CN115189740A

CN115189740A - Channel state information feedback system and method

Info

Publication number: CN115189740A
Application number: CN202210745916.4A
Authority: CN
Inventors: 赵顾良
Original assignee: Shandong Inspur Science Research Institute Co Ltd
Current assignee: Shandong Inspur Science Research Institute Co Ltd
Priority date: 2022-06-27
Filing date: 2022-06-27
Publication date: 2022-10-14

Abstract

The invention provides a channel state information feedback system and a method, wherein the system comprises: the terminal side comprises a first CSI-RS feedback reasoning network intelligent agent deployed on the network side and a second CSI-RS feedback reasoning network intelligent agent deployed on the terminal side; the first CSI-RS feedback reasoning network intelligent agent is used for sending a feedback reasoning network model to the second CSI-RS feedback reasoning network intelligent agent and carrying out on-line training on the feedback reasoning network model based on a federal learning mechanism and matched with the second CSI-RS feedback reasoning network intelligent agent; and the second CSI-RS feedback reasoning network intelligent body is used for performing on-line training on the feedback reasoning network model by matching with the first CSI-RS feedback reasoning network intelligent body based on a federal learning mechanism. According to the invention, the intelligent agents are respectively deployed at the network side and the terminal side, and the federal learning of the specific feedback reasoning network model is realized, so that the accuracy of channel state information feedback can be improved.

Description

Channel state information feedback system and method

Technical Field

The present invention relates to the field of wireless communication technologies, and in particular, to a channel state information feedback system and method.

Background

Massive MIMO (Multiple-Input Multiple-Output) is one of the core technologies of wireless communication. With the rapid increase of the demand of wireless mobile communication technology for bandwidth resources and the rapid development of multi-antenna technology, the advantages of the massive MIMO technology are gradually shown, and a larger-scale antenna array can multiply the communication capacity, optimize the energy transmission efficiency of information, and shorten the air interface transmission delay.

The most critical determining factor of the performance, especially the capacity, of the massive MIMO transmission link is the proximity between the downlink CSI (Channel State Information) report obtained by the base station and the downlink CSI during actual transmission. If the CSI can reflect the link condition at the actual transmission time more accurately, the base station scheduler can schedule wireless resources according to the CSI and send the information to the receiving terminal in a mode closer to the maximum capacity of an actual channel. The CSI information fed back by the Multiple terminals may further assist the base station to perform MU-MIMO (Multi-User Multiple-Input Multiple-Output ) pairing transmission.

For an FDD (Frequency Division duplex) system, since a downlink carrier and an uplink carrier are not in the same Frequency band, channel reciprocity cannot be used, and a base station cannot optimize transmission on the downlink carrier by using channel estimation information of an uplink signal transmitted by a UE (User Equipment). The common method is that UE measures downlink channels, and then reports the measured channel state information CSI to a base station through an uplink channel, and after the base station receives the measured channel state information CSI, the base station optimizes and schedules downlink signals by using the reported information in a next downlink transmission window, so that the purposes of improving downlink transmission efficiency, improving downlink experience of users and optimizing various performances are achieved.

The CSI reporting configuration of the existing 5G NR (New Radio, new wireless air interface) is generally a CSI feedback method based on a codebook, but as the number of antennas of a massive MIMO system increases, the size and computational complexity of the codebook may significantly increase. For a terminal in broadband wireless communication to report CSI Information to a base station, a related technology compresses CSI Information through a deep learning technology, and analyzes and recovers the CSI Information after the base station receives the CSI Information, but for terminals in different cells, different antenna panel/array combinations used by the terminal and the base station ("UE-gNB pair") when "CSI-RS (Channel State Information-Reference Signal, channel State Information Reference Signal) feedback inference networks" are used, because different Channel environments of different cells are different, and when a multi-Transmission and Reception Point (multi-Transmission and Reception node) is used, there are differences in terms of weight values in a network model and the same model, and it is necessary for both the UE and the gNB to infer that the CSI-RS feedback network "is locally carried by the UE, a decoder (carried by the gNB) is updated, and the existing Channel State Information feedback method based on artificial intelligence analysis is difficult to correspondingly update the" CSI-RS feedback network "used by the terminal and the base station, resulting in high accuracy of the CSI-RS feedback network side and non-recovery of the CSI Information.

Therefore, how to improve the accuracy of the channel state information feedback based on artificial intelligence becomes an urgent problem to be solved in the industry.

Disclosure of Invention

Aiming at the problems in the prior art, the invention provides a channel state information feedback system and a channel state information feedback method.

In a first aspect, the present invention provides a channel state information feedback system, including:

the terminal side comprises a first CSI-RS feedback reasoning network intelligent agent deployed on the network side and a second CSI-RS feedback reasoning network intelligent agent deployed on the terminal side;

the first CSI-RS feedback reasoning network intelligent agent is used for sending a CSI-RS feedback reasoning network model to a second CSI-RS feedback reasoning network intelligent agent corresponding to a terminal side in a cell, and is matched with the second CSI-RS feedback reasoning network intelligent agent to carry out online training on the CSI-RS feedback reasoning network model based on a federal learning mechanism;

the second CSI-RS feedback reasoning network intelligent body is used for performing online training on the CSI-RS feedback reasoning network model based on a federal learning mechanism and matched with the first CSI-RS feedback reasoning network intelligent body.

Optionally, a channel state information feedback system provided in the present invention further includes: a third CSI-RS feedback reasoning network agent deployed on the core network side;

and the third CSI-RS feedback reasoning network agent is used for managing and evaluating the performance of the whole network candidate CSI-RS feedback reasoning network model.

In a second aspect, the present invention further provides a channel state information feedback method based on any one of the channel state information feedback systems in the first aspect, which is applied to a first CSI-RS feedback inference network agent, and includes:

under the condition that the network side is determined to receive a connection establishment request message sent by a target terminal side, establishing connection with a second CSI-RS feedback reasoning network intelligent agent corresponding to the target terminal side;

judging whether the code word of the first CSI-RS feedback inference network encoder model currently used by the target terminal side is the same as the code word of the first CSI-RS feedback inference network decoder model currently used by the network side through information interaction with the second CSI-RS feedback inference network agent;

and under the condition that the first CSI-RS feedback inference network encoder model and the first CSI-RS feedback inference network decoder model are determined to be the same, establishing a first CSI-RS feedback inference network model between the network side and the target terminal side.

Optionally, according to the channel state information feedback method provided by the present invention, after the first CSI-RS feedback inference network model between the network side and the target terminal side is established, the method further includes:

judging whether the first CSI-RS feedback inference network encoder model is available in the cell;

and sending first indication information carrying a first model updating message to the second CSI-RS feedback reasoning network intelligent agent under the condition that the second CSI-RS feedback reasoning network intelligent agent is determined to be unavailable, wherein the first indication information is used for indicating the second CSI-RS feedback reasoning network intelligent agent to update the first CSI-RS feedback reasoning network encoder model based on the first model updating message.

Optionally, a method for feeding back channel state information provided by the present invention further includes:

under the condition that the first weight value and the second weight value are different, judging whether a first weight value corresponding to the first CSI-RS feedback inference network encoder model and a first weight value corresponding to the first CSI-RS feedback inference network encoder model are available in the cell;

and sending second indication information carrying a second model updating message to the second CSI-RS feedback inference network intelligent agent under the condition that the first CSI-RS feedback inference network encoder model is determined to be available in the cell and part of or all of the first weights are unavailable, wherein the second indication information is used for indicating the second CSI-RS feedback inference network intelligent agent to update the first CSI-RS feedback inference network encoder model based on the second model updating message.

Optionally, the method for feeding back channel state information according to the present invention further includes:

and sending abnormal backspacing information to the second CSI-RS feedback reasoning network intelligent agent under the condition that the two CSI-RS feedback reasoning network intelligent agents are different, wherein the abnormal backspacing information is used for indicating the second CSI-RS feedback reasoning network intelligent agent to perform abnormal backspacing on the first CSI-RS feedback reasoning network encoder model.

Optionally, according to the channel state information feedback method provided by the present invention, after sending the abnormal fallback information to the second CSI-RS feedback inference network agent, the method further includes:

determining a second CSI-RS feedback reasoning network encoder model obtained after the second CSI-RS feedback reasoning network agent performs exception backspacing on the first CSI-RS feedback reasoning network encoder model;

and loading a second CSI-RS feedback inference network decoder model matched with the second CSI-RS feedback inference network encoder model, and establishing a second CSI-RS feedback inference network model between the network side and the target terminal side based on the second CSI-RS feedback inference network encoder model and the second CSI-RS feedback inference network decoder model.

on the basis of a federal learning mechanism, the second CSI-RS feedback reasoning network model is trained on line by matching with the second CSI-RS feedback reasoning network agent;

and optimizing a second weight corresponding to the second CSI-RS feedback inference network model based on a federal learning mechanism under the condition that the performance of the second CSI-RS feedback inference network model is determined to meet a first preset threshold.

and under the condition that the performance of the second CSI-RS feedback reasoning network model meets a second preset threshold, determining a candidate CSI-RS feedback reasoning network model provided by a third CSI-RS feedback reasoning network agent, and carrying out federal learning on the candidate CSI-RS feedback reasoning network model based on a federal learning mechanism.

counting training result information of a local CSI-RS feedback inference network model;

and reporting the training result information to the third CSI-RS feedback reasoning network intelligent agent so that the third CSI-RS feedback reasoning network intelligent agent can manage the full-network candidate CSI-RS feedback reasoning network model based on the model training result information.

In a third aspect, the present invention also provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and when the processor executes the computer program, the channel state information feedback method according to the second aspect is implemented.

In a fourth aspect, the present invention also provides a non-transitory computer-readable storage medium having stored thereon a computer program which, when executed by a processor, implements the channel state information feedback method according to the second aspect.

In a fifth aspect, the present invention also provides a computer program product comprising a computer program which, when executed by a processor, implements the channel state information feedback method according to the second aspect.

According to the channel state information feedback system and method provided by the invention, the CSI-RS feedback reasoning network intelligent bodies are respectively deployed at the network side and the terminal side, and the specific feedback reasoning network model is subjected to federal learning based on the CSI-RS feedback reasoning network intelligent bodies at the network side and the terminal side, so that the feedback reasoning network model with better performance is obtained, and the accuracy of channel state information feedback can be improved.

Drawings

In order to more clearly illustrate the technical solutions of the present invention or the prior art, the drawings needed for the description of the embodiments or the prior art will be briefly described below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and those skilled in the art can also obtain other drawings according to the drawings without creative efforts.

Fig. 1 is a schematic structural diagram of a channel state information feedback system provided by the present invention;

fig. 2 is a schematic flow chart of a channel state information feedback method provided by the present invention;

fig. 3 is a schematic physical structure diagram of an electronic device provided in the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is obvious that the described embodiments are some, but not all embodiments of the present invention. All other embodiments, which can be obtained by a person skilled in the art without inventive step based on the embodiments of the present invention, are within the scope of protection of the present invention.

In order to facilitate a clearer understanding of embodiments of the present invention, some relevant background information is first presented below.

The existing CSI reporting configuration of 5G NR includes two parts: the configuration of the CSI-RS related resources (indicating the UE to measure at the time-frequency resource) and the configuration of the UE to report the CSI (indicating the UE to send a CSI measurement report at the time-frequency resource). The resource configuration is used for configuring reference signals for calculating the CSI, and the report configuration is used for configuring the behavior of reporting the CSI. The RRC (Radio Resource Control ) layer signaling CSI-reporting configuration IE indicates resources used for Channel measurement and interference measurement, and further includes codebook configuration, including Type I, type II, or enhanced Type II codebooks and codebook restriction subsets, and indicates that a periodic manner, a semi-continuous manner based on a PUSCH (Physical Uplink Shared Channel), or an aperiodic manner is adopted for time domain CSI reporting. The setting content further includes: frequency domain wideband and sub-band granularity of CQI (Channel Quality Indication) and PMI (Precoding Matrix Indicator); a limitation on channel measurements and a limitation on interference measurements; the information type that the CSI reported by the UE needs to indicate includes CQI, PMI, CRI (CSI-RS Resource Indicator), SSBRI (SS/PBCH Block Resource Indicator, SSB Resource Indicator), LI (Lay Indicator), RI (Rank Indicator), L1-RSRP (Layer 1Signal Received Power), or L1-SINR (Layer 1Signal Interference plus Noise Ratio), etc.

According to the CSI feedback method based on the codebook, as the number of antennas of the massive MIMO system increases, the size and the calculation complexity of the codebook are obviously increased.

The method comprises the steps that by means of an artificial intelligence method, all or part of channel state information measured by UE is subjected to information coding and compression through a neural network encoder deployed on the side of the UE to form a special code word, the code word is sent to a base station through an air interface uplink, and the base station side recovers all channel characteristics with high similarity to the UE measurement information through the received code word and priori knowledge learned and stored in a network by the base station side neural network through the neural network matched with the UE side. The following problems need to be solved to realize this function:

(1) Terminals in different cells, the "CSI-RS feedback inference network" used by the "UE-gNB pair" can have differences in network models and weights in the same model due to different channel environments of each cell and different antenna panel/array combinations when multi-TRP is used;

(2) When the UE is started to register or is switched to a new cell, the difference requires that the UE and the gNB respectively update an encoder (carried by the UE) and a decoder (carried by the gNB) in a local CSI-RS feedback reasoning network, how the gNB efficiently updates the encoder of the UE part of the CSI-RS feedback reasoning network is solved, and meanwhile, the problem that too many system downlink wireless resources are not occupied is solved;

(3) After the UE obtains the encoder of the UE part of the CSI-RS feedback inference network from the current gNB to update, the performance of a local encoder needs to be optimized in a certain mode while inference is carried out by utilizing the existing network, and the method is more universal in that the local encoder is learned online while inference is carried out; considering that the calculation capacity of the UE is limited, learning and optimizing the weight of the encoder at the UE side in a certain mode, namely learning the weight of the encoder at the UE side on line by adopting a mechanism of 'UE-gNB to distributed federal learning', and after certain conditions are met, the UE can feed back the weight combination with better effect to the gNB through an uplink and then update the suitable UE in the cell by the gNB;

(4) Because the number of the UE in the cell is large, the gNB can also adopt a mode of configuring different encoder models for different UE and adopting different decoders on a base station side for joint training to carry out performance monitoring and on-line training on different 'CSI-RS feedback inference network' models;

(5) In order to prevent the network security risk aiming at artificial intelligence, a set of efficient and credible privacy computing security prevention mechanism based on federal learning needs to be designed for the CSI-RS feedback reasoning network.

In order to overcome the above-mentioned drawbacks, the present invention provides a system and method for feeding back channel state information. The following describes the channel state information feedback system and method provided by the present invention with reference to fig. 1-3.

Fig. 1 is a schematic structural diagram of a channel state information feedback system provided in the present invention, and as shown in fig. 1, the system includes:

the second CSI-RS feedback reasoning network intelligent body is used for performing on-line training on the CSI-RS feedback reasoning network model by matching with the first CSI-RS feedback reasoning network intelligent body based on a federal learning mechanism.

Specifically, in the embodiment of the present invention, the channel state information feedback system may include a first CSI-RS feedback inference network agent deployed on a network side and a second CSI-RS feedback inference network agent deployed on a terminal side, where the first CSI-RS feedback inference network agent may be configured to send a CSI-RS feedback inference network model to the second CSI-RS feedback inference network agent corresponding to the terminal side in a cell, and perform online training on the CSI-RS feedback inference network model in cooperation with the second CSI-RS feedback inference network agent based on a federal learning mechanism, and the second CSI-RS feedback inference network agent is configured to perform online training on the CSI-RS feedback inference network model in cooperation with the first CSI-RS feedback inference network agent based on the federal learning mechanism.

Optionally, as shown in fig. 1, the first CSI-RS feedback inference network agent deployed on the network side may include the following functional modules: the CSI-RS feedback reasoning model updating module comprises a CSI-RS feedback reasoning model updating module, a CSI-RS reasoning module and a CSI-RS feedback reasoning model federal learning management module, wherein the CSI-RS feedback reasoning model updating module can comprise a plurality of model performance evaluation sub-modules, a model distribution sub-module and a model training sub-module.

Optionally, as shown in fig. 1, the second feedback inference network agent deployed at the terminal side may include the following functional modules: the CSI-RS feedback reasoning module comprises a CSI-RS feedback reasoning module updating module, a CSI-RS reasoning module and a CSI-RS feedback reasoning module federal learning management module.

Specifically, the work flow of each functional module described above is described below.

Optionally, in the embodiment of the present invention, the network side may be a base station.

Optionally, the model distribution submodule on the network side may be configured to be responsible for managing the terminals in the cell, configure which CSI-RS feedback inference network model is used by the terminals in the cell for CQI feedback by using a certain policy, and perform online federal learning at the same time.

Optionally, the model training sub-module may be a management control unit for performing online training and federal learning on a "CSI-RS feedback inference network model" used by different UEs on the base station side, and is responsible for configuring various parameters and training methods of an encoder on the terminal side and a corresponding decoder on the base station side. The model training submodule determines the effect of downlink scheduling and transmission based on the information according to the CQI code word information reported after each time of terminal compression by the encoder and the downlink channel information recovered by the decoder at the base station side, and the specific method comprises the following steps: and in each training period, the accuracy and reasoning ability of the model are evaluated on line by calling a preset training set.

Optionally, the CSI-RS inference module corresponds to an execution mechanism, and when the neural network model is trained, the neural network model is loaded into an execution environment, and the neural network operating in the execution environment is a module for performing an inference function. The CSI-RS reasoning modules at the base station side and the terminal side are integrated models, the two parts of networks are mutually matched, and code words formed by compressing a coder at the terminal side are transmitted to a decoder at the base station side through an air interface, so that all information channel information is recovered to reduce air interface transmission overhead, and meanwhile, a relatively accurate downlink channel estimation value can be obtained.

For example, the CSI-RS inference process may be understood that the network at the terminal side and the network at the base station side have learned the rule and knowledge of the current downlink channel, and the base station side can perform better downlink resource scheduling according to the indication by only indicating the mode of the current downlink channel through the code words carried by the fewer bits. This principle is similar to codebook-based CSI feedback, but the neural network, taking advantage of its nonlinear capabilities, can achieve a more accurate indication of the current channel state.

It can be understood that, a more general model in the AI (Artificial Intelligence) -based CSI feedback network is to deploy a joint AI model on the terminal side and the base station side, generally, a certain neural network is used as a CSI Information coding and compressing module, so as to perform compression coding on all Channel characteristics or part of Channel characteristics measured and Channel estimated by the UE on the CSI-RS (Channel State Information-Reference Signal) sent by the gbb (base station), so as to form a codeword, report the codeword to the base station through an Uplink Channel such as PUCCH (Physical Uplink Control Channel) or PUSCH, the base station side uses an Artificial Intelligence decoding network matched with the codeword, send the compression coding of all Channel characteristics or part of Channel characteristics received into a decoding network, and recover all Channel characteristics or part of Channel characteristics from the received codeword through a priori knowledge stored in the decoding network.

The implementation of the above functions requires a network architecture with relatively stable, safe and acceptable computation amount of the gNB granularity for support. The channel state information feedback system provided by the invention can ensure the performance and the safety of an artificial intelligent compressed reporting system of the channel state information, and realize the online training, reasoning and updating of an artificial intelligent network model of a UE encoder and a gNB decoder, and the safety protection and management of a distributed intelligent agent.

The channel state information feedback system provided by the embodiment of the invention supports that a plurality of UE under the same base station respectively utilize the wireless environment, the service condition and the self-contained AI computing power of the UE to cooperate with the base station to carry out online AI compression CSI and decompression after the CSI is transmitted to the base station.

According to the channel state information feedback system provided by the invention, the CSI-RS feedback reasoning network intelligent bodies are respectively arranged on the network side and the terminal side, and the specific feedback reasoning network model is subjected to federal learning based on the CSI-RS feedback reasoning network intelligent bodies on the network side and the terminal side, so that the feedback reasoning network model with better performance is obtained, and the accuracy of channel state information feedback can be improved.

Optionally, the channel state information feedback system provided in the present invention further includes: a third CSI-RS feedback reasoning network agent deployed on the core network side;

Specifically, as shown in fig. 1, in the embodiment of the present invention, the channel state information feedback system may further include a third CSI-RS feedback inference network agent deployed on the core network side, where the third CSI-RS feedback inference network agent is configured to manage and perform performance evaluation on the full-network candidate CSI-RS feedback inference network model.

Optionally, as shown in fig. 1, a third CSI-RS feedback inference network agent deployed on the core network side may include the following functional modules: and the CSI-RS feedback reasoning model updating module comprises a plurality of model performance evaluation sub-modules and a model distribution sub-module.

Optionally, the model distribution submodule on the core network side may control the model strategy of each gNB according to the performance evaluation condition of the whole network on the multiple CSI feedback neural network models.

Specifically, the work flow of the channel state information feedback system provided by the invention comprises the following steps:

(1) Establishing a connection between target UE and gNB, establishing a connection between a target UE side and a gNB model updating module when the target UE is started to register or is switched to a new cell, updating a model in a target UE side CSI-RS feedback reasoning network agent, and turning to the step (2);

(2) A model distribution submodule in the gNB is connected with a model updating module of the target UE, through information interaction, the model updating module on the gNB side evaluates the CSI-RS feedback reasoning network encoder model currently used by the UE, the weight value and other conditions, if updating is needed, the step (4) is carried out, and if available, the step (3) is carried out;

(3) If the current model and the weight of the UE are available, informing a CSI-RS reasoning module in the gNB to load a corresponding decoder model and the weight, establishing a CSI-RS feedback reasoning network between the gNB and the UE, simultaneously bringing the network into a model federal learning management module of the gNB, monitoring and managing the network, and turning to the step (5); if the feedback reasoning network of the current UE is abnormal, the gNB model federal learning management module indicates the UE to execute abnormal rollback, and if the abnormal rollback fails, the UE is controlled to go to the step (1); meanwhile, reporting the abnormity to a gNB model training submodule by the gNB federal learning management module, and turning to the step (6);

(4) The model updating module at the gNB side evaluates the CSI-RS feedback encoder model and the weight value and other conditions currently used by the UE, if the model is available in the cell but part of the weight value or all the weight value in the encoder weight value can not be used, then a part of weight value or all the weight value updating process in the weight value corresponding to the UE encoder model is executed, and the method specifically comprises the following steps: and informing a model updating module of the UE intelligent agent to execute partial weight value or all weight value updating in the weight value corresponding to the encoder model, and the UE intelligent agent receives the updated weight value information by utilizing a unicast or multicast channel to update the weight value of the local CSI-RS reasoning module of the UE. When the CSI-RS feedback encoder model and the weight value at the UE side are available, turning to the step (3), and if the model is not available, turning to the step (5);

(5) A model updating module at the gNB side evaluates the CSI-RS feedback encoder model and the weight value and other conditions currently used by the UE, and if the model and the weight value are not available in the cell, immediately executes the UE encoder model and the weight value updating process, and specifically comprises the following steps: informing a model updating module of the UE intelligent agent to execute the encoder model and weight updating, wherein the UE intelligent agent receives updated weight information by using a unicast or multicast channel, updates the weight of a local CSI-RS reasoning module of the UE, and goes to the step (2);

(6) After loading a CSI-RS decoder model and a weight corresponding to the current UE, the gNB intelligent agent starts and monitors a 'CSI-RS reasoning module' between the gNB and the UE to work; training encoders and decoders by using an online learning mechanism according to a Federal learning rule by using a model Federal learning management module on the gNB and the UE side, sending a notification to a model training submodule in a local (gNB) model updating module when a CSI-RS reasoning network model pair is continuously trained for a period of time and a better performance threshold is found to be reached, managing online learning of the gNB-UE pair using the same model in the local (gNB) by using the model training submodule, optimizing a weight of the CSI-RS reasoning network model by using a Federal learning method, and if a group of weights with better performance is obtained by evaluating the gNB model training submodule at a certain moment or the performance deterioration of the model reaches a certain degree, notifying the gNB model performance evaluation submodule to go to the step (7);

(7) The performance evaluation submodule selects different UE in the cell based on relevant rules according to a candidate 'CSI-RS feedback reasoning network model' provided by a model performance evaluation submodule in a core network, and performs federal learning on a plurality of different 'CSI-RS feedback reasoning network models' respectively; when the report of the model training submodule is received, after local summarization and relevant processing are carried out, the various models and relevant information of the model training environment are reported to the model performance evaluation submodule at the core network side, and the step (8) is carried out;

optionally, since the CSI-RS feedback inference network model is generally related to the terminal type in the network, the related rules may include: for terminals with strong reasoning capabilities and strong GPU computing resources, a more complex neural network may be used.

Optionally, different UEs may be selected for federal learning according to different CSI-RS feedback inference network models. For example, for an indoor UE with a low mobility rate, due to low mobility and stable channel environment, a network with a high compression rate may be used for CSI feedback. For the UE with high mobility, the fed-back codeword cannot be too short, and the requirement on timeliness is also high.

(8) The model performance evaluation submodule on the core network side manages and analyzes various CSI-RS feedback reasoning network models and model applicable environment information, manages 'CSI-RS feedback reasoning network models' of different gNBs in the whole network, and goes to the step (9);

(9) According to the feedback of each gNB, the core network evaluation submodule controls the core network model distribution submodule to initiate all or part of the updating of the model and the corresponding weight, the updating mode comprises a unicast, multicast or broadcast session mode, the model or the weight is updated according to the UE grouped by using different models in the gNB, and the step (10) is carried out;

it can be understood that, by using the multicast broadcast channel of 5G NR and the multicast mode at the same wave velocity, it is possible to send the local encoder weight of the UE to the terminals in the same group by using one time-frequency resource, and such a mechanism improves the model updating efficiency and saves the time-frequency resource.

(10) And the gNB model distribution submodule controls the UE in the cell to update the model and update all or part of the model parameters according to the rule, and the step (2) is carried out.

It can be understood that, the channel state information feedback system provided by the present invention manages the "CSI-RS feedback inference network model" used between the base station side and the terminal side through the agents distributed at the core network side, the base station side and the terminal side.

In the embodiment of the invention, a core network intelligent agent manages a whole network candidate 'CSI-RS feedback inference network model'; selecting a proper UE group for model distribution, reasoning application and online learning by utilizing intelligent agents distributed on the base station side according to the prefabrication conditions of different models; when a certain UE-gNB achieves a set threshold after performing certain online training on a used model and obtains a better effect, information such as a current weight of the model and external conditions related to the model is collected and reported to the intelligent agent of the base station and the core network, the effect of the model is evaluated on the level of the base station and the level of the whole network respectively, the intelligent agent of the base station can expand the evaluation range of the model, and more UEs are used for evaluating the performance of the model.

And performing on-line training on the model by using a plurality of UEs in a federal learning mode. The core network side is used as a multi-model management and control and model source service providing end and supports different gNBs to carry out modeling, online training and reasoning work suitable for the cell model. When a certain UE fails to load a model, firstly, relevant information is collected by the current gNB, and the UE is controlled to carry out model rollback; on the other hand, the gNB informs the core network to perform model initialization again for the UE according to the current UE related conditions.

In the process of carrying out the Federal learning of the CSI-RS feedback inference network model, the Federal learning management module in the gNB intelligent agent adopts a method based on horizontal Federal learning in the application, simultaneously monitors the Federal learning process of a plurality of UE using the same model, processes and integrates model training results, and realizes the online learning and performance optimization of a specific model.

And the model performance evaluation submodule on the core network side is used for carrying out online learning and performance analysis on the candidate model imported by the background, and optimizing and evaluating the model reasoning capability from the aspects of different layers, different scenes and the like.

It can be understood that the channel state information feedback system provided by the invention manages and maintains a structure of the CSI-RS feedback inference network based on federal learning by deploying the CSI-RS feedback inference network intelligent agents on the network side and the terminal side, and realizes the safety protection and behavior tracing of the artificial intelligent system of the CSI-RS feedback inference network of the cell.

In the embodiment of the invention, an intelligent agent of a 'CSI-RS feedback inference network' is used as a core, and a federal learning mechanism jointly participated by terminals in a cell is adopted to realize the on-line training and performance monitoring of various 'CSI-RS feedback inference network' models. Through the management and maintenance of the intelligent agent of the CSI-RS feedback reasoning network, the functions of model updating, total and partial weight updating of the existing model, online model learning, performance evaluation, reasoning and the like are performed on a coder of the CSI-RS feedback reasoning network used by the UE in the cell and a decoder matched with the coder on the network side.

It can be understood that, in the embodiment of the present invention, each base station may be responsible for organizing terminals within the coverage area of the base station itself to perform the "AI-based CSI feedback" function. Due to the adoption of an on-line training and federal learning mechanism, the model is continuously trained through real service data each time in the using process of the model, and the weight of the current model is optimized according to the set federal learning mechanism. After a period of federal learning, the model performance evaluation submodule on the base station side finds that the performance of the current model is obviously improved compared with that of the original model, the model weight is updated, and meanwhile, the new model is reported to an intelligent management entity on the core network side. In addition, the model performance evaluation submodule on the base station side can set various different models to a plurality of different terminals in a cell, respectively learn the models and evaluate the effectiveness of each model according to the obtained result.

According to the channel state information feedback system, the CSI-RS feedback reasoning network intelligent bodies are respectively arranged on the network side and the terminal side, and the CSI-RS feedback reasoning network intelligent bodies on the network side and the terminal side are used for realizing federal learning of a specific feedback reasoning network model and obtaining the feedback reasoning network model with better performance, so that the accuracy of channel state information feedback can be improved.

The following describes the channel state information feedback method provided by the present invention, and the channel state information feedback method described below and the channel state information feedback system described above may be referred to correspondingly.

Fig. 2 is a schematic flow chart of a channel state information feedback method provided by the present invention, and as shown in fig. 2, the channel state information feedback method provided by the present invention is applied to a first CSI-RS feedback inference network agent, and may include:

200, under the condition that a network side is determined to receive a connection establishment request message sent by a target terminal side, establishing connection with a second CSI-RS feedback inference network intelligent agent corresponding to the target terminal side;

step 210, judging whether the code word of the first CSI-RS feedback inference network encoder model currently used by the target terminal side is the same as the code word of the first CSI-RS feedback inference network decoder model currently used by the network side through information interaction with the second CSI-RS feedback inference network agent;

and step 230, under the condition that the first CSI-RS feedback inference network encoder model and the first CSI-RS feedback inference network decoder model are the same, establishing a first CSI-RS feedback inference network model between the network side and the target terminal side.

Specifically, in the embodiment of the present invention, when a first CSI-RS feedback inference network agent determines that a network side receives a connection establishment request message sent by a target terminal side, the first CSI-RS feedback inference network agent may establish a connection with a second CSI-RS feedback inference network agent corresponding to the target terminal side, then, through information interaction with the second CSI-RS feedback inference network agent, determine whether a code word of a first CSI-RS feedback inference network encoder model currently used by the target terminal side is the same as a code word of a first CSI-RS feedback inference network decoder model currently used by the network side, and establish a first CSI-RS feedback inference network model between the network side and the target terminal side based on the first CSI-RS feedback inference network encoder model and the first CSI-RS feedback inference network decoder model when the determination is the same.

For example, the UE establishes connection with the gNB, when the UE is started to register or is switched to a new cell, a second CSI-RS feedback inference network intelligent body on the UE side establishes connection with a first CSI-RS feedback inference network intelligent body on the gNB side, and a CSI-RS feedback inference network model in the CSI-RS feedback inference network intelligent body on the UE side is updated.

According to the channel state information feedback method, the first CSI-RS feedback reasoning network intelligent body is connected with the second CSI-RS feedback reasoning network intelligent body, whether the code word of the first CSI-RS feedback reasoning network encoder model currently used by the target terminal side is matched with the first CSI-RS feedback reasoning network decoder model currently used by the network side or not is determined through information interaction, and the first CSI-RS feedback reasoning network encoder model and the first CSI-RS feedback reasoning network decoder model are used for establishing the first CSI-RS feedback reasoning network model between the network side and the target terminal side under the condition that the code word is matched with the first CSI-RS feedback reasoning network encoder model and the first CSI-RS feedback reasoning network decoder model, so that the accuracy of channel state information feedback can be improved.

Optionally, after the first CSI-RS feedback inference network model between the network side and the target terminal side is established, the method further includes:

Specifically, after a first CSI-RS feedback inference network model between a network side and a target terminal side is established, the first CSI-RS feedback inference network agent may determine whether the first CSI-RS feedback inference network encoder model is available in a cell, and send first indication information carrying a first model update message to the second CSI-RS feedback inference network agent when the first CSI-RS feedback inference network encoder model is determined to be unavailable, where the first indication information is used to indicate the second CSI-RS feedback inference network agent to update the first CSI-RS feedback inference network encoder model based on the first model update message.

Optionally, the first model update message may include weight information that updates the CSI-RS feedback inference network model.

Optionally, the channel state information feedback method provided by the present invention further includes:

and sending second indication information carrying a second model update message to the second CSI-RS feedback inference network intelligent agent under the condition that the first CSI-RS feedback inference network encoder model is determined to be available in the cell and part or all of the first weights are unavailable, wherein the second indication information is used for indicating the second CSI-RS feedback inference network intelligent agent to update the first CSI-RS feedback inference network encoder model based on the second model update message.

Specifically, in the embodiment of the present invention, the first CSI-RS feedback inference network agent may determine whether the first weight corresponding to the first CSI-RS feedback inference network encoder model and the first CSI-RS feedback inference network encoder model is available in the cell or not, when it is determined that the code word of the first CSI-RS feedback inference network encoder model currently used by the target terminal side is different from the code word of the first CSI-RS feedback inference network decoder model currently used by the network side; and sending second indication information carrying a second model updating message to the second CSI-RS feedback inference network intelligent agent under the condition that the first CSI-RS feedback inference network encoder model is available in the cell and part of or all of the first weights are unavailable, wherein the second indication information is used for indicating the second CSI-RS feedback inference network intelligent agent to update the first CSI-RS feedback inference network encoder model based on the second model updating message.

Optionally, the second model update message may include weight information that updates the CSI-RS feedback inference network model.

Optionally, the second indication information may be used to indicate that the second CSI-RS feedback inference network agent updates part or all of the first weights corresponding to the first CSI-RS feedback inference network encoder model based on the second model update message.

Specifically, in the embodiment of the present invention, when it is determined that the coding codeword of the first CSI-RS feedback inference network encoder model currently used by the target terminal side is different from the decoding codeword of the first CSI-RS feedback inference network decoder model currently used by the network side, the first CSI-RS feedback inference network agent may send exception fallback information to the second CSI-RS feedback inference network agent, where the exception fallback information may be used to instruct the second CSI-RS feedback inference network agent to perform exception fallback on the first CSI-RS feedback inference network encoder model.

It can be understood that when the CSI inference network model is not available, the decoder model at the network side and the encoder model at the terminal side need to be updated simultaneously. In the embodiment of the invention, the CSI inference neural network model can be replaced or not replaced, and only the weight parameter corresponding to the CSI inference neural network model is returned to the last available version.

Optionally, after sending the abnormal fallback information to the second CSI-RS feedback inference network agent, the method further includes:

determining a second CSI-RS feedback inference network encoder model obtained after the second CSI-RS feedback inference network agent performs exception backspacing on the first CSI-RS feedback inference network encoder model;

Specifically, after the first CSI-RS feedback inference network agent sends the exception rollback information to the second CSI-RS feedback inference network agent, it may be determined that the second CSI-RS feedback inference network encoder model is obtained after the second CSI-RS feedback inference network agent performs the exception rollback on the first CSI-RS feedback inference network encoder model; and then loading a second CSI-RS feedback inference network decoder model matched with the second CSI-RS feedback inference network encoder model, and establishing a second CSI-RS feedback inference network model between the network side and the target terminal side based on the second CSI-RS feedback inference network encoder model and the second CSI-RS feedback inference network decoder model.

Specifically, in the embodiment of the present invention, the first CSI-RS feedback inference network agent may perform online training on the second CSI-RS feedback inference network model in cooperation with the second CSI-RS feedback inference network agent based on a federal learning mechanism; after the period of training, optimizing a second weight corresponding to the second CSI-RS feedback inference network model based on a federal learning mechanism under the condition that the performance of the second CSI-RS feedback inference network model is determined to meet a first preset threshold.

Optionally, the first preset threshold may be that the accuracy of the inference times of the second CSI-RS feedback inference network model reaches a certain preset value, or a certain preset value is increased.

Specifically, in the embodiment of the present invention, the first CSI-RS feedback inference network agent determines the candidate CSI-RS feedback inference network model provided by the third CSI-RS feedback inference network agent on the condition that it is determined that the performance of the second CSI-RS feedback inference network model satisfies the second preset threshold, and performs federated learning on the candidate CSI-RS feedback inference network model based on a federated learning mechanism.

Specifically, in the embodiment of the present invention, while the first CSI-RS feedback inference network agent performs the federal online training process, the training result information of the local CSI-RS feedback inference network model may be counted; and then reporting the training result information to a third CSI-RS feedback reasoning network intelligent agent so that the third CSI-RS feedback reasoning network intelligent agent can manage the full-network candidate CSI-RS feedback reasoning network model based on the model training result information.

According to the channel state information feedback method, the first CSI-RS feedback reasoning network intelligent body is connected with the second CSI-RS feedback reasoning network intelligent body, whether the code word of the first CSI-RS feedback reasoning network encoder model currently used by the target terminal side is matched with the first CSI-RS feedback reasoning network decoder model currently used by the network side is determined through information interaction, and the first CSI-RS feedback reasoning network encoder model and the first CSI-RS feedback reasoning network decoder model are used for establishing the first CSI-RS feedback reasoning network model between the network side and the target terminal side under the condition that the code word is matched with the first CSI-RS feedback reasoning network decoder model currently used by the network side, so that the accuracy of channel state information feedback can be improved.

Fig. 3 is a schematic physical structure diagram of an electronic device provided in the present invention, and as shown in fig. 3, the electronic device may include: a processor (processor) 310, a communication Interface (communication Interface) 320, a memory (memory) 330 and a communication bus 340, wherein the processor 310, the communication Interface 320 and the memory 330 communicate with each other via the communication bus 340. The processor 310 may call logic instructions in the memory 330 to execute the channel state information feedback method provided by the above methods, which includes:

In addition, the logic instructions in the memory 330 may be implemented in the form of software functional units and stored in a computer readable storage medium when the software functional units are sold or used as independent products. Based on such understanding, the technical solution of the present invention may be embodied in the form of a software product, which is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and other various media capable of storing program codes.

In another aspect, the present invention also provides a computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions which, when executed by a computer, enable the computer to perform the channel state information feedback method provided by the above methods, the method comprising:

In yet another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored, the computer program being implemented by a processor to perform the above-provided channel state information feedback method, the method including:

The above-described embodiments of the apparatus are merely illustrative, and the units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the present embodiment. One of ordinary skill in the art can understand and implement it without inventive effort.

Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented by software plus a necessary general hardware platform, and certainly can also be implemented by hardware. With this understanding in mind, the above-described technical solutions may be embodied in the form of a software product, which can be stored in a computer-readable storage medium such as ROM/RAM, magnetic disk, optical disk, etc., and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the methods described in the embodiments or some parts of the embodiments.

Finally, it should be noted that: the above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.

Claims

1. A channel state information feedback system, comprising:

the first CSI-RS feedback reasoning network intelligent agent is used for sending a CSI-RS feedback reasoning network model to a second CSI-RS feedback reasoning network intelligent agent corresponding to a terminal side in a cell, and on the basis of a federal learning mechanism, the first CSI-RS feedback reasoning network intelligent agent is matched with the second CSI-RS feedback reasoning network intelligent agent to carry out on-line training on the CSI-RS feedback reasoning network model;

2. The csi feedback system of claim 1, further comprising: a third CSI-RS feedback reasoning network agent deployed on the core network side;

3. A channel state information feedback method applied to the channel state information feedback system of claim 1 or 2, wherein the method applied to the first CSI-RS feedback inference network agent comprises:

judging whether the coding code word of the first CSI-RS feedback inference network encoder model currently used by the target terminal side is the same as the decoding code word of the first CSI-RS feedback inference network decoder model currently used by the network side through information interaction with the second CSI-RS feedback inference network agent;

under the condition that the determination is the same, establishing a first CSI-RS feedback inference network model between the network side and the target terminal side based on the first CSI-RS feedback inference network encoder model and the first CSI-RS feedback inference network decoder model.

4. The channel state information feedback method according to claim 3, wherein after establishing the first CSI-RS feedback inference network model between the network side and the target terminal side, the method further comprises:

and sending first indication information carrying a first model updating message to the second CSI-RS feedback inference network agent under the condition that the information is determined to be unavailable, wherein the first indication information is used for indicating the second CSI-RS feedback inference network agent to update the first CSI-RS feedback inference network encoder model based on the first model updating message.

5. The channel state information feedback method according to claim 3, further comprising:

6. The channel state information feedback method of claim 3, further comprising:

7. The method of claim 6, wherein after sending the abnormal fallback information to the second CSI-RS feedback inference network agent, further comprising:

8. The channel state information feedback method according to claim 7, further comprising:

on the basis of a federal learning mechanism, the second CSI-RS feedback reasoning network intelligent agent is matched to perform on-line training on the second CSI-RS feedback reasoning network model;

9. The channel state information feedback method according to claim 8, further comprising:

and under the condition that the performance of the second CSI-RS feedback reasoning network model is determined to meet a second preset threshold, determining a candidate CSI-RS feedback reasoning network model provided by a third CSI-RS feedback reasoning network agent, and carrying out federal learning on the candidate CSI-RS feedback reasoning network model based on a federal learning mechanism.

10. The channel state information feedback method according to claim 9, further comprising: