WO2023125985A1 - 模型的数据处理方法及装置 - Google Patents

模型的数据处理方法及装置 Download PDF

Info

Publication number
WO2023125985A1
WO2023125985A1 PCT/CN2022/144185 CN2022144185W WO2023125985A1 WO 2023125985 A1 WO2023125985 A1 WO 2023125985A1 CN 2022144185 W CN2022144185 W CN 2022144185W WO 2023125985 A1 WO2023125985 A1 WO 2023125985A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
data
training
output
input
Prior art date
Application number
PCT/CN2022/144185
Other languages
English (en)
French (fr)
Inventor
柴晓萌
孙琰
吴艺群
Original Assignee
华为技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 华为技术有限公司 filed Critical 华为技术有限公司
Publication of WO2023125985A1 publication Critical patent/WO2023125985A1/zh

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/14—Network analysis or design
    • H04L41/145—Network analysis or design involving simulating, designing, planning or modelling of a network
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/08—Configuration management of networks or network elements
    • H04L41/0803—Configuration setting
    • H04L41/0813—Configuration setting characterised by the conditions triggering a change of settings
    • H04L41/082—Configuration setting characterised by the conditions triggering a change of settings the condition being updates or upgrades of network functionality
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/08—Configuration management of networks or network elements
    • H04L41/0896—Bandwidth or capacity management, i.e. automatically increasing or decreasing capacities
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/14—Network analysis or design
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/16—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks using machine learning or artificial intelligence

Definitions

  • the present application relates to the field of communication technology, and in particular to a method and device for processing model data.
  • a wireless communication network such as a mobile communication network
  • services supported by the network are becoming more and more diverse, and therefore requirements to be met are becoming more and more diverse.
  • the network needs to be able to support ultra-high speed, ultra-low latency, and/or very large connections.
  • This feature makes network planning, network configuration, and/or resource scheduling increasingly complex.
  • functions of the network become more and more powerful, such as supporting higher and higher spectrum, supporting high-order multiple input multiple output (MIMO) technology, supporting beamforming, and/or supporting new technologies such as beam management, etc. technology, making network energy saving a hot research topic.
  • MIMO multiple input multiple output
  • These new requirements, new scenarios, and new features have brought unprecedented challenges to network planning, O&M, and efficient operations.
  • artificial intelligence technology can be introduced into the wireless communication network to realize network intelligence. Based on this, how to effectively implement artificial intelligence in the network is a problem worth studying.
  • the application provides a model data processing method and device, which can use one model to complete inference tasks under different wireless resource configurations, reducing model training and storage costs.
  • a data processing method for a model is provided.
  • the execution subject of the method is the first node, and the first node is a model training node or a model reasoning node, etc., and can also be configured as a component (processor, chip) in the first node or other), or may be a software module, the method includes: determining the data processing method of the model; according to the data processing method of the model, at least one of the following is realized: performing model training, or performing model reasoning.
  • the data processing method of the model includes at least one of the following: the input data filling method of the model, or the output data interception method of the model; or, the data processing method of the model includes at least one of the following : The input data interception method of the model, or the output data filling method of the model.
  • the model reasoning node can use the same model to complete reasoning tasks under different wireless resource configurations according to the data processing method. Compared with configuring corresponding models for each wireless resource configuration, the overhead of model training and model storage can be reduced.
  • the determining the data processing method of the model includes: receiving indication information from the second node, where the indication information is used to indicate the data processing method of the model.
  • the first node is different from the second node.
  • the first node is a model reasoning node or a model training node, and can receive indication information from other nodes (such as a second node), the indication information is used to indicate the data processing method of the model, for example, the second node is OAM wait.
  • the second node can be a model training node, and the model inference node can receive instruction information from the model training node, etc.
  • determine the data processing method of the model is a model reasoning node or a model training node, and can receive indication information from other nodes (such as a second node).
  • the indication information is used to indicate the data processing method of the model
  • the second node is OAM wait.
  • the second node can be a model training node, and the model inference node can receive instruction information from the model training node
  • the model training node or the model reasoning node can obtain the data processing method of the model through agreement agreement. Or, receive instructions from other nodes, and determine the data processing method of the model according to the instructions of other nodes. Model training node or model inference node can flexibly determine the processing method of model data.
  • the input data filling method or the output data filling method includes at least one of the following: length of filled data, type of filled data, or rule of filled data.
  • the output data interception method or the input data interception method includes at least one of the following: the length of the intercepted data, or a rule for intercepting the data.
  • the performing model reasoning according to the data processing method of the model includes: performing data filling on the first input data according to the input data filling method to obtain second input data; Two input data and the model, determining first output data; performing data interception on the first output data according to the output data interception method.
  • the model reasoning node fills the input data and intercepts the output data, which can make use of one model to complete reasoning tasks under different wireless resource configurations, reducing the training and storage costs of the model.
  • the input length of the model is greater than or equal to the length of the input data with the largest input data length among all potential radio resource configurations
  • the output length is greater than or equal to the length of the largest output data length among all potential radio resource configurations The length of the output data.
  • the performing model training according to the data processing method of the model includes: performing data filling on the first input training data according to the input data filling method to obtain second input training data;
  • the second input training data and the model are used to determine the first output training data; according to the output data interception method, the first output training data is intercepted to obtain the second output training data; according to the second output training data, and adjust the parameters of the model.
  • model training collect training data under various wireless resource configurations as much as possible; use data processing methods to process the training data under each wireless resource configuration; use the processed training data Train the AI model until the AI model satisfies the conditions, then end the model training.
  • the learning method in the model training process is not limited, such as supervised learning (label learning), unsupervised learning (non-label learning), or reinforcement learning.
  • the model trained by this method can meet the reasoning tasks under various wireless resource configurations and reduce the model training overhead.
  • the performing model reasoning according to the data processing method of the model includes: performing data interception on first input data according to the input data interception method to obtain second input data; data and the model, determine first output data; and perform data filling on the first output data according to the output data filling method.
  • the model reasoning node intercepts the input data and fills the output data, which can make use of one model to complete reasoning tasks under different wireless resource configurations, reducing the training and storage costs of the model.
  • the model training according to the data processing method of the model includes: performing data interception on the first input training data according to the input data interception method to obtain the second input training data; according to the first input data interception method Two input training data and the model, determine the first output training data; according to the output data filling method, perform data filling on the first output training data to obtain second output training data; according to the second output training data to adjust the parameters of the model.
  • the model parameters are adjusted, including adjusting at least one of the following: the number of layers of the neural network, the width of the neural network, the connection relationship of the layers, the weight value of the neuron, the neuron The activation function, or the bias in the activation function, so that the difference between the output of the neural network cell and the ideal target value is as small as possible.
  • the model includes a first sub-model
  • performing model inference according to the data processing method of the model includes: determining the first output data of the first sub-model; intercepting the output data according to the method , performing data interception on the first output data.
  • the model includes a second sub-model
  • performing model inference according to the data processing method of the model includes: performing data filling on the first input data according to the input data filling method to obtain second input data; determining first output data according to the second input data and the second sub-model.
  • multiple paired sub-models can be trained simultaneously, for example, the above-mentioned first sub-model can be deployed on the terminal side, and the second sub-model can be deployed on the base station side.
  • air interface overhead can be reduced by intercepting the output data of the first sub-model.
  • the model includes a first sub-model and a second sub-model
  • performing model training according to the data processing method of the model includes: input training data according to the first sub-model and the first sub-model model, determining the first output training data of the first sub-model; according to the output data interception method, performing data interception on the first output training data to obtain the first input training data of the second sub-model; according to the input A data filling method, performing data filling on the first input training data of the second sub-model to obtain second input training data; according to the second input training data and the second sub-model, determining the output training data; according to the input training data of the first sub-model and the output training data of the second sub-model, adjust at least one of the following: model parameters of the first sub-model, or, the second sub-model Model parameters for the second submodel.
  • a data processing method for a model is provided.
  • the execution subject of the method is the first node, and the first node is a model training node or a model reasoning node, etc., and can also be configured as a component (processor, chip) in the first node or other), or may be a software module, including: sending instruction information to the second node, where the instruction information is used to indicate the data processing method of the model.
  • the data processing method of the model includes at least one of the following: the input data filling method of the model, or the output data interception method of the model; or, the data processing method of the model includes at least one of the following: A method for intercepting input data of the model, or a method for filling output data of the model.
  • the first node is a model training node
  • the model training node can determine a data processing method according to the data processing method during model training, and indicate the data processing method to the model reasoning node.
  • the model reasoning node executes model reasoning, it processes data in the same way as during training, so that the model reasoning node can correctly use the model for model reasoning, improving the accuracy and reasoning speed of model reasoning output results .
  • the input data filling method or the output data filling method includes at least one of the following: length of filled data, type of filled data, or rule of filled data.
  • the output data interception method or the input data interception method includes at least one of the following: the length of the intercepted data, or a rule for intercepting the data.
  • a device is provided.
  • the device may be a first node, and the first node may be a model training node or a model reasoning node, or a device configured in the first node, or A device that can be used in conjunction with the first node.
  • the device includes a one-to-one unit for performing the method/operation/step/action described in the first aspect, and the unit may be a hardware circuit, or software, or a combination of hardware circuit and software.
  • the device may include a processing unit, and the processing unit may perform the corresponding functions in any design example of the first aspect above, specifically:
  • the processing unit is used to determine the data processing method of the model, and, according to the data processing method of the model, implement at least one of the following: perform model training, or perform model reasoning;
  • the data processing method of the model includes at least one of the following: the input data filling method of the model, or the output data interception method of the model; or, the data processing method of the model includes at least one of the following: A method for intercepting input data of the model, or a method for filling output data of the model.
  • the device includes a processor, configured to implement the method described in the first aspect above.
  • the apparatus may also include memory for storing instructions and/or data.
  • the memory is coupled to the processor, and when the processor executes the program instructions stored in the memory, the method described in the first aspect above can be implemented.
  • the device includes:
  • the processor is used to determine the data processing method of the model, and implement at least one of the following according to the data processing method of the model: perform model training, or perform model reasoning;
  • the data processing method of the model includes at least one of the following: the input data filling method of the model, or the output data interception method of the model; or, the data processing method of the model includes at least one of the following: A method for intercepting input data of the model, or a method for filling output data of the model.
  • a device is provided.
  • the device may be the first node, and the first node may be a model training node, or a device configured in the first node, or capable of communicating with the first node.
  • the node matches the device used, etc.
  • the device includes a one-to-one unit for performing the methods/operations/steps/actions described in the second aspect.
  • the unit may be a hardware circuit, or software, or a combination of hardware circuit and software.
  • the device may include a communication unit, and the communication unit may perform the corresponding functions in any design example of the second aspect above, specifically:
  • a communication unit configured to send instruction information to the second node, where the instruction information is used to indicate a data processing method of the model
  • the data processing method of the model includes at least one of the following: the input data filling method of the model, or the output data interception method of the model; or, the data processing method of the model includes at least one of the following: A method for intercepting input data of the model, or a method for filling output data of the model.
  • the device includes a processor, configured to control the communication interface to implement the method described in the second aspect above.
  • the apparatus may also include memory for storing instructions and/or data.
  • the memory is coupled to the processor, and when the processor executes the program instructions stored in the memory, the method described in the second aspect above can be implemented.
  • the device may also include a communication interface for the device to communicate with other devices.
  • the communication interface may be a transceiver, a circuit, a bus, a module, a pin or other types of communication interfaces, and other devices may be model inference nodes and the like.
  • the device includes:
  • a processor configured to control the communication interface to send instruction information to the second node, where the instruction information is used to indicate a data processing method of the model
  • the data processing method of the model includes at least one of the following: the input data filling method of the model, or the output data interception method of the model; or, the data processing method of the model includes at least one of the following: A method for intercepting input data of the model, or a method for filling output data of the model.
  • the present application further provides a computer-readable storage medium, including instructions, which, when run on a computer, cause the computer to execute the method of any one of the first aspect or the second aspect.
  • the present application further provides a system-on-a-chip, which includes a processor and may further include a memory, for implementing the method in any one of the first aspect or the second aspect.
  • the system-on-a-chip may consist of chips, or may include chips and other discrete devices.
  • the present application further provides a computer program product, including instructions, which, when run on a computer, cause the computer to execute the method of any one of the first aspect or the second aspect.
  • the present application further provides a system, which includes the device of the third aspect and the device of the fourth aspect.
  • FIG. 1 is a schematic diagram of a communication system provided by the present application.
  • FIG. 1 and Figure 3 are schematic diagrams of the deployment of the AI model provided by this application.
  • FIG. 4a and Figure 4b are schematic diagrams of the architecture of the communication system provided by the present application.
  • Figure 4c is a schematic diagram of training an AI model for each bandwidth provided by the present application.
  • FIG. 5 is a schematic diagram of the application architecture of the AI model provided by the present application.
  • Fig. 6 is the schematic diagram of the neuron provided by the present application.
  • Fig. 7 is the schematic diagram of the neural network provided by the present application.
  • FIG 8 and Figure 9A are flow charts of the model data processing method provided by the present application.
  • Figure 9B and Figure 9C are the data processing process of the model training phase and the model reasoning phase respectively;
  • FIG. 10 and FIG. 11 are schematic diagrams of a type 1 reference signal and a type 2 reference signal, respectively;
  • Figure 12 is a schematic diagram of the CSI feedback process provided by this application.
  • FIG. 13 is a schematic diagram of data processing in the CSI feedback process provided by this application.
  • Figure 14 is a schematic diagram of the Transformer model provided by the present application.
  • Figure 15 and Figure 16 are schematic diagrams of splitting raw data into vectors provided by the present application.
  • FIG. 17 and FIG. 18 are schematic structural diagrams of a communication device provided in the present application.
  • FIG. 1 is a schematic structural diagram of a communication system 1000 to which the present application can be applied.
  • the communication system includes a radio access network 100 and a core network 200 , and optionally, the communication system 1000 may also include the Internet 300 .
  • the radio access network 100 may include at least one access network device (such as 110a and 110b in FIG. 1 ), and may also include at least one terminal device (such as 120a-120j in FIG. 1 ).
  • the terminal device is connected to the access network device in a wireless manner, and the access network device is connected to the core network in a wireless or wired manner.
  • the core network device and the access network device can be independent and different physical devices, or the functions of the core network device and the logical functions of the access network device can be integrated on the same physical device, or they can be integrated on one physical device Part of the functions of the core network device and part of the functions of the access network device are specified.
  • the terminal equipment and the terminal equipment and between the access network equipment and the access network equipment may be connected to each other through wired or wireless methods.
  • FIG. 1 is only a schematic diagram.
  • the communication system may also include other network devices, such as wireless relay devices and wireless backhaul devices, which are not shown in FIG. 1 .
  • the access network equipment can be a base station (base station), an evolved base station (evolved NodeB, eNodeB), a transmission reception point (transmission reception point, TRP), and a next-generation base station in the fifth generation (5th generation, 5G) mobile communication system (next generation NodeB, gNB), access network equipment in the open radio access network (open radio access network, O-RAN), next-generation base stations in the sixth generation (6th generation, 6G) mobile communication system, future mobile
  • DU distributed unit
  • CU control plane centralized unit control plane
  • CU-CP centralized unit control plane
  • CU user plane centralized unit user plane
  • the access network device may be a macro base station (such as 110a in Figure 1), a micro base station or an indoor station (such as 110b in Figure 1), or a relay node or a donor node.
  • a macro base station such as 110a in Figure 1
  • a micro base station such as 110b in Figure 1
  • a relay node or a donor node.
  • the device used to realize the function of the access network device may be the access network device; it may also be a device capable of supporting the access network device to realize the function, such as a chip system, a hardware circuit, a software module, or a hardware A circuit plus a software module, the device can be installed in the access network equipment or can be matched with the access network equipment for use.
  • a system-on-a-chip may consist of a chip, or may include a chip and other discrete devices.
  • the technical solutions provided by the present application are described below by taking the apparatus for realizing the functions of the access network equipment as the access network equipment and the access network equipment as the base station as an example.
  • the protocol layer structure may include a control plane protocol layer structure and a user plane protocol layer structure.
  • the control plane protocol layer structure may include a radio resource control (radio resource control, RRC) layer, a packet data convergence protocol (packet data convergence protocol, PDCP) layer, a radio link control (radio link control, RLC) layer, a media The access control (media access control, MAC) layer and the function of the protocol layer such as the physical layer.
  • the user plane protocol layer structure may include the functions of the PDCP layer, the RLC layer, the MAC layer, and the physical layer.
  • the PDCP layer may also include a service data adaptation protocol (service data adaptation protocol). protocol, SDAP) layer.
  • the protocol layer structure between the access network device and the terminal device may further include an artificial intelligence (AI) layer, which is used to transmit data related to the AI function.
  • AI artificial intelligence
  • Access devices may include CUs and DUs. Multiple DUs can be centrally controlled by one CU.
  • the interface between the CU and the DU may be referred to as an F1 interface.
  • the control plane (control panel, CP) interface may be F1-C
  • the user plane (user panel, UP) interface may be F1-U. This application does not limit the specific names of each interface.
  • CU and DU can be divided according to the protocol layer of the wireless network: for example, the functions of the PDCP layer and above protocol layers are set in the CU, and the functions of the protocol layers below the PDCP layer (such as RLC layer and MAC layer, etc.) are set in the DU; another example, PDCP The functions of the protocol layer above the layer are set in the CU, and the functions of the PDCP layer and the protocol layer below are set in the DU, without restriction.
  • the CU or DU may be divided into functions having more protocol layers, and for example, the CU or DU may also be divided into part processing functions having protocol layers.
  • part of the functions of the RLC layer and the functions of the protocol layers above the RLC layer are set in the CU, and the rest of the functions of the RLC layer and the functions of the protocol layers below the RLC layer are set in the DU.
  • the functions of the CU or DU can also be divided according to the business type or other system requirements, for example, according to the time delay, and the functions whose processing time needs to meet the time delay requirement are set in the DU, which does not need to meet the time delay
  • the required feature set is in the CU.
  • the CU may also have one or more functions of the core network.
  • the CU can be set on the network side to facilitate centralized management.
  • the wireless unit (radio unit, RU) of the DU is set remotely.
  • the RU may have a radio frequency function.
  • DUs and RUs can be divided in a physical layer (physical layer, PHY).
  • the DU can implement high-level functions in the PHY layer
  • the RU can implement low-level functions in the PHY layer.
  • the functions of the PHY layer may include at least one of the following: adding a cyclic redundancy check (cyclic redundancy check, CRC) code, channel coding, rate matching, scrambling, modulation, layer mapping, precoding, Resource mapping, physical antenna mapping, or radio frequency transmission functions.
  • CRC cyclic redundancy check
  • the functions of the PHY layer may include at least one of the following: CRC check, channel decoding, de-rate matching, descrambling, demodulation, de-layer mapping, channel detection, resource de-mapping, physical antenna de-mapping, or RF receiving function.
  • the high-level functions in the PHY layer may include a part of the functions of the PHY layer, for example, this part of the functions is closer to the MAC layer, and the lower-level functions in the PHY layer may include another part of the functions of the PHY layer, for example, this part of the functions is closer to the radio frequency function.
  • high-level functions in the PHY layer may include adding CRC codes, channel coding, rate matching, scrambling, modulation, and layer mapping
  • low-level functions in the PHY layer may include precoding, resource mapping, physical antenna mapping, and radio transmission functions
  • high-level functions in the PHY layer may include adding CRC codes, channel coding, rate matching, scrambling, modulation, layer mapping, and precoding
  • low-level functions in the PHY layer may include resource mapping, physical antenna mapping, and radio frequency send function.
  • the high-level functions in the PHY layer may include CRC check, channel decoding, de-rate matching, decoding, demodulation, and de-layer mapping
  • the low-level functions in the PHY layer may include channel detection, resource de-mapping, physical antenna de-mapping, and RF receiving functions
  • the high-level functions in the PHY layer may include CRC check, channel decoding, de-rate matching, decoding, demodulation, de-layer mapping, and channel detection
  • the low-level functions in the PHY layer may include resource de-mapping , physical antenna demapping, and RF receiving functions.
  • the function of the CU may be implemented by one entity, or may also be implemented by different entities.
  • the functions of the CU can be further divided, that is, the control plane and the user plane are separated and realized by different entities, namely, the control plane CU entity (ie, the CU-CP entity) and the user plane CU entity (ie, the CU-UP entity) .
  • the CU-CP entity and CU-UP entity can be coupled with the DU to jointly complete the functions of the access network equipment.
  • any one of the foregoing DU, CU, CU-CP, CU-UP, and RU may be a software module, a hardware structure, or a software module+hardware structure, without limitation.
  • the existence forms of different entities may be different, which is not limited.
  • DU, CU, CU-CP, and CU-UP are software modules
  • RU is a hardware structure.
  • the access network device includes CU-CP, CU-UP, DU and RU.
  • the execution subject of this application includes DU, or includes DU and RU, or includes CU-CP, DU and RU, or includes CU-UP, DU and RU, without limitation.
  • the methods executed by each module are also within the protection scope of the present application.
  • the terminal equipment may also be called a terminal, user equipment (user equipment, UE), mobile station, mobile terminal equipment, and the like.
  • Terminal devices can be widely used in communication in various scenarios, including but not limited to at least one of the following scenarios: device-to-device (device-to-device, D2D), vehicle-to-everything (V2X), machine-type communication ( machine-type communication (MTC), Internet of Things (IOT), virtual reality, augmented reality, industrial control, automatic driving, telemedicine, smart grid, smart furniture, smart office, smart wear, smart transportation, or intelligence city etc.
  • the terminal device can be a mobile phone, a tablet computer, a computer with wireless transceiver function, a wearable device, a vehicle, a drone, a helicopter, an airplane, a ship, a robot, a mechanical arm, or a smart home device, etc.
  • This application does not limit the specific technology and specific equipment form adopted by the terminal equipment.
  • the device used to realize the function of the terminal device may be a terminal device; it may also be a device capable of supporting the terminal device to realize the function, such as a chip system, a hardware circuit, a software module, or a hardware circuit plus a software module.
  • the device can be installed in the terminal equipment or can be matched with the terminal equipment for use.
  • the technical solution provided by the present application will be described below by taking the apparatus for realizing the functions of the terminal equipment as the terminal equipment and the terminal equipment as the UE as an example.
  • Base stations and terminal equipment can be fixed or mobile.
  • Base stations and/or terminal equipment can be deployed on land, including indoors or outdoors, handheld or vehicle-mounted; they can also be deployed on water; they can also be deployed on aircraft, balloons and artificial satellites in the air.
  • This application does not limit the application scenarios of the base station and terminal equipment.
  • the base station and the terminal device can be deployed in the same scene or in different scenarios. For example, the base station and the terminal device are deployed on land at the same time; or, the base station is deployed on land and the terminal device is deployed on water.
  • the helicopter or drone 120i in FIG. Device 120i is a base station; however, for base station 110a, 120i is a terminal device, that is, communication between 110a and 120i is performed through a wireless air interface protocol. Communication between 110a and 120i may also be performed through an interface protocol between base stations. In this case, relative to 110a, 120i is also a base station. Therefore, both the base station and the terminal equipment can be collectively referred to as a communication device, 110a and 110b in FIG. 1 can be referred to as a communication device with a base station function, and 120a-120j in FIG. 1 can be referred to as a communication device with a terminal device function.
  • an independent network element such as AI network element, or AI node, etc.
  • AI network element can be introduced into the communication system shown in Figure 1 to realize AI-related operations, and the AI network element can communicate with the interface in the communication system
  • Network access devices are directly connected, or can be indirectly connected through third-party network elements and access network devices.
  • the third-party network element can be a core network element such as authentication management function (authentication management function, AMF) or user plane function (user plane function, UPF); or, AI can be configured in other network elements in the communication system Functions, AI modules or AI entities to implement AI-related operations, for example, the other network elements can be access network equipment (such as gNB), core network equipment, or network management (operation, administration and maintenance, OAM), etc., in this
  • the network element performing AI-related operations is a network element with a built-in AI function.
  • the above-mentioned OAM is used to operate, manage and maintain the access network equipment and/or the core network equipment and the like.
  • At least one of the core network equipment, access network equipment, terminal equipment, or OAM can be deployed with an AI model, and use the AI model to implement corresponding functions.
  • the AI models deployed in different nodes can be the same or different, and the different models include at least one of the following differences: the structural parameters of the models are different, for example, the number of layers and/or weights of the models are different; the input reference of the models different; or the output reference of the model is different, etc.
  • different input parameters of the model and/or different output parameters of the model can be described as different functions of the model. Different from the above-mentioned FIG. 2, in FIG.
  • the function of the access network device is divided into CU and DU.
  • the CU and DU may be CU and DU under the O-RAN architecture.
  • One or more AI models can be deployed in the CU. And/or, one or more AI models may be deployed in the DU.
  • the CU in FIG. 3 may be further split into CU-CP and CU-UP.
  • one or more AI models may be deployed in the CU-CP.
  • one or more AI models may be deployed in the CU-UP.
  • the OAM of the access network device and the OAM of the core network device may be distributed and deployed independently.
  • FIG. 4a is a communication system architecture of the present application.
  • the access network equipment includes a near real-time access network intelligent controller (RAN intelligent controller, RIC) module for model training and inference.
  • RAN intelligent controller RIC
  • near real-time RIC can be used to train an AI model and use that AI model for inference.
  • the near real-time RIC can obtain network-side and/or terminal-side information from at least one of CU, DU, or RU, and this information can be used as training data or inference data.
  • the near real-time RIC may submit the reasoning result to at least one of CU, DU, RU or terminal device.
  • the inference results can be exchanged between the CU and the DU.
  • the reasoning results can be exchanged between the DU and the RU, for example, the near real-time RIC submits the reasoning result to the DU, and the DU forwards it to the RU.
  • the access network device includes a non-real-time RIC (optionally, the non-real-time RIC can be located in the OAM or in the core network device) for model training and reasoning.
  • non-real-time RIC is used to train an AI model and use that model for inference.
  • the non-real-time RIC can obtain network-side and/or terminal-side information from at least one of CU, DU, or RU, which can be used as training data or inference data, and the inference results can be submitted to CU, DU, or RU or at least one of the terminal devices.
  • the inference results can be exchanged between the CU and the DU.
  • the reasoning results can be exchanged between the DU and the RU, for example, the non-real-time RIC submits the reasoning result to the DU, and the DU forwards it to the RU.
  • the access network equipment includes a near-real-time RIC, and the access network equipment includes a non-real-time RIC (optionally, the non-real-time RIC can be located in the OAM or the core network device).
  • the non-real-time RIC can be used for model training and reasoning.
  • the near real-time RIC can be used for model training and reasoning.
  • the non-real-time RIC performs model training, and the near-real-time RIC can obtain AI model information from the non-real-time RIC, and obtain network-side and/or terminal-side information from at least one of CU, DU or RU, and use the information And the AI model information to get the reasoning result.
  • the near real-time RIC may submit the reasoning result to at least one of CU, DU, RU or terminal device.
  • the inference results can be exchanged between the CU and the DU.
  • the reasoning results can be exchanged between the DU and the RU, for example, the near real-time RIC submits the reasoning result to the DU, and the DU forwards it to the RU.
  • near real-time RIC is used to train model A and use model A for inference.
  • non-real-time RIC is used to train Model B and utilize Model B for inference.
  • the non-real-time RIC is used to train the model C, and the information of the model C is sent to the near-real-time RIC, and the near-real-time RIC uses the model C for inference.
  • Fig. 4b is an architecture of another communication system of the present application. Compared with Figure 4a, CU is separated into CU-CP and CU-UP in Figure 4b.
  • the bandwidth of the wireless signal is flexibly scheduled by the base station.
  • the bandwidth scheduled in one time slot may be 4 resource blocks (resource blocks, RBs), and the bandwidth scheduled in the next time slot may become 8 RBs.
  • the AI model The input format is fixed and cannot change with the bandwidth of the signal. For example, if the input format of the AI model corresponds to 4 RBs, the AI model can only support signals with an input of 4 RBs, but cannot process signals with an input of 8 RBs.
  • an AI model is trained for each bandwidth. As shown in Figure 4c, for a signal with an input bandwidth of 4 RB and a signal with an input bandwidth of 8 RB, an AI model is trained separately, which results in a large training and storage overhead.
  • a solution is provided in this application, in which: an AI model is trained, and the AI model is applicable to input data and output data in various formats. Specifically, during training, data in different formats (such as input data and/or output data) are converted into data in the same format by means of filling or interception, and an AI model is jointly trained. This AI model works on this multiple formats. When the AI model is used or reasoned, the input data and/or output data are converted into input data and/or output data that match the AI model through the same data processing method as that used for training, so that using the same The AI model can complete inference tasks under different wireless resource configurations, reducing the overhead of training and storage. It should be noted that at least two examples are included in the scheme of the present application.
  • the scheduling bandwidth in the wireless network includes 4RB, 8RB, 16RB and 32RB.
  • An AI model is trained for the above four scheduling bandwidths.
  • this example is mainly used as an example for illustration.
  • one AI model is designed for input data and/or output data in some formats, and another AI model is designed for input data and/or output data in other formats.
  • the first AI model is trained, that is, the first AI model is suitable for inputting 4RB and 8RB signals.
  • the second AI model is trained, that is, the second AI model is suitable for input of 16RB and 32RB signals, and the like.
  • the AI model is the specific realization of the AI function, and the AI model represents the mapping relationship between the input and output of the model.
  • the AI model can be a neural network, a linear regression model, a decision tree model, a support vector machine (SVM), a Bayesian network, a Q-learning model, or other machine learning models.
  • the AI function may include at least one of the following: data collection (collection of training data and/or reasoning data), data preprocessing, model training (or model learning), model information publishing (configuration model information), Model verification, model inference, or release of inference results.
  • reasoning can also be called prediction.
  • the AI model may be referred to as a model for short.
  • the design of the AI model mainly includes the data collection link (for example, the data source can collect training data and/or inference data), model training link and model inference link. It may further include the link of application of reasoning results.
  • a model testing link may also be included.
  • Figure 5 is a schematic diagram of an application architecture of the AI model.
  • Data source is used to provide training data and inference data.
  • the model training host obtains the AI model by analyzing or training the training data provided by the data source. Among them, learning the AI model through the model training node is equivalent to using the training data to learn the mapping relationship between the input and output of the model by the model training node.
  • the model training node can also update the AI model deployed on the model inference node.
  • the model inference node can also feed back information about the deployed model to the model training node, so that the model training node can optimize or update the deployed AI model.
  • the model inference node uses the AI model to perform inference based on the inference data provided by the data source, and obtains inference results.
  • This method can be realized as follows: the model inference node inputs the inference data into the AI model, and obtains an output through the AI model, and the output is the inference result.
  • the inference result may indicate: configuration parameters used (executed) by the execution object, and/or operations performed by the execution object.
  • the reasoning result can be uniformly planned by the execution (actor) entity, and sent to one or more execution objects (for example, network entities) for execution.
  • the execution entity or execution object can also feed back the performance of the model to the data source as training data, so as to facilitate subsequent model update training.
  • AI models can be neural networks or other machine learning models.
  • neural network is a specific implementation form of machine learning technology. According to the general approximation theorem, the neural network can theoretically approximate any continuous function, so that the neural network has the ability to learn any mapping. Therefore, neural networks can perform accurate abstract modeling of complex high-dimensional problems.
  • Each neuron performs a weighted sum operation on its input values, and passes the weighted sum result through an activation function to generate an output.
  • FIG. 6 it is a schematic diagram of a neuron structure.
  • the bias of the weighted sum for b The form of the activation function can be diversified.
  • the output of the neuron is:
  • the output of the neuron is: w i , x i and b can be various possible values such as decimals, integers (including 0, positive integers or negative integers, etc.), or complex numbers.
  • the activation functions of different neurons in a neural network can be the same or different.
  • a neural network generally includes a multi-layer structure, and each layer may include one or more neurons. Increasing the depth and/or width of a neural network can improve the expressive power of the neural network, providing more powerful information extraction and abstract modeling capabilities for complex systems.
  • the depth of the neural network may refer to the number of layers included in the neural network, and the number of neurons included in each layer may be referred to as the width of the layer.
  • Figure 7 it is a schematic diagram of the layer relationship of the neural network.
  • a neural network includes an input layer and an output layer. The input layer of the neural network processes the input received by neurons, and then passes the result to the output layer, and the output layer obtains the output result of the neural network.
  • a neural network in another implementation, includes an input layer, a hidden layer, and an output layer.
  • the input layer of the neural network processes the input received by neurons, and then passes the result to the hidden layer in the middle, and the hidden layer then passes the calculation result to the output layer or the adjacent hidden layer, and finally the output layer obtains the result of the neural network. Output the result.
  • a neural network may include one or more sequentially connected hidden layers without limitation.
  • a loss function can be defined.
  • the loss function describes the gap or difference between the output value of the neural network and the ideal target value, and the application does not limit the specific form of the loss function.
  • the training process of the neural network is to make the value of the loss function less than the threshold value or The process of meeting the target needs.
  • a loss function can be defined.
  • the loss function describes the gap or difference between the output value of the AI model and the ideal target value, and this application does not limit the specific form of the loss function.
  • the training process of the AI model is to adjust some or all parameters of the AI model so that for one or more training data, the value of the loss function or the weighted sum value (such as the average value) is less than the threshold value or meets the target requirements.
  • the AI model is a neural network
  • one or more of the following parameters can be adjusted during the training process: the number of layers of the neural network, the width of the neural network, the connection relationship of the layers, the weight value of the neuron, the activation function of the neuron, or A bias in the activation function such that the difference between the output of the neural network cell and the ideal target value is as small as possible.
  • the AI model is a neural network f ⁇ ( ⁇ ) as an example.
  • the model training node collects training data, which includes training samples and labels. For example, the training sample x is used as input, after being processed by the neural network f ⁇ ( ⁇ ), the inference result f ⁇ (x) is output, and the loss function calculates the difference between the inference result f ⁇ (x) and the label of the training sample x.
  • the model training node can use the model optimization algorithm to optimize the network parameters based on the difference obtained by the loss function.
  • the difference between the output of the neural network and the label of each training sample is less than the threshold value or meets the target Requirement, or, the weighted sum (for example, the average value) of the difference between the output of the neural network of all training samples and the label is less than the threshold value or meets the target requirement, so as to complete the training of the neural network.
  • the training of the AI model can also adopt unsupervised learning, and use the algorithm to learn the internal mode of the training samples, so as to complete the training of the AI model based on the training samples.
  • the training of the AI model can also use reinforcement learning to obtain the excitation signal of the environmental feedback through interaction with the environment, so as to learn the strategy to solve the problem and realize the optimization of the AI model.
  • the method of model training is not limited.
  • the execution body of the method is a model training node or a model inference node. It can be understood that the model training node and the model inference node can be the same node, or different nodes, etc. , without limitation.
  • the process includes at least:
  • Step 801 the first node determines the data processing method of the AI model.
  • the first node is a model training node or a model reasoning node.
  • the first node receives indication information from other nodes (such as the second node), where the indication information is used to indicate the data processing method of the AI model.
  • the first node determines the data processing method of the AI model according to the agreement.
  • the first node determines the data processing method of the AI model by itself.
  • the model inference node can determine the data processing method of the AI model according to the format of the collected training data and the format of the AI model to be trained. For details, please refer to the following description.
  • Step 802 According to the data processing method of the AI model, the first node implements at least one of the following: perform model training, or perform model inference.
  • the process of providing the data processing method includes at least the following steps:
  • Step 901 Collect training data.
  • the subject that collects training data may be a model training node, or other AI entities, or AI modules.
  • the collected training data includes but not limited to data in the wireless network, such as wireless channel information, received signal, reference signal, reference signal receiving power (reference signal receiving power, RSRP) and so on.
  • the wireless channel information may include estimated channel response, or channel characteristics, and the like.
  • a direct acquirer (or called an estimator, a measurer, or a collector, etc.) of these data may be a UE or a base station.
  • the direct acquirer of data sends the data to the entity or module performing data collection.
  • the collected training data may include measurement data of different UEs or measurement data collected by UEs in different geographic locations and in different channel environments.
  • the collected data may be real data obtained by the UE or the base station in an actual network, or may be virtual data generated through a simulation platform or a simulation platform, without limitation.
  • the collected training data includes data under multiple wireless resource configurations. For example, received signals under multiple bandwidths or multiple reference signal patterns, or wireless channel information, and the like.
  • Data under multiple wireless resource configurations is also generated from data under the same wireless resource configuration through a data enhancement method, for example, wireless channels under the same bandwidth are intercepted into wireless channels under different bandwidths.
  • the specific type of training data collected is related to the function of the AI model.
  • the training data required by the AI model includes at least the received reference signal Y and the original reference signal S.
  • the reference signal Y and the original reference signal S under various no-resource configurations can be collected as training data.
  • reference signal Y and original reference signal S under different bandwidths and/or reference signal types may be collected.
  • the training data required by the AI model includes at least the channel response H or the channel characteristic W, and different bandwidths and/or different numbers of antenna ports can be collected The following channel response H or channel characteristics W.
  • the training data required by the AI model includes at least the channel response H, the received signal Y, or the RSRP of the received signal Y, which can be collected under different bandwidths and/or different numbers of antenna ports Channel response H or receive signal Y or RSRP.
  • the training data required by the AI model includes at least a channel response H, and channel responses H under different bandwidths and/or different numbers of antenna ports may be collected.
  • Step 902 model training, including: processing the collected training data.
  • the training data collected in the above-mentioned data collection stage contains multiple formats
  • the model training stage it is necessary to process the collected training data and unify the format of the training data.
  • the input data of the AI model is called the input training data
  • the output data of the AI model is called the output training data.
  • the originally collected training data ie, raw data
  • the processed training data called the second input training data
  • the second input training data can be used as the input training data of the AI model
  • the output data of the AI model is called the first output training data
  • the data after processing the output data of the AI model It is called the second output training data, or called the target data.
  • the data processing method of the present application includes at least one of the following: an input data processing method, or an output data processing method.
  • the model training node can process the input data of the collected first input training data to obtain the second input training data; obtain the first output training data according to the second input training data and the AI model; First output training data, perform output data processing to obtain second output training data; determine whether the objective function satisfies requirements according to the second output training data. If it is satisfied, the AI model is output, and the training of the AI model is completed. Otherwise, update the parameters of the AI model and continue training the AI model.
  • the objective function may also be referred to as a loss function.
  • the adjusted parameters can include one or more of the following: the number of layers of the neural network, the width of the neural network, the connection relationship of the layers, the weight value of the neuron, the activation function of the neuron, Or bias in activation functions etc.
  • the above training process can adopt supervised learning, and the model training node can determine the value of the loss function according to the second output training data and the corresponding label; if the value of the loss function is less than the threshold value or the target requirement, the training of the AI model is ended ; otherwise continue training the AI model.
  • Different from supervised learning in the process of unsupervised learning, there is no label, so the model training node can determine the value of the loss function according to the second output training data. In supervised learning and unsupervised learning, the design of the loss function is different. It can be understood that, in the process of model training, other model training methods can also be used, such as reinforcement learning, etc., without limitation.
  • the process of the model training node performing input data processing on the first input training data includes: determining the input format and output format of the AI model; according to the input format of the AI model, processing the first input training data to obtain the second input training data.
  • the format of the second input training data is the same as that of the AI model.
  • the first output training data is determined.
  • the format of the first output training data is the same as the output format of the AI model.
  • the process of performing output data processing on the first output training data includes: processing the first output training data to obtain second output training data.
  • the first output training data may be processed according to the format of the label in the training data to obtain the second output training data.
  • the format of the second output training data is the same as that of the corresponding label.
  • the first output training data can be processed according to the wireless resource configuration to obtain the second output training data, and the second output training data satisfies the requirement of the corresponding wireless resource. configuration, or, according to the loss function of the AI model, process the first output data to obtain the second output training data and so on.
  • wireless resource configuration if the AI model needs to feed back channel state information with a length of B bits, the length of the first output training data output by the AI model is truncated to a length of B bits.
  • the parameters of the AI model are adjusted.
  • the model training node can calculate the value of the loss function according to the second output training data and labels in the training data. If the value of the loss function is less than the threshold value or meets the target requirements, the training of the AI model is completed; otherwise, the parameters of the AI model are adjusted and the training of the AI model is continued.
  • the above-mentioned process of processing the first input training data and the process of processing the first output training data do not necessarily need to be executed. Since the collected training data includes multiple formats, for the input training data of a certain format, if the format of the input training data is the same as the input format of the AI model, the input training data of this format can be directly input into the AI model, No need to process the input training data, the first input training data may not be processed, and the first output training data may be directly determined according to the first input training data and the AI model. Similarly, the output of the first AI model is called the first output training data. If the format of the first output training data meets the requirements, for example, it is the same as the format of the label, or meets the format of the radio resource configuration, etc., then no further processing may be performed on the first output training data.
  • the input format and output format of the AI model may be related to the configuration of wireless resources.
  • the length of the input data of the AI model can be greater than or equal to the length of the largest training sample among all potential wireless resource configurations in the application scenario of the AI model
  • the output data length of the AI model may be greater than or equal to the length of the tag with the largest length among all potential wireless resource configurations in the application scenario of the AI model.
  • the format may include two meanings such as dimension and/or length.
  • the object to be processed is usually a wireless signal or a wireless channel, etc.
  • the dimensions of the wireless signal or wireless channel usually include dimensions such as time domain, frequency domain, and/or space frequency.
  • the dimensions of the originally collected first input training data can be transformed to meet the input dimension requirements of the AI model, and/or the length of each dimension signal can be processed to meet the input requirements of the AI model Requirements for length in each dimension.
  • the focus is on the process of processing the length of input data and output data, and the processing of the length of one-dimensional data is taken as an example.
  • the length description is all is the length of the one-dimensional data. Processing the length of multi-dimensional data can be directly extended by processing the length of one-dimensional data.
  • Step 903 The model training node sends the AI model and instruction information to the model inference node.
  • the instruction information is used to indicate the data processing method of the AI model.
  • the data processing method of the AI model can also be called the usage method of the AI model.
  • the model training node when the model training node is different from the model inference node, model deployment is required, that is, the model training node needs to send the trained AI model to the model inference node.
  • the model training node may send AI model information to the model inference node, and the AI model information includes at least one of the following: model parameters, model input format, or model output format.
  • the parameters of the model include at least one of the following: the number of layers of the neural network, the width of the neural network, the connection relationship between layers, the weight of neurons, the activation function of neurons, or the bias in the activation function value etc.
  • the model reasoning node recovers or determines the AI model according to the information of the AI model.
  • the method of using the AI model (that is, the data processing method of the AI model) can also be indicated to the model reasoning node.
  • the usage method of the AI model can be predefined. This application does not limit the manner in which the model inference node acquires the usage method of the AI model.
  • the data processing method of the AI model includes at least one of the following: an input data processing method, or an output data processing method.
  • the input data processing method includes at least one of the following: a data filling method, or a data interception method.
  • the output data processing method includes at least one of the following: a data filling method, or a data interception method.
  • the data filling method includes at least one of the following: filling rules (or called filling positions, that is, where to fill and how long data is filled), length after filling, and type of filled data (ie, filled value).
  • the data interception method includes at least one of the following: interception rules (or called interception positions, that is, at which positions and lengths of data are intercepted), or the length after interception.
  • multiple data processing methods can be predetermined, and the model training node can indicate one of the above multiple data processing methods to the model inference node.
  • n types of data processing methods are predefined, and the n types of data processing methods correspond to different indexes, and the model training node can specifically indicate the index of a certain data processing method to the model reasoning node.
  • the model training node may directly indicate to the model inference node the parameters of the corresponding data processing method, for example, filling values for input data, filling positions, and/or intercepting positions for output data, and the like.
  • Step 904 The model reasoning node performs model reasoning according to the data processing method of the AI model.
  • the model inference node can perform model inference according to the data processing method of the AI model indicated by the model training node, and complete the inference tasks under different wireless resource configurations, that is, the model inference node uses the same rules as the training data processing method , to process data during model inference.
  • the initial input data is referred to as the first input data, or raw data
  • the input data of the AI model after the first input data is processed is referred to as the second input data
  • the output data of the AI model is called first output data
  • the data after processing the first output data is called second output data, or target data.
  • the data processing method of the AI model indicated by the model reasoning node includes at least one of the following: an input data processing method, or an output data processing method.
  • the input data processing method is the same as, or corresponding to, the input training data processing method in the model training phase.
  • the output data processing method is the same as, or corresponding to, the output training data processing method in the model reasoning stage.
  • the model reasoning node can process the first input data according to the input data processing method in the data processing method of the AI model to obtain the second input data, the format of the second input data is the same as that of the AI model in the same format.
  • the first output data is determined, and the format of the first output data is the same as the output format of the AI model.
  • the first output data is processed to determine the second output data.
  • the second output data can be regarded as the reasoning result of the AI model.
  • the process of the model inference node processing the first input data to determine the second input data, and the process of processing the first output data to obtain the second output data are not necessarily executed, or may not be executed .
  • the first input data matches the input format of the AI model, or is the same as the length of the filled data indicated in the input data processing method, the first input data may not be processed any more, and the first input data may be directly as the input data of the AI model.
  • the output data of the AI model satisfies the length of the wireless resource configuration, or is the same as the length of the corresponding tag, or the output data of the AI model is the same as the length of the intercepted data indicated in the output data processing method, etc., Then there is no need to process the first output data, and the first output data can be directly used as the reasoning result of the AI reasoning.
  • an AI model can be used to complete inference tasks under different wireless resource configurations, and there is no need to train and deploy an AI model for each wireless resource configuration, saving training and storage costs.
  • the data processing method of the AI model includes at least one of the following: the input data processing method of the AI model, or the output data processing method of the AI model.
  • the input data processing method of the AI model is an input data filling method
  • the output data processing method of the AI model is a process of output data interception method.
  • the data processing method of the AI model includes at least one of the following: a method for filling input data of the AI model, or a method for intercepting output data of the AI model.
  • the model training node performs the model training process according to the data processing method of the AI model, including: the model training node fills the first input training data according to the input data filling method to obtain the second input training data; according to the second input training data and the AI model, determine the first output training data; according to the output data interception method, intercept the first output training data to obtain the second output training data. Perform parameter adjustment on the AI model according to the second output training data.
  • the input data filling method may include at least one of the following:
  • the length of the data after padding is the same as the input data length of the AI model, that is, it matches the input format of the AI model.
  • the padding data may be 0, or may be a large positive number, or a small negative number, etc.
  • padding in front of the input training data padding in the back, padding in the middle, padding on both sides, padding with equal intervals, or padding with non-equal intervals, etc.
  • the output data interception method may include at least one of the following:
  • the length of the intercepted data is the same as the length of the label, that is, it matches the format of the label.
  • a model training node can collect enough training data.
  • the training data includes training samples and labels, and the training samples can also be referred to as first input training data, which can refer to data that needs to be input into the AI model for model training.
  • Determine the length of the input data and the length of the output data of the AI model the length of the input data of the AI model is greater than or equal to the length of the largest training sample among the training samples of the collected training data, and the output length of the AI model is greater than or equal to the length of the collected training data The length of the largest label among the labels of the training data.
  • the input data length of the AI model is greater than or equal to the length of the original data with the largest length among all potential wireless resource configurations in the application scenario of the AI model
  • the output length of the AI model is greater than or equal to the length of all potential wireless resource configurations in the application scenario of the AI model.
  • the model training node fills the first input training data according to the input data processing method to obtain the second input training data.
  • the second input training data can be called the filled first input training data, and the second input training data
  • the length is the same as the input data length of the AI model.
  • the input training data may be filled according to the filling data type and filling data rule in the input data filling method.
  • the frequency domain length of the input data of the AI model is 8RB
  • the frequency domain length of the first input training data is 4RB
  • the frequency domain length of the first input training data may be filled from 4RB to 8RB.
  • How to specifically fill the frequency domain length of the first input training data from 4RB to 8RB can be based on the type of filling data (for example, a specific filling value) and the rule of filling data in the input data filling method (for example, in Front padding, back padding, or equal interval padding, etc.) is determined.
  • the first output training data is determined.
  • the output data intercepting method the first output training data is intercepted to obtain the second output training data.
  • the length of the second output training data is equal to the length of the label.
  • how to intercept the first output data can be determined according to the interception rules in the output data interception method, for example, intercept before, after, or at equal intervals, etc., the first output training data.
  • the parameters of the AI model are adjusted.
  • the value of the loss function can be calculated according to the second output training data and labels. If the value of the loss function is less than the threshold value or meets the target requirements, the training of the AI model is completed; otherwise, the parameters of the AI model are adjusted and the training of the AI model is continued.
  • the model training node can send the above-mentioned trained AI model to the model reasoning node.
  • the model inference node can perform data inference on the AI model according to the data processing method of the AI model.
  • the model reasoning node performs the process of model reasoning according to the data processing method of the AI model, including:
  • the first input data is filled to obtain the second input data; according to the second input data and the AI model, the first output data is determined.
  • the output data interception method the first output data is intercepted, the interception may also be called extraction, and the first output data after interception may be called second output data.
  • the second output data is the reasoning result of the AI model.
  • the process of filling the input data and intercepting the output data is the same as the process of filling the input training data and intercepting the output training data in the model training phase.
  • the channel estimation using the same AI model under different bandwidths and different reference signal patterns is introduced.
  • the input data of the AI model is filled, and the output data of the AI model is intercepted.
  • Y represents the received signal
  • S represents the transmitted signal
  • S represents the reference signal in the channel estimation scenario
  • n indicates noise
  • the length of the sum S is L
  • L is the number of REs occupied by the reference signal.
  • the process of channel estimation is: by receiving the signal Y and the reference signal S, the channel response H in the frequency domain is estimated.
  • H is the channel on all REs within the scheduled bandwidth, as part of H.
  • the input data of the AI model is the received signal Y and the reference signal S
  • the reference signal S is known
  • the output data of the AI model is the channel response H in the frequency domain.
  • the frequency domain channels scheduled by the base station may have different bandwidths.
  • the bandwidth of the frequency domain channel scheduled by the base station may be 4 resource blocks (resource block, RB), 8 RBs, or 16 RBs.
  • the UE can send different types of reference signals to the base station. For example, type 1 reference signal and type 2 reference signal.
  • the length of H depends on the scheduling bandwidth, Y, The lengths of and S depend on the scheduling bandwidth and the pattern of the reference signal. For example, it is set that the bandwidth scheduled by the base station is k RBs. Since one RB includes 12 REs, the length of H is 12k. As shown in FIG.
  • the type 1 reference signal occupies 1 RE at intervals of 1 RE, and each RB occupies 6 REs.
  • the type 2 reference signal occupies 2 REs every 4 REs, and each RB, the type 2 reference signal occupies 4 REs.
  • the scheduled bandwidth is k RBs
  • the Y of the type 1 reference signal, and S of length 6k, type 2 reference signal Y, and the length of S is 4k.
  • the input data of the AI model are Y and S.
  • the lengths of Y and S are 6k
  • the lengths of Y and S are 4k
  • k is the frequency domain channel bandwidth scheduled by the base station.
  • the training data set can include training data under all possible wireless resource configurations in the application scenario as much as possible.
  • Each training data includes a training sample and a label.
  • the training sample may also be referred to as the above-mentioned first input training data, including a received signal Y and an original signal S, and the label is a channel response H in the frequency domain.
  • the training data includes type 1 reference signals with scheduling bandwidth of 4RB
  • Each training data includes training samples (Y, S) and labels (H).
  • the training data processing includes filling the first input training data.
  • the filled input training data is called the second input training data, and the length of the second input training data is the same as the length of the input data of the AI model.
  • the first output training data is determined.
  • the first output training data is intercepted to determine the second output training data, and the length of the second output training data is the same as the length of the label.
  • the process of padding the first input training data includes: first, for different types of reference signals, padding can be performed according to their patterns, and Y and S are padded so that the length of H is the same. After filling, only the real values of Y and S are in the position corresponding to the RE where the reference signal is located, and the filling values are in the other RE positions. After that, according to the input data length of the AI model, that is, the frequency domain bandwidth supported by the AI model, the filled Y and S are filled again.
  • the filling rules can be filled in the front, in the back, or in the middle. , padding on both sides, etc., are not limited.
  • the padding value can be 0, or a large positive number, or a small negative number, etc.
  • the length of the filled data that is, the length of the second input training data, is the same as the length of the input data of the AI model.
  • the training samples in the collected training data that is, the first input training data include:
  • Received signal corresponding to reference signal type 1 under scheduling bandwidth 4RB Y i represents the received signal Y with the number i, Indicates the jth element in the received signal Y numbered i;
  • the first input training data is filled, the filled data may be 0, and the length of the filled input training data is 96RE.
  • the scheduling bandwidth is at most 8RB, that is, 96RE, it may be considered to design the input data length of the AI model to be 96RE.
  • the length of the training samples in the training data is padded to be the same as the input data length 96RE of the AI model.
  • Y of different types of reference signals may be filled first so that the length of the label H is the same, and then Y may be filled again, specifically, Y may be filled after Y.
  • the filled first input training data that is, the second input training data are respectively:
  • the Y 1 corresponding to the reference signal type 1 is filled with
  • the Y 1 corresponding to the reference signal type 2 is filled with
  • the reference signal type 1 corresponds to Y 3 after padding is
  • the reference signal type 2 corresponds to Y 4 after padding is
  • the filling method of S is similar to the filling method of Y mentioned above.
  • the filled S and Y may be referred to as the second input training data.
  • the filled Y and S are input into the AI model, the output of the AI model is called the first output training data, and the length of the first output training data is the same as the output data length of the AI model.
  • the first output training data is intercepted, and the intercepted output training data may be called second output training data, and the length of the second output training data is the same as that of the label (frequency domain channel response H).
  • the interception rule for the first output data may be one or more of interception at the front, interception at the back, interception at the middle, and interception at both sides, etc., which is not limited.
  • the truncation rule can be independent from the filling rule, or can be matched with the filling rule. For example, if the first input training data is filled at the back, then the first output training data is intercepted at the front, and so on.
  • the output data length of the AI model is 96RE
  • the length of its label (frequency domain channel response H) is 48RB
  • the first 48 elements in the output data of the AI model can be intercepted, and 4RB label H for comparison.
  • the scheduling bandwidth of 8RB there is no need to intercept the output data of the AI model, and the output data of the AI model can be directly compared with the label H of 8RB.
  • the model training node can determine the structure of the AI model, which includes the input format and output format of the AI model.
  • the length of the input data in the input format of the AI model is greater than or equal to the length of the maximum length of the training sample in the training data.
  • the output data length in the output format of the AI model is greater than or equal to the length of the maximum-length label in the training data.
  • the first output training data of the AI model is intercepted, and the length of the intercepted output training data is the same as the length of the corresponding label.
  • the model training node can determine the value of the loss function according to the second output training data and labels; if the value of the loss function is less than the threshold value or meets the target requirements, the training of the AI model is completed; otherwise, adjust the parameters of the AI model and continue to AI model for training.
  • the model training node needs to send the trained AI model to the model reasoning node. It is necessary to indicate the data processing method of the AI model to the model inference node.
  • the model deployment node may send instruction information to the model inference node, the instruction information is used to indicate the data processing method of the AI model, and the instruction information may indicate input data filling rules, output data interception rules, and the like.
  • the model inference node performs model inference according to the data processing method of the AI model indicated by the model training node, and completes channel estimation tasks under different bandwidths and reference signal types. For example, use the same filling rules as in the model training process to fill the input data, use the same interception rules as in the model training process to intercept the output data, etc.
  • channel estimation in the frequency domain is taken as an example to describe the process of channel estimation using an AI model. It should be understood that channel estimation also includes other scenarios, and in each scenario, input data and output data are different. For example, in the scenario where further channel estimation is required for a channel that is roughly estimated using traditional methods, the input data is the noisy channel on the RE occupied by the reference signal, and the output data is the channel on all REs within the scheduled bandwidth. The input data needs to be Filling, intercepting the output data, and obtaining the inference result. For another example, in the scenario of channel estimation in the delay domain, the input data is the received signal and the reference signal, and the output data is the channel in the delay domain. The input data needs to be filled, and the output data does not need to be intercepted.
  • one AI model can be used to complete the channel estimation task under the conditions of various bandwidths and different reference signal patterns.
  • this article introduces the use of the same AI model to perform CSI feedback under different bandwidths and/or different antenna ports (channel characteristic lengths).
  • the input data of the AI model is filled, and the output data of the AI model is intercepted.
  • the UE acquires the downlink channel response H or the characteristic information W of the downlink channel, and a first sub-model is deployed in the UE, and the first sub-model may be called a sub-model f.
  • the above-mentioned downlink channel response H or characteristic information W of the downlink channel is used as the input of the sub-model f, and the output of the sub-model f is the feedback bit B corresponding to the CSI, and the feedback bit B is sent to the base station.
  • a second sub-model is deployed in the base station, and the second sub-model can be called sub-model g, and the feedback bit B is used as the input of the sub-model g, and the output of the sub-model g is the recovered downlink channel Or the characteristic information of the downlink channel
  • the dimension of the downlink channel response H is bandwidth*number of antenna ports, where the antenna ports include base station antenna ports and/or UE antenna ports.
  • the bandwidth of the CSI feedback configured by the base station may be different, for example, it may be 4RB or 8RB. Taking 1RB granularity as an example, the length of the downlink channel bandwidth dimension may be 4 or 8.
  • the number of antenna ports for CSI feedback configured by the base station may be different, for example, it may be 16 ports or 32 ports, and the length of the downlink channel antenna port dimension may be 4 or 8.
  • the dimension of the feature information W of the downlink channel is the number of subbands*the length of the feature vector (the number of base station antenna ports).
  • the number of subbands configured by the base station may be different, for example, it can be 6 subbands or 12 subbands.
  • the length of the tape dimension can be 6 or 12.
  • the eigenvector length of the CSI feedback configured by the base station may be different, for example, it may be 16 or 32, and the length of the eigenvector dimension of the feature information of the downlink channel may be 16 or 32.
  • the characteristic information W of the downlink channel is calculated according to the downlink channel response H, and the specific calculation method is as follows:
  • H i is the channel on the i-th unit bandwidth
  • the projected C may also be called feature information W of the downlink channel.
  • the subband dimension of V is transformed into the delay dimension
  • the eigenvector (base station antenna port) dimension of V is transformed into the beam (angle) dimension, but the length of the dimension remains unchanged. Therefore, no matter the characteristic information W of the downlink channel is Whether before or after projection, we describe its dimension as the number of subbands * length of eigenvector (number of base station antenna ports).
  • both the training sample and the label are the downlink channel response H or the feature information W of the downlink channel.
  • the training data can include the downlink channel response H or the feature information W of the downlink channel in different dimensions. The specific dimensions are as shown above.
  • the training data processing includes filling the first input training data.
  • the filled input training data is called the second input training data, and the length of the second input training data is the same as the length of the input data of the AI model.
  • the first output training data is determined.
  • the first output training data is intercepted to determine the second output training data, and the length of the second output training data is the same as the length of the label.
  • the first input training data is the downlink channel response H, its bandwidth dimension and/or antenna port dimension need to be filled; if the first input training data is the characteristic information W of the downlink channel, then it is necessary to fill in The sub-band dimension and/or feature vector dimension are filled to obtain the second input training data, and the length of the second input training data is the same as the input data length of the AI model. According to the second input training data and the AI model, the first output training data is determined.
  • the label is the downlink channel response H, it is necessary to intercept the bandwidth dimension and/or antenna port dimension of the first output training data, and if the label is the characteristic information W of the downlink channel, then it is necessary to intercept the subband dimension of the first output training data and/or feature vector dimensions to obtain the second output training data, the length of the second output training data is the same as the length of the label.
  • the model training node can determine the structure of the AI model, which includes the input format and output format of the AI model.
  • the length of the input data in the input format of the AI model is greater than or equal to the length of the maximum length of the training sample in the training data.
  • the output data length in the output format of the AI model is greater than or equal to the length of the maximum-length label in the training data.
  • the training sample that is, the first input training data is filled, and the length of the input training data after filling is the same as the length of the input data of the AI model.
  • the first output training data of the AI model is intercepted, and the length of the intercepted output training data is the same as the length of the corresponding label.
  • the model training node can determine the value of the loss function according to the second output training data and labels; if the value of the loss function is less than the threshold value or meets the target requirements, the training of the AI model is completed; otherwise, adjust the parameters of the AI model and continue to AI model for training.
  • the model training node not only needs to send the trained AI model to the model inference node, but also needs to indicate the data processing method of the AI model to the model inference node.
  • the model deployment node may send instruction information to the model inference node, the instruction information is used to indicate the data processing method of the AI model, and the instruction information may indicate input data filling rules, output data interception rules, and the like.
  • the model inference node performs model inference according to the data processing method of the AI model indicated by the model training node, and completes the CSI feedback task under different bandwidths and/or antenna port numbers, or completes the CSI feedback task under different subband numbers and/or feature vector lengths.
  • CSI Feedback Task For example, use the same filling rules as in the model training process to fill the input data, use the same interception rules as in the model training process to intercept the output data, etc.
  • the input data processing method of the AI model can also be used to indicate whether the input data of the AI model is the downlink channel or the feature information of the downlink channel, and whether the feature information of the downlink channel is the feature matrix of the downlink channel or the feature matrix after sparse projection, etc. .
  • the output data processing method of the AI model can also be used to indicate whether the output data of the AI model is downlink channel or feature information of the downlink channel, and whether the feature information of the downlink channel is the feature matrix of the downlink channel or the feature matrix after sparse projection.
  • the data processing method of the AI model includes at least one of the following: the input data processing method of the AI model, or the output data processing method of the AI model.
  • the input data processing method of the AI model is an input data interception method
  • the output data processing method of the AI model is an output data filling method.
  • the data processing method of the AI model includes at least one of the following: a method for intercepting input data of the AI model, or a method for filling output data of the AI model.
  • the model training node performs model training according to the data processing method of the AI model, including: intercepting the first input training data according to the input data interception method to obtain the second input training data; Two input training data and the AI model, determine the first output training data; according to the output data filling method, fill the first output training data to obtain the second output training data; according to the second output training data, and adjust the parameters of the AI model.
  • the model inference node performs model inference according to the data processing method of the AI model, including: intercepting the first input data according to the input data interception method to obtain the second input data; data and the AI model, determine first output data; and fill the first output data according to the output data filling method.
  • the filled first output data may be referred to as second output data, and the second output data may be referred to as an inference result of the AI model.
  • the input data interception method may include at least one of the following: length after interception, or rules for intercepting data, and the like.
  • the output data padding method may include at least one of the following: length after padding, type of padding data, or rules of padding data, etc.
  • interception and filling instructions please refer to the above design.
  • the input data can be intercepted during model reasoning or model training.
  • the input data of different lengths is intercepted into a uniform length, and the length of the input data after interception is the same as that of the AI model.
  • the downlink channel feature information W is a feature matrix after sparse projection
  • W is sparse in the angle domain and/or delay domain, that is, although the dimension of W is the number of subbands*the length of the feature vector
  • the values of most of the elements of this matrix are very small, and its total energy is mainly concentrated on several angles and delay paths, so the characteristic information W of the downlink channel can be intercepted, and only a few elements with larger values are kept, and the second Input data.
  • an AI model by intercepting the input data and filling the output data, an AI model can be used to complete task reasoning in different scenarios.
  • the AI model includes a first sub-model and a second sub-model.
  • the data processing method of the AI model includes at least one of the following: performing data interception on the output of the first sub-model, or performing data filling on the input of the second sub-model. Or, it can also be described as: the data processing method of the AI model, including at least one of the following: a method for intercepting the output data of the AI model (for example, the first sub-model), or, for the AI model (for example, , the input data filling method for the second submodel).
  • the model training node performs model training according to the data processing method of the AI model, including: determining the first output training data of the first sub-model according to the input training data of the first sub-model and the first sub-model ;
  • the output data interception method perform data interception on the first output training to obtain the second output training data of the first sub-model, and obtain the second output training data of the second sub-model according to the second output training data of the first sub-model Input training data;
  • the input data filling method perform data filling on the first input training of the second sub-model to obtain second input training data; according to the second input training data and the second sub-model model, to determine the output training data of the second sub-model; according to the input training data of the first sub-model and the output training data of the second sub-model, at least one of the following is adjusted: the first sub-model A model parameter, or a model parameter of the second sub-model.
  • the model inference node performs model inference according to the data processing method of the AI model, including: determining first output data of the first sub-model; and performing data interception on the first output data according to the output data interception method.
  • the model reasoning node performs model reasoning according to the data processing method of the AI model, including: filling the first input data according to the input data filling method to obtain second input data; according to the second input data and the The second sub-model is used to determine the first output data.
  • the specific input data filling method and output data interception method please refer to the description of the aforementioned design.
  • the model inference node of the first sub-model and the model inference node of the second sub-model may be the same or different nodes, without limitation.
  • the inference node of the first sub-model is the UE
  • the inference node of the second sub-model is the base station
  • the inference node of the second sub-model is the UE and so on.
  • the CSI feedback using the same AI model with different feedback bit lengths is introduced.
  • the length of the feedback bit B is determined by the base station, and is usually not fixed. For example, if the base station wants to obtain higher CSI recovery accuracy, it can allow the UE to feed back more feedback bits, and at this time the length of the feedback bits B is longer. Alternatively, if the base station wants to reduce the overhead of CSI feedback, it can allow the UE to feed back fewer feedback bits, and at this time the length of the feedback bits B is shorter. In one design, since the length of the feedback bit B is not fixed, a pair of corresponding sub-model f and sub-model g needs to be trained for each feedback bit B, which makes the training and storage overhead of the AI model larger.
  • an AI model can be designed, and the AI model can be applied to different feedback bit B lengths.
  • this application will be introduced from the aspects of collecting training data, training data processing, model training, model deployment and model inference.
  • the feedback bits belong to the intermediate result of the overall CSI feedback AI model, and can also be regarded as the output data of the sub-model f and the input data of the sub-model g.
  • the feedback bits B as training data.
  • the training data set is determined.
  • the training data set includes all possible feedback bit B length training data in the application scenario.
  • each training data includes training samples, feedback bits B and labels, wherein the training samples and labels are the same, and both are downlink channels or feature information H of downlink channels.
  • the training data includes feedback bit B of length 20, 40, and 60.
  • the training data also needs to include the corresponding training samples and labels under the feedback bits B of each length.
  • the training data needs to include a training sample, a feedback bit B (20) and a label.
  • the training sample can be the downlink channel obtained by the UE or the feature information H of the downlink channel.
  • the label can be the base station Correctly recovered downlink channel or characteristic information of downlink channel
  • the training data processing includes intercepting the output data of the sub-model f to make its length the same as that of the feedback bit B, and padding the feedback bit B so that its length is the same as that of the sub-model
  • the input data length of g is the same.
  • the interception rules may be one or more of interception at the front, interception at the back, interception at the middle, interception at both sides, equal interval interception or unequal interval interception, etc., without limitation.
  • the training data After interception, the training data includes feedback bits of different lengths, which need to be filled, and the filled data is used as the input of the sub-model g.
  • the padding rule can be one or more of padding in the front, padding in the back, padding in the middle, padding on both sides, equal interval filling or non-equal interval filling, there is no limit; the filling value can be 0, or one A large positive number, or a small negative number, etc., are not limited.
  • the input length of sub-model f is 60
  • the filling rule can be independent from the interception rule, or can match with the interception rule. For example, if the interception of the output data of the sub-model f is to intercept the front elements, then the filling of the feedback bits can be filling in the back, etc.
  • the feedback bit B is an intermediate result generated by the AI model, that is, the feedback bit is only generated during the training or inference process B, the above-mentioned truncation and filling operation on the feedback bit B occurs during the training or inference process.
  • the length of the output data of the sub-model f is greater than or equal to the length of the feedback bit B with the largest length in the training data
  • the length of the input data of the sub-model g is greater than or equal to the length of the longest feedback bit B in the training data.
  • the length of the feedback bit B is the length of the feedback bit B.
  • the sub-model f and the sub-model g are trained using the above-mentioned processed training data containing various lengths of feedback bits B.
  • the training data includes training samples, labels and feedback bits B.
  • the process of model training may include: inputting input training data (such as training samples) into sub-model f to obtain first output training data of sub-model f. Intercept the first output training data of the sub-model f, and intercept its length as the target length, which is all potential feedback bit lengths in the application scenario of the AI model, and obtain the second output data of the sub-model f, according to The second output data of sub-model f results in the first input training data of sub-model g.
  • the first input training data is filled, the length of the filled first input training data is the same as the input length of the sub-model g, and the filled first input training data is called the second input training data.
  • the output training data of the sub-model g is obtained.
  • the input training data of the sub-model f and the output training data of the sub-model g adjust at least one of the following: the parameters of the sub-model f, or the parameters of the sub-model g. For example, determine the input training data (i.e.
  • model training node needs to send the trained AI model to the model inference node.
  • the model training node needs to send the AI model to the model reasoning node, and also needs to send the data of the AI model
  • the processing method is indicated to the model inference node.
  • the data processing method of the AI model including data filling rules and interception rules, etc.
  • the model inference node can correctly use the AI model to complete the CSI feedback task under different feedback bit B lengths according to the data processing method of the AI model indicated by the model training node, that is, the model inference node uses the same rules as the training data processing method to process model inference time data.
  • the UE uses the sub-model f
  • the UE obtains the downlink channel response H, inputs the downlink channel response H into the sub-model f, and obtains the output data of the sub-model f;
  • the UE adopts The output data of sub-model f has the same interception rules, and the output data of sub-model f during inference is intercepted to obtain the feedback bit B of inference.
  • the UE feeds back the feedback bits B to the base station, and the base station uses the same filling method as the feedback bits B during training to fill in the feedback bits during inference, and then uses the sub-model g to recover the downlink channel response H.
  • an AI model can be trained by intercepting the output of the sub-model f into feedback bits B of different lengths, and filling the feedback bits B of different lengths into the input data of the sub-model g of the same length.
  • the AI model can be used for CSI feedback in the case of various feedback bit B lengths.
  • the first output data of the first sub-model is intercepted and the input data of the second sub-model is filled as an example, which is not intended to limit the application.
  • the output data of the first sub-model may be filled, the input data of the second sub-model may be intercepted, and so on.
  • the process of processing the output data of the first sub-model and the input data of the second sub-model is mainly described. In this application, there is no limitation on whether to process the input data of the first sub-model and the output data of the second sub-model.
  • the first input data of the first sub-model can be processed to obtain the second input data of the first sub-model; according to the second input data of the first sub-model and the first sub-model, determine The first output data of the first sub-model; intercepting (or filling) the first output data of the first sub-model to obtain the second output data of the first sub-model.
  • the first input data of the second sub-model is obtained from the second output data of the first sub-model.
  • the first input data of the second sub-model is filled (or intercepted) to obtain the second input data of the second sub-model. Based on the second input data of the second sub-model and the second sub-model, the first output data of the second sub-model is obtained.
  • the first output data of the second sub-model is processed to obtain the second output data of the second sub-model.
  • the manner of processing the first input data of the first sub-model and the manner of processing the first output data of the second sub-model are not limited.
  • this embodiment can be combined with the aforementioned process of processing input data and output data, for example, filling or intercepting the first input data of the first sub-model, and processing the first output data of the second sub-model For interception or filling, etc.
  • the data processing method of the AI model includes at least one of the following: a method for splitting input data, or a method for reorganizing output data.
  • the model training node performs model training according to the data processing method of the AI model, including: performing data splitting on the first input training data according to the splitting method of the input data to obtain the second input training data data; according to the second input training data and the AI model, determine the first output training data; according to the reorganization method of the output data, perform data reorganization on the first output training data to obtain the second output training data; according to the The second output is training data, and parameter adjustment is performed on the AI model.
  • the model inference node performs model inference according to the data processing method of the AI model, including: performing data splitting on the first input data according to the splitting method of the input data to obtain second input data;
  • the input data and the AI model are used to obtain first output data; according to the output data reorganization method, data reorganization is performed on the first output data to obtain second output data.
  • an AI model that needs to process input data and/or output data of the AI model such as Transformer, is described as an example.
  • Transformer is a sequence-to-sequence AI model that was first used in the field of natural language processing, such as translation. Since the length of the sentence is various, one of its characteristics is to support the input of any length. Therefore, in theory, Transformer can be applied to signal processing under different wireless resource configurations. As shown in Figure 14, the input and output of the Transformer are a collection of vectors. In wireless signal processing, the objects to be processed are usually wireless signals or wireless channels, and the dimensions of wireless signals or wireless channels usually include time domain, frequency domain and airspace etc. Therefore, applying the Transformer to wireless signal processing requires processing the original data, converting it into a vector set, inputting it into the Transformer, and then converting the vector set output by the Transformer into target data.
  • the 12*14 original data can be directly split into 14 vectors with a dimension of 12 in the time domain dimension.
  • the 12*14 original data can also be split into 12 2*7 matrices first, and then each 2*7 matrix is converted into a vector with a dimension of 14.
  • other splitting methods can also be used.
  • the model training node needs to send the trained AI model to the model inference node , it is also necessary to notify the model inference node of the splitting method of the original data and the method of reorganizing the target data during training.
  • the model inference node uses the same method as the training time to split the original data and output data to the AI model. to reorganize.
  • the model training node can indicate to the model inference node the splitting method of the original data and the reorganization method of the AI model output data, so that the AI model can Correctly handle signals in wireless networks.
  • the model training node and the model use node include hardware structures and/or software modules corresponding to each function.
  • the application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a certain function is executed by hardware or computer software drives the hardware depends on the specific application scenario and design constraints of the technical solution.
  • FIG. 17 and FIG. 18 are schematic structural diagrams of possible communication devices provided in this application. These communication devices can be used to implement the functions of the model training node or the model inference node in the above method, and thus can also realize the beneficial effects of the above method.
  • a communication device 1700 includes a processing unit 1710 and a transceiver unit 1720 .
  • the communication device 1700 is used to implement the function of the first node in the method shown in FIG. 8 above, and the first node may be a model training node or a model inference node.
  • the communication device 1700 is used to realize the function of the first node in the method shown in FIG. 8:
  • the processing unit 1710 is configured to determine the data processing method of the model; and, according to the data processing method of the model, implement at least one of the following: perform model training, or perform model inference;
  • the data processing method of the model includes at least one of the following: the input data filling method of the model, or the output data interception method of the model; or, the data processing method of the model includes at least one of the following: A method for intercepting input data of the model, or a method for filling output data of the model.
  • the transceiver unit 1720 is configured to send indication information to the second node, where the indication information is used to indicate a data processing method of the model;
  • the data processing method of the model includes at least one of the following: the input data filling method of the model, or the output data interception method of the model; or, the data processing method of the model includes at least one of the following: A method for intercepting input data of the model, or a method for filling output data of the model.
  • processing unit 1710 and the transceiver unit 1720 can be directly obtained by referring to related descriptions in the method shown in FIG. 8 , and details are not repeated here.
  • a communication device 1800 includes a processor 1810 and an interface circuit 1820 .
  • the processor 1810 and the interface circuit 1820 are coupled to each other.
  • the interface circuit 1820 may be a transceiver or an input-output interface.
  • the communication device 1800 may further include a memory 1830 for storing instructions executed by the processor 1810 or storing input data required by the processor 1810 to execute the instructions or storing data generated by the processor 1810 after executing the instructions.
  • the processor 1810 is used to implement the functions of the processing unit 1710
  • the interface circuit 1820 is used to implement the functions of the transceiver unit 1720 .
  • the first node module realizes the function of the first node in the above method.
  • the module of the first node receives information from other modules (such as radio frequency modules or antennas) in the first node, and the information is sent to the first node by the second node; or, the first node module sends information to the first node Other modules (such as radio frequency modules or antennas) send information, which is sent by the first node to the second node.
  • the first node module here may be the baseband chip of the first node, or other modules.
  • processor in this application can be a central processing unit (central processing unit, CPU), and can also be other general processors, digital signal processors (digital signal processor, DSP), application specific integrated circuits (application specific integrated circuit, ASIC), field programmable gate array (field programmable gate array, FPGA) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof.
  • CPU central processing unit
  • DSP digital signal processor
  • ASIC application specific integrated circuit
  • FPGA field programmable gate array
  • a general-purpose processor can be a microprocessor, or any conventional processor.
  • the memory in this application can be random access memory, flash memory, read-only memory, programmable read-only memory, erasable programmable read-only memory, electrically erasable programmable read-only memory, register, hard disk, mobile hard disk, CD-ROM or any other form of storage media known in the art.
  • An exemplary storage medium is coupled to the processor such the processor can read information from, and write information to, the storage medium.
  • a storage medium may also be an integral part of the processor.
  • the processor and storage medium can be located in the ASIC.
  • the ASIC can be located in the base station or the terminal.
  • the processor and the storage medium may also exist in the base station or the terminal as discrete components.
  • the methods in this application may be fully or partially implemented by software, hardware, firmware or any combination thereof.
  • software 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 comprises one or more computer programs or instructions.
  • the processes or functions described in the present application are executed in whole or in part.
  • the computer may be a general computer, a special computer, a computer network, a network device, a user device, a core network device, an OAM or other programmable devices.
  • the computer program or instructions may be stored in or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer program or instructions may be downloaded from a website, computer, A server or data center transmits to another website site, computer, server or data center by wired or wireless means.
  • 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 integrating 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; it may also be an optical medium, such as a digital video disk; or it may be a semiconductor medium, such as a solid state disk.
  • the computer readable storage medium may be a volatile or a nonvolatile storage medium, or may include both volatile and nonvolatile types of storage media.
  • “at least one” means one or more, and “multiple” means two or more.
  • “And/or” describes the association relationship of associated objects, indicating that there may be three types of relationships, for example, A and/or B, which can mean: A exists alone, A and B exist at the same time, and B exists alone, where A, B can be singular or plural.
  • the character “/” generally indicates that the contextual objects are an “or” relationship; in the formulas of this application, the character “/” indicates that the contextual objects are a “division” Relationship.
  • “Including at least one of A, B or C” may mean: including A; including B; including C; including A and B; including A and C; including B and C; including A, B, and C.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Databases & Information Systems (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

模型的数据处理方法及装置，该方法包括：第一节点确定模型的数据处理方法；第一节点根据所述模型的数据处理方法，实现以下至少一项：进行模型训练，或进行模型推理；利用该方法及装置，可实现利用一个模型，完成不同无线资源配置下的推理任务，减少模型训练和存储开销。

Description

模型的数据处理方法及装置
相关申请的交叉引用
本申请要求在2021年12月31日提交中国专利局、申请号为202111658528.4、申请名称为“模型的数据处理方法及装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及通信技术领域，尤其涉及模型的数据处理方法及装置。
背景技术
在无线通信网络中，例如在移动通信网络中，网络支持的业务越来越多样，因此需要满足的需求越来越多样。例如，网络需要能够支持超高速率、超低时延、和/或超大连接。该特点使得网络规划、网络配置、和/或资源调度越来越复杂。此外，由于网络的功能越来越强大，例如支持的频谱越来越高、支持高阶多入多出(multiple input multiple output，MIMO)技术、支持波束赋形、和/或支持波束管理等新技术，使得网络节能成为了热门研究课题。这些新需求、新场景和新特性给网络规划、运维和高效运营带来了前所未有的挑战。为了迎接该挑战，可以将人工智能技术引入无线通信网络中，从而实现网络智能化。基于此，如何在网络中有效地实现人工智能是一个值得研究的问题。
发明内容
本申请提供模型的数据处理方法及装置，可利用一个模型，完成不同无线资源配置下的推理任务，减少模型训练和存储开销。
第一方面，提供模型的数据处理方法，该方法的执行主体为第一节点，第一节点为模型训练节点或模型推理节点等，还可以为配置为第一节点中的部件(处理器、芯片或其它)，或者可以为软件模块等，该方法包括：确定模型的数据处理方法；根据所述模型的数据处理方法，实现以下至少一项：进行模型训练，或进行模型推理。
可选的，所述模型的数据处理方法包括以下至少一项：所述模型的输入数据填充方法，或所述模型的输出数据截取方法；或者，所述模型的数据处理方法包括以下至少一项：所述模型的输入数据截取方法，或所述模型的输出数据填充方法。
通过上述方法，当第一节点为模型推理节点，模型推理节点可根据该数据处理方法，利用同一个模型，完成不同无线资源配置下的推理任务。相对于针对每种无线资源配置，分别配置对应的模型，可减少模型训练和模型存储的开销。
在一种设计中，所述确定模型的数据处理方法，包括：接收来自第二节点的指示信息，所述指示信息用于指示所述模型的数据处理方法。其中，第一节点与第二节点不同。例如，第一节点为模型推理节点或模型训练节点，可接收来自其它节点(例如第二节点)的指示信息，该指示信息用于指示所述模型的数据处理方法，例如，第二节点为OAM等。或者，当第一节点为模型推理节点时，第二节点可以为模型训练节点，则模型推理节点可以接收 来自模型训练节点的指示信息等。或者，根据协议约定，确定所述模型的数据处理方法。
通过上述方法，模型训练节点或模型推理节点可以通过协议约定，获取所述模型的数据处理方法。或者，接收来自其它节点的指示，根据其它节点的指示，确定所述模型的数据处理方法等。模型训练节点或模型推理节点，可灵活的确定模型数据的处理方法。
在一种设计中，所述输入数据填充方法或所述输出数据填充方法，包括以下至少一项：填充后数据的长度、填充数据的类型、或填充数据的规则。
在一种设计中，所述输出数据截取方法或所述输入数据截取方法，包括以下至少一项：截取后数据的长度、或截取数据的规则。
在一种设计中，所述根据所述模型的数据处理方法，进行模型推理，包括：根据所述输入数据填充方法，对第一输入数据进行数据填充，得到第二输入数据；根据所述第二输入数据和所述模型，确定第一输出数据；根据所述输出数据截取方法，对所述第一输出数据进行数据截取。
通过上述方法，模型推理节点对输入数据进行填充，对输出数据截取的方法，可使得利用一个模型，完成不同无线资源配置下的推理任务，减小模型的训练和存储开销。可选的，在该设计中，所述模型的输入长度大于或等于所有潜在无线资源配置中输入数据长度最大的输入数据的长度，输出长度大于或等于所有潜在无线资源配置中输出数据长度最大的输出数据的长度。
在一种设计中，所述根据所述模型的数据处理方法，进行模型训练，包括：根据所述输入数据填充方法，对第一输入训练数据进行数据填充，得到第二输入训练数据；根据所述第二输入训练数据和所述模型，确定第一输出训练数据；根据所述输出数据截取方法，对所述第一输出训练进行数据截取，得到第二输出训练数据；根据所述第二输出训练数据，对所述模型进行参数调整。
通过上述方法，在模型训练过程中，尽可能多的收集多种无线资源配置下的训练数据；对每种无线资源配置下的训练数据，采用数据处理方法分别进行处理；利用处理后的训练数据对AI模型进行训练，直至该AI模型满足条件，则结束模型训练。对所述模型训练过程中的学习方法不作限定，例如监督学习(标签学习)、非监督学习(非标签学习)、或强化学习等方法。采用该方法所训练出的模型，可满足各种无线资源配置下的推理任务，减少模型训练开销。
在一种设计中，所述根据所述模型的数据处理方法，进行模型推理，包括：根据所述输入数据截取方法，对第一输入数据进行数据截取，得到第二输入数据；根据第二输入数据和所述模型，确定第一输出数据；根据所述输出数据填充方法，对所述第一输出数据进行数据填充。
通过上述方法，模型推理节点对输入数据进行截取，对输出数据填充的方法，可使得利用一个模型，完成不同无线资源配置下的推理任务，减小模型的训练和存储开销。
在一种设计中，所述根据所述模型的数据处理方法，进行模型训练，包括：根据所述输入数据截取方法，对第一输入训练数据进行数据截取，得到第二输入训练数据；根据第二输入训练数据和所述模型，确定第一输出训练数据；根据所述输出数据填充方法，对所述第一输出训练数据进行数据填充，得到第二输出训练数据；根据所述第二输出训练数据，对所述模型进行参数调整。
通过上述设计，在模型训练过程中，对模型参数进行调整，包括对以下至少一项进行 调整：神经网络的层数、神经网络的宽度、层的连接关系、神经元的权重值、神经元的激活函数、或激活函数中的偏置，以使得神经网元的输出与理想目标值之间的差异尽可能小。
在一种设计中，所述模型包括第一子模型，所述根据所述模型的数据处理方法，进行模型推理，包括：确定第一子模型的第一输出数据；根据所述输出数据截取方法，对所述第一输出数据进行数据截取。
在一种设计中，所述模型包括第二子模型，所述根据所述模型的数据处理方法，进行模型推理，包括：根据所述输入数据填充方法，对第一输入数据进行数据填充，得到第二输入数据；根据所述第二输入数据和所述第二子模型，确定第一输出数据。
通过上述方法，可对多种配对使用的子模型同时进行训练，例如上述第一子模型可部署在终端侧，第二子模型可部署在基站侧。同时，通过对第一子模型的输出数据进行截取，可减少空口开销。
在一种设计中，所述模型包括第一子模型和第二子模型，所述根据所述模型的数据处理方法，进行模型训练，包括：根据第一子模型的输入训练数据和第一子模型，确定第一子模型的第一输出训练数据；根据所述输出数据截取方法，对所述第一输出训练数据进行数据截取，得到第二子模型的第一输入训练数据；根据所述输入数据填充方法，对所述第二子模型的第一输入训练数据进行数据填充，得到第二输入训练数据；根据所述第二输入训练数据和所述第二子模型，确定第二子模型的输出训练数据；根据所述第一子模型的输入训练数据和所述第二子模型的输出训练数据，对以下至少一项进行调整：所述第一子模型的模型参数，或，所述第二子模型的模型参数。
第二方面，提供模型的数据处理方法，该方法的执行主体为第一节点，第一节点为模型训练节点或模型推理节点等，还可以为配置为第一节点中的部件(处理器、芯片或其它)，或者可以为软件模块等，包括：向第二节点发送指示信息，所述指示信息用于指示模型的数据处理方法。
其中，所述模型的数据处理方法包括以下至少一项：所述模型的输入数据填充方法，或所述模型的输出数据截取方法；或者，所述模型的数据处理方法包括以下至少一项：所述模型的输入数据截取方法，或所述模型的输出数据填充方法。
通过上述方法，在一种示例中，第一节点为模型训练节点，模型训练节点可根据模型训练时，数据的处理方式，确定数据处理方法，且将该数据处理方法指示给模型推理节点。该模型推理节点在执行模型推理时，使用与训练时相同的方式，对数据进行处理，使得该模型推理节点可正确的使用该模型进行模型推理，提高了模型推理输出结果的准确性和推理速度。
在一种设计中，所述输入数据填充方法或者所述输出数据填充方法，包括以下至少一项：填充后数据的长度、填充数据的类型、或填充数据的规则。
在一种设计中，所述输出数据截取方法或所述输入数据截取方法，包括以下至少一项：截取后数据的长度、或截取数据的规则。
第三方面，提供装置，有益效果可参见第一方面的记载，该装置可以是第一节点，该第一节点可以是模型训练节点或模型推理节点、或者配置于第一节点中的装置，或者能够和第一节点匹配使用的装置。在一种设计中，该装置包括执行第一方面中所描述的方法/操作/步骤/动作一一对应的单元，该单元可以是硬件电路，也可是软件，也可以是硬件电路结合软件实现。
示例性地，该装置可以包括处理单元，且处理单元可以执行上述第一方面任一种设计示例中的相应功能，具体的：
处理单元，用于确定模型的数据处理方法，以及，根据所述模型的数据处理方法，实现以下至少一项：进行模型训练，或进行模型推理；
其中，所述模型的数据处理方法包括以下至少一项：所述模型的输入数据填充方法，或所述模型的输出数据截取方法；或者，所述模型的数据处理方法包括以下至少一项：所述模型的输入数据截取方法，或所述模型的输出数据填充方法。
关于上述处理单元的具体执行过程可以参考第一方面，这里不再赘述。
示例性地，所述装置包括处理器，用于实现上述第一方面描述的方法。所述装置还可以包括存储器，用于存储指令和/或数据。所述存储器与所述处理器耦合，所述处理器执行所述存储器中存储的程序指令时，可以实现上述第一方面描述的方法。在一种可能的设计中，该装置包括：
存储器，用于存储程序指令；
处理器，用于确定模型的数据处理方法，以及根据所述模型的数据处理方法，实现以下至少一项：进行模型训练，或进行模型推理；
其中，所述模型的数据处理方法包括以下至少一项：所述模型的输入数据填充方法，或所述模型的输出数据截取方法；或者，所述模型的数据处理方法包括以下至少一项：所述模型的输入数据截取方法，或所述模型的输出数据填充方法。
关于处理器的具体执行过程，可参见上述第一方面的记载，不再赘述。
第四方面，提供装置，有益效果可参见第二方面的记载，该装置可以是第一节点，该第一节点可以为模型训练节点，或者配置于第一节点中的装置，或者能够和第一节点匹配使用的装置等。一种设计中，该装置包括执行第二方面中所描述的方法/操作/步骤/动作一一对应的单元，该单元可以是硬件电路，也可是软件，也可以是硬件电路结合软件实现。
示例性地，该装置可以包括通信单元，且通信单元可以执行上述第二方面任一种设计示例中的相应功能，具体的：
通信单元，用于向第二节点发送指示信息，所述指示信息用于指示模型的数据处理方法；
其中，所述模型的数据处理方法包括以下至少一项：所述模型的输入数据填充方法，或所述模型的输出数据截取方法；或者，所述模型的数据处理方法包括以下至少一项：所述模型的输入数据截取方法，或所述模型的输出数据填充方法。
上述通信单元的具体执行过程可以参考第二方面，这里不再赘述。
示例性地，所述装置包括处理器，用于控制通信接口实现上述第二方面描述的方法。所述装置还可以包括存储器，用于存储指令和/或数据。所述存储器与所述处理器耦合，所述处理器执行所述存储器中存储的程序指令时，可以实现上述第二方面描述的方法。所述装置还可以包括通信接口，所述通信接口用于该装置和其它设备进行通信。示例性地，通信接口可以是收发器、电路、总线、模块、管脚或其它类型的通信接口，其它设备可以为模型推理节点等。在一种可能的设计中，该装置包括：
存储器，用于存储程序指令；
处理器，用于控制通信接口向第二节点发送指示信息，所述指示信息用于指示模型的数据处理方法；
其中，所述模型的数据处理方法包括以下至少一项：所述模型的输入数据填充方法，或所述模型的输出数据截取方法；或者，所述模型的数据处理方法包括以下至少一项：所述模型的输入数据截取方法，或所述模型的输出数据填充方法。
关于通信接口与处理器的具体执行过程，可参见上述第二方面的记载，不再赘述。
第五方面，本申请还提供计算机可读存储介质，包括指令，当其在计算机上运行时，使得计算机执行第一方面或第二方面任一方面的方法。
第六方面，本申请还提供芯片系统，该芯片系统包括处理器，还可以包括存储器，用于实现第一方面或第二方面任一方面的方法。该芯片系统可以由芯片构成，也可以包含芯片和其他分立器件。
第七方面，本申请还提供计算机程序产品，包括指令，当其在计算机上运行时，使得计算机执行第一方面或第二方面任一方面的方法。
第八方面，本申请还提供系统，该系统中包括第三方面的装置，和第四方面的装置。
附图说明
图1为本申请提供的通信系统的示意图；
图2和图3为本申请提供的AI模型的部署示意图；
图4a和图4b为本申请提供的通信系统的架构示意图；
图4c为本申请提供的针对每个带宽，分别训练一个AI模型的示意图；
图5为本申请提供的AI模型的应用架构示意图；
图6为本申请提供的神经元的示意图；
图7为本申请提供的神经网络的示意图；
图8和图9A为本申请提供的模型数据处理方法的流程图；
图9B和图9C分别为模型训练阶段和模型推理阶段的数据处理过程；
图10和图11分别为类型1参考信号和类型2参考信号的示意图；
图12为本申请提供的CSI反馈流程的示意图；
图13为本申请提供的CSI反馈流程中对数据处理的示意图；
图14为本申请提供的Transformer模型的示意图；
图15和图16为本申请提供的将原始数据拆分为向量的示意图；
图17和图18为本申请提供的通信装置的结构示意图。
具体实施方式
图1是本申请能够应用的通信系统1000的架构示意图。如图1所示，该通信系统包括无线接入网100和核心网200，可选的，通信系统1000还可以包括互联网300。其中，无线接入网100可以包括至少一个接入网设备(如图1中的110a和110b)，还可以包括至少一个终端设备(如图1中的120a-120j)。终端设备通过无线的方式与接入网设备相连，接入网设备通过无线或有线方式与核心网连接。核心网设备与接入网设备可以是独立的不同的物理设备，或者可以是将核心网设备的功能与接入网设备的逻辑功能集成在同一个物理设备上，或者可以是一个物理设备上集成了部分核心网设备的功能和部分的接入网设备的功能。终端设备和终端设备之间以及接入网设备和接入网设备之间可以通过有线或无线的 方式相互连接。图1只是示意图，该通信系统中还可以包括其它网络设备，如还可以包括无线中继设备和无线回传设备等，在图1中未画出。
接入网设备可以是基站(base station)、演进型基站(evolved NodeB，eNodeB)、发送接收点(transmission reception point，TRP)、第五代(5th generation，5G)移动通信系统中的下一代基站(next generation NodeB，gNB)、开放无线接入网(open radio access network，O-RAN)中的接入网设备、第六代(6th generation，6G)移动通信系统中的下一代基站、未来移动通信系统中的基站或无线保真(wireless fidelity，WiFi)系统中的接入节点等；或者可以是完成基站部分功能的模块或单元，例如，可以是集中式单元(central unit，CU)、分布式单元(distributed unit，DU)、集中单元控制面(CU control plane，CU-CP)模块、或集中单元用户面(CU user plane，CU-UP)模块。接入网设备可以是宏基站(如图1中的110a)，也可以是微基站或室内站(如图1中的110b)，还可以是中继节点或施主节点等。本申请中对接入网设备所采用的具体技术和具体设备形态不做限定。
在本申请中，用于实现接入网设备的功能的装置可以是接入网设备；也可以是能够支持接入网设备实现该功能的装置，例如芯片系统、硬件电路、软件模块、或硬件电路加软件模块，该装置可以被安装在接入网设备中或可以与接入网设备匹配使用。在本申请中，芯片系统可以由芯片构成，也可以包括芯片和其他分立器件。为了便于描述，下文以用于实现接入网设备的功能的装置是接入网设备，接入网设备为基站为例，描述本申请提供的技术方案。
(1)协议层结构。
接入网设备和终端设备之间的通信遵循一定的协议层结构。该协议层结构可以包括控制面协议层结构和用户面协议层结构。例如，控制面协议层结构可以包括无线资源控制(radio resource control，RRC)层、分组数据汇聚层协议(packet data convergence protocol，PDCP)层、无线链路控制(radio link control，RLC)层、媒体接入控制(media access control，MAC)层和物理层等协议层的功能。例如，用户面协议层结构可以包括PDCP层、RLC层、MAC层和物理层等协议层的功能，在一种可能的实现中，PDCP层之上还可以包括业务数据适配协议(service data adaptation protocol，SDAP)层。
可选的，接入网设备和终端设备之间的协议层结构还可以包括人工智能(artificial intelligence，AI)层，用于传输AI功能相关的数据。
(2)集中单元(central unit，CU)和分布单元(distributed unit，DU)。
接入设备可以包括CU和DU。多个DU可以由一个CU集中控制。作为示例，CU和DU之间的接口可以称为F1接口。其中，控制面(control panel，CP)接口可以为F1-C，用户面(user panel，UP)接口可以为F1-U。本申请不限制各接口的具体名称。CU和DU可以根据无线网络的协议层划分：比如，PDCP层及以上协议层的功能设置在CU，PDCP层以下协议层(例如RLC层和MAC层等)的功能设置在DU；又比如，PDCP层以上协议层的功能设置在CU，PDCP层及以下协议层的功能设置在DU，不予限制。
上述对CU和DU的处理功能按照协议层的划分仅仅是一种举例，也可以按照其他的方式进行划分。例如可以将CU或者DU划分为具有更多协议层的功能，又例如将CU或DU还可以划分为具有协议层的部分处理功能。在一种设计中，将RLC层的部分功能和RLC层以上的协议层的功能设置在CU，将RLC层的剩余功能和RLC层以下的协议层的功能设置在DU。在另一种设计中，还可以按照业务类型或者其他系统需求对CU或者DU 的功能进行划分，例如按时延划分，将处理时间需要满足时延要求的功能设置在DU，不需要满足该时延要求的功能设置在CU。在另一种设计中，CU也可以具有核心网的一个或多个功能。示例性的，CU可以设置在网络侧方便集中管理。在另一种设计中，将DU的无线单元(radio unit，RU)拉远设置。可选的，RU可以具有射频功能。
可选的，DU和RU可以在物理层(physical layer，PHY)进行划分。例如，DU可以实现PHY层中的高层功能，RU可以实现PHY层中的低层功能。其中，用于发送时，PHY层的功能可以包括以下至少一项：添加循环冗余校验(cyclic redundancy check，CRC)码、信道编码、速率匹配、加扰、调制、层映射、预编码、资源映射、物理天线映射、或射频发送功能。用于接收时，PHY层的功能可以包括以下至少一项：CRC校验、信道解码、解速率匹配、解扰、解调、解层映射、信道检测、资源解映射、物理天线解映射、或射频接收功能。其中，PHY层中的高层功能可以包括PHY层的一部分功能，例如该部分功能更加靠近MAC层，PHY层中的低层功能可以包括PHY层的另一部分功能，例如该部分功能更加靠近射频功能。例如，PHY层中的高层功能可以包括添加CRC码、信道编码、速率匹配、加扰、调制、和层映射，PHY层中的低层功能可以包括预编码、资源映射、物理天线映射、和射频发送功能；或者，PHY层中的高层功能可以包括添加CRC码、信道编码、速率匹配、加扰、调制、层映射和预编码，PHY层中的低层功能可以包括资源映射、物理天线映射、和射频发送功能。例如，PHY层中的高层功能可以包括CRC校验、信道解码、解速率匹配、解码、解调、和解层映射，PHY层中的低层功能可以包括信道检测、资源解映射、物理天线解映射、和射频接收功能；或者，PHY层中的高层功能可以包括CRC校验、信道解码、解速率匹配、解码、解调、解层映射、和信道检测，PHY层中的低层功能可以包括资源解映射、物理天线解映射、和射频接收功能。
示例性的，CU的功能可以由一个实体来实现，或者也可以由不同的实体来实现。例如，可以对CU的功能进行进一步划分，即将控制面和用户面分离并通过不同实体来实现，分别为控制面CU实体(即CU-CP实体)和用户面CU实体(即CU-UP实体)。该CU-CP实体和CU-UP实体可以与DU相耦合，共同完成接入网设备的功能。
可选的，上述DU、CU、CU-CP、CU-UP和RU中的任一个可以是软件模块、硬件结构、或者软件模块+硬件结构，不予限制。其中，不同实体的存在形式可以是不同的，不予限制。例如DU、CU、CU-CP、CU-UP是软件模块，RU是硬件结构。这些模块及其执行的方法也在本申请的保护范围内。
一种可能的实现中，接入网设备包括CU-CP、CU-UP、DU和RU。例如，本申请的执行主体包括DU，或者包括DU和RU，或者包括CU-CP、DU和RU，或者包括CU-UP、DU和RU，不予限制。各模块所执行的方法也在本申请的保护范围内。
终端设备也可以称为终端、用户设备(user equipment，UE)、移动台、移动终端设备等。终端设备可以广泛应用于各种场景中的通信，例如包括但不限于以下至少一个场景：设备到设备(device-to-device，D2D)、车物(vehicle to everything，V2X)、机器类通信(machine-type communication，MTC)、物联网(internet of things，IOT)、虚拟现实、增强现实、工业控制、自动驾驶、远程医疗、智能电网、智能家具、智能办公、智能穿戴、智能交通、或智慧城市等。终端设备可以是手机、平板电脑、带无线收发功能的电脑、可穿戴设备、车辆、无人机、直升机、飞机、轮船、机器人、机械臂、或智能家居设备等。本申请对终端设备所采用的具体技术和具体设备形态不做限定。
在本申请中，用于实现终端设备的功能的装置可以是终端设备；也可以是能够支持终端设备实现该功能的装置，例如芯片系统、硬件电路、软件模块、或硬件电路加软件模块，该装置可以被安装在终端设备中或可以与终端设备匹配使用。为了便于描述，下文以用于实现终端设备的功能的装置是终端设备，终端设备为UE为例，描述本申请提供的技术方案。
基站和终端设备可以是固定位置的，也可以是可移动的。基站和/或终端设备可以部署在陆地上，包括室内或室外、手持或车载；也可以部署在水面上；还可以部署在空中的飞机、气球和人造卫星上。本申请对基站和终端设备的应用场景不做限定。基站和终端设备可以部署在相同的场景或不同的场景，例如，基站和终端设备同时部署在陆地上；或者，基站部署在陆地上，终端设备部署在水面上等，不再一一举例。
基站和终端设备的角色可以是相对的，例如，图1中的直升机或无人机120i可以被配置成移动基站，对于那些通过120i接入到无线接入网100的终端设备120j来说，终端设备120i是基站；但对于基站110a来说，120i是终端设备，即110a与120i之间是通过无线空口协议进行通信的。110a与120i之间也可以是通过基站与基站之间的接口协议进行通信的，此时，相对于110a来说，120i也是基站。因此，基站和终端设备都可以统一称为通信装置，图1中的110a和110b可以称为具有基站功能的通信装置，图1中的120a-120j可以称为具有终端设备功能的通信装置。
在本申请中，可在前述图1所示的通信系统中引入独立的网元如称为AI网元、或AI节点等)来实现AI相关的操作，AI网元可以和通信系统中的接入网设备之间直接连接，或者可以通过第三方网元和接入网设备实现间接连接。其中，第三方网元可以是认证管理功能(authentication management function，AMF)、或用户面功能(user plane function，UPF)等核心网网元；或者，可以在通信系统中的其他网元内配置AI功能、AI模块或AI实体来实现AI相关的操作，例如该其他网元可以是接入网设备(如gNB)、核心网设备、或网管(operation,administration and maintenance，OAM)等，在这种情况下，执行AI相关的操作的网元为内置AI功能的网元。其中，上述OAM用于对接入网设备和/或核心网设备等进行操作、管理和维护等。
在本申请中，如图2或图3所示，核心网设备、接入网设备、终端设备或OAM等中的至少一个设备可以部署有AI模型，利用该AI模型实现相应的功能。在本申请中，不同节点中部署的AI模型可以相同或不同，模型不同包括以下至少一项不同：模型的结构参数不同，例如，模型的层数和/或权值等不同；模型的输入参考不同；或模型的输出参考不同等。其中，模型的输入参数和/或模型的输出参数不同可以描述为模型的功能不同。与上述图2不同的是，在图3中，将接入网设备的功能拆分为CU和DU。可选的，CU和DU可以为O-RAN架构下的CU和DU。CU中可以部署有一个或多个AI模型。和/或，DU中可以部署有一个或多个AI模型。可选的，还可以进一步，将图3中的CU拆分为CU-CP和CU-UP。可选的，CU-CP中可以部署有一个或多个AI模型。和/或，CU-UP中可以部署有一个或多个AI模型。可选的，图2或图3中，接入网设备的OAM和核心网设备的OAM可以分布独立部署。
可选的，图4a为本申请的一种通信系统的架构。如图4a所示，在第一种设计中，接入网设备中包括近实时接入网智能控制(RAN intelligent controller，RIC)模块，用于进行模型训练和推理。例如，近实时RIC可以用于训练AI模型，利用该AI模型进行推理。例 如，近实时RIC可以从CU、DU或RU中的至少一项获得网络侧和/或终端侧的信息，该信息可以作为训练数据或者推理数据。可选的，近实时RIC可以将推理结果递交至CU、DU、RU或终端设备中的至少一项。可选的，CU和DU之间可以交互推理结果。可选的，DU和RU之间可以交互推理结果，例如近实时RIC将推理结果递交至DU，由DU转发给RU。
或者，在第二种设计中，如图4a所示，接入网设备之外包括非实时RIC(可选的，非实时RIC可以位于OAM中或者核心网设备中)，用于进行模型训练和推理。例如，非实时RIC用于训练AI模型，利用该模型进行推理。例如，非实时RIC可以从CU、DU或RU中的至少一项获得网络侧和/或终端侧的信息，该信息可以作为训练数据或者推理数据，该推理结果可以被递交至CU、DU、RU或终端设备中的至少一项。可选的，CU和DU之间可以交互推理结果。可选的，DU和RU之间可以交互推理结果，例如非实时RIC将推理结果递交至DU，由DU转发给RU。
或者，在第三种设计中，如图4a所示，接入网设备中包括近实时RIC，接入网设备之外包括非实时RIC(可选的，非实时RIC可以位于OAM中或者核心网设备中)。同上述第二种设计，非实时RIC可以用于进行模型训练和推理。和/或，同上述第一种设计，近实时RIC可以用于进行模型训练和推理。和/或，非实时RIC进行模型训练，近实时RIC可以从非实时RIC获得AI模型信息，并从CU、DU或RU中的至少一项获得网络侧和/或终端侧的信息，利用该信息和该AI模型信息得到推理结果。可选的，近实时RIC可以将推理结果递交至CU、DU、RU或终端设备中的至少一项。可选的，CU和DU之间可以交互推理结果。可选的，DU和RU之间可以交互推理结果，例如近实时RIC将推理结果递交至DU，由DU转发给RU。例如，近实时RIC用于训练模型A，利用模型A进行推理。例如，非实时RIC用于训练模型B，利用模型B进行推理。例如，非实时RIC用于训练模型C，将模型C的信息发送给近实时RIC，近实时RIC利用模型C进行推理。
图4b为本申请的另一种通信系统的架构。相对图4a，图4b中将CU分离成为了CU-CP和CU-UP等。
在无线网络中，使用AI模型，需要解决的一个问题是无线网络中灵活可变的信号格式与AI模型固定的输入格式和输出格式之间的矛盾。例如，无线信号的带宽是基站灵活调度的，在一个时隙可能调度的带宽为4个资源块(resource block，RB)，在下一个时隙调度的带宽可能变成8个RB，但AI模型的输入格式是固定的，无法随着信号的带宽而改变。例如，AI模型的输入格式对应4个RB，则该AI模型仅能支持输入为4RB的信号，无法处理输入为8RB的信号。为了解决上述问题，提供一种解决方案：针对每种带宽均训练一个AI模型。如图4c所示，针对输入带宽为4RB的信号和输入带宽为8RB的信号，分别训练一个AI模型，这样导致训练和存储开销都很大。
在本申请提供一种解决方案，在该解决方案中：训练一个AI模型，该AI模型可适用于多种格式的输入数据和输出数据。具体的，在训练时，将不同格式的数据(如输入数据和/或输出数据)等，通过填充或截取等方式，转化为格式相同的数据，共同训练一个AI模型。该AI模型适用于该多种格式。在该AI模型在使用或推理时，将输入数据和/或输出数据，通过与训练时相同的数据处理方法，转化为与该AI模型匹配的输入数据和/或输出数据，从而使得使用同一个AI模型，完成不同无线资源配置下的推理任务，减少训练和存储的开销。应当指出，在本申请的方案中至少包括两种示例。例如，针对所有格式的输 入数据和/或输出数据训练一个AI模型。比如，无线网络中的调度带宽包括4RB、8RB、16RB和32RB。则针对上述4种调度带宽，训练出一个AI模型。后续描述中，主要以该示例为例说明。或者，针对部分格式的输入数据和/或输出数据设计一个AI模型，针对其它格式的输入数据和/或输出数据再设计另外其它的AI模型等。例如，针对调度带宽为4RB和8RB，训练第一AI模型，也就是第一AI模型适用于对4RB和8RB信号的输入。针对调度带宽为16RB和32RB，训练第二AI模型，也就是第二AI模型适用于16RB和32RB信号的输入等。
为了便于理解，首先对本申请涉及的名词或概念进行说明。
1、AI模型。
AI模型是AI功能的具体实现，AI模型表征了模型的输入和输出之间的映射关系。AI模型可以是神经网络、线性回归模型、决策树模型、支持向量机(support vector machine，SVM)、贝叶斯网络、Q学习模型或者其他机器学习模型等。本申请中，AI功能可以包括以下至少一项：数据收集(收集训练数据和/或推理数据)、数据预处理、模型训练(或称为，模型学习)、模型信息发布(配置模型信息)、模型校验、模型推理、或推理结果发布。其中，推理又可以称为预测。本申请中，可以将AI模型简称为模型。
AI模型的设计主要包括数据收集环节(例如数据源可收集训练数据和/或推理数据)、模型训练环节以及模型推理环节。进一步地还可以包括推理结果应用环节。可选的，还可以包括模型测试环节。如图5所示为AI模型的一种应用架构示意图。数据源(data source)用于提供训练数据和推理数据。在模型训练环节中，模型训练节点(model trainning host)通过对数据源提供的训练数据(training data)进行分析或训练，得到AI模型。其中，通过模型训练节点学习得到AI模型，相当于由模型训练节点利用训练数据学习得到模型的输入和输出之间的映射关系。将AI模型部署在模型推理节点(model inference host)中。可选的，模型训练节点还可以对已部署在模型推理节点的AI模型进行更新。模型推理节点还可以向模型训练节点反馈已部署模型的相关信息，以使得模型训练节点对已部署的AI模型进行优化或更新等。
在模型推理环节中，模型推理节点使用AI模型，基于数据源提供的推理数据进行推理，得到推理结果。该方法可以实现为：模型推理节点将推理数据输入到AI模型，通过AI模型得到输出，该输出即为推理结果。该推理结果可以指示：由执行对象使用(执行)的配置参数、和/或由执行对象执行的操作。推理结果可以由执行(actor)实体统一规划，并发送给一个或多个执行对象(例如，网络实体)去执行。可选的，执行实体或执行对象还可以反馈模型的性能给数据源作为训练数据，便于后续实施模型的更新训练。
2、神经网络(neural network,NN)。
AI模型可以是神经网络或其它机器学习模型。以神经网络为例，神经网络是机器学习技术的一种具体实现形式。根据通用近似定理，神经网络在理论上可以逼近任意连续函数，从而使得神经网络具备学习任意映射的能力。因此神经网络可以对复杂的高维度问题进行准确的抽像建模。
神经网络的思想来源于大脑组织的神经元结构。每个神经元都对其输入值做加权求和运算，将加权求和结果通过一个激活函数产生输出。如图6所示，为神经元结构示意图。 假设神经元的输入为x＝[x 0,x 1,…,x n],与各输入对应的权值分别为w＝[w,w 1,…,w n],加权求和的偏置为b。激活函数的形式可以多样化，假设一个神经元的激活函数为：y＝f(z)＝max(0,z)，则该神经元的输出为：
Figure PCTCN2022144185-appb-000001
Figure PCTCN2022144185-appb-000002
再例如一个神经元的激活函数为：y＝f(z)＝z，则该神经元的输出为：
Figure PCTCN2022144185-appb-000003
Figure PCTCN2022144185-appb-000004
w i、x i和b可以为小数、整数(包括0、正整数或负整数等)、或复数等各种可能的取值。神经网络中不同神经元的激活函数可以相同或不同。
神经网络一般包括多层结构，每层可包括一个或多个神经元。增加神经网络的深度和/或宽度可以提高该神经网络的表达能力，为复杂系统提供更强大的信息提取和抽象建模能力。其中，神经网络的深度可以指神经网络包括的层数，每层包括的神经元个数可以称为该层的宽度。如图7所示，为神经网络的层关系示意图。一种实现中，神经网络包括输入层和输出层。神经网络的输入层将接收到的输入经过神经元处理后，将结果传递给输出层，由输出层得到神经网络的输出结果。另一种实现中，神经网络包括输入层、隐藏层和输出层。神经网络的输入层将接收到的输入经过神经元处理后，将结果传递给中间的隐藏层，隐藏层再将计算结果传递给输出层或者相邻的隐藏层，最后由输出层得到神经网络的输出结果。一个神经网络可以包括一层或多层依次连接的隐藏层，不予限制。神经网络的训练过程中，可以定义损失函数。损失函数描述了神经网络的输出值和理想目标值之间的差距或差异，本申请不限制损失函数的具体形式。神经网络的训练过程就是通过调整神经网络参数，如神经网络的层数、宽度、神经元的权值、和/或神经元的激活函数中的参数等，使得损失函数的值小于门限值或者满足目标需求的过程。
3、AI模型的训练。
AI模型的训练过程中，可定义损失函数。损失函数描述了AI模型的输出值和理想目标值之间的差距或差异，本申请不限制损失函数的具体形式。AI模型的训练过程就是通过调整AI模型的部分或全部参数，使得针对一个或多个训练数据，损失函数的值或者加权求和值(例如平均值)小于门限值或者满足目标需求的过程。如AI模型为神经网络，则训练过程中可以调整以下一项或多项参数：神经网络的层数、神经网络的宽度、层的连接关系、神经元的权重值、神经元的激活函数、或激活函数中的偏置，使得神经网元的输出与理想目标值之间的差异尽可能小。
在一种设计中，以AI模型为神经网络f θ(·)为例。模型训练节点收集训练数据，该训练数据中包括训练样本和标签。例如，训练样本x作为输入，经过神经网络f θ(·)处理后输出推理结果f θ(x)，损失函数计算得到的推理结果f θ(x)与训练样本x的标签之间的差异。模型训练节点可基于损失函数得到的差异，采用模型优化算法，优化网络参数。通过利用大量训练数据对神经网络进行训练，可使得针对一组训练样本(例如一个或多个训练样本)，各训练样本的神经网络的输出与标签之间的差异都小于门限值或者满足目标需求，或者，所有训练样本的神经网络的输出与标签之间的差异的加权求和(例如平均值)小于门限值或者满足目标需求，从而完成神经网络的训练。
需要说明的是，上述过程描述了监督学习的训练方式，并不作为对本申请的限定。在本申请中，AI模型的训练也可以采用非监督学习，利用算法学习训练样本的内在模式，实现基于训练样本完成AI模型的训练。AI模型的训练还可以采用强化学习，通过与环境进行交互获取环境反馈的激励信号，从而学习解决问题的策略，实现AI模型的优化等。本申请中，对模型训练的方法不作限定。
如图8所示，提供模型的数据处理方法的流程，该方法的执行主体为模型训练节点或模型推理节点，可以理解的是，模型训练节点和模型推理节点可以为同一节点，或不同节点等，不作限定。该流程至少包括：
步骤801：第一节点确定AI模型的数据处理方法。
其中，第一节点为模型训练节点或模型推理节点等。例如，第一节点接收来自其它节点(例如第二节点)的指示信息，该指示信息用于指示所述AI模型的数据处理方法。或者，第一节点根据协议约定，确定AI模型的数据处理方法。或者，第一节点自行确定AI模型的数据处理方法。例如，模型推理节点可根据收集的训练数据的格式和待训练AI模型的格式等，确定AI模型的数据处理方法等，具体可参见下述说明。
步骤802：第一节点根据所述AI模型的数据处理方法，实现以下至少一项：进行模型训练，或进行模型推理。
如图9A所示，以模型训练节点和模型推理节点分别部署，且模型训练节点向模型推理节点指示AI模型的数据处理方法为例，提供数据处理方法的流程，至少包括以下步骤：
步骤901：收集训练数据。
在本申请中，执行收集训练数据的主体可以是模型训练节点，或者其它AI实体，或者AI模块等。在本申请中，所收集的训练数据包括但不限于无线网络中的数据，例如无线信道信息、接收信号、参考信号、参考信号接收功率(reference signal receiving power，RSRP)等。无线信道信息可以包括估计的信道响应、或信道特征等。这些数据的直接获取者(或称为估计者、测量者、或收集者等)可以是UE或基站。当执行数据收集的实体或模块与数据的直接获取者不同时，由数据的直接获取者将数据发送给执行数据收集的实体或模块。收集到的训练数据可以包括不同的UE的测量数据或UE在不同地理位置、不同信道环境收集的测量数据。收集到的数据可以是UE或者基站在实际网络中获取到的真实数据，也可以是通过仿真平台或者模拟平台生成的虚拟数据等，不作限定。
本申请的目的是训练一个适配多种无线资源配置的AI模型，因此，收集到的训练数据包括多种无线资源配置下的数据。例如，多种带宽或多种参考信号图样下的接收信号、或无线信道信息等。多种无线资源配置下的数据也以是同一种无线资源配置下的数据通过数据增强方法生成的，例如，将同一带宽下的无线信道截取成不同带宽下的无线信道。
可选的，在本申请中，具体的收集训练数据的类型与AI模型的功能有关。例如，对于频域信道估计的AI模型，该AI模型所需要的训练数据至少包括所接收到的参考信号Y和原始的参考信号S。则对于该AI模型，可收集各种无资源配置下的参考信号Y和原始的参考信号S，作为训练数据。例如，可收集不同带宽和/或参考信号类型下的参考信号Y和原始的参考信号S等。
又例如，对于用于信道状态信息(channel state information，CSI)反馈的AI模型，该AI模型所需要的训练数据至少包括信道响应H或信道特征W，可收集不同带宽和/或不同天线端口数下的信道响应H或信道特征W。
又例如，对于用于波束管理的AI模型，该AI模型所需要的训练数据至少包括信道响应H、接收信号Y、或接收信号Y的RSRP，可收集不同带宽和/或不同天线端口数下的信道响应H或接收信号Y或RSRP。
又例如，对于定位相关的AI模型，该AI模型所需要的训练数据至少包括信道响应H，可收集不同带宽和/或不同天线端口数下的信道响应H。
步骤902：模型训练，包括：对收集的训练数据进行处理。
在本申请中，由于在上述数据收集阶段所收集到的训练数据包含多种格式，在模型训练阶段，需要对收集到的训练数据进行处理，统一训练数据的格式。
为了区别在AI模型推理阶段，AI模型的输入数据和输出数据，将AI模型训练阶段，AI模型的输入数据称为输入训练数据，将AI模型的输出数据称为输出训练数据。进一步，由于在本申请中，涉及到训练数据的处理过程，可将原始收集的训练数据(即原始数据)，即处理前的训练数据，称为第一输入训练数据，将处理后的训练数据，称为第二输入训练数据，该第二输入训练数据可作为AI模型的输入训练数据，将AI模型的输出数据称为第一输出训练数据，将对AI模型的输出数据处理后的数据，称为第二输出训练数据，或称为目标数据。
示例的，本申请的数据处理方法包括以下至少一项：输入数据处理方法、或输出数据处理方法。如图9B所示，模型训练节点可对收集的第一输入训练数据，进行输入数据处理，得到第二输入训练数据；根据第二输入训练数据和AI模型，得到第一输出训练数据；对第一输出训练数据，进行输出数据处理，得到第二输出训练数据；根据第二输出训练数据，确定目标函数是否满足要求。如果满足，则输出该AI模型，对AI模型的训练完成。否则，更新AI模型的参数，继续对AI模型进行训练。所述目标函数还可称为损失函数。以AI模型为神经网络为例，则调整的参数可包括以下一项或多项：神经网络的层数、神经网络的宽度、层的连接关系、神经元的权重值、神经元的激活函数，或激活函数中的偏置等。上述训练过程可采用监督学习，模型训练节点可根据第二输出训练数据与对应的标签，确定损失函数的值；如果损失函数的值的小于门限值或目标需求，则结束对AI模型的训练；否则继续对AI模型进行训练。与监督学习不同，在非监督学习的过程中，不存在标签，则模型训练节点可根据第二输出训练数据，确定损失函数的值。在监督学习和非监督学习中，对损失函数的设计不同。可以理解的是，在模型训练的过程中，还可以采用其它模型训练方法，例如强化学习等，不作限定。
例如，模型训练节点对第一输入训练数据进行输入数据处理的过程，包括：确定AI模型的输入格式和输出格式；根据AI模型的输入格式，对第一输入训练数据进行处理，得到第二输入训练数据。可选的，第二输入训练数据的格式与AI模型的输入格式相同。根据第二输入训练数据和AI模型，确定第一输出训练数据。可选的，第一输出训练数据的格式与AI模型的输出格式相同。对第一输出训练数据进行输出数据处理的过程，包括：对第一输出训练数据进行处理，获得第二输出训练数据。例如，可根据训练数据中的标签的格式，对第一输出训练数据进行处理，获得第二输出训练数据。可选的，第二输出训练数据的格式与对应标签的格式相同。或者，在非监督学习场景下，训练数据中不存在标签，则可根据无线资源配置，对第一输出训练数据进行处理，获得第二输出训练数据，该第二输出训练数据满足对应无线资源的配置，或者，根据AI模型的损失函数，对第一输出数据进行处理，获得第二输出训练数据等。例如，在无线资源配置中，需要AI模型反馈长度为B比特的信道状态信息，则将AI模型输出的第一输出训练数据的长度，截取为B比特的长度等。根据第二输出训练数据，对AI模型的参数进行调整。例如，以监督学习为例，模型训练节点可根据第二输出训练数据和训练数据中的标签，计算损失函数的值。若损失函数的值小于门限值或者满足目标需求，则完成对AI模型的训练；否则，调整AI模型的参数，继续对AI模型训练。
应当理解，上述对第一输入训练数据处理的过程，和对第一输出训练数据处理的过程，并非一定需要执行。由于收集的训练数据包括多种格式，对于某一格式的输入训练数据，若该输入训练数据的格式与AI模型的输入格式相同，则可将该格式的输入训练数据直接输入到AI模型中，无需再对输入训练数据进行处理，即可对第一输入训练数据不进行处理，直接根据第一输入训练数据和AI模型，确定第一输出训练数据。同理，第一AI模型的输出称为第一输出训练数据。若该第一输出训练数据的格式满足要求，例如，与标签的格式相同，或者满足无线资源配置的格式等，则可对该第一输出训练数据不再作进一步处理。
在本申请中，AI模型的输入格式和输出格式，可能与无线资源的配置相关。以对AI模型的输入数据进行填充，对AI模型的输出数据进行截取为例，AI模型的输入数据长度可以大于或等于该AI模型应用场景下所有潜在无线资源配置中长度最大的训练样本的长度，AI模型的输出数据长度可以大于或等于该AI模型应用场景下所有潜在无线资源配置中的长度最大的标签的长度等。
需要说明的是，在本申请的描述中，格式可包括维度和/或长度等两层含义。在无线信号处理中，通常需要处理的对象是无线信号或无线信道等，无线信号或无线信道的维度通常包括时域、频域和/或空频等维度。在本申请中，可对原始采集的第一输入训练数据的维度进行变换，使其满足AI模型的输入维度要求，和/或，对每维信号的长度进行处理，使其满足AI模型的输入对每维度中长度的要求。在本申请中的描述中，重点描述对输入数据和输出数据的长度进行处理的过程，且以对一维数据的长度进行处理为例，在后续描述中，如无特殊说明，长度描述的都是该一维数据的长度。对多维数据的长度进行处理，可由对一维数据长度的处理直接扩展得到。
步骤903：模型训练节点向模型推理节点发送AI模型，以及指示信息，该指示信息用于指示该AI模型的数据处理方法，AI模型的数据处理方法还可称为AI模型的使用方法。
在本申请中，当模型训练节点与模型推理节点不同时，需要进行模型部署，即模型训练节点需要将训练好的AI模型发送给模型推理节点。例如，模型训练节点可以向模型推理节点发送AI模型的信息，该AI模型的信息中包括以下至少一项：模型的参数、模型的输入格式、或模型的输出格式等。以神经网络为例，模型的参数中包括以下至少一项：神经网络的层数、神经网络的宽度、层间的连接关系、神经元的权重、神经元的激活函数、或激活函数中的偏值等。模型推理节点根据该AI模型的信息，恢复或确定该AI模型。
进一步的，在本申请中，为了使模型推理节点可以正确使用该AI模型完成推理任务，还可以将该AI模型的使用方法(即AI模型的数据处理方法)，指示给模型推理节点。或者，可以预定义该AI模型的使用方法。本申请不限制模型推理节点获取AI模型的使用方法的方式。
在本申请中，AI模型的数据处理方法包括以下至少一项：输入数据处理方法、或输出数据处理方法。输入数据处理方法包括以下至少一项：数据填充方法、或数据截取方法。输出数据处理方法包括以下至少一项：数据填充方法、或数据截取方法。数据填充方法包括以下至少一项：填充规则(或称为填充位置，即在哪些位置填充多长数据)、填充后的长度、填充数据的类型(即填充的数值)。数据截取方法包括以下至少一项：截取规则(或称为截取位置，即在哪些位置截取多长数据)、或截取后的长度。
在本申请中，可预定多种数据处理方法，模型训练节点可向模型推理节点指示上述多 种数据处理方法中的一种。例如，预定义n种数据处理方法，该n种数据处理方法对应不同的索引，模型训练节点可具体向模型推理节点指示某一种数据处理方法的索引。或者，模型训练节点可直接向模型推理节点指示对应数据处理方法的参数，例如，对输入数据的填充值、填充位置、和/或对输出数据的截取位置等。
步骤904：模型推理节点根据该AI模型的数据处理方法，进行模型推理。
在本申请中，模型推理节点可根据模型训练节点所指示的AI模型的数据处理方法，进行模型推理，完成不同无线资源配置下的推理任务，即模型推理节点使用与训练数据处理方法相同的规则，处理模型推理时的数据。
为了便于描述，在模型推理阶段，将初始的输入数据称为第一输入数据，或原始数据，将对第一输入数据处理后，AI模型的输入数据称为第二输入数据。AI模型的输出数据称为第一输出数据，将对第一输出数据处理后的数据，称为第二输出数据，或目标数据。
在一种设计中，所述模型推理节点所指示的AI模型的数据处理方法，包括以下至少一项：输入数据处理方法、或输出数据处理方法。该输入数据处理方法与模型训练阶段的输入训练数据的处理方法相同，或相对应。该输出数据处理方法与模型推理阶段的输出训练数据的处理方法相同，或相对应。例如，如图9C所示，模型推理节点可根据AI模型的数据处理方法中的输入数据处理方法，对第一输入数据进行处理，获得第二输入数据，该第二输入数据的格式与AI模型的格式相同。根据第二输入数据和AI模型，确定第一输出数据，该第一输出数据的格式与AI模型的输出格式相同。根据AI模型的数据处理方法中的输出数据处理方法，对第一输出数据进行处理，确定第二输出数据。该第二输出数据可认为是该AI模型的推理结果。
与上述训练过程相似，模型推理节点对第一输入数据进行处理，确定第二输入数据的过程，与对第一输出数据进行处理，获得第二输出数据的过程，并非一定执行，也可以不执行。比如，第一输入数据与AI模型的输入格式相匹配，或者与输入数据处理方法中所指示的填充后的数据长度相同等，则可以不再对第一输入数据进行处理，第一输入数据可直接作为AI模型的输入数据。或者，AI模型的输出数据满足无线资源配置的长度，或者，与对应标签的长度相同，或者，该AI模型的输出数据与所述输出数据处理方法中所指示的截取后的数据长度相同等，则无需对第一输出数据进行处理，第一输出数据可直接作为该AI推理的推理结果等。
在上述设计中，通过一个AI模型，可完成不同无线资源配置下的推理任务，无需针对每种无线资源配置各训练和部署一个AI模型，节省了训练和存储开销等。
前已述，所述AI模型的数据处理方法，包括以下至少一项：所述AI模型的输入数据处理方法，或者所述AI模型的输出数据处理方法。在一种设计中，重点介绍，所述AI模型的输入数据处理方法为输入数据填充方法，所述AI模型的输出数据处理方法为输出数据截取方法的过程。或者，还可以描述成：所述AI模型的数据处理方法，包括以下至少一项：所述AI模型的输入数据填充方法，或者所述AI模型的输出数据截取方法。
在该设计中，模型训练节点根据所述AI模型的数据处理方法，进行模型训练的过程，包括：模型训练节点根据所述输入数据填充方法，对第一输入训练数据进行填充，得到第二输入训练数据；根据第二输入训练数据和所述AI模型，确定第一输出训练数据；根据所述输出数据截取方法，对所述第一输出训练数据进行截取，得到第二输出训练数据。根 据所述第二输出训练数据，对所述AI模型进行参数调整。
其中，输入数据填充方法可包括以下至少一项：
填充后数据的长度。如未配置该项，默认的，填充后的输入训练数据的长度与AI模型的输入数据长度相同，即与AI模型的输入格式匹配。
填充数据的类型。例如，填充数据可以是0、或者可以是很大的正数，或者很小的负数等。
或填充数据的规则。例如，在输入训练数据的前面填充、后面填充、中间填充、两边填充、等间隔填充、或非等间隔填充等。
其中，输出数据截取方法可包括以下至少一项：
截取后数据的长度。如未配置该项，默认的，截取后的输出数据的长度与标签的长度相同，即与标签的格式匹配。
或截取数据的规则。例如，在输出数据的前面截取、后面截取、中间截取、两边截取、等间隔截取，或非等间隔截取等。
例如，模型训练节点可以收集足够多的训练数据。以监督学习为例，该训练数据中包括训练样本和标签，训练样本还可称为第一输入训练数据，所述第一输入训练数据可指需要输入到AI模型中进行模型训练的数据。确定AI模型的输入数据长度和输出数据长度，该AI模型的输入数据长度大于或等于所收集的训练数据的训练样本中长度最大的训练样本的长度，该AI模型的输出长度大于或等于所收集的训练数据的标签中长度最大标签的长度。或者说，该AI模型的输入数据长度大于或等于该AI模型应用场景下所有潜在无线资源配置中长度最大的原始数据的长度，该AI模型的输出长度大于或等于该AI模型应用场景下所有潜在无线资源配置中长度最大的目标数据的长度。模型训练节点，根据输入数据处理方法，对第一输入训练数据进行填充，得到第二输入训练数据，该第二输入训练数据可称为填充后的第一输入训练数据，第二输入训练数据的长度与AI模型的输入数据长度相同。具体的，可根据输入数据填充方法中的填充数据类型和填充数据规则等，对输入训练据进行填充。例如，AI模型的输入数据频域长度为8RB，第一输入训练数据的频域长度为4RB。在本申请中，可将第一输入训练数据的频域长度由4RB填充至8RB。具体如何将上述第一输入训练数据的频域长度由4RB填充至8RB的过程，可根据输入数据填充方法中的填充数据的类型(例如，具体的填充值)和填充数据的规则(例如，在前面填充、后面填充或等间隔填充等)确定。根据第二输入训练数据和AI模型，确定第一输出训练数据。根据所述输出数据截取方法，对第一输出训练数据截取，得到第二输出训练数据。例如，第二输出训练数据的长度与标签的长度相等。具体的，如何对第一输出数据进行截取，可根据输出数据截取方法中的截取规则所确定，例如在第一输出训练数据的前面截取、后面截取、或等间隔截取等。根据第二输出训练数据，对AI模型进行参数调整。例如，在本申请中，可根据第二输出训练数据和标签，计算损失函数的值。若损失函数的值小于门限值或者满足目标需求，则完成对AI模型的训练；否则，调整AI模型的参数，继续对AI模型训练。
在本申请中，模型训练节点可将上述训练好的AI模型发送给模型推理节点。模型推理节点可根据AI模型的数据处理方法，对AI模型进行数据推理。在该设计中，模型推理节点根据所述AI模型的数据处理方法，进行模型推理的过程，包括：
根据所述输入数据填充方法，对第一输入数据进行填充，得到第二输入数据；根据所 述第二输入数据和所述AI模型，确定第一输出数据。根据所述输出数据截取方法，对所述第一输出数据进行截取，所述截取还可称为抽取，截取后的第一输出数据可称为第二输出数据。该第二输出数据为AI模型的推理结果。
可选的，在模型推理过程中，对输入数据进行填充和输出数据进行截取的过程，与模型训练阶段中，对输入训练数据进行填充和对输出训练数据进行截取的过程相同。
以频域信道估计场景为例，介绍使用同一个AI模型进行不同带宽、不同参考信号图样下的信道估计。在该场景中，对AI模型的输入数据进行填充，对AI模型的输出数据进行截取。
为了便于理解，首先介绍信道估计的过程：单UE单天线单正交频分复用(orthogonal frequency division multiplexing，OFDM)符号的无线信号的传播模型为
Figure PCTCN2022144185-appb-000005
其中，Y表示接收的信号；S表示发送信号，在信道估计场景S表示参考信号；
Figure PCTCN2022144185-appb-000006
表示参考信号所在资源元素(resource element，RE)的频域信道(响应)；n表示噪声；·表示元素乘法，即两个向量或矩阵中索引相同的元素相乘。Y，
Figure PCTCN2022144185-appb-000007
和S的长度为L，L为参考信号所占用RE的数量。信道估计的过程为：通过接收信号Y和参考信号S，估计频域信道响应H。H为调度的带宽内所有RE上的信道，
Figure PCTCN2022144185-appb-000008
为H的一部分。在该场景中，在模型训练和模型推理过程中，AI模型的输入数据为接收信号Y和参考信号S，参考信号S是已知的，AI模型的输出数据为频域信道响应H。
在本申请中，基站调度的频域信道的带宽可能不同。例如，基站调度的频域信道的带宽可以为4个资源块(resource block，RB)、8个RB、或16个RB等。针对同一个调度带宽，UE可向基站发送不同类型的参考信号。例如，类型1的参考信号和类型2的参考信号。在本申请中，H的长度取决于调度带宽，Y、
Figure PCTCN2022144185-appb-000009
和S的长度取决于调度带宽和参考信号的图样。比如，设定基站调度的带宽为k个RB。由于一个RB包括12个RE，则H的长度为12k。如图10所示，类型1的参考信号为每间隔1个RE占用1个RE，每个RB，类型1的参考信号占用6个RE。如图11所示，类型2的参考信号每间隔4个RE占用2个RE，每个RB，类型2的参考信号占用4个RE。在调度的带宽为k个RB时，类型1的参考信号的Y，
Figure PCTCN2022144185-appb-000010
和S的长度为6k，类型2的参考信号的Y，
Figure PCTCN2022144185-appb-000011
和S的长度为4k。
如果利用AI方式进行频域信道估计，则AI模型的输入数据为Y和S。而对于类型1的参考信号，其Y和S的长度为6k，类型2的参考信号，其Y和S的长度为4k，k为基站调度的频域信道带宽。可以看出，在设计AI模型，需要根据调度带宽和参考信号类型的不同，分别设计AI模型。在该设计中，设计一个AI模型，该AI可适用于不同调度带宽和不同类型参考信号下的信道估计，从而节省AI模型的训练和存储开销。
在后续，将从收集训练数据、训练数据处理、模型训练、模型部署、和模型推理等方面，对信道估计的过程进行介绍。
收集训练数据。
根据AI模型的应用场景，确定训练数据集。为了使AI模型在各种无线资源配置下都能取得较好的性能，训练数据集中可尽可能地包括该应用场景下所有可能无线资源配置下的训练数据。每个训练数据中包括训练样本和标签，训练样本还可称为上述第一输入训练数据，包括接收信号Y和原始信号S，标签为频域信道响应H。例如，如果想到训练一个AI模型、可以在调度带宽为4RB和8RB的带宽下，对类型1和类型2的参考信号进行信道估计，则训练数据中包括调度带宽为4RB下，类型1的参考信号对应的训练数据和类型 2的参考信号对应的训练数据，和调度带宽为8RB下，类型1的参考信号对应的训练数据和类型2的参考信号对应的训练数据。每个训练数据中包括训练样本(Y、S)和标签(H)。
训练数据处理。
训练数据处理，包括对第一输入训练数据进行填充，填充后的输入训练数据称为第二输入训练数据，第二输入训练数据的长度与AI模型的输入数据长度相同。根据第二输入训练数据和AI模型，确定第一输出训练数据。对第一输出训练数据进行截取，确定第二输出训练数据，第二输出训练数据的长度与标签的长度相同。
在本申请中，对第一输入训练数据进行填充的过程，包括：首先针对不同类型的参考信号，可以按照其图样进行填充，将Y和S填充为H的长度相同。填充后，仅在参考信号所在RE对应的位置上为Y和S的真实值，其余RE位置上为填充的值。之后，根据AI模型的输入数据长度，即该AI模型所支持的频域带宽，对填充后的Y和S，再次进行填充，该次填充的规则可以在前面填充、在后面填充、在中间填充、在两边填充等中的一种或者多种等，不做限定。填充值可以是0，或者一个很大的正数，或者一个很小的负数等。填充后数据的长度，即第二输入训练数据的长度，与AI模型的输入数据长度相同。
例如，收集的训练数据中的训练样本，即第一输入训练数据中包括：
在调度带宽4RB下、参考信号类型1对应的接收信号
Figure PCTCN2022144185-appb-000012
Y i表示编号为i的接收信号Y，
Figure PCTCN2022144185-appb-000013
表示编号为i的接收信号Y中的第j个元素；
在调度带宽4RB下、参考信号类型2对应的接收信号
Figure PCTCN2022144185-appb-000014
在调度带宽8RB下、参考信号类型1对应的接收信号
Figure PCTCN2022144185-appb-000015
在调度带宽8RB下、参考信号类型2对应的接收信号
Figure PCTCN2022144185-appb-000016
对第一输入训练数据进行填充，填充的数据可以为0，填充后的输入训练数据的长度为96RE。可选的，由于在该训练场景下，调度带宽最大为8RB，即96RE，因此，可以考虑将AI模型的输入数据长度设计为96RE。在本申请中，将训练数据中训练样本的长度，填充为与AI模型的输入数据长度96RE相同。
在本申请中，可首先将不同类型的参考信号的Y填充为标签H的长度相同，之后对Y再次填充，具体的可在Y的后面进行填充。填充后的第一输入训练数据，即第二输入训练数据分别为：
在4RB的调度带宽下、参考信号类型1对应的Y 1填充后为
Figure PCTCN2022144185-appb-000017
在4RB的调度带宽下、参考信号类型2对应的Y 1填充后为
Figure PCTCN2022144185-appb-000018
在8RB的调度带宽下、参考信号类型1对应Y 3填充后为
Figure PCTCN2022144185-appb-000019
在8RB的调度带宽下、参考信号类型2对应Y 4填充后为
Figure PCTCN2022144185-appb-000020
其中，S的填充方法与上述Y的填充方法相似。填充后的S和Y，可称为第二输入训练数据。将填充后的Y和S，输入到AI模型中，该AI模型的输出称为第一输出训练数据，第一输出训练数据的长度与AI模型的输出数据长度相同。对第一输出训练数据进行截取，截取后的输出训练数据可称为第二输出训练数据，第二输出训练数据的长度与标签(频域信道响应H)的长度相同。对第一输出数据的截取规则可以是在前面截取、在后面截取、在中间截取，在两边截取等中的一种或者多种，不作限定。可选的，截取规则可以与填充 规则独立，也可以与填充规则相匹配。例如，对第一输入训练数据在后面填充，则对第一输出训练数据的截取是在前面截取等。
例如，假设AI模型的输出数据长度是96RE，在4RB的调度带宽下，其标签(频域信道响应H)的长度是48RB，则可以截取AI模型的输出数据中的前48个元素，与4RB的标签H进行比较。对于8RB的调度带宽下，则不需要对AI模型的输出数据截取，可以直接将AI模型的输出数据，与8RB的标签H进行比较等。
模型训练。
在模型训练过程中，模型训练节点可确定AI模型的结构，该结构中包括AI模型的输入格式和输出格式。该AI模型的输入格式中的输入数据长度大于或等于训练数据中的最大长度的训练样本的长度。该AI模型的输出格式中的输出数据长度大于或等于训练数据中的最大长度的标签的长度。使用上述收集的包含多种带宽、不同参考信号类型的处理后的训练数据，训练AI模型。其中，在AI模型的训练过程中，对训练样本，即第一输入训练数据进行填充，填充后输入训练数据的长度与AI模型的输入数据长度相同。对AI模型的第一输出训练数据进行截取，截取后的输出训练数据的长度与对应标签的长度相同。模型训练节点可根据第二输出训练数据和标签，确定损失函数的值；如果损失函数的值小于门限值或满足目标需求，则AI模型的训练完成；否则，调整AI模型的参数，继续对AI模型进行训练。
模型部署。
在本申请中，为了使模型推理节点可以正确使用同一个AI模型完成不同带宽，不同参考信号类型下的信道估计任务，模型训练节点除了需要将训练好的AI模型发送给模型推理节点外，还需要将AI模型的数据处理方法指示给模型推理节点。例如，模型部署节点可以向模型推理节点发送指示信息，该指示信息用于指示所述AI模型的数据处理方法，该指示信息可指示输入数据的填充规则和输出数据的截取规则等。
模型推理。
模型推理节点根据模型训练节点指示的AI模型的数据处理方法，进行模型推理，完成不同带宽、参考信号类型下的信道估计任务。例如，采用与模型训练过程相同的填充规则，对输入数据进行填充，采用与模型训练过程中相同的截取规则，对输出数据进行截取等。
在本申请中，以频域信道估计为例，描述利用AI模型信道估计的过程。应理解，信道估计还包括其它场景，每个场景下，输入数据和输出数据都不相同。例如，对使用传统方法粗估的信道需要进一步信道估计的场景下，输入数据为参考信号所占RE上的含噪的信道，输出数据为调度的带宽内所有RE上的信道，需要对输入数据进行填充，对输出数据进行截取，获得推理结果。又例如，时延域信道估计的场景下，输入数据为接收信号和参考信号，输出数据为时延域信道，需要对输入数据进行填充，无需对输出进行数据截取等。
在本申请中，通过对AI模型的输入数据进行填充，对AI模型的输出数据进行截取，可使用一个AI模型，完成多种带宽、不同参考信号图样情况下的信道估计任务。
以CSI反馈场景为例，介绍使用同一个AI模型进行不同带宽和/或不同天线端口(信道特征长度)下的CSI反馈。在该场景中，对AI模型的输入数据进行填充，对AI模型的 输出数据进行截取。
为了便于理解，首先对基于AI的CSI反馈流程进行介绍，如图12所示：
UE获取到下行信道响应H或下行信道的特征信息W，UE中部署有第一子模型，该第一子模型可称为子模型f。将上述下行信道响应H或下行信道的特征信息W，作为子模型f的输入，该子模型f的输出为CSI对应的反馈比特B，将反馈比特B发送给基站。基站中部署有第二子模型，该第二子模型可称为子模型g，将反馈比特B作为子模型g的输入，该子模型g的输出为恢复出的下行信道
Figure PCTCN2022144185-appb-000021
或下行信道的特征信息
Figure PCTCN2022144185-appb-000022
下行信道响应H的维度是带宽*天线端口数，天线端口包括基站天线端口和/或UE天线端口。基站配置的CSI反馈的带宽可能不同，例如，可以为4RB或8RB，以1RB粒度为例，则下行信道带宽维度的长度可以为4或8。基站配置的CSI反馈的天线端口数可能不同，例如，可以为16端口或32端口，则下行信道天线端口维度的长度可以为4或8。
下行信道的特征信息W的维度为子带数*特征向量长度(基站天线端口数)，基站配置的子带数可能不同，例如可以为6个子带或者12个子带，则下行信道的特征信息子带维度的长度可以为6或12。基站配置的CSI反馈的特征向量长度可能不同，例如，可以为16或32，则下行信道的特征信息特征向量维度的长度可以为16或32。
下行信道的特征信息W是根据下行信道响应H计算得到的，具体计算方法如下：
对每一个子带的下行信道响应H，计算协方差矩阵R UU，
Figure PCTCN2022144185-appb-000023
H i为第i个单位带宽上的信道，M为子带粒度，即每个子带包括多少个单位带宽，例如，单位带宽为RB，则H i为第i个RB上的信道，子带粒度为4RB，则M＝4。
对每个子带的协方差矩阵进行特征值分解，
Figure PCTCN2022144185-appb-000024
其中，λ包括R UU的至少1个特征值，
Figure PCTCN2022144185-appb-000025
为每个特征值对应的特征向量组成的矩阵。取最大特征值对应的特征向量为V m，即为该子带信道的特征向量，将所有子带的特征向量按顺序拼成矩阵，即为下行信道的特征矩阵V，也可称为下行信道的特征信息W。
可选的，也可对下行信道的特征矩阵V进行进一步处理，投影到更稀疏的空间，例如，通过离散傅里叶变换(discrete fourier transform，DFT)公式产生两组DFT基底，分别是空域基底S和频域基底F，空频联合投影参考如下公式：C＝S H*V*F,其中S H是S的共轭转置矩阵。投影后的C，也可称为下行信道的特征信息W。投影后，V的子带维度转化为时延维度，V的特征向量(基站天线端口)维度转化为波束(角度)维度，但维度的长度保持不变，因此，无论下行信道的特征信息W是投影前还是投影后，我们都将其维度描述为子带数*特征向量长度(基站天线端口数)。
收集训练数据。
在该场景下，训练样本和标签都为下行信道响应H或下行信道的特征信息W，训练数据中可以包括不同维度的下行信道响应H或下行信道的特征信息W，具体维度如上所示。
训练数据处理。
训练数据处理，包括对第一输入训练数据进行填充，填充后的输入训练数据称为第二输入训练数据，第二输入训练数据的长度与AI模型的输入数据长度相同。根据第二输入训练数据和AI模型，确定第一输出训练数据。对第一输出训练数据进行截取，确定第二输出训练数据，第二输出训练数据的长度与标签的长度相同。
在本实施例中，如果第一输入训练数据为下行信道响应H，则需要对其带宽维度和/或天线端口维度进行填充，如果第一输入训练数据为下行信道的特征信息W，则需要对其 子带维度和/或特征向量维度进行填充，得到第二输入训练数据，第二输入训练数据的长度与AI模型的输入数据长度相同。根据第二输入训练数据和AI模型，确定第一输出训练数据。如果标签是下行信道响应H，则需要对第一输出训练数据的带宽维度和/或天线端口维度进行截取，如果标签是下行信道的特征信息W，则需要对第一输出训练数据的子带维度和/或特征向量维度进行截取，得到第二输出训练数据，第二输出训练数据的长度与标签的长度相同。
具体填充方法和截取方法与之前的实施例类似，不再赘述。
模型训练。
在模型训练过程中，模型训练节点可确定AI模型的结构，该结构中包括AI模型的输入格式和输出格式。该AI模型的输入格式中的输入数据长度大于或等于训练数据中的最大长度的训练样本的长度。该AI模型的输出格式中的输出数据长度大于或等于训练数据中的最大长度的标签的长度。使用上述收集的包含多种带宽和/或天线端口数的原始训练数据处理后的训练数据，或者，使用上述收集的包含多种子带数和/或特征向量长度的原始训练数据处理后的训练数据，训练AI模型。其中，在AI模型的训练过程中，对训练样本，即第一输入训练数据进行填充，填充后输入训练数据的长度与AI模型的输入数据长度相同。对AI模型的第一输出训练数据进行截取，截取后的输出训练数据的长度与对应标签的长度相同。模型训练节点可根据第二输出训练数据和标签，确定损失函数的值；如果损失函数的值小于门限值或满足目标需求，则AI模型的训练完成；否则，调整AI模型的参数，继续对AI模型进行训练。
模型部署。
在本申请中，为了使模型推理节点可以正确使用同一个AI模型完成不同带宽和/或天线端口数下的CSI反馈任务，或者，使用同一个AI模型完成不同子带数和/或特征向量长度下的CSI反馈任务，模型训练节点除了需要将训练好的AI模型发送给模型推理节点外，还需要将AI模型的数据处理方法指示给模型推理节点。例如，模型部署节点可以向模型推理节点发送指示信息，该指示信息用于指示所述AI模型的数据处理方法，该指示信息可指示输入数据的填充规则和输出数据的截取规则等。
模型推理。
模型推理节点根据模型训练节点指示的AI模型的数据处理方法，进行模型推理，完成不同带宽和/或天线端口数下的CSI反馈任务，或者，完成不同子带数和/或特征向量长度下的CSI反馈任务。例如，采用与模型训练过程相同的填充规则，对输入数据进行填充，采用与模型训练过程中相同的截取规则，对输出数据进行截取等。
对于CSI反馈场景，AI模型的输入数据处理方法还可以用于指示AI模型的输入数据是下行信道还是下行信道特征信息，以及下行信道特征信息是下行信道的特征矩阵还是稀疏投影后的特征矩阵等。AI模型的输出数据处理方法还可以用于指示AI模型的输出数据是下行信道还是下行信道特征信息，以及下行信道特征信息是下行信道的特征矩阵还是稀疏投影后的特征矩阵等。
前已述，所述AI模型的数据处理方法，包括以下至少一项：所述AI模型的输入数据处理方法，或者所述AI模型的输出数据处理方法。在该设计中，重点介绍，所述AI模型的输入数据处理方法为输入数据截取方法，所述AI模型的输出数据处理方法为输出数据 填充方法。或者，还可以描述成：所述AI模型的数据处理方法包括以下至少一项：所述AI模型的输入数据截取方法，或所述AI模型的输出数据填充方法。
在该设计中，模型训练节点根据所述AI模型的数据处理方法，进行模型训练，包括：根据所述输入数据截取方法，对第一输入训练数据进行截取，得到第二输入训练数据；根据第二输入训练数据和所述AI模型，确定第一输出训练数据；根据所述输出数据填充方法，对所述第一输出训练数据进行填充，得到第二输出训练数据；根据所述第二输出训练数据，对所述AI模型进行参数调整。
在该设计中，模型推理节点根据所述AI模型的数据处理方法，进行模型推理，包括：根据所述输入数据截取方法，对第一输入数据进行截取，得到第二输入数据；根据第二输入数据和所述AI模型，确定第一输出数据；根据所述输出数据填充方法，对所述第一输出数据进行填充。填充后的第一输出数据可称为第二输出数据，该第二输出数据可称为AI模型的推理结果。
在该设计中，所述输入数据截取方法可包括以下至少一项：截取后的长度、或截取数据的规则等。所述输出数据填充方法可包括以下至少一项：填充后的长度、填充数据的类型、或填充数据的规则等。具体的截取与填充的说明，可参见上述设计。
例如，输入数据中的部分数据是有效数据的，剩余部分是无效数据的，或输入数据的能量主要集中在部分数据中，那么在模型推理或模型训练的过程中，可以对输入数据进行截取。将不同长度的输入数据截取成统一的长度，截取后输入数据的长度与AI模型的输入数据长度相同。根据截取后的输入数据和AI模型，确定输出数据。再对AI模型的输出数据进行填充。
以CSI反馈场景为例，当下行信道特征信息W是稀疏投影后的特征矩阵时，W在角度域和/或时延域是稀疏的，即虽然W的维度为子带数*特征向量长度，但该矩阵大部分的元素的值都很小，其总能量主要集中在若干角度和时延径上，则可以对下行信道特征信息W进行截取，仅保留数值较大的若干元素，得到第二输入数据。根据第二输入数据和AI模型，确定第一输出数据；再对第一输出数据进行填充，得到第二输出数据，即AI模型恢复的下行信道特征信息
Figure PCTCN2022144185-appb-000026
采用上述设计，通过对输入数据进行截取，对输出数据进行填充，可利用一个AI模型，完成对不同场景下的任务推理。
在另一种设计中，所述AI模型包括第一子模型和第二子模型。所述AI模型的数据处理方法，包括以下至少一项：对第一子模型的输出进行数据截取，或对第二子模型的输入进行数据填充。或者，还可以描述成：所述AI模型的数据处理方法，包括以下至少一项：对所述AI模型(例如，第一子模型)的输出数据截取方法，或者，对所述AI模型(例如，第二子模型)的输入数据填充方法。
在该设计中，模型训练节点根据所述AI模型的数据处理方法，进行模型训练，包括：根据第一子模型的输入训练数据和第一子模型，确定第一子模型的第一输出训练数据；根据所述输出数据截取方法，对所述第一输出训练进行数据截取，得到第一子模型的第二输出训练数据，根据第一子模型的第二输出训练数据得到第二子模型的第一输入训练数据；根据所述输入数据填充方法，对所述第二子模型的第一输入训练进行数据填充，得到第二输入训练数据；根据所述第二输入训练数据和所述第二子模型，确定第二子模型的输出训 练数据；根据所述第一子模型的输入训练数据和所述第二子模型的输出训练数据，对以下至少一项进行调整：所述第一子模型的模型参数，或所述第二子模型的模型参数。
模型推理节点根据所述AI模型的数据处理方法，进行模型推理，包括：确定第一子模型的第一输出数据；根据所述输出数据截取方法，对所述第一输出数据进行数据截取。模型推理节点根据所述AI模型的数据处理方法，进行模型推理，包括：根据所述输入数据填充方法，对第一输入数据进行填充，得到第二输入数据；根据所述第二输入数据和所述第二子模型，确定第一输出数据。关于具体的输入数据填充方法，和输出数据截取方法，可参见前述设计的说明。
在本申请中，第一子模型的模型推理节点和第二子模型的模型推理节点，可以为相同或不同的节点，不作限定。例如，在下行信道测量的场景中，第一子模型的推理节点为UE，第二子模型的推理节点为基站等。或者，在上行信道测量的场景中，第一子模型的推理节点为基站，第二子模型的推理节点为UE等。
以下行信道的CSI反馈场景为例，介绍使用同一AI模型进行不同反馈比特长度情况的CSI反馈。
在无线网络中，反馈比特B的长度是由基站确定的，通常不是固定的。例如，基站想要获得更高的CSI恢复精度，则可以让UE反馈更多的反馈比特，此时反馈比特B的长度较长。或者，基站想要降低CSI反馈的开销，则可以让UE反馈更少的反馈比特，此时反馈比特B的长度较短。在一种设计中，由于反馈比特B的长度不是固定的，那针对每个反馈比特B，都需要训练一对对应的子模型f和子模型g，使得AI模型的训练和存储开销都较大。
在本申请中，可以设计一个AI模型，该AI模型可适用于不同反馈比特B长度的情况。后续，将从收集训练数据、训练数据处理、模型训练、模型部署和模型推理等方面对本申请进行介绍。
收集训练数据。
在CSI反馈场景，反馈比特属于整体CSI反馈AI模型的中间结果，同时，也可以看做是子模型f的输出数据和子模型g的输入数据。在本申请中，我们将反馈比特B也看作训练数据。
根据AI模型的应用场景，确定训练数据集，为了使得AI模型在各种反馈比特B的长度下都能取得较好的性能，训练数据集中包括该应用场景下所有可能反馈比特B长度的训练数据。在本申请中，每个训练数据包括训练样本、反馈比特B和标签，其中训练样本和标签相同，都是下行信道或者是下行信道的特征信息H。例如，如果想要训练一个AI模型，可适用于反馈比特B的长度为20,30和40的CSI反馈，则训练数据中包括长度为20,40和60的反馈比特B。当然训练数据中还需要包括在各个长度的反馈比特B下，对应的训练样本和标签。例如，对于长度为20的反馈比特B，训练数据中需要包括训练样本、反馈比特B(20)和标签，该训练样本可以在UE获取的下行信道或下行信道的特征信息H，标签可以是基站正确恢复的下行信道或下行信道的特征信息
Figure PCTCN2022144185-appb-000027
训练数据处理。
在本申请，如图13所示，对训练数据处理包括对子模型f的输出数据进行截取，使其长度和反馈比特B的长度相同，以及对反馈比特B进行填充，使其长度与子模型g的输入数据长度相同。
在本申请中，截取的规则可以是在前面截取，在后面截取，在中间截取，在两边截取，等间隔截取或非等间隔截取等中的一种或者多种，不做限定。
例如，子模型f的输出数据为A＝[A 1,A 2,…,A 60]：
若反馈比特B的长度为20，则可以在子模型f的输出数据A中截取20个元素。例如，可以截取A中的前20个元素，截取后的输出数据为反馈比特B＝[A 1,A 2,…,A 20]，或者，在A中，每间隔3个元素截取1个元素，反馈比特B＝[A 1,A 4,…,A 58]。
若反馈比特B的长度为40，则可以在A中截取40个元素。例如，截取A中的前40个元素，反馈比特B＝[A 1,A 2,…,A 40]。或者，在A中，每间隔3个元素截取2个元素，反馈比特B＝[A 1,A 2,A 4,A 5,…,A 58,A 59]。
若反馈比特B的长度为60，可以不对A进行数据截取，反馈比特B＝A＝[A 1,A 2,…,A 60]。
在截取后，训练数据中包括了不同长度的反馈比特，需要对其进行填充，填充后的数据作为子模型g的输入。其中，填充规则可以是在前面填充，在后面填充，在中间填充，在两边填充，等间隔填充或非等间隔填充中的一种或者多种，不做限定；填充值可以是0，或者一个很大的正数，或者一个很小的负数等，不做限定。
例如，若反馈比特B长度为20，子模型g的第一输入数据为反馈比特B＝[A 1,A 2,…,A 20]，子模型f的输入长度为60，则填充后的子模型g的输入数据为C＝[A 1,A 2,…,A 20,0,0,…,0]。
若反馈比特B长度为40，子模型g的第一输入数据为反馈比特B＝[A 1,A 2,…,A 40]，子模型f的输入长度为60，则填充后的子模型g的输入数据为C＝[A 1,A 2,…,A 40,0,0,…,0]。
若反馈比特B长度为60，子模型f的输入数据长度为60，则不再对反馈比特B进行填充，子模型g的输入为C＝B＝[A 1,A 2,…,A 60]。
其中，填充规则可以与截取规则独立，也可以与截取规则匹配，例如，如果对子模型f的输出数据的截取是截取前面的元素，则对反馈比特的填充可以为在后面填充等。
需要注意的是，虽然在本申请中，我们将反馈比特B看作是训练数据，但实际上，反馈比特B是AI模型产生的中间结果，即只有在训练或推断过程中才会生成反馈比特B，上述对反馈比特B的截取和填充操作，是发生在训练或推断过程中的。
模型训练。
确定AI模型的输入数据长度和输出数据长度，子模型f的输出数据长度大于或等于训练数据中长度最大的反馈比特B的长度，子模型g的输入数据长度大于或等于训练数据中长度最大的反馈比特B的长度。
使用上述包含多种反馈比特B长度的处理后的训练数据训练子模型f和子模型g。前已述，训练数据中包括训练样本、标签和反馈比特B。模型训练的过程可包括：将输入训练数据(例如训练样本)输入到子模型f，得到子模型f的第一输出训练数据。对子模型f的第一输出训练数据进行截取，将其长度截取为目标长度，该目标长度为该AI模型应用场景中所有潜在可能的反馈比特长度，得到子模型f的第二输出数据，根据子模型f的第二输出数据得到子模型g的第一输入训练数据。对第一输入训练数据进行填充，填充后的第一输入训练数据的长度与子模型g的输入长度相同，填充后的第一输入训练数据称为第二输入训练数据。根据第二输入训练数据和子模型g，得到子模型g的输出训练数据。根据所述子模型f的输入训练数据和子模型g的输出训练数据，调整以下至少一项：所述子 模型f的参数，或所述子模型g的参数。例如，确定第一子模型f的某一项输入训练数据(即训练样本)和对应的标签；根据该标签与子模型g的输出训练数据，确定损失函数的值；若损失函数的值小于门限值或满足目标需求，则对子模型f和子模型g的训练完成；否则调整子模型f的参数和/或子模型g的参数，继续对子模型f和子模型g训练。
模型部署。
当执行模型训练和模型推理的节点不同时，需要进行模型部署，即模型训练节点需要将训练好的AI模型发送给模型推理节点。在本申请中，为了使模型推理节点正确使用同一个AI模型完成不同反馈比特B长度下的CSI反馈任务，模型训练节点除了需要将AI模型发送给模型推理节点外，还需要将AI模型的数据处理方法指示给模型推理节点。该AI模型的数据处理方法，包括数据的填充规则和截取规则等。
模型推理。
模型推理节点可根据模型训练节点指示的AI模型的数据处理方法，正确的使用AI模型完成不同反馈比特B长度下的CSI反馈任务，即模型推理节点使用与训练数据处理方法相同的规则处理模型推理时的数据。例如，UE使用子模型f时，UE获取下行信道响应H，将下行信道响应H输入到子模型f，获得子模型f的输出数据；UE根据基站配置的反馈比特B的长度，采用与训练时子模型f输出数据相同的截取规则，截取推理时子模型f的输出数据，获得推理的反馈比特B。UE将反馈比特B反馈给基站，基站采用与训练时对反馈比特B相同的填充方式填充推理时的反馈比特，然后使用子模型g恢复下行信道响应H。
在该设计中，通过将子模型f的输出截取成不同长度的反馈比特B，以及将不同长度的反馈比特B再填充成长度相同的子模型g的输入数据，使得可以训练一个AI模型，该AI模型可用于多种反馈比特B长度情况下的CSI反馈。
在本申请中，是以对第一子模型的第一输出数据进行截取，对第二子模型的输入数据进行填充为例描述，并不作为对本申请的限定。例如，在一种设计中，可以对第一子模型的输出数据进行填充，对第二子模型的输入数据进行截取等。此外，在本申请中，重点描述了，对第一子模型的输出数据和第二子模型的输入数据进行处理的过程。在本申请中，对第一子模型的输入数据和第二子模型的输出数据是否进行处理不作限定。例如，在一种设计中，可以对第一子模型的第一输入数据进行处理，获得第一子模型的第二输入数据；根据第一子模型的第二输入数据和第一子模型，确定第一子模型的第一输出数据；对第一子模型的第一输出数据进行截取(或填充)，得到第一子模型的第二输出数据。根据第一子模型的第二输出数据得到第二子模型的第一输入数据。对第二子模型的第一输入数据进行填充(或截取)，获得第二子模型的第二输入数据。根据第二子模型的第二输入数据和第二子模型，获得第二子模型的第一输出数据。对第二子模型的第一输出数据进行处理，获得第二子模型的第二输出数据。关于对于第一子模型的第一输入数据进行处理的方式，和对第二子模型的第一输出数据进行处理的方式，不作限定。例如，可将该实施例，与前述对输入数据和输出数据处理的过程相结合，比如，对第一子模型的第一输入数据进行填充或截取等，对第二子模型的第一输出数据进行截取或填充等。
在另一种设计中，所述AI模型的数据处理方法，包括以下至少一项：输入数据的拆分方法，或输出数据的重组方法。
在该设计中，模型训练节点根据所述AI模型的数据处理方法，进行模型训练，包括： 根据所述输入数据的拆分方法，对第一输入训练数据进行数据拆分，得到第二输入训练数据；根据第二输入训练数据和所述AI模型，确定第一输出训练数据；根据所述输出数据的重组方法，对第一输出训练数据进行数据重组，得到第二输出训练数据；根据所述第二输出训练数据，对所述AI模型进行参数调整。
模型推理节点根据所述AI模型的数据处理方法，进行模型推理，包括：根据所述输入数据的拆分方法，对第一输入数据进行数据拆分，得到第二输入数据；根据所述第二输入数据和AI模型，得到第一输出数据；根据所述输出数据的重组方法，对第一输出数据进行数据重组，得到第二输出数据。
例如，以Transformer等需要对AI模型的输入数据和/或输出数据进行处理的AI模型为例进行描述。
Transformer是一种序列到序列的AI模型，最早应用于自然语言处理领域，例如翻译等。由于句子的长短是各种各样的，它的特点之一是支持任意长度的输入。因此，理论上，Transformer可以适用于不同无线资源配置下的信号处理。如图14所示，Transformer的输入和输出都是向量的集合，在无线信号处理中，通常需要处理的对象是无线信号或无线信道等，而无线信号或无线信道的维度通常包括时域、频域和空域等。因此，将Transformer应用于无线信号处理中，需要对原始数据进行处理，将其转化为向量集合，输入到Transformer，再将Transformer输出的向量集合转化为目标数据。
将原始数据转化为向量集合、以及将向量集合转化为目标数据的方式可以有很多种。例如，以原始数据为12个子载波、14个OFDM符号的无线信道为例，如图15所示，可以在时域维度直接将12*14的原始数据拆分为14个维度为12的向量。或者，如图16所示，也可以先将12*14的原始数据拆分为12个2*7的矩阵，再将每个2*7的矩阵转化为维度为14的向量。当然，还可以采用其它拆分方法。同理，将Transformer输出的多个向量重组为目标数据的方法也有很多。
由于存在多种原始数据的拆分方法和多种目标数据的重组方法，为了保证模型推理节点可以正确使用该AI模型完成模型推理，模型训练节点除了需要将训练好的AI模型发送给模型推理节点，还需要将训练时对原始数据的拆分方法和对目标数据重组的方法通知给模型推理节点，模型推理节点使用与训练时相同的方法，对原始数据进行拆分，和对AI模型输出数据进行重组。
在本申请中，针对如Transformer等对AI的输入和输出有一定要求的AI模型，通过模型训练节点向模型推理节点指示原始数据的拆分方法和AI模型输出数据的重组方法，可以使AI模型正确处理无线网络中的信号。
可以理解的是，为了实现上述方法中的功能，模型训练节点和模型使用节点包括了执行各个功能相应的硬件结构和/或软件模块。本领域技术人员应该很容易意识到，结合本申请描述的各示例的单元及方法步骤，本申请能够以硬件或硬件和计算机软件相结合的形式来实现。某个功能究竟以硬件还是计算机软件驱动硬件的方式来执行，取决于技术方案的特定应用场景和设计约束条件。
图17和图18为本申请提供的可能的通信装置的结构示意图。这些通信装置可以用于实现上述方法中模型训练节点或模型推理节点的功能，因此也能实现上述方法所具备的有益效果。
如图17所示，通信装置1700包括处理单元1710和收发单元1720。通信装置1700用 于实现上述图8中所示的方法中第一节点的功能，该第一节点可以为模型训练节点或模型推理节点。当通信装置1700用于实现图8所示的方法中第一节点的功能时：
在一种设计中，处理单元1710用于确定模型的数据处理方法；以及，根据所述模型的数据处理方法，实现以下至少一项：进行模型训练，或进行模型推理；
其中，所述模型的数据处理方法包括以下至少一项：所述模型的输入数据填充方法，或所述模型的输出数据截取方法；或者，所述模型的数据处理方法包括以下至少一项：所述模型的输入数据截取方法，或所述模型的输出数据填充方法。
在另一种设计中，收发单元1720，用于向第二节点发送指示信息，所述指示信息用于指示模型的数据处理方法；
其中，所述模型的数据处理方法包括以下至少一项：所述模型的输入数据填充方法，或所述模型的输出数据截取方法；或者，所述模型的数据处理方法包括以下至少一项：所述模型的输入数据截取方法，或所述模型的输出数据填充方法。
有关处理单元1710和收发单元1720更详细的描述可以直接参考图8所示的方法中相关描述直接得到，这里不加赘述。
如图18所示，通信装置1800包括处理器1810和接口电路1820。处理器1810和接口电路1820之间相互耦合。可以理解的是，接口电路1820可以为收发器或输入输出接口。可选的，通信装置1800还可以包括存储器1830，用于存储处理器1810执行的指令或存储处理器1810运行指令所需要的输入数据或存储处理器1810运行指令后产生的数据。
当通信装置1800用于实现上述方法时，处理器1810用于实现上述处理单元1710的功能，接口电路1820用于实现上述收发单元1720的功能。
当上述通信装置为应用于第一节点的模块时，该第一节点模块实现上述方法中第一节点的功能。该第一节点的模块从第一节点中的其它模块(如射频模块或天线)接收信息，该信息是第二节点发送给第一节点的；或者，该第一节点模块向第一节点中的其它模块(如射频模块或天线)发送信息，该信息是第一节点发送给第二节点的。这里的第一节点模块可以是第一节点的基带芯片，也可以其他模块。
可以理解的是，本申请中的处理器可以是中央处理单元(central processing unit，CPU)，还可以是其它通用处理器、数字信号处理器(digital signal processor，DSP)、专用集成电路(application specific integrated circuit，ASIC)、现场可编程门阵列(field programmable gate array，FPGA)或者其它可编程逻辑器件、晶体管逻辑器件，硬件部件或者其任意组合。通用处理器可以是微处理器，也可以是任何常规的处理器。
本申请中的存储器可以是随机存取存储器、闪存、只读存储器、可编程只读存储器、可擦除可编程只读存储器、电可擦除可编程只读存储器、寄存器、硬盘、移动硬盘、CD-ROM或者本领域熟知的任何其它形式的存储介质。
一种示例性的存储介质耦合至处理器，从而使处理器能够从该存储介质读取信息，且可向该存储介质写入信息。存储介质也可以是处理器的组成部分。处理器和存储介质可以位于ASIC中。另外，该ASIC可以位于基站或终端中。当然，处理器和存储介质也可以作为分立组件存在于基站或终端中。
本申请中的方法可以全部或部分地通过软件、硬件、固件或者其任意组合来实现。当使用软件实现时，可以全部或部分地以计算机程序产品的形式实现。所述计算机程序产品包括一个或多个计算机程序或指令。在计算机上加载和执行所述计算机程序或指令时，全 部或部分地执行本申请所述的流程或功能。所述计算机可以是通用计算机、专用计算机、计算机网络、网络设备、用户设备、核心网设备、OAM或者其它可编程装置。所述计算机程序或指令可以存储在计算机可读存储介质中，或者从一个计算机可读存储介质向另一个计算机可读存储介质传输，例如，所述计算机程序或指令可以从一个网站站点、计算机、服务器或数据中心通过有线或无线方式向另一个网站站点、计算机、服务器或数据中心进行传输。所述计算机可读存储介质可以是计算机能够存取的任何可用介质或者是集成一个或多个可用介质的服务器、数据中心等数据存储设备。所述可用介质可以是磁性介质，例如，软盘、硬盘、磁带；也可以是光介质，例如，数字视频光盘；还可以是半导体介质，例如，固态硬盘。该计算机可读存储介质可以是易失性或非易失性存储介质，或可包括易失性和非易失性两种类型的存储介质。
在本申请中，如果没有特殊说明以及逻辑冲突，不同的实施例之间的术语和/或描述具有一致性、且可以相互引用，不同的实施例中的技术特征根据其内在的逻辑关系可以组合形成新的实施例。
本申请中，“至少一个”是指一个或者多个，“多个”是指两个或两个以上。“和/或”，描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B的情况，其中A，B可以是单数或者复数。在本申请的文字描述中，字符“/”，一般表示前后关联对象是一种“或”的关系；在本申请的公式中，字符“/”，表示前后关联对象是一种“相除”的关系。“包括A，B或C中的至少一个”可以表示：包括A；包括B；包括C；包括A和B；包括A和C；包括B和C；包括A、B和C。
可以理解的是，在本申请中涉及的各种数字编号仅为描述方便进行的区分，并不用来限制本申请的范围。上述各过程的序号的大小并不意味着执行顺序的先后，各过程的执行顺序应以其功能和内在逻辑确定。

Claims (21)

  1. 一种模型的数据处理方法，其特征在于，包括：
    确定模型的数据处理方法；
    根据所述模型的数据处理方法，实现以下至少一项：进行模型训练，或进行模型推理；
    其中，所述模型的数据处理方法包括以下至少一项：所述模型的输入数据填充方法，或所述模型的输出数据截取方法；或者，
    所述模型的数据处理方法包括以下至少一项：所述模型的输入数据截取方法，或所述模型的输出数据填充方法。
  2. 如权利要求1所述的方法，其特征在于，所述确定模型的数据处理方法，包括：
    接收来自第一节点的指示信息，所述指示信息用于指示所述模型的数据处理方法；或者，
    根据协议约定，确定所述模型的数据处理方法。
  3. 如权利要求1或2所述的方法，其特征在于，所述输入数据填充方法或所述输出数据填充方法，包括以下至少一项：填充后数据的长度、填充数据的类型、或填充数据的规则。
  4. 如权利要求1至3中任一项所述的方法，其特征在于，所述输出数据截取方法或所述输入数据截取方法，包括以下至少一项：截取后数据的长度、或截取数据的规则。
  5. 如权利要求1至4中任一项所述的方法，其特征在于，所述根据所述模型的数据处理方法，进行模型推理，包括：
    根据所述输入数据填充方法，对第一输入数据进行数据填充，得到第二输入数据；
    根据所述第二输入数据和所述模型，确定第一输出数据；
    根据所述输出数据截取方法，对所述第一输出数据进行数据截取。
  6. 如权利要求1至5中任一项所述的方法，其特征在于，所述根据所述模型的数据处理方法，进行模型训练，包括：
    根据所述输入数据填充方法，对第一输入训练数据进行数据填充，得到第二输入训练数据；
    根据所述第二输入训练数据和所述模型，确定第一输出训练数据；
    根据所述输出数据截取方法，对所述第一输出训练进行数据截取，得到第二输出训练数据；
    根据所述第二输出训练数据，对所述模型进行参数调整。
  7. 如权利要求1至4中任一项所述的方法，其特征在于，所述根据所述模型的数据处理方法，进行模型推理，包括：
    根据所述输入数据截取方法，对第一输入数据进行数据截取，得到第二输入数据；
    根据第二输入数据和所述模型，确定第一输出数据；
    根据所述输出数据填充方法，对所述第一输出数据进行数据填充。
  8. 如权利要求1至4或7中任一项所述的方法，其特征在于，所述根据所述模型的数据处理方法，进行模型训练，包括：
    根据所述输入数据截取方法，对第一输入训练数据进行数据截取，得到第二输入训练数据；
    根据第二输入训练数据和所述模型，确定第一输出训练数据；
    根据所述输出数据填充方法，对所述第一输出训练数据进行数据填充，得到第二输出训练数据；
    根据所述第二输出训练数据，对所述模型进行参数调整。
  9. 如权利要求1至4中任一项所述的方法，其特征在于，所述模型包括第一子模型，所述根据所述模型的数据处理方法，进行模型推理，包括：
    确定第一子模型的第一输出数据；
    根据所述输出数据截取方法，对所述第一输出数据进行数据截取。
  10. 如权利要求1至4、或9中任一项所述的方法，其特征在于，所述模型包括第二子模型，所述根据所述模型的数据处理方法，进行模型推理，包括：
    根据所述输入数据填充方法，对第一输入数据进行数据填充，得到第二输入数据；
    根据所述第二输入数据和所述第二子模型，确定第一输出数据。
  11. 如权利要求1至4、9或10中任一项所述的方法，其特征在于，所述模型包括第一子模型和第二子模型，所述根据所述模型的数据处理方法，进行模型训练，包括：
    根据第一子模型的输入训练数据和第一子模型，确定第一子模型的第一输出训练数据；
    根据所述输出数据截取方法，对所述第一输出训练数据进行数据截取，得到第二子模型的第一输入训练数据；
    根据所述输入数据填充方法，对所述第二子模型的第一输入训练数据进行数据填充，得到第二输入训练数据；
    根据所述第二输入训练数据和所述第二子模型，确定第二子模型的输出训练数据；
    根据所述第一子模型的输入训练数据和所述第二子模型的输出训练数据，对以下至少一项进行调整：所述第一子模型的模型参数，或，所述第二子模型的模型参数。
  12. 一种模型的数据处理方法，其特征在于，包括：
    向第二节点发送指示信息，所述指示信息用于指示模型的数据处理方法；
    其中，所述模型的数据处理方法包括以下至少一项：所述模型的输入数据填充方法，或所述模型的输出数据截取方法；或者，
    所述模型的数据处理方法包括以下至少一项：所述模型的输入数据截取方法，或所述模型的输出数据填充方法。
  13. 如权利要求12所述的方法，其特征在于，所述输入数据填充方法或者所述输出数据填充方法，包括以下至少一项：填充后数据的长度、填充数据的类型、或填充数据的规则。
  14. 如权利要求12或13所述的方法，其特征在于，所述输出数据截取方法或所述输入数据截取方法，包括以下至少一项：截取后数据的长度、或截取数据的规则。
  15. 一种通信装置，其特征在于，包括用于实现权利要求1至11中任一项所述方法的单元。
  16. 一种通信装置，其特征在于，包括处理器和存储器，所述处理器和存储器耦合，所述处理器用于实现权利要求1至11中任一项所述的方法。
  17. 一种通信装置，其特征在于，包括用于实现权利要求12至14中任一项所述方法的单元。
  18. 一种通信装置，其特征在于，包括处理器和存储器，所述处理器和存储器耦合，所 述处理器用于实现权利要求12至14中任一项所述的方法。
  19. 一种通信系统，其特征在于，包括权利要求15或16所述的通信装置，和权利要求17或18所述的通信装置。
  20. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质中存储有指令，当所述指令在计算机上运行时，使得计算机执行权利要求1至11中任一项所述的方法，或者权利要求12至14中任一项所述的方法。
  21. 一种计算机程序产品，其特征在于，包括指令，当所述指令在计算机上运行时，使得计算机执行权利要求1至11中任一项所述的方法，或者权利要求12至14中任一项所述的方法。
PCT/CN2022/144185 2021-12-31 2022-12-30 模型的数据处理方法及装置 WO2023125985A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111658528.4A CN116418686A (zh) 2021-12-31 2021-12-31 模型的数据处理方法及装置
CN202111658528.4 2021-12-31

Publications (1)

Publication Number Publication Date
WO2023125985A1 true WO2023125985A1 (zh) 2023-07-06

Family

ID=86998196

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/144185 WO2023125985A1 (zh) 2021-12-31 2022-12-30 模型的数据处理方法及装置

Country Status (2)

Country Link
CN (1) CN116418686A (zh)
WO (1) WO2023125985A1 (zh)

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109885832A (zh) * 2019-02-14 2019-06-14 平安科技（深圳）有限公司 模型训练、语句处理方法、装置、计算机设备及存储介质
CN110855474A (zh) * 2019-10-21 2020-02-28 广州杰赛科技股份有限公司 Kqi数据的网络特征提取方法、装置、设备及存储介质
CN111382557A (zh) * 2020-03-31 2020-07-07 字节跳动有限公司 非定长输入数据的批量处理方法、装置、终端及存储介质
CN111985243A (zh) * 2019-05-23 2020-11-24 中移(苏州)软件技术有限公司 情感模型的训练方法、情感分析方法、装置及存储介质
CN112149736A (zh) * 2020-09-22 2020-12-29 腾讯科技（深圳）有限公司 一种数据处理方法、装置、服务器及介质
CN112215238A (zh) * 2020-10-29 2021-01-12 支付宝(杭州)信息技术有限公司 一种通用特征提取模型构建方法、系统及装置
US20210097383A1 (en) * 2019-09-30 2021-04-01 International Business Machines Corporation Combined Data Pre-Process And Architecture Search For Deep Learning Models
CN112884092A (zh) * 2021-04-28 2021-06-01 深圳索信达数据技术有限公司 Ai模型生成方法、电子设备及存储介质

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109885832A (zh) * 2019-02-14 2019-06-14 平安科技（深圳）有限公司 模型训练、语句处理方法、装置、计算机设备及存储介质
CN111985243A (zh) * 2019-05-23 2020-11-24 中移(苏州)软件技术有限公司 情感模型的训练方法、情感分析方法、装置及存储介质
US20210097383A1 (en) * 2019-09-30 2021-04-01 International Business Machines Corporation Combined Data Pre-Process And Architecture Search For Deep Learning Models
CN110855474A (zh) * 2019-10-21 2020-02-28 广州杰赛科技股份有限公司 Kqi数据的网络特征提取方法、装置、设备及存储介质
CN111382557A (zh) * 2020-03-31 2020-07-07 字节跳动有限公司 非定长输入数据的批量处理方法、装置、终端及存储介质
CN112149736A (zh) * 2020-09-22 2020-12-29 腾讯科技（深圳）有限公司 一种数据处理方法、装置、服务器及介质
CN112215238A (zh) * 2020-10-29 2021-01-12 支付宝(杭州)信息技术有限公司 一种通用特征提取模型构建方法、系统及装置
CN112884092A (zh) * 2021-04-28 2021-06-01 深圳索信达数据技术有限公司 Ai模型生成方法、电子设备及存储介质

Also Published As

Publication number Publication date
CN116418686A (zh) 2023-07-11

Similar Documents

Publication Publication Date Title
Lin et al. Relay-assisted cooperative federated learning
WO2023036268A1 (zh) 一种通信方法及装置
WO2022172198A1 (en) Deep generative models for downlink channel estimation in fdd massive mimo systems
Zhu et al. OFDM-based massive grant-free transmission over frequency-selective fading channels
Cui et al. On the Physical Layer of Digital Twin: An Integrated Sensing and Communications Perspective
WO2023125985A1 (zh) 模型的数据处理方法及装置
WO2023126007A1 (zh) 信道信息传输方法及装置
WO2023098860A1 (zh) 通信方法和通信装置
Chen et al. Federated learning: sum power constraints optimization design
WO2023016508A1 (zh) 一种信道信息反馈、恢复方法及装置
Jing et al. Transceiver beamforming for over-the-air computation in massive MIMO systems
Yi et al. 3-D Positioning and Resource Allocation for Multi-UAV Base Stations Under Blockage-Aware Channel Model
WO2023231934A1 (zh) 一种通信方法及装置
Moorthy et al. Swarm UAV networking with collaborative beamforming and automated ESN learning in the presence of unknown blockages
WO2024027424A1 (zh) 一种数据质量度量方法及装置
WO2023093777A1 (zh) 一种通信方法及装置
WO2023125996A1 (zh) 一种上行预编码方法及装置
WO2023231881A1 (zh) 一种模型应用方法及装置
WO2023066346A1 (zh) 一种通信方法及装置
WO2023093725A1 (zh) 一种校准方法及装置
Sande et al. Fast convergence for efficient beamforming in massive MIMO heterogeneous networks
WO2024051789A1 (zh) 一种波束管理方法
WO2023036280A1 (zh) 一种模型测试方法及装置
WO2024045888A1 (zh) 一种处理装置及控制方法
WO2024109682A1 (zh) 一种用于定位的方法及装置

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22915232

Country of ref document: EP

Kind code of ref document: A1