The terminal device according to claim 196, wherein,

The value of the ijkth position in the three-dimensional matrix is used to represent the i-th first granularity in the fourth dimension, the j-th second granularity in the fifth dimension, and the k-th in the sixth dimension channel quality at the third granularity; i, j and k are all positive integers.
The terminal device according to any one of claims 192-197, wherein,

The fourth dimension is a frequency domain dimension; the first granularity includes one of the following: L1 resource blocks RB, L2 subcarriers; L1 and L2 are positive integers.
The terminal device according to any one of claims 192-197, wherein,

The fourth dimension is a time domain dimension; the first granularity includes one of the following: K1 microseconds, K2 symbol length, K3 number of sampling points of symbols; K1, K2 and K3 are positive integers.
The terminal device of claim 199, wherein the symbols are OFDM symbols.
The terminal device according to any one of claims 192-200, wherein the fifth dimension is a spatial domain dimension;

The second granularity is an interval between a pair of transmitting and receiving antennas or an angle of arrival.
The terminal device according to any one of claims 195-201, wherein,

The sixth dimension is a complex dimension; the third granularity is 1, and the quantity W of the third granularity under the complex dimension is 2.
The terminal device according to claim 202, wherein when the k is the first value, the value of the ijk-th position in the three-dimensional matrix is used to represent the i-th position in the fourth dimension A granularity, the real part of the channel quality at the jth second granularity of the fifth dimension;

When the k is the second value, the value of the ijkth position in the three-dimensional matrix is used to represent the i-th first granularity in the fourth dimension and the j-th granularity in the fifth dimension The imaginary part of the channel quality at the second granularity.
The terminal device according to any one of claims 171-203, wherein the first communication unit is configured to send the second model.
The terminal device according to any one of claims 171-204, wherein the first communication unit is configured to send the first model.
The terminal device according to any one of claims 171-204, wherein the first communication unit is configured to send the estimation sub-model and the compression sub-model in the first model.
The terminal device according to any one of claims 171-204, wherein the first communication unit is configured to send the estimation sub-model, the compression sub-model and the channel generation sub-model in the first model.
The terminal device according to any one of claims 171-207, wherein the first communication unit is configured to send the third model.
The terminal device according to any one of claims 204-208, wherein the first model is carried by at least one of the following: uplink control signaling, RRC messages, uplink data transmission, and transmission requirements for artificial intelligence services Uplink data transmission;

The second model is carried by at least one of the following: uplink control signaling, RRC messages, uplink data transmission, and uplink data transmission for artificial intelligence business transmission requirements;

The third model is carried by at least one of the following: uplink control signaling, RRC messages, uplink data transmission, and uplink data transmission for artificial intelligence business transmission requirements;

The estimation sub-model is carried by at least one of the following: uplink control signaling, RRC message, uplink data transmission, uplink data transmission for artificial intelligence business type transmission requirements;

The compressed sub-model is carried by at least one of the following: uplink control signaling, RRC messages, uplink data transmission, and uplink data transmission for artificial intelligence business transmission requirements;

The channel generation sub-model is carried by at least one of the following: uplink control signaling, RRC messages, uplink data transmission, and uplink data transmission for the transmission requirements of artificial intelligence services.
A network device comprising:

The second communication unit is configured to send the first information; receive the second information; wherein, the second information is obtained by processing the first information through the first model;

The second processing unit is configured to process the second information based on a second model to obtain channel information; wherein, the first model and the second model are obtained through joint training.
The network device according to claim 210, wherein the second information is channel compression information;

The second model is used to decompress the channel compressed information to obtain channel information.
The network device according to claim 211, wherein the second processing unit is configured to input the channel compression information into the second model, and obtain the channel information output by the second model.
The network device according to claim 210, wherein the second information is channel compression information; the channel compression information includes eigenvector information of compressed channel estimation information; the channel information is eigenvector information of channel information;

The second model is used to decompress the compressed eigenvector information of the channel estimation information to obtain the eigenvector information of the channel information.
The network device of claim 213, wherein:

The eigenvector information of the channel information includes R groups of eigenvector sequence information; R is a positive integer.
The network device according to claim 213 or 214, wherein the second processing unit is configured to input the eigenvector information of the compressed channel estimation information into the second model, and obtain the output of the second model Eigenvector information of the channel information.
The network device according to any one of claims 210-215, wherein the first information is a reference signal.
The network device according to any one of claims 210-216, wherein the second communication unit is configured to receive the second model.
The network device of claim 217, wherein,

The second model is carried by at least one of the following: uplink control signaling, RRC message, uplink data transmission, and uplink data transmission for the transmission requirements of artificial intelligence services.
The network device according to claim 217 or 218, wherein the second communication unit is configured to receive the first model.
The network device of claim 219, wherein,

The first model is carried by at least one of the following: uplink control signaling, RRC message, uplink data transmission, and uplink data transmission for the transmission requirements of artificial intelligence services.
The network device of claim 219, wherein the first model comprises: an estimation sub-model and a compression sub-model;

Wherein, the estimation sub-model is used to perform channel estimation on the first information to obtain channel estimation information;

The compression sub-model is used to compress the channel estimation information to obtain the second information.
The network device according to claim 219, wherein the first model comprises: an estimation submodel, a channel generation submodel and a compression submodel;

Wherein, the estimation sub-model is used to perform channel estimation on the first information to obtain channel estimation information;

The channel generation sub-model is used to perform eigendecomposition on the channel estimation information to obtain eigenvector information of the channel estimation information;

The compression sub-model is used to compress the eigenvector information of the channel estimation information to obtain the second information.
The network device according to claim 217 or 218, wherein the second communication unit is configured to receive the estimated sub-model and the compressed sub-model;

The second processing unit is configured to generate the first model based on the estimated sub-model and the compressed sub-model.
The network device according to claim 223, wherein the estimation sub-model is used to perform channel estimation on the first information to obtain channel estimation information;

The compression sub-model is used to compress the channel estimation information to obtain the second information.
The network device according to claim 217 or 218, wherein the second communication unit is configured to receive the estimation sub-model, the compression sub-model, and the channel generation sub-model;

The second processing unit is configured to generate the first model based on the estimation sub-model, the compression sub-model and the channel generation sub-model.
The network device of claim 225, wherein,

The estimation sub-model is used to perform channel estimation on the first information to obtain channel estimation information;

The channel generation sub-model is used to perform eigendecomposition on the channel estimation information to obtain eigenvector information of the channel estimation information;

The compression sub-model is used to compress the eigenvector information of the channel estimation information to obtain the second information.
The network device according to any one of claims 223-226, wherein the estimation sub-model is carried by one of the following: uplink control signaling, RRC message, uplink data transmission, and uplink transmission requirements for artificial intelligence services data transmission;

The compressed sub-model is carried by one of the following: uplink control signaling, RRC messages, uplink data transmission, and uplink data transmission for artificial intelligence business transmission requirements;

The channel generation sub-model is carried by one of the following: uplink control signaling, RRC message, uplink data transmission, and uplink data transmission for the transmission requirements of artificial intelligence services.
The network device according to any one of claims 217-227, wherein the second communication unit is configured to receive the third model.
The network device according to claim 228, wherein the third model is used to input the second information output by the first model into the second model after data conversion processing;

The first model, the second model and the third model are obtained through joint training.
The network device according to claim 229, wherein said data transformation processing comprises: convolution processing or Fourier transform processing.
The network device according to any one of claims 210-216, wherein the second processing unit is configured to use training samples to jointly train the first preset model and the second preset model to obtain the trained said first model and said second model;

Wherein, the first model is the first preset model after training, and the second model is the second preset model after training.
The network device according to claim 231, wherein the loss function used in the training is a first loss function;

The first loss function is constructed based on the degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model.
The network device of claim 232, wherein the degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model is determined based on distance, Or determined based on the degree of similarity.
The network device according to claim 232 or 233, wherein the second processing unit is configured to input the first training sample into the estimated preset sub-model of the first preset model to obtain the estimated preset sub-model The initial information output by the model;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into the second preset model to obtain restoration information output by the second preset model;

determining the first loss function based on the degree of difference between the restoration information and the initial information;

performing reverse conduction according to the first loss function to update the first preset model and the second preset model.
The network device according to claim 232 or 233, wherein the second processing unit is configured to input the first training sample into the estimated preset sub-model of the first preset model to obtain the estimated preset sub-model The initial information output by the model;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

inputting the compressed feature vector information into the second preset model to obtain restored feature vector information output by the second preset model;

determining a first loss function based on the degree of difference between the restored eigenvector information and the eigenvector information of the initial information;

performing reverse conduction according to the first loss function to update the first preset model and the second preset model.
The network device according to claim 232 or 233, wherein the second processing unit is configured to use training samples to jointly train the first preset model, the second preset model, and the third preset model to obtain the training After the first model, the second model and the third model; wherein, the third model is the third preset model after training.
The network device according to claim 236, wherein the second processing unit is configured to input a first training sample into an estimated preset sub-model of the first preset model, and obtain an output of the estimated preset sub-model the initial information;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into a third preset model to obtain transformed information output by the third preset model;

inputting the transformed information into the second preset model to obtain restoration information output by the second preset model;

determining a first loss function based on the degree of difference between the restoration information and the initial information;

performing reverse conduction according to the first loss function to update the first preset model, the second preset model, and the third preset model.
The network device according to claim 236, wherein the second processing unit is configured to input a first training sample into an estimated preset sub-model of the first preset model, and obtain an output of the estimated preset sub-model the initial information;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

Inputting the compressed feature vector information into a third preset model to obtain transformed feature vector information output by the third preset model;

inputting the transformed eigenvector information into the second preset model to obtain restored eigenvector information output by the second preset model;

determining a first loss function based on the degree of difference between the restored eigenvector information and the eigenvector information of the initial information;

performing reverse conduction according to the first loss function to update the first preset model, the second preset model, and the third preset model.
The network device according to claim 231, wherein the loss function used in the training is a second loss function;

The second loss function is based on the first degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model and the first preset model The second difference degree between the output information of the estimated preset sub-model and the second training sample is constructed; wherein, the second training sample corresponds to the first training sample input into the estimated preset sub-model.
The network device according to claim 239, wherein the first degree of difference is determined based on a distance, or is determined based on a degree of similarity; and/or,

The second degree of difference is determined based on distance, or determined based on similarity.
The network device according to claim 239 or 240, wherein the second processing unit is configured to input the first training sample into the estimated preset sub-model of the first preset model to obtain the estimated preset sub-model The initial information output by the model;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into the second preset model to obtain restoration information output by the second preset model;

Determine a second loss function based on a first degree of difference between the restored information and the initial information, and based on a second degree of difference between the initial information and a second training sample; the second training sample and The first training sample corresponds to;

performing reverse conduction according to the second loss function to update the first preset model and the second preset model.
The network device according to claim 239 or 240, wherein the second processing unit is configured to input the first training sample into the estimated preset sub-model of the first preset model to obtain the estimated preset sub-model The initial information output by the model;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

inputting the compressed feature vector information into the second preset model to obtain restored feature vector information output by the second preset model;

determining a second loss function based on a first degree of difference between the restored feature vector information and the feature vector information of the initial information, and based on a second degree of difference between the initial information and a second training sample; The second training sample corresponds to the first training sample;

performing reverse conduction according to the second loss function to update the first preset model and the second preset model.
The network device according to claim 239 or 240, wherein the second processing unit is configured to use training samples to jointly train the first preset model, the second preset model, and the third preset model to obtain the training After the first model, the second model and the third model; wherein, the third model is the third preset model after training.
The network device according to claim 243, wherein the second processing unit is configured to input a first training sample into an estimated preset sub-model of the first preset model, and obtain an output of the estimated preset sub-model the initial information;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into a third preset model to obtain transformed information output by the third preset model;

inputting the transformed information into the second preset model to obtain restoration information output by the second preset model;

Determine a second loss function based on a first degree of difference between the restored information and the initial information, and based on a second degree of difference between the initial information and a second training sample; the second training sample and The first training sample corresponds to;

performing reverse conduction according to the second loss function to update the first preset model, the second preset model, and the third preset model.
The network device according to claim 243, wherein the second processing unit is configured to input a first training sample into an estimated preset sub-model of the first preset model, and obtain an output of the estimated preset sub-model the initial information;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

Inputting the compressed feature vector information into a third preset model to obtain transformed feature vector information output by the third preset model;

inputting the transformed eigenvector information into the second preset model to obtain restored eigenvector information output by the second preset model;

determining a second loss function based on a first degree of difference between the restored feature vector information and the feature vector information of the initial information, and based on a second degree of difference between the initial information and a second training sample; The second training sample corresponds to the first training sample;

performing reverse conduction according to the second loss function to update the first preset model, the second preset model, and the third preset model.
The network device according to any one of claims 231-245, wherein the training samples include a first training sample.
The network device of claim 246, wherein:

The first training samples are distributed in the first dimension and/or the second dimension.
The network device of claim 247, wherein,

The first dimension is a time domain dimension;

The first training samples include first information samples distributed in m time units in the time domain dimension; m is a positive integer.
The network device of claim 247, wherein,

The second dimension is a frequency domain dimension;

The first training samples include first information samples distributed on x frequency domain resources in the frequency domain dimension; x is a positive integer.
The network device according to any one of claims 247-249, wherein the first training samples are also distributed in a third dimension;

The third dimension is a complex dimension; the first training samples include the real part of the first information sample and the imaginary part of the first information sample.
The network device according to any one of claims 246-250, wherein the training samples further include a second training sample corresponding to the first training sample;

The second training sample is composed of a matrix of T dimensions; T is an integer greater than or equal to 2.
The network device of claim 251, wherein,

The T dimensions include a fourth dimension and a fifth dimension.
The network device of claim 252, wherein:

The matrix of the T dimensions is a two-dimensional matrix of M×N; wherein, M represents the quantity of the first granularity under the fourth dimension, and N represents the quantity of the second granularity under the fifth dimension; both M and N are is a positive integer.
The network device of claim 253, wherein:

The value at the ijth position in the two-dimensional matrix is used to represent the channel quality at the i-th first granularity in the fourth dimension and the j-th second granularity in the fifth dimension; both i and j are is a positive integer.
The network device according to any one of claims 252-254, wherein the T dimensions further include a sixth dimension.
The network device of claim 255, wherein,

The matrix of T dimensions is a three-dimensional matrix of M×N×W; wherein, M represents the number of first granularities in the fourth dimension, N represents the number of second granularities in the fifth dimension, and W represents the number of granularities in the fifth dimension. The quantity of the third granularity under the sixth dimension; M, N and W are all positive integers.
The network device of claim 256, wherein:

The value of the ijkth position in the three-dimensional matrix is used to represent the i-th first granularity in the fourth dimension, the j-th second granularity in the fifth dimension, and the k-th in the sixth dimension channel quality at the third granularity; i, j and k are all positive integers.
The network device according to any one of claims 252-257, wherein,

The fourth dimension is a frequency domain dimension; the first granularity includes one of the following: L1 resource blocks RB, L2 subcarriers; L1 and L2 are positive integers.
The network device according to any one of claims 252-257, wherein,

The fourth dimension is a time domain dimension; the first granularity includes one of the following: K1 microseconds, K2 symbol length, K3 number of sampling points of symbols; K1, K2 and K3 are positive integers.
The network device of claim 259, wherein the symbols are OFDM symbols.
The network device according to any one of claims 252-260, wherein the fifth dimension is a spatial domain dimension;

The second granularity is an interval between a pair of transmitting and receiving antennas or an angle of arrival.
The network device according to any one of claims 255-261, wherein,

The sixth dimension is a complex dimension; the third granularity is 1, and the quantity W of the third granularity under the complex dimension is 2.
The network device according to claim 262, wherein, when the k is the first value, the value of the ijk-th position in the three-dimensional matrix is used to represent the i-th position in the fourth dimension A granularity, the real part of the channel quality at the jth second granularity of the fifth dimension;

When the k is the second value, the value of the ijkth position in the three-dimensional matrix is used to represent the i-th first granularity in the fourth dimension and the j-th granularity in the fifth dimension The imaginary part of the channel quality at the second granularity.
The network device according to any one of claims 231-263, wherein the second communication unit is configured to send the second model.
The network device according to any one of claims 231-264, wherein the second communication unit is configured to send the first model.
The network device according to any one of claims 231-264, wherein the second communication unit is used for the estimation sub-model and the compression sub-model.
The network device according to any one of claims 231-264, wherein the second communication unit is configured to send the estimation sub-model, the compression sub-model and the channel generation sub-model.
The network device according to any one of claims 231-267, wherein the second communication unit is configured to send the third model.
The network device according to any one of claims 265-268, wherein the first model is carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, manual Downlink data transmission required for intelligent business transmission;

The second model is carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements;

The third model is carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements;

The estimated sub-model is carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements;

The compressed sub-model is carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements;

The channel generation sub-model is carried by at least one of the following: downlink control signaling, MAC CE message, RRC message, broadcast message, downlink data transmission, downlink data transmission for artificial intelligence business transmission requirements.
An electronic device comprising:

The third processing unit is configured to use training samples to jointly train the first preset model and the second preset model to obtain the trained first model and the second model;

Wherein, the first model is the first preset model after training, and the second model is the second preset model after training; the first model is used to process the first information to obtain Second information; the second model is used to process the second information to obtain channel information.
The electronic device according to claim 270, wherein the loss function used in the training is a first loss function;

The first loss function is constructed based on the degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model.
The electronic device according to claim 271, wherein the degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model is determined based on a distance, or determined based on a degree of similarity of.
The electronic device according to claim 271 or 272, wherein the third processing unit is configured to input a first training sample into the estimated preset sub-model of the first preset model to obtain the estimated preset sub-model The initial information output by the model;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into the second preset model to obtain restoration information output by the second preset model;

determining the first loss function based on the degree of difference between the restoration information and the initial information;

performing reverse conduction according to the first loss function to update the first preset model and the second preset model.
The electronic device according to claim 271 or 272, wherein the third processing unit is configured to input a first training sample into the estimated preset sub-model of the first preset model to obtain the estimated preset sub-model The initial information output by the model;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

inputting the compressed feature vector information into the second preset model to obtain restored feature vector information output by the second preset model;

determining a first loss function based on the degree of difference between the restored eigenvector information and the eigenvector information of the initial information;

performing reverse conduction according to the first loss function to update the first preset model and the second preset model.
The electronic device according to claim 271 or 272, wherein the third processing unit is configured to use training samples to jointly train the first preset model, the second preset model and the third preset model to obtain the training After the first model, the second model and the third model; wherein, the third model is the third preset model after training.
The electronic device according to claim 275, wherein the third processing unit is configured to input a first training sample into an estimated preset sub-model of the first preset model, and obtain an output of the estimated preset sub-model the initial information;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into a third preset model to obtain transformed information output by the third preset model;

inputting the transformed information into the second preset model to obtain restoration information output by the second preset model;

determining a first loss function based on the degree of difference between the restoration information and the initial information;

performing reverse conduction according to the first loss function to update the first preset model, the second preset model and the third preset model.
The electronic device according to claim 275, wherein the third processing unit is configured to input a first training sample into an estimated preset sub-model of the first preset model, and obtain an output of the estimated preset sub-model the initial information;

Inputting the initial information into the channel generation preset sub-model of the first preset model to obtain the eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

Inputting the compressed feature vector information into a third preset model to obtain transformed feature vector information output by the third preset model;

inputting the transformed eigenvector information into the second preset model to obtain restored eigenvector information output by the second preset model;

determining a first loss function based on the degree of difference between the restored eigenvector information and the eigenvector information of the initial information;

performing reverse conduction according to the first loss function to update the first preset model, the second preset model and the third preset model.
The electronic device according to claim 270, wherein the loss function used in the training is a second loss function;

The second loss function is based on the first degree of difference between the output information of the second preset model and the input information of the compressed preset sub-model of the first preset model and the first preset model The second difference degree between the output information of the estimated preset sub-model and the second training sample is constructed; wherein, the second training sample corresponds to the first training sample input into the estimated preset sub-model.
The electronic device according to claim 278, wherein the first degree of difference is determined based on a distance, or determined based on a degree of similarity; and/or,

The second degree of difference is determined based on distance, or determined based on similarity.
The electronic device according to claim 278 or 279, wherein the third processing unit is configured to input the first training sample into the estimated preset sub-model of the first preset model to obtain the estimated preset sub-model The initial information output by the model;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into the second preset model to obtain restoration information output by the second preset model;

Determine a second loss function based on a first degree of difference between the restored information and the initial information, and based on a second degree of difference between the initial information and a second training sample; the second training sample and The first training sample corresponds to;

performing reverse conduction according to the second loss function to update the first preset model and the second preset model.
The electronic device according to claim 278 or 279, wherein the third processing unit is configured to input the first training sample into the estimated preset sub-model of the first preset model to obtain the estimated preset sub-model The initial information output by the model;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

inputting the compressed feature vector information into the second preset model to obtain restored feature vector information output by the second preset model;

determining a second loss function based on a first degree of difference between the restored feature vector information and the feature vector information of the initial information, and based on a second degree of difference between the initial information and a second training sample; The second training sample corresponds to the first training sample;

performing reverse conduction according to the second loss function to update the first preset model and the second preset model.
The electronic device according to claim 278 or 279, wherein the third processing unit is configured to use training samples to jointly train the first preset model, the second preset model and the third preset model to obtain the training After the first model, the second model and the third model; wherein, the third model is the third preset model after training.
The electronic device according to claim 282, wherein the third processing unit is configured to input a first training sample into an estimated preset sub-model of the first preset model, and obtain an output of the estimated preset sub-model the initial information;

inputting the initial information into the compressed preset sub-model of the first preset model to obtain compressed information output by the compressed preset sub-model;

inputting the compressed information into a third preset model to obtain transformed information output by the third preset model;

inputting the transformed information into the second preset model to obtain restoration information output by the second preset model;

Determine a second loss function based on a first degree of difference between the restored information and the initial information, and based on a second degree of difference between the initial information and a second training sample; the second training sample and The first training sample corresponds to;

performing reverse conduction according to the second loss function to update the first preset model, the second preset model, and the third preset model.
The electronic device according to claim 282, wherein the third processing unit is configured to input a first training sample into an estimated preset sub-model of the first preset model, and obtain an output of the estimated preset sub-model the initial information;

inputting the initial information into the channel generation preset sub-model of the first preset model, and obtaining eigenvector information of the initial information output by the channel generation preset sub-model;

Inputting the eigenvector information of the initial information into the compressed preset sub-model of the first preset model to obtain the compressed eigenvector information output by the compressed preset sub-model;

Inputting the compressed feature vector information into a third preset model to obtain transformed feature vector information output by the third preset model;

inputting the transformed eigenvector information into the second preset model to obtain restored eigenvector information output by the second preset model;

determining a second loss function based on a first degree of difference between the restored feature vector information and the feature vector information of the initial information, and based on a second degree of difference between the initial information and a second training sample; The second training sample corresponds to the first training sample;

performing reverse conduction according to the second loss function to update the first preset model and the second preset model.
The electronic device according to any one of claims 270-284, wherein the training samples include a first training sample.
The electronic device of claim 285, wherein:

The first training samples are distributed in the first dimension and/or the second dimension.
The electronic device of claim 285, wherein:

The first dimension is a time domain dimension; the first training samples include first information samples distributed in m time units in the time domain dimension; m is a positive integer.
The electronic device of claim 285, wherein:

The second dimension is a frequency domain dimension; the first training sample includes first information samples distributed on x frequency domain resources in the frequency domain dimension; x is a positive integer.
The electronic device according to any one of claims 286-288, wherein the first training samples are further distributed in a third dimension;

The third dimension is a complex dimension; the first training samples include the real part of the first information sample and the imaginary part of the first information sample.
The electronic device according to any one of claims 285-289, wherein the training samples further include a second training sample corresponding to the first training sample;

The second training sample is composed of a matrix of T dimensions; T is an integer greater than or equal to 2.
The electronic device according to claim 290, wherein,

The T dimensions include a fourth dimension and a fifth dimension.
The electronic device of claim 291, wherein:

The matrix of the T dimensions is a two-dimensional matrix of M×N; wherein, M represents the quantity of the first granularity under the fourth dimension, and N represents the quantity of the second granularity under the fifth dimension; both M and N are is a positive integer.
The electronic device of claim 292, wherein:

The value at the ijth position in the two-dimensional matrix is used to represent the channel quality at the i-th first granularity in the fourth dimension and the j-th second granularity in the fifth dimension; both i and j are is a positive integer.
The electronic device according to any one of claims 291-293, wherein the T dimensions further include a sixth dimension.
The electronic device of claim 294, wherein:

The matrix of T dimensions is a three-dimensional matrix of M×N×W; wherein, M represents the number of first granularities in the fourth dimension, N represents the number of second granularities in the fifth dimension, and W represents the number of granularities in the fifth dimension. The quantity of the third granularity under the sixth dimension; M, N and W are all positive integers.
The electronic device of claim 295, wherein:

The value of the ijkth position in the three-dimensional matrix is used to represent the i-th first granularity in the fourth dimension, the j-th second granularity in the fifth dimension, and the k-th in the sixth dimension channel quality at the third granularity; i, j and k are all positive integers.
The electronic device according to any one of claims 291-296, wherein,

The fourth dimension is a frequency domain dimension; the first granularity includes one of the following: L1 resource blocks RB, L2 subcarriers; L1 and L2 are positive integers.
The electronic device according to any one of claims 291-296, wherein,

The fourth dimension is a time domain dimension; the first granularity includes one of the following: K1 microseconds, K2 symbol length, K3 number of sampling points of symbols; K1, K2 and K3 are positive integers.
The electronic device of claim 298, wherein the symbols are OFDM symbols.
The electronic device according to any one of claims 291-299, wherein the fifth dimension is a spatial domain dimension;

The second granularity is an interval between a pair of transmitting and receiving antennas or an angle of arrival.
The electronic device according to any one of claims 294-300, wherein,

The sixth dimension is a complex dimension; the third granularity is 1, and the quantity W of the third granularity under the complex dimension is 2.
The electronic device according to claim 301, wherein, when the k is the first value, the value of the ijkth position in the three-dimensional matrix is used to represent the i-th position in the fourth dimension A granularity, the real part of the channel quality at the jth second granularity of the fifth dimension;

When the k is the second value, the value of the ijkth position in the three-dimensional matrix is used to represent the i-th first granularity in the fourth dimension and the j-th granularity in the fifth dimension The imaginary part of the channel quality at the second granularity.
A terminal device, comprising: a processor and a memory, the memory is used to store a computer program, and the processor is used to call and run the computer program stored in the memory, so that the terminal device executes claims 1 to 58 any one of the methods described.
A network device, comprising: a processor and a memory, the memory is used to store a computer program, and the processor is used to invoke and run the computer program stored in the memory, so that the network device performs the tasks described in claims 59 to 118 any one of the methods described.
An electronic device, comprising: a processor and a memory, the memory is used to store a computer program, the processor is used to call and run the computer program stored in the memory, so that the network device performs the any one of the methods described.
A chip, comprising: a processor, configured to invoke and run a computer program from a memory, so that a device equipped with the chip executes the method as claimed in any one of claims 1 to 58.
A chip, comprising: a processor for invoking and running a computer program from a memory, so that a device equipped with the chip executes the method as claimed in any one of claims 59 to 118.
A chip, comprising: a processor, configured to invoke and run a computer program from a memory, so that a device equipped with the chip executes the method as claimed in any one of claims 119 to 151.
A computer-readable storage medium for storing a computer program, which causes the device to execute the method according to any one of claims 1 to 151 when the computer program is executed by the device.
A computer program product comprising computer program instructions for causing a computer to perform the method as claimed in any one of claims 1 to 151.
A computer program that causes a computer to perform the method according to any one of claims 1 to 151.
A communication system comprising:

A terminal device, configured to perform the method according to any one of claims 1 to 58;

A network device configured to execute the method according to any one of claims 59-118.