Data processing method and apparatus for models

Abstract

A data processing method and an apparatus for a model are provided. The method includes: a first node determines a data processing method for the model; and the first node performs at least one of the following: performing model training or performing model inference according to the data processing method for the model. Based on the method and the apparatus, inference tasks for different radio resource configurations can be completed by using one model, so that the overhead for training and maintaining the model is reduced.

Description

[Technical Field]

[0001] The present application relates to the field of communication technology, and in particular to a data processing method and apparatus for models. [Background technology]

[0002] In wireless communication networks, e.g., mobile communication networks, the services supported by the network are becoming increasingly diverse, and therefore the requirements that need to be met are also becoming increasingly diverse. For example, networks need to be capable of supporting ultra-high speeds, ultra-low latency, and / or extremely large connections. This characteristic makes network planning, network configuration, and / or resource scheduling increasingly complex. Furthermore, as network capabilities become increasingly powerful, e.g., to support increasingly higher spectrum and new technologies such as high-order multiple-input multiple-output (MIMO) technology, beamforming, and / or beam management, network energy conservation has become a hot research topic. These new requirements, new scenarios, and new features pose unprecedented challenges to network planning, operation, and maintenance, as well as efficient operation. To meet this challenge, artificial intelligence techniques may be introduced into wireless communication networks to implement network intelligence. Based on this, how to effectively implement artificial intelligence in networks is a worthy research topic. Summary of the Invention

[0003] The present application provides a data processing method and apparatus for models so as to complete inference tasks for different radio resource configurations by using one model, thereby reducing the overhead for training and maintaining the model.

[0004] According to a first aspect, there is provided a data processing method for a model. The method is executed by a first node, which is a model training node, a model inference node, etc. Alternatively, the method may be executed by a component (a processor, a chip, or other component) configured in the first node, or may be executed by a software module or the like. The method includes determining a data processing method for the model, and performing at least one of the following: performing model training or performing model inference according to the data processing method for the model.

[0005] Optionally, the data processing method for the model includes at least one of the following: an input data padding method for the model, or an output data truncation method for the model; alternatively, the data processing method for the model includes at least one of the following: an input data truncation method for the model, or an output data padding method for the model.

[0006] According to the above method, when the first node is a model inference node, the model inference node can complete inference tasks for different radio resource configurations by using the same model according to the data processing method. Compared with a method in which a corresponding model is set separately for each radio resource configuration, this method can reduce the overhead for training and maintaining the model.

[0007] In the design, determining a data processing method for the model includes receiving, from a second node, instruction information instructing the data processing method for the model. The first node is different from the second node. For example, the first node may be a model inference node or a model training node and may receive instruction information instructing the data processing method for the model from another node (e.g., the second node). For example, the second node is an OAM. Alternatively, when the first node is a model inference node, the second node may be a model training node. In this case, the model inference node may receive instruction information, etc., from the model training node. Alternatively, the data processing method for the model is determined as defined in the protocol.

[0008] According to the above method, the model training node or the model inference node may obtain the data processing method for the model as defined in the protocol. Alternatively, the model training node or the model inference node may receive instructions from other nodes and determine the data processing method for the model based on the instructions from the other nodes. The model training node or the model inference node may flexibly determine the data processing method for the model.

[0009] In the design, the input data padding method or the output data padding method includes at least one of the following: a length of padded data, a padding data type, or a data padding rule.

[0010] In the design, the output data truncation method or the input data truncation method includes at least one of the following: length of data after truncation or data truncation rule.

[0011] In the design, performing model inference according to a data processing method for the model includes performing data padding on first input data according to an input data padding method to obtain second input data, determining first output data based on the second input data and the model, and performing data truncation on the first output data according to an output data truncation method.

[0012] According to the above method in which the model inference node pads input data and truncates output data, inference tasks for different radio resource configurations can be completed by using one model, such that the overhead for training and maintaining the model is reduced. Optionally, in this design, the input length of the model is longer than or equal to the length of the input data with the longest input data length among all potential radio resource configurations, and the output length of the model is longer than or equal to the length of the output data with the longest output data length among all potential radio resource configurations.

[0013] In the design, performing model training according to the data processing method for the model includes performing data padding on first input training data according to an input data padding method to obtain second input training data; determining first output training data based on the second input training data and the model; performing data truncation on the first output training data according to an output data truncation method to obtain second output training data; and performing parameter adjustment on the model based on the second output training data.

[0014] According to the above method, in the model training process, training data under multiple radio resource settings is collected as much as possible, the training data under each radio resource setting is processed separately by using a data processing method, and the AI ​​model is trained based on the processed training data until the AI ​​model satisfies a condition, in which case the model training ends. The learning method in the model training process is not limited. For example, the learning method may be a supervised learning (label learning) method, an unsupervised learning (non-label learning) method, or a reinforcement learning method. By using this method, the trained model can satisfy inference tasks under various radio resource settings, thereby reducing the overhead for model training.

[0015] In the design, performing model inference according to a data processing method for the model includes performing data truncation on first input data according to an input data truncation method to obtain second input data, determining first output data based on the second input data and the model, and performing data padding on the first output data according to an output data padding method.

[0016] According to the above method in which the model inference node truncates input data and pads output data, inference tasks for different radio resource configurations can be completed by using one model, so that the overhead for training and maintaining the model is reduced.

[0017] In the design, performing model training according to the data processing method for the model includes performing data truncation on first input training data according to an input data truncation method to obtain second input training data; determining first output training data based on the second input training data and the model; performing data padding on the first output training data according to an output data padding method to obtain second output training data; and performing parameter adjustment on the model based on the second output training data.

[0018] According to the above design, adjusting the model parameters in the model training process includes adjusting 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 values ​​of the neurons, the activation functions of the neurons, or the offsets in the activation functions, so that the difference between the outputs of the neural network elements and the ideal target values ​​is as small as possible.

[0019] In the design, the model includes a first submodel, and performing model inference according to a data processing method for the model includes determining first output data of the first submodel and performing data truncation on the first output data according to an output data truncation method.

[0020] In the design, the model includes a second sub-model, and performing model inference according to the data processing method for the model includes performing data padding on the first input data according to the input data padding method to obtain second input data, and determining first output data based on the second input data and the second sub-model.

[0021] According to the above method, multiple types of sub-models used in pairs can be trained simultaneously. For example, the first sub-model can be deployed on the terminal side, and the second sub-model can be deployed on the base station side. Moreover, the output data of the first sub-model can be truncated so that the interface overhead can be reduced.

[0022] In the design, the model includes a first sub-model and a second sub-model, and performing model training according to a data processing method for the model includes determining first output training data of the first sub-model based on the input training data of the first sub-model and the first sub-model, performing data truncation on the first output training data according to an output data truncation method to obtain first input training data of the second sub-model, performing data padding on the first input training data of the second sub-model according to an input data padding method to obtain second input training data, determining output training data of the second sub-model based on the second input training data and the second sub-model, and adjusting at least one of model parameters of the first sub-model or model parameters of the second sub-model based on the input training data of the first sub-model and the output training data of the second sub-model.

[0023] According to a second aspect, there is provided a data processing method for a model. The method is performed by a first node, which may be a model training node, a model inference node, or the like. Alternatively, the method may be performed by a component (a processor, chip, or other component) configured in the first node, or may be performed by a software module or the like. The method includes sending instruction information to a second node instructing a data processing method for the model.

[0024] The data processing method for the model includes at least one of the following: an input data padding method for the model, or an output data truncation method for the model; or the data processing method for the model includes at least one of the following: an input data truncation method for the model, or an output data padding method for the model.

[0025] According to the above method, in the example, the first node is a model training node. The model training node can determine a data processing method based on the data processing method used during model training and instruct the model inference node to use the data processing method. When performing model inference, the model inference node processes the data in the same way as used during training, so that the model inference node can accurately use the model to perform model inference, thereby improving the accuracy of the output result of model inference and the inference speed.

[0026] In the design, the input data padding method or the output data padding method includes at least one of the following: a length of padded data, a padding data type, or a data padding rule.

[0027] In the design, the output data truncation method or the input data truncation method includes at least one of the following: length of data after truncation or data truncation rule.

[0028] According to a third aspect, an apparatus is provided. For advantageous effects, see the description of the first aspect. The apparatus may be a first node, which may be a model training node or a model inference node. Alternatively, the apparatus may be configured in the first node or may be usable in conjunction with the first node. In design, the apparatus may include units in one-to-one correspondence with the methods / operations / steps / acts described in the first aspect. The units may be implemented using hardware circuits, software, or a combination of hardware circuits and software.

[0029] For example, the device may include a processing unit, and the processing unit may perform the corresponding functions in any design example of the first aspect. Specifically, it is as follows.

[0030] The processing unit is configured to determine a data processing method for the model, and perform at least one of the following: performing model training or performing model inference according to the data processing method for the model; The data processing method for the model includes at least one of the following: an input data padding method for the model, or an output data truncation method for the model; or the data processing method for the model includes at least one of the following: an input data truncation method for the model, or an output data padding method for the model.

[0031] For the specific execution process of the processing unit, please refer to the first aspect, and the details will not be described again here.

[0032] For example, the apparatus may include a processor configured to perform the method described in the first aspect. The apparatus may further include a memory configured to store instructions and / or data. The memory is coupled to the processor. When the processor executes program instructions stored in the memory, the method described in the first aspect may be performed. In a possible design, the apparatus may include: a memory configured to store program instructions; a processor configured to determine a data processing method for the model and perform at least one of the following in accordance with the data processing method for the model: performing model training; or performing model inference; The data processing method for the model includes at least one of the following: an input data padding method for the model, or an output data truncation method for the model; or the data processing method for the model includes at least one of the following: an input data truncation method for the model, or an output data padding method for the model.

[0033] For the specific execution process of the processor, please refer to the description of the first aspect, and the details will not be described again.

[0034] According to a fourth aspect, an apparatus is provided. For advantageous effects, see the description of the second aspect. The apparatus may be a first node, and the first node may be a model training node. Alternatively, the apparatus may be configured in the first node, or may be usable in conjunction with the first node, etc. In design, the apparatus may include units in one-to-one correspondence with the methods / operations / steps / acts described in the second aspect. The units may be implemented using hardware circuits, software, or a combination of hardware circuits and software.

[0035] For example, the device may include a communication unit, and the communication unit may perform corresponding functions in any design example of the second aspect. Specifically, it is as follows.

[0036] the communication unit is configured to send instruction information to the second node instructing how to process data for the model; The data processing method for the model includes at least one of the following: an input data padding method for the model, or an output data truncation method for the model; or the data processing method for the model includes at least one of the following: an input data truncation method for the model, or an output data padding method for the model.

[0037] For the specific execution process of the communication unit, please refer to the second aspect, and the details will not be described again here.

[0038] For example, the apparatus includes a processor configured to control a communication interface to perform the method described in the first aspect. The apparatus may further include a memory configured to store instructions and / or data. The memory is coupled to the processor. When the processor executes program instructions stored in the memory, the method described in the second aspect may be performed. In a possible design, the apparatus may further include a communication interface, which is used by the apparatus to communicate with other devices. For example, the communication interface may be a transceiver, a circuit, a bus, a module, a pin, or other type of communication interface, and the other device may be a model inference node, etc. In a possible design, the apparatus includes: a memory configured to store program instructions; a processor configured to control the communication interface to send instruction information to the second node instructing how to process data for the model; The data processing method for the model includes at least one of the following: an input data padding method for the model, or an output data truncation method for the model; or the data processing method for the model includes at least one of the following: an input data truncation method for the model, or an output data padding method for the model.

[0039] For the specific execution process of the communication interface and the processor, please refer to the description of the second aspect, and the details will not be described again.

[0040] According to a fifth aspect, the present application further provides a computer-readable storage medium comprising instructions which, when executed on a computer, enable the computer to carry out the method of the first aspect or the second aspect.

[0041] According to a sixth aspect, the present application further provides a chip system, the chip system including a processor, and optionally a memory, configured to perform the method of the first or second aspect, the chip system may include a chip, or may include a chip and other discrete components.

[0042] According to a seventh aspect, the present application further provides a computer program product comprising instructions which, when executed on a computer, enable the computer to carry out the method of the first aspect or the second aspect.

[0043] According to an eighth aspect, the present application further provides a system, the system comprising the device of the third aspect and the device of the fourth aspect. [Brief explanation of the drawings]

[0044] [Figure 1] 1 is a diagram of a communication system according to the present application; [Figure 2] FIG. 1 is a diagram illustrating the deployment of an AI model according to the present application. [Figure 3] FIG. 1 is a diagram illustrating the deployment of an AI model according to the present application. [Figure 4a] 1 is a diagram of the architecture of a communication system according to the present application; [Figure 4b] 1 is a diagram of the architecture of a communication system according to the present application; [Figure 4c] FIG. 1 is a diagram of training an AI model by bandwidth according to the present application. [Figure 5] FIG. 1 is a diagram of the application architecture of the AI ​​model according to the present application. [Figure 6] FIG. 1 is a diagram of a neuron according to the present application. [Figure 7] FIG. 1 is a diagram of a neuronal network according to the present application. [Figure 8] 1 is a flowchart of a data processing method for a model according to the present application. [Figure 9A] 1 is a flowchart of a data processing method for a model according to the present application. [Figure 9B]1 shows the data processing process in the model training phase. [Figure 9C] 1 shows the data processing process in the model inference phase. [Figure 10] FIG. 1 is a diagram of a type-1 reference signal. [Figure 11] FIG. 1 is a diagram of a type-2 reference signal. [Figure 12] FIG. 1 is a diagram of a CSI feedback procedure according to the present application. [Figure 13] FIG. 1 is a diagram of data processing in the CSI feedback procedure according to the present application. [Figure 14] FIG. 1 is a diagram of a Transformer model according to the present application. [Figure 15] FIG. 1 is a diagram illustrating the division of raw data into vectors according to the present application. [Figure 16] FIG. 1 is a diagram illustrating the division of raw data into vectors according to the present application. [Figure 17] 1 is a diagram of the structure of a communication device according to the present application; [Figure 18] 1 is a diagram of the structure of a communication device according to the present application; DETAILED DESCRIPTION OF THE INVENTION

[0045] FIG. 1 is a diagram illustrating the architecture of a communication system 1000 to which the present application is applied. As shown in FIG. 1, the communication system includes a wireless access network 100 and a core network 200. Optionally, the communication system 1000 may further include the Internet 300. The wireless access network 100 may include at least one access network device (e.g., 110a and 110b in FIG. 1) and may further include at least one terminal (e.g., 120a-120j in FIG. 1). The terminal 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 may be different physical devices independent of each other, or the functionality of the core network device and the logical functionality of the access network device may be integrated into the same physical device, or some of the functionality of the core network device and some of the functionality of the access network device may be integrated into one physical device. The terminal devices may be connected to each other in a wired or wireless manner, and the access network devices may be connected to each other in a wired or wireless manner. FIG. 1 is merely a diagram. The communication system may further include other network devices, such as wireless relay devices, wireless backhaul devices, etc., which are not shown in FIG.

[0046] The access network device may be a base station, an evolved NodeB (eNodeB), a transmission reception point (TRP), a next generation NodeB (gNB) in a fifth generation (5G) mobile communication system, an access network device in an open radio access network (O-RAN), a next generation base station in a sixth generation (6G) mobile communication system, a base station in a future mobile communication system, an access node in a wireless fidelity (Wi-Fi) system, etc., or may be a module or unit, such as a central unit (CU), a distributed unit (DU), a central unit control plane (CU-CP) module, or a central unit user plane (CU-UP) module that realizes part of the functions of a base station. The access network device may be a macro base station (e.g., 110a in FIG. 1), or may be a micro base station or an indoor base station (e.g., 110b in FIG. 1), or may be a relay node, a donor node, etc. The specific technology used by the access network device and the specific device form are not limited by this application.

[0047] In the present application, an apparatus configured to implement the functions of an access network device may be an access network device, or may be, for example, an apparatus that can support an access network device to implement the functions, such as a chip system, a hardware circuit, a software module, or a combination of a hardware circuit and a software module, where the apparatus may be installed in the access network device or adapted for use with the access network device. In the present application, a chip system may include a chip, or may include a chip and other discrete components. For ease of description, the following describes the technical solution provided in the present application by using an example in which the apparatus configured to implement the functions of an access network device is an access network device, and the access network device is a base station.

[0048] (1) Protocol layer structure

[0049] Communications between the access network device and the terminal device follow a specific protocol layer structure. The protocol layer structure may include a control plane protocol layer structure and a user plane protocol layer structure. For example, the control plane protocol layer structure may include protocol layer functions such as a radio resource control (RRC) layer, a packet data convergence protocol (PDCP) layer, a radio link control (RLC) layer, a media access control (MAC) layer, and a physical layer. For example, the user plane protocol layer structure may include protocol layer functions such as a PDCP layer, an RLC layer, a MAC layer, and a physical layer. In a possible implementation, a service layer adaptation protocol (SDAP) layer may be further included above the PDCP layer.

[0050] Optionally, the protocol layer structure between the access network device and the terminal device may further include an artificial intelligence (AI) layer configured to transmit data related to AI functions.

[0051] (2) Central Unit (CU) and Distributed Unit (DU)

[0052] An access device may include a CU and a DU. Multiple DUs may be controlled by one CU in a centralized manner. For example, the interface between a CU and a DU may be called an F1 interface. A control plane (CP) interface may be F1-C, and a user plane (UP) interface may be F1-U. The specific names of each interface are not limited in this application. The CU and DU may be obtained by division based on the protocol layers of a wireless network. For example, the functions of the PDCP layer and the protocol layers above the PDCP layer are set in the CU, and the functions of the protocol layers below the PDCP layer (e.g., the RLC layer and the MAC layer) are set in the DU. As another example, the functions of the protocol layers above the PDCP layer are set in the CU, and the functions of the PDCP layer and the protocol layers below the PDCP layer are set in the DU. This is not a limitation.

[0053] The above division of the processing functions of the CU and DU based on protocol layers is merely an example, and the processing functions of the CU and DU may alternatively be divided in other manners. For example, the CU or DU may be divided to include more protocol layer functions. As another example, the CU or DU may alternatively include some of the processing functions of the protocol layers. In one design, some of the RLC layer functions and the functions of the protocol layers above the RLC layer are set in the CU, and the remaining RLC layer functions and the functions of the protocol layers below the RLC layer are set in the DU. In another design, the division of the functions of the CU or DU may alternatively be performed based on service type or other system requirements. For example, the division may be performed based on delay. Functions whose processing time must satisfy delay requirements are set in the DU, and functions whose processing time does not need to satisfy delay requirements are set in the CU. In another design, the CU may alternatively have one or more functions of the core network. For example, the CU may be deployed on the network side to facilitate centralized management. In another design, the radio unit (RU) of the DU is deployed remotely. Optionally, the RU may have radio frequency capabilities.

[0054] Optionally, the DU and RU may be divided by the physical layer (PHY). For example, the DU may implement upper layer functions of the PHY layer, and the RU may implement lower layer functions of the PHY layer. When the PHY layer is used for transmission, the PHY layer functions may include at least one of the following functions: adding a cyclic redundancy check (CRC) code, channel coding, rate matching, scrambling, modulation, layer mapping, precoding, resource mapping, physical antenna mapping, or radio frequency transmission. When the PHY layer is used for reception, the PHY layer functions may include at least one of the following functions: CRC, channel decoding, de-rate matching, descrambling, demodulation, layer demapping, channel detection, resource demapping, physical antenna demapping, or radio frequency reception. The upper layer functions of the PHY layer may include some of the functions of the PHY layer. For example, some of the functions are closer to the MAC layer. The lower layer functions of the PHY layer may include other parts of the functions of the PHY layer. For example, some of the functions are closer to the radio frequency functions. For example, the upper layer functions of the PHY layer may include CRC code addition, channel coding, rate matching, scrambling, modulation, and layer mapping, and the lower layer functions of the PHY layer may include precoding, resource mapping, physical antenna mapping, and radio frequency transmission functions. Alternatively, the upper layer functions of the PHY layer may include CRC code addition, channel coding, rate matching, scrambling, modulation, layer mapping, and precoding, and the lower layer functions of the PHY layer may include resource mapping, physical antenna mapping, and radio frequency transmission functions. For example, the upper layer functions of the PHY layer may include CRC, channel decoding, de-rate matching, decoding, demodulation, and layer demapping, and the lower layer functions of the PHY layer may include channel detection, resource demapping, physical antenna demapping, and radio frequency reception functions.Alternatively, the upper layer functions of the PHY layer may include CRC, channel decoding, de-rate matching, decoding, demodulation, layer demapping, and channel detection, and the lower layer functions of the PHY layer may include resource demapping, physical antenna demapping, and radio frequency reception functions.

[0055] For example, the functionality of the CU may be implemented by one entity, or may be implemented by different entities. For example, the functionality of the CU may be further divided, i.e., the control plane and the user plane are separated and implemented by different entities, which are a control plane CU entity (i.e., a CU-CP entity) and a user plane CU entity (i.e., a CU-UP entity). The CU-CP entity and the CU-UP entity may be coupled to the DU to together realize the functionality of an access network device.

[0056] Optionally, any one of the DU, CU, CU-CP, CU-UP, and RU may be a software module, a hardware structure, or a combination of a software module and a hardware structure. This is not limited thereto. Different entities may exist in different forms. This is not limited thereto. For example, the DU, CU, CU-CP, and CU-UP are software modules, and the RU is a hardware structure. These modules and the methods performed by these modules also fall within the scope of protection of the present disclosure.

[0057] In a possible implementation, the access network device includes a CU-CP, a CU-UP, a DU, and an RU. For example, the present application may be performed by a DU, or a DU and an RU, or a CU-CP, a DU, and an RU, or a CU-UP, a DU, and an RU. This is not limited thereto. Methods performed by the modules also fall within the scope of protection of the present application.

[0058] A terminal device may alternatively be referred to as a terminal, user equipment (UE), mobile station, mobile terminal device, etc. The terminal device may be widely used for communication in various scenarios, including, but not limited to, at least one of the following scenarios: device-to-device (D2D), vehicle-to-everything (V2X), machine-type communication (MTC), Internet of Things (IoT), virtual reality, augmented reality, industrial control, autonomous driving, telemedicine, smart grid, smart furniture, smart office, smart wearable, smart transportation, smart city, etc. The terminal device may be a mobile phone, a tablet computer, a computer with wireless transceiver capabilities, a wearable device, a vehicle, an unmanned aerial vehicle, a helicopter, an airplane, a ship, a robot, a robotic arm, a smart home device, etc. The specific technology used by the terminal device and the specific device form are not limited herein.

[0059] In the present application, an apparatus configured to implement the functions of a terminal device may be a terminal device, or may be an apparatus capable of supporting a terminal device to implement the functions, such as a chip system, a hardware circuit, a software module, or a combination of a hardware circuit and a software module. The apparatus may be installed in a terminal device or adapted for use with a terminal device. For ease of description, the following describes the technical solution provided in the present application by using an example in which an apparatus configured to implement the functions of a terminal device is a terminal device, and the terminal device is a UE.

[0060] The base station and the terminal device may be fixed or mobile. The base station and / or the terminal device may be deployed on the ground, including indoor or outdoor scenarios, and handheld or vehicle-mounted scenarios, or on water, or may be deployed in the air on airplanes, balloons, and satellites. The application scenarios of the base station and the terminal device are not limited by this application. The base station and the terminal device may be deployed in the same scenario or in different scenarios. For example, both the base station and the terminal device may be deployed on the ground, or the base station may be deployed on the ground and the terminal device may be deployed on water. Examples will not be described one by one.

[0061] The roles of a base station and a terminal device may be relative. For example, helicopter or unmanned aerial vehicle 120i in FIG. 1 may be configured as a mobile base station. In the case of terminal device 120j accessing wireless access network 100 via 120i, terminal device 120i is a base station. However, in the case of base station 110a, 120i is a terminal device. In other words, 110a and 120i communicate with each other based on a wireless air interface protocol. 110a and 120i may alternatively communicate with each other based on an inter-base station interface protocol. In this case, 120i is also a base station for 110a. Therefore, both base stations and terminal devices may be collectively referred to as communication devices. 110a and 110b in FIG. 1 may be referred to as communication devices having base station functionality, and 120a to 120j in FIG. 1 may be referred to as communication devices having terminal device functionality.

[0062] In this application, an independent network element (e.g., referred to as an AI network element or AI node) may be introduced into the communication system shown in FIG. 1 to implement AI-related operations. The AI ​​network element may be directly connected to an access network device within the communication system, or may be indirectly connected to the access network device via a third-party network element. The third-party network element may be a core network element, such as an authentication management function (AMF) or a user plane function (UPF). Alternatively, an AI function, AI module, or AI entity may be configured in another network element within the communication system to implement AI-related operations. For example, the other network element may be an access network device (e.g., a gNB), a core network device, or an operation, administration, and maintenance (OAM) device. In this case, the network element that performs AI-related operations is a network element with built-in AI functions. The OAM is configured to perform operation, administration, maintenance, etc. on the access network device and / or the core network device.

[0063] In the present application, as shown in FIG. 2 or FIG. 3, an AI model may be deployed in at least one of a core network device, an access network device, a terminal device, an OAM, etc., and corresponding functions are implemented by using the AI ​​model. In the present application, the AI ​​models deployed in different nodes may be the same or different. Differences in the models include differences in at least one of the following: differences in the model structural parameters, such as the number of layers and / or the model weight values; differences in the model input parameters; differences in the model output parameters; etc. Differences in the model input parameters and / or the model output parameters can be described as differences in the model functions. Unlike FIG. 2, in FIG. 3, the access network device is functionally divided into a CU and a DU. Optionally, the CU and the DU may be the CU and the DU in an O-RAN architecture. One or more AI models may be deployed in the CU. Additionally / alternatively, one or more AI models may be deployed in the DU. Optionally, the CU in FIG. 3 may be further divided into a CU-CP and a CU-UP. Optionally, one or more AI models may be deployed in the CU-CP. Additionally / alternatively, one or more AI models may be deployed in the CU-UP. Optionally, in Figure 2 or Figure 3, the OAM of the access network device and the OAM of the core network device may be deployed separately and independently.

[0064] Optionally, FIG. 4a is an architecture of a communication system according to the present application. As shown in FIG. 4a, in a first design, an access network device includes a near-real-time access network intelligent controller (RAN intelligent controller, RIC) module configured to train a model and perform inference. For example, the near-real-time RIC may be configured to train an AI model and perform inference by using the AI ​​model. For example, the near-real-time RIC may obtain network-side and / or terminal-side information from at least one of a CU, a DU, or an RU, and the information may be used as training data or inference data. Optionally, the near-real-time RIC may present the inference result to at least one of a CU, a DU, an RU, or a terminal device. Optionally, the CU and the DU may exchange inference results. Optionally, the DU and the RU may exchange the inference result. For example, the near-real-time RIC presents the inference result to the DU, and the DU forwards the inference result to the RU.

[0065] Alternatively, in a second design, as shown in FIG. 4a, a non-real-time RIC (optionally, the non-real-time RIC may be located in an OAM or core network device) is included outside the access network device and configured to train a model and perform inference. For example, the non-real-time RIC is configured to train an AI model and perform inference by using the model. For example, the non-real-time RIC may obtain network-side and / or terminal-side information from at least one of a CU, a DU, or an RU. The information may be used as training data or inference data, and inference results may be presented to at least one of a CU, a DU, an RU, or at least one of a terminal device. Optionally, the CU and the DU may exchange the inference results. Optionally, the DU and the RU may exchange the inference results. For example, the non-real-time RIC presents the inference results to the DU, and the DU forwards the inference results to the RU.

[0066] Alternatively, in a third design, as shown in FIG. 4a, the access network device includes a near-real-time RIC, and a non-real-time RIC (optionally, the non-real-time RIC may be located in an OAM or core network device) is included outside the access network device. As with the second design above, the non-real-time RIC may be configured to train the model and perform inference. Additionally / alternatively, as with the first design above, the near-real-time RIC may be configured to train the model and perform inference. Additionally / alternatively, the non-real-time RIC may train the model, and the near-real-time RIC may obtain AI model information from the non-real-time RIC, obtain network-side and / or terminal-side information from the CU, DU, or RU, and obtain inference results by using the information and the AI ​​model information. Optionally, the near-real-time RIC may present the inference results to at least one of the CU, DU, RU, or terminal device. Optionally, the CU and DU may exchange the inference results. Optionally, the DU and RU may exchange the inference results. For example, the near-real-time RIC presents the inference results to the DU, and the DU forwards the inference results to the RU. For example, the near real-time RIC is configured to train model A and perform inference by using model A. For example, the non-real-time RIC is configured to train model B and perform inference by using model B. For example, the non-real-time RIC is configured to train model C and send information about model C to the near real-time RIC. The near real-time RIC performs inference by using model C.

[0067] Fig. 4b shows another architecture of a communication system according to the present invention. Compared with the CU in Fig. 4a, the CU in Fig. 4b is separated into a CU-CP, a CU-UP, and so on.

[0068] When an AI model is used in a wireless network device, a problem that needs to be solved is the conflict between the flexible and variable signal format in the wireless network and the fixed input and output formats of the AI ​​model. For example, the bandwidth of a wireless signal is flexibly scheduled by a base station. The bandwidth that can be scheduled in one slot may be four resource blocks (RBs), and the bandwidth that can be scheduled in the next slot may change to eight RBs. However, the input format of the AI ​​model is fixed and cannot change according to the signal bandwidth. For example, if the input format of the AI ​​model corresponds to four RBs, the AI ​​model only supports signals with an input bandwidth of four RBs and cannot process signals with an input bandwidth of eight RBs. To solve the above problem, a solution is provided: the AI ​​model is trained for each bandwidth. As shown in Figure 4c, the AI ​​model is trained separately for signals with an input bandwidth of four RBs and signals with an input bandwidth of eight RBs. This results in high overhead for both training and maintenance.

[0069] This application provides a solution. In the solution, an AI model is trained, and the AI ​​model is applicable to input data and output data in multiple formats. Specifically, during training, data (e.g., input data and / or output data) in different formats is converted into data in the same format by padding, truncation, etc., to jointly train the AI ​​model. The AI ​​model is applicable to multiple formats. When the AI ​​model is used or performs inference, the input data and / or output data is converted into input data and / or output data compatible with the AI ​​model by using the same data processing method used during training. Thus, inference tasks for different radio resource configurations can be completed by using the same AI model, and training and maintenance overhead is reduced. Note that at least two examples are included in the solution of this application. For example, the AI ​​model is trained for input data and / or output data in all formats. For example, scheduled bandwidths in a wireless network include 4 RBs, 8 RBs, 16 RBs, and 32 RBs. In this case, the AI ​​model is trained for the above four scheduled bandwidths. In the following description, this example is used for explanation purposes only. Alternatively, an AI model may be designed for input data and / or output data in some formats, and another AI model may be designed for input data and / or output data in other formats. For example, a first AI model may be trained for scheduled bandwidths of 4 RBs and 8 RBs. In other words, the first AI model may be applicable to signal inputs of 4 RBs and 8 RBs. A second AI model may be trained for scheduled bandwidths of 16 RBs and 32 RBs. In other words, the second AI model may be applicable to signal inputs of 16 RBs and 32 RBs, etc.

[0070] For ease of understanding, terms or concepts in this application are first described.

[0071] 1. AI model

[0072] An AI model is a specific implementation of an AI function. An AI model represents a mapping relationship between the input and output of the model. An AI model may be a neural network, a linear regression model, a decision tree model, a support vector machine (SVM), a Bayesian network, a Q-learning model, other machine learning models, etc. In this application, an AI function may include at least one of the following: data collection (collection of training data and / or inference data), data preprocessing, model training (also called model learning), model information release (setting model information), model validation, model inference, or inference result release. Inference may alternatively be called prediction. In this application, an AI model may be referred to as model for short.

[0073] An AI model design mainly includes a data collection section (e.g., a data source may collect training data and / or inference data), a model training section, and a model inference section. Furthermore, the AI ​​model design may further include an inference result application section. Optionally, the AI ​​model design may further include a model testing section. FIG. 5 is a diagram of an application architecture of an AI model. A data source is configured to provide training data and inference data. In the model training section, a model training node (model training host) analyzes or trains the training data provided by the data source to obtain an AI model. Optionally, the AI ​​model trained by the model training node is equivalent to obtaining a mapping relationship between the input and output of the model through learning by the model training node by using the training data. The AI ​​model is deployed by a model inference node (model inference host). Optionally, the model training node may further update the AI ​​model deployed by the model inference node. The model inference node may further feed back relevant information of the deployed model to the model training node, so that the model training node optimizes, updates, etc. the deployed AI model.

[0074] In the model inference section, the model inference node uses the AI ​​model to make inferences based on the inference data provided by the data source and obtains inference results. The method may be implemented as follows: the model inference node inputs the inference data to the AI ​​model and obtains output by using the AI ​​model. The output is the inference result. The inference result may indicate configuration parameters executed (activated) by the action subject and / or operations performed by the action subject. The inference result may indicate configuration parameters used (activated) by the action subject and / or operations performed by the action subject. The inference result may be centrally planned by an actor entity and sent to one or more action subjects (e.g., network entities) for action. Optionally, the actor entity or the action subject may further feed back model performance to the data source, where the model performance is used as training data to aid in subsequent model update training.

[0075] 2. Neural network (NN)

[0076] The AI ​​model may be a neural network or other machine learning model. A neural network is used as an example. A neural network is a specific embodiment of machine learning technology. According to the universal approximation theorem, a neural network can theoretically approximate any continuous function, thereby allowing the neural network to learn any mapping. Therefore, a neural network can accurately perform abstract modeling for complex high-level problems.

[0077] The idea of ​​neural networks comes from the neuron structure of the brain. Each neuron performs a weighted sum operation on the neuron's input values ​​and outputs the result of the weighted sum through an activation function. Figure 6 shows the structure of a neuron. The neuron's inputs are x=[x0,x1,...,x n ] and the weights corresponding to the inputs are w=[w,w1, ,w n] and the bias of the weighted sum is assumed to be b. The form of the activation function can be varied. If the activation function of a neuron is y = f(z) = max(0,z), the output of the neuron is

number

number

[0078] Neural networks generally include a multi-layer structure, with each layer containing one or more neurons. Increasing the depth and / or width of a neural network improves the neural network's representational capabilities, providing more powerful information extraction and abstract modeling capabilities for complex systems. The depth of a neural network can refer to the number of layers included in the neural network, and the number of neurons included in each layer can be referred to as the layer width. Figure 7 illustrates the layer relationships of a neural network. In one implementation, a neural network includes an input layer and an output layer. After performing neuronal processing on the received input, the input layer of the neural network forwards the results to the output layer, which obtains the neural network's output result. In another implementation, a neural network includes an input layer, a hidden layer, and an output layer. After performing neuronal processing on the received input, the input layer of the neural network forwards the results to an intermediate hidden layer. The hidden layer then forwards the calculation results to the output layer or an adjacent hidden layer. Finally, the output layer obtains the neural network's output result. A neural network can include one hidden layer or multiple hidden layers connected in series. This is not limited to this. In the neural network training process, a loss function may be defined. The loss function represents the gap or difference between the neural network output value and an ideal target value. The specific form of the loss function is not limited by this application. The neural network training process is a process of adjusting neural network parameters, such as the number and width of neural network layers, neuron weight values, and / or parameters in neuron activation functions, so that the value of the loss function is less than a threshold or meets the target requirement.

[0079] 3. AI model training

[0080] In the training process of an AI model, a loss function may be defined. The loss function represents the gap or difference between the output value of the AI ​​model and an ideal target value. The specific form of the loss function is not limited by this application. The training process of an AI model is a process in which some or all of the parameters of the AI ​​model are adjusted so that the value of the loss function or the weighted sum value (e.g., the average value) for one or more training data is smaller than a threshold or can meet the target requirement. If the AI ​​model is a neural network, one or more of the following parameters may be adjusted in the training process: the number of layers of the neural network, the width of the neural network, the connection relationship between layers, the weight values ​​of neurons, the activation function of neurons, or the offset in the activation function, so that the difference between the output of the neural network and the ideal target value is as small as possible.

[0081] In design, the AI ​​model is a neural network θ (·) is used as an example. The model training node collects training data, where the training data includes training samples and labels. For example, the training sample x is used as input, and the inference result f θ (x) is the neural network f θ (·) is the output after processing the training sample x, and the inference result f θ The difference between (x) and the label of training sample x is obtained by calculation using a loss function. The model training node may optimize network parameters by using a model optimization algorithm based on the difference obtained by using the loss function. The neural network is trained by using a large amount of training data, so that for a group of training samples (e.g., one or more training samples), the difference between the neural network output of each training sample and the label of the training sample is smaller than a threshold or satisfies a target requirement, or a weighted sum (e.g., average value) of the differences between the neural network output of all training samples and the labels of all training samples is smaller than a threshold or satisfies a target requirement, and the training of the neural network is completed.

[0082] Although the supervised learning training method is described in the above process, it is not intended to limit the present application. In the present application, AI model training may alternatively be performed by using unsupervised learning. An internal mode of the training samples is learned by using an algorithm to complete AI model training based on the training samples. AI model training may also be performed by using reinforcement learning. An excitation signal fed back by the environment is obtained through interaction with the environment to learn a problem-solving policy, optimize the AI ​​model, etc. In the present application, the model training method is not limited.

[0083] As shown in Figure 8, a procedure of a data processing method for a model is provided. The method is executed by 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, different nodes, etc. This is not limited. The procedure includes at least the following steps:

[0084] Step 801: The first node determines a data processing method for the AI ​​model.

[0085] The first node may be a model training node, a model inference node, etc. For example, the first node may receive instruction information from another node (e.g., the second node), where the instruction information instructs a data processing method for the AI ​​model. Alternatively, the first node may determine a data processing method for the AI ​​model as defined in a protocol. Alternatively, the first node may determine a data processing method for the AI ​​model. For example, the model inference node may determine a data processing method for the AI ​​model based on the format of the collected training data, the format of the AI ​​model to be trained, etc. See the description below for details.

[0086] Step 802: The first node performs at least one of the following according to the data processing method for the AI ​​model: performing model training, or performing model inference.

[0087] 9A, an example in which a model training node and a model inference node are deployed separately, and the model training node indicates a data processing method for an AI model to the model inference node, is used to provide a procedure of the data processing method. The procedure includes at least the following steps:

[0088] Step 901: Collect training data.

[0089] In this application, the entity that collects training data may be a model training node, another AI entity, an AI module, etc. In this application, the collected training data includes, but is not limited to, data in a wireless network, such as wireless channel information, a received signal, a reference signal, and reference signal received power (RSRP). The wireless channel information may include estimated channel responses, channel characteristics, etc. The direct acquirer of data (also referred to as an estimator, measurer, collector, etc.) may be a UE or a base station. When the entity or module that collects data is different from the direct acquirer of data, the direct acquirer of data sends the data to the entity or module that collects data. The collected training data may include measurement data of different UEs, or measurement data acquired by a UE in different geographic locations and different channel environments. The collected data may be actual data acquired by a UE or a base station in a real network, or may be virtual data generated by using an emulation platform or a simulation platform, etc. This is not limited to this.

[0090] The objective of this application is to train an AI model that can adapt to multiple radio resource configurations. Therefore, the collected training data includes data for multiple radio resource configurations. For example, the collected training data includes received signals or radio channel information for multiple bandwidths or multiple reference signal patterns. Alternatively, the data for multiple radio resource configurations may be generated by using a data enhancement method on data for the same radio resource configuration. For example, a radio channel with the same bandwidth is truncated to a radio channel with a different bandwidth.

[0091] Optionally, in this application, the specific type of collected training data is related to the function of the AI ​​model. For example, in the case of an AI model for frequency-domain channel estimation, the training data required by the AI ​​model includes at least a received reference signal Y and an original reference signal S. For the AI ​​model, the reference signal Y and the original reference signal S in various radio resource configurations may be collected as training data. For example, the reference signal Y and the original reference signal S in different bandwidths and / or different reference signal types may be collected.

[0092] As another example, in the case of an AI model used for channel state information (CSI) feedback, the training data required by the AI ​​model includes at least the channel response H and the channel characteristics W, and the channel response H or the channel characteristics W at different bandwidths and / or different numbers of antenna ports may be collected.

[0093] As another example, in the case of an AI model used for beam management, 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, and the channel response H, the received signal Y, or the RSRP at different bandwidths and / or different numbers of antenna ports may be collected.

[0094] As another example, in the case of a positioning-related AI model, the training data required by the AI ​​model includes at least the channel response H, and the channel responses H at different bandwidths and / or different numbers of antenna ports may be collected.

[0095] Step 902: Perform model training, which includes processing the collected training data.

[0096] In the present application, since the training data collected in the data collection phase includes multiple formats, in the model training phase, the collected training data needs to be processed to unify the formats of the training data.

[0097] To distinguish between the input data and output data of an AI model in the AI ​​model inference phase, the input data of an AI model in the AI ​​model training phase is referred to as input training data, and the output data of an AI model in the AI ​​model training phase is referred to as output training data. Furthermore, this application includes a process of processing training data. Therefore, the originally collected training data (i.e., raw data), i.e., the training data before processing, may be referred to as first input training data, and the processed training data may be referred to as second input training data. The second input training data may be used as the input training data of the AI ​​model. The output data of the AI ​​model is referred to as first output training data, and the data obtained after the output data of the AI ​​model is processed is referred to as second output training data or target data.

[0098] For example, 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. As shown in FIG. 9B , the model training node may perform input data processing on the collected first input training data to obtain second input training data; obtain first output training data based on the second input training data and the AI ​​model; perform output data processing on the first output training data to obtain second output training data; and determine whether the objective function satisfies a requirement based on the second output training data. If the objective function satisfies the requirement, the AI ​​model is output, and training of the AI ​​model is completed. Alternatively, if the objective function does not satisfy the requirement, parameters of the AI ​​model are updated, and training of the AI ​​model continues. The objective function may alternatively be referred to as a loss function. For example, the AI ​​model is a neural network. In this case, the adjusted parameters may include one or more of the following: the number of layers of the neural network, the width of the neural network, the connection relationship between layers, neuron weight values, neuron activation functions, activation function offsets, etc. Supervised learning may be used in the above training process. The model training node may determine the value of the loss function based on the second output training data and the corresponding label. If the value of the loss function is less than a threshold or satisfies the target requirement, the training of the AI ​​model ends. Alternatively, if the value of the loss function is greater than or equal to a threshold or does not satisfy the requirement, the training of the AI ​​model continues. Unlike that of supervised learning, in the unsupervised learning process, there is no label. In this case, the model training node may determine the value of the loss function based on the second output training data. In supervised learning and unsupervised learning, the loss function has different designs. It may be understood that other model training methods, such as reinforcement learning, may alternatively be used in the model training process. This is not limited.

[0099] For example, the process of the model training node performing input data processing on first input training data includes determining an input format and an output format of the AI ​​model, and processing the first input training data based on the input format of the AI ​​model to obtain second input training data. Optionally, the format of the second input training data is the same as the input format of the AI ​​model. The first output training data is determined based on the second input training data and the AI ​​model. Optionally, 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. For example, processing may be performed on the first output training data based on a format of labels in the training data to obtain second output training data. Optionally, the format of the second output training data is the same as the format of the corresponding labels. Alternatively, in an unsupervised learning scenario, no labels exist in the training data. In this case, processing may be performed on the first output training data based on a radio resource configuration to obtain second output training data, where the second output training data satisfies the corresponding radio resource configuration. Alternatively, the first output training data is processed based on the loss function of the AI ​​model to obtain second output training data, etc. For example, in radio resource configuration, the AI ​​model needs to feed back channel state information with a length of B bits. In this case, 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 based on the second output training data. For example, supervised learning is used as an example. The model training node may calculate the value of the loss function based on the second output training data and the labels in the training data. If the value of the loss function is less than a threshold or satisfies the target requirement, the training of the AI ​​model is completed. Alternatively, if the value of the loss function is greater than or equal to the threshold or does not satisfy the target requirement, the AI ​​model parameters are adjusted, and the training of the AI ​​model continues.

[0100] It should be understood that the above process of processing the first input training data and the process of processing the first output training data do not necessarily need to be performed. The collected training data includes multiple formats. Therefore, for input training data in a specific format, if the format of the input training data is the same as the input format of the AI ​​model, the input training data in that format may be directly input to the AI ​​model, and the input training data no longer needs to be processed. In other words, the first input training data may not be processed, and the first output training data may be directly determined based on the first input training data and the AI ​​model. Similarly, the output of the first AI model is referred to as first output training data. If the format of the first output training data meets requirements, for example, is the same as the format of the label or meets the format of the radio resource configuration, the first output training data may no longer be processed.

[0101] In this application, the input format and output format of the AI ​​model may be related to the radio resource configuration. For example, the input data of the AI ​​model may be padded, and the output data of the AI ​​model may be truncated. The length of the input data of the AI ​​model may be longer than or equal to the length of the longest training sample among all potential radio resource configurations in the application scenario of the AI ​​model, and the length of the output data of the AI ​​model may be longer than or equal to the length of the longest label among all potential radio resource configurations in the application scenario of the AI ​​model.

[0102] In the present description, the term "format" may have two meanings, such as "dimension" and / or "length." In wireless signal processing, the object to be processed is typically a wireless signal, a wireless channel, etc. The dimensions of the wireless signal or wireless channel typically include dimensions such as the time domain, the frequency domain, and / or spatial frequency. In the present description, the dimensions of the originally collected first input training data may be transformed so that the dimensions meet the input dimension requirements of the AI ​​model, and / or the signal length in each dimension is processed so that the length meets the input requirements of the AI ​​model for the length in each dimension. In the present description, the processes for processing the input data length and the output data length are mainly described, and processing of one-dimensional data length is used as an example. In the following description, unless otherwise specified, the lengths described are one-dimensional data lengths. A method for processing multi-dimensional data lengths can be obtained by directly extending the method for processing one-dimensional data lengths.

[0103] Step 903: The model training node sends the AI ​​model and instruction information to the model inference node, where the instruction information indicates a data processing method for the AI ​​model, and the data processing method for the AI ​​model may alternatively be referred to as an AI model usage.

[0104] In this application, if the model training node is different from the model inference node, model deployment needs to be performed. Specifically, the model training node needs to send the trained AI model to the model inference node. For example, the model training node may send information about the AI ​​model to the model inference node. The information about the AI ​​model includes at least one of the following: model parameters, model input format, model output format, etc. A neural network is used as an example. The model parameters 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, neuron weights, neuron activation functions, activation function offsets, etc. The model inference node recovers or determines the AI ​​model based on the information about the AI ​​model.

[0105] Furthermore, in the present application, the AI ​​model usage (i.e., the data processing method for the AI ​​model) may be further instructed to the model inference node so that the model inference node can correctly use the AI ​​model to complete the inference task. Alternatively, the AI ​​model usage may be predefined. The method by which the model inference node obtains the AI ​​model usage is not limited in the present application.

[0106] In this application, a data processing method for an 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 padding method or a data truncation method. The output data processing method includes at least one of the following: a data padding method or a data truncation method. The data padding method includes at least one of the following: a padding rule (or padding position, i.e., a specific position where data of a specific length is padded), a length obtained after padding, and a padding data type (i.e., a padding value). The data truncation method includes at least one of the following: a truncation rule (or truncation position, i.e., a specific position where data of a specific length is truncated), or a length obtained after truncation.

[0107] In the present application, multiple data processing methods may be predetermined, and the model training node may instruct the model inference node to select one of the multiple data processing methods. For example, n data processing methods may be predetermined, and the n data processing methods correspond to different indexes. The model training node may specifically instruct the model inference node to select the index of a specific data processing method. Alternatively, the model training node may directly instruct the model inference node on parameters corresponding to the data processing method, such as padding values ​​and padding positions for input data and / or truncation positions for output data.

[0108] Step 904: The model inference node performs model inference according to the data processing method for the AI ​​model.

[0109] In this application, the model inference node may perform model inference according to the data processing method instructed by the model training node for the AI ​​model to complete inference tasks for different radio resource configurations. In other words, the model inference node processes data during model inference by using the same rules as those of the training data processing method.

[0110] For ease of description, in the model inference phase, the initial input data is referred to as first input data or raw data, the input data of the AI ​​model obtained after the first input data is processed is referred to as second input data, the output data of the AI ​​model is referred to as first output data, and the data obtained after the first output data is processed is referred to as second output data or target data.

[0111] In the design, the data processing method indicated by the model inference node for 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 is the same as or corresponds to the input training data processing method in the model training phase. The output data processing method is the same as or corresponds to the output training data processing method in the model inference phase. For example, as shown in FIG. 9C , the model inference node may process first input data according to the input data processing method in the data processing method for the AI ​​model to obtain second input data. The format of the second input data is the same as the format of the AI ​​model. The first output data is determined based on the second input data and the AI ​​model. The format of the first output data is the same as the output format of the AI ​​model. The first output data is processed according to the output data processing method in the data processing method for the AI ​​model to determine second output data. The second output data can be considered as the inference result of the AI ​​model.

[0112] Similar to the above-described training process, the process in which the model inference node processes the first input data to determine the second input data and the process in which the model inference node processes the first output data to obtain the second output data are not necessarily performed, and may alternatively be omitted. For example, if the first input data matches the input format of the AI ​​model or has the same length as the padding data specified in the input data processing method, the first input data no longer needs to be processed and can be directly used as the input data of the AI ​​model. If the output data of the AI ​​model meets the length of the radio resource configuration or is the same as the length of the corresponding label, or if the output data of the AI ​​model has the same length as the truncated data specified in the output data processing method, the first output data does not need to be processed and can be directly used as the inference result of the AI ​​inference.

[0113] In the above design, inference tasks for different radio resource configurations can be completed by using one AI model, eliminating the need to train and deploy one AI model for each radio resource configuration, thereby reducing overhead for training and maintenance, etc.

[0114] As described above, the data processing method for an AI model includes at least one of the following: an input data processing method for an AI model, or an output data processing method for an AI model. In the design, a process is mainly described in which the input data processing method for an AI model is an input data padding method, and the output data processing method for an AI model is an output data truncation method. Alternatively, it may be described as follows: the data processing method for an AI model includes at least one of the following: an input data padding method for an AI model, or an output data truncation method for an AI model.

[0115] In this design, the process of the model training node performing model training according to the data processing method for the AI ​​model includes the model training node padding first input training data according to the input data padding method to obtain second input training data, determining first output training data based on the second input training data and the AI ​​model, and truncating the first output training data according to the output data truncation method to obtain second output training data, and performing parameter adjustment on the AI ​​model based on the second output training data.

[0116] The input data padding method may include one of the following items: Padded data length: If this item is not set, the length of the padded input training data is the same as the length of the AI ​​model's input data by default, that is, it conforms to the AI ​​model's input format; a padding data type, e.g., the padding data type may be 0, or a very large positive number, or a very small negative number; or Data padding rules, for example, padding can be done at equal or unequal intervals before, after, in the middle, or on both sides of the input training data.

[0117] The output data truncation method may include at least one of the following items: The length of the data after truncation. If this item is not set, the length of the truncated output data defaults to the same as the label length, i.e., it conforms to the label format; or Data truncation rules, for example, truncation can occur at equal or unequal intervals before, after, in the middle, or on both sides of the output data.

[0118] For example, the model training node may collect sufficient training data. Supervised learning is used as an example. The training data includes training samples and labels. The training samples may alternatively be referred to as first input training data, which may be data that needs to be input to the AI ​​model for model training. The length of the input data of the AI ​​model and the length of the output data of the AI ​​model are determined. The length of the input data of the AI ​​model is longer than or equal to the length of the longest training sample among the training samples of the collected training data, and the length of the output data of the AI ​​model is longer than or equal to the length of the longest label among the labels of the collected training data. In other words, the length of the input data of the AI ​​model is longer than or equal to the length of the longest original data among all potential radio resource settings in the application scenario of the AI ​​model, and the length of the output data of the AI ​​model is longer than or equal to the length of the longest target data among all potential radio resource settings in the application scenario of the AI ​​model. The model training node pads the first input training data according to an input data processing method to obtain second input training data. The second input training data may be referred to as padded first 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. Specifically, padding may be performed on the input training data based on the type of padding data, the data padding rule, etc. in the input data padding method. For example, the frequency domain length of the input data of the AI ​​model is 8 RBs, and the frequency domain length of the first input training data is 4 RBs. In the present application, the frequency domain length of the first input training data may be padded from 4 RBs to 8 RBs. The specific process of padding the frequency domain length of the first input training data from 4 RBs to 8 RBs may be determined based on the padding data type (e.g., a specific padding value) and the data padding rule (e.g., front padding, rear padding, or equal-interval padding) in the input data padding method. The first output training data is determined based on the second input training data and the AI ​​model.The first output training data is truncated according to an output data truncation method to obtain second output training data. For example, the length of the second output training data is equal to the length of the label. Specifically, the method for truncating the first output training data can be determined based on a truncation rule in the output data truncation method. For example, truncation can be performed before or after the first output training data, or at equal intervals. Parameter adjustment is performed on the AI ​​model based on the second output training data. For example, in the present application, a value of a loss function can be calculated based on the second output training data and the label. If the value of the loss function is less than a threshold or meets the target requirement, training of the AI ​​model is completed. Alternatively, if the value of the loss function is greater than or equal to the threshold or does not meet the target requirement, the parameters of the AI ​​model are adjusted, and training of the AI ​​model continues.

[0119] In this application, the model training node may send the trained AI model to the model inference node. The model inference node may perform data inference on the AI ​​model according to the data processing method for the AI ​​model. In this design, the process by which the model inference node performs model inference based on the data processing method for the AI ​​model includes: padding the first input data according to an input data padding method to obtain second input data; determining first output data based on the second input data and the AI ​​model; truncating the first output data according to an output data truncation method; The truncation may alternatively be referred to as extraction, and the truncated first output data may be referred to as second output data, which is the inference result of the AI ​​model.

[0120] Optionally, the process of padding the input data and truncating the output data in the model inference process is the same as the process of padding the input training data and truncating the output training data in the model training phase.

[0121] To describe channel estimation under different bandwidths and different reference signal patterns by using the same AI model, a frequency-domain channel estimation scenario is used as an example, in which the input data of the AI ​​model is padded and the output data of the AI ​​model is truncated.

[0122] For ease of understanding, the channel estimation process is described first. The propagation model for a wireless signal with single UE single antenna orthogonal frequency division multiplexing (OFDM) symbols is given by:

number

[0123] In this application, the bandwidth of the frequency domain channel scheduled by the base station may vary. For example, the bandwidth of the frequency domain channel scheduled by the base station may be 4 resource blocks (RBs), 8 RBs, 16 RBs, etc. For the same scheduled bandwidth, the UE may transmit different types of reference signals, such as type-1 reference signals and type-2 reference signals, to the base station. In this application, the length of H depends on the scheduled bandwidth, and the lengths of Y, H, and S depend on the scheduled bandwidth and the reference signal pattern. For example, the bandwidth scheduled by the base station is set to k RBs. Since one RB includes 12 REs, the length of H is 12k. As shown in Figure 10, the type-1 reference signal occupies one RE at an interval of 1 RE, so the type-1 reference signal occupies six REs in each RB. As shown in Figure 11, the type-2 reference signal occupies two REs at an interval of 4 REs, so the type-2 reference signal occupies four REs in each RB. When the scheduled bandwidth is k RBs, the length of each of the type-1 reference signals Y, H, and S is 6k, and the length of each of the type-1 reference signals Y, H, and S is 4k.

[0124] When frequency-domain channel estimation is performed using the AI ​​method, the input data of the AI ​​model are Y and S. The length of Y and S of the type-1 reference signal is 6k, and the length of Y and S of the type-2 reference signal is 4k, where k is the frequency-domain channel bandwidth scheduled by the base station. During AI model design, it can be seen that the AI ​​model needs to be designed separately based on different scheduled bandwidths and different types of reference signals. In this design, the AI ​​model is designed so that the AI ​​can be applicable to channel estimation under different scheduled bandwidths and different types of reference signals, thereby reducing the overhead for training and maintaining the AI ​​model.

[0125] Subsequently, the channel estimation process is described in terms of training data collection, training data processing, model training, model deployment, model inference, etc.

[0126] Training Data Collection

[0127] The training data set is determined based on application scenarios of the AI ​​model. The training data set may include training data for all possible radio resource settings in as many application scenarios as possible so that the AI ​​model can achieve excellent performance under various radio resource settings. Each training data set includes a training sample and a label. The training sample may alternatively be referred to as first input training data and includes a received signal Y and an original signal S. The label is a frequency-domain channel response. For example, when an AI model is expected to be trained, channel estimation may be performed for type-1 reference signals and type-2 reference signals within a scheduled bandwidth of 4 RBs and 8 RBs. In this case, the training data includes training data corresponding to type-1 reference signals and type-2 reference signals within a scheduled bandwidth of 4 RBs, and training data corresponding to type-1 reference signals and type-2 reference signals within a scheduled bandwidth of 8 RBs. Each training data set includes a training sample (Y, S) and a label (H).

[0128] Training data processing

[0129] The training data processing includes padding the first input training data, where the padded input training data is referred to as 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 based on the second input training data and the AI ​​model. The first output training data is truncated to determine second output training data, where the length of the second output training data is the same as the length of the labels.

[0130] In this application, the process of padding the first input training data includes the following: First, for each different type of reference signal, padding may be performed based on the pattern of the reference signal, and Y and S are padded to have the same length as H. After padding, the actual values ​​of Y and S are only in the positions corresponding to the REs where the reference signals are located, and padding values ​​are in the remaining RE positions. Then, the padded Y and S are padded again based on the length of the input data of the AI ​​model, i.e., the frequency domain bandwidth supported by the AI ​​model. At this time, the padding rule may be one or more of front padding, rear padding, middle padding, double padding, etc., but is not limited thereto. The padding value may be 0, a very large positive number, or a very small negative number. The length of the padded data, i.e., the length of the second input training data, is the same as the length of the input data of the AI ​​model.

[0131] For example, the training data collected, i.e. the training samples in the first input training data, are: Received signal corresponding to type-1 reference signal within the scheduled bandwidth of 4RB

number

number

number

number

[0132] The first input training data is padded, where the padding data may be 0, and the length of the padded input training data is 96 RE. Optionally, in the training scenario, since the scheduled maximum bandwidth is 8 RB, that is, 96 RE, it can be considered that the length of the input data of the AI ​​model is designed as 96 RE. In this application, the length of the training samples in the training data is padded to be the same as the length of the input data of the AI ​​model, 96 RE.

[0133] In this application, each Y of different types of reference signals may be first padded to have the same length as the length of the label H, and then Y is padded again. Specifically, padding may be performed after Y. The multiple padded first input training data, i.e., the multiple second input training data, are respectively as follows: Y corresponding to type-1 reference signals within the scheduled bandwidth of 4RBs 1 is obtained after being padded

number

number

number

number

[0134] The method for padding S is similar to the above method for padding Y. The padded S and padded Y may be referred to as second input training data. The padded Y and padded S are input to the AI ​​model. The output of the AI ​​model is referred to as first output training data, and the length of the first output training data is the same as the length of the output data of the AI ​​model. The first output training data is truncated. The truncated output training data may be referred to as second output training data, and the length of the second output training data is the same as the length of the label (frequency-domain channel response H). The truncation rule for the first output data may be one or more of truncation at the front, truncation at the back, truncation at the middle, truncation at both ends, etc. This is not limited. Optionally, the truncation rule may be unrelated to or coincident with the padding rule. For example, if the first input training data is padded at the back of the first input training data, the first output training data is truncated at the front of the first output training data.

[0135] For example, in a scheduled bandwidth of 4 RBs, it is assumed that the length of the output data of the AI ​​model is 96 RE and the length of the label (frequency-domain channel response H) of the AI ​​model is 48 RE. In this case, the first 48 elements in the output data of the AI ​​model may be truncated to be compared with the 4-RB label H. In a scheduled bandwidth of 8 RBs, the output data of the AI ​​model does not need to be truncated and the output data of the AI ​​model may be directly compared with the 8-RB label H, etc.

[0136] Model training

[0137] In the model training process, the model training node may determine the structure of the AI ​​model, including 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 longer than or equal to the length of the longest training sample in the training data. The length of the output data in the output format of the AI ​​model is longer than or equal to the length of the longest label in the training data. The AI ​​model is trained by using training data obtained after the collected training data, including multiple bandwidths and multiple different types of reference signals, is processed. In the AI ​​model training process, the training sample, i.e., the first input training data, is padded. The length of the padded input training data is the same as the length of the input data of the AI ​​model. The first output training data of the AI ​​model is truncated, and the length of the truncated output training data is the same as the length of the corresponding label. The model training node may determine the value of a loss function based on the second output training data and the label. If the value of the loss function is smaller than a threshold or satisfies the target requirement, the training of the AI ​​model is completed. Alternatively, if the value of the loss function is greater than or equal to the threshold or does not satisfy the target requirement, the parameters of the AI ​​model are adjusted, and the training of the AI ​​model continues.

[0138] Model deployment

[0139] In this application, the model training node needs to send the trained AI model to the model inference node, and further needs to instruct the model inference node on the data processing method for the AI ​​model so that the model inference node can correctly use the same AI model to complete channel estimation tasks with different bandwidths and different types of reference signals. For example, the model deployment host may send instruction information to the model inference node, where the instruction information instructs the data processing method for the AI ​​model, and the instruction information may instruct the input data padding rule, the output data truncation rule, etc.

[0140] Model Inference

[0141] The model inference node performs model inference according to the data processing method instructed by the model training node for the AI ​​model to complete the channel estimation task under different bandwidths and different types of reference signals, for example, the input data is padded by using the same padding rule as that in the model training process, and the output data is truncated by using the same truncation rule as that in the model training process.

[0142] In this application, frequency-domain channel estimation is used as an example to describe the process of performing channel estimation by using an AI model. It should be understood that channel estimation also includes other scenarios. In each scenario, the input data and output data are different. For example, in a scenario where further channel estimation needs to be performed on a channel roughly estimated by using a conventional method, the input data is a noisy channel on REs occupied by a reference signal, and the output data is a channel on all REs within the scheduled bandwidth. To obtain an inference result, the input data needs to be padded and the output data needs to be truncated. As another example, in a delay-domain channel estimation scenario, the input data is a received signal and a reference signal, and the output data is a delay-domain channel. The input data needs to be padded, and no data truncation needs to be performed on the output data, etc.

[0143] In this application, the input data of the AI ​​model is padded and the output data of the AI ​​model is truncated, so that the channel estimation task in the case of multiple bandwidths and different reference signal patterns can be completed by using one AI model.

[0144] To describe CSI feedback performed on different bandwidths and / or different antenna ports (channel characteristic lengths) by using the same AI model, a CSI feedback scenario is used as an example, in which the input data of the AI ​​model is padded and the output data of the AI ​​model is truncated.

[0145] For ease of understanding, the AI-based CSI feedback procedure is described first, as shown in FIG.

[0146] The UE obtains a downlink channel response H or downlink channel characteristic information W. A first submodel is deployed in the UE, and the first submodel may be referred to as submodel f. The downlink channel response H or downlink channel characteristic information W is used as an input of submodel f, and the output of submodel f is feedback bits B corresponding to CSI. The feedback bits B are transmitted to the base station. A second submodel is deployed in the base station, and the second submodel may be referred to as submodel g. The feedback bits B are used as an input of submodel g, and the output of submodel g is

number

[0147] The dimension of the downlink channel response H is bandwidth × number of antenna ports, where the antenna ports include the antenna ports of the base station and / or the antenna ports of the UE. The bandwidth configured by the base station for CSI feedback may vary, for example, 4 RBs or 8 RBs. A granularity of 1 RB is used as an example. In this case, the length of the downlink channel in the bandwidth dimension is 4 or 8. The number of antenna ports configured by the base station for CSI feedback may vary, for example, 16 ports or 32 ports. In this case, the length of the downlink channel in the antenna port dimension is 16 or 32 becomes.

[0148] The dimension of the downlink channel characteristic information W is the number of subbands × eigenvector length (the number of antenna ports of the base station). The number of subbands configured by the base station may vary, for example, 6 subbands or 12 subbands. In this case, the length of the downlink channel in the subband dimension, which is the characteristic information, is 6 or 12. The eigenvector length configured by the base station for CSI feedback may vary, for example, 16 or 32. In this case, the length of the downlink channel in the eigenvector dimension, which is the characteristic information, is 16 or 32.

[0149] The downlink channel characteristic information W is obtained by calculation based on the downlink channel response H. The specific calculation method is as follows:

[0150] covariance matrix R UU is calculated for the downlink channel response H of each subband,

number

[0151] An eigenvalue decomposition is performed on the covariance matrix of each subband;

number

number

[0152] Optionally, the downlink channel feature matrix V can be further processed and projected into a sparser space. For example, two groups of DFT bases, the space-domain basis S and the frequency-domain basis F, are generated by using the discrete Fourier transform (DFT) formula. For joint space-frequency projection, the following formula is used: C=S H Please refer to *V*F, and at this time, S H is the conjugate transpose matrix of S. C obtained after projection may alternatively be referred to as downlink channel characteristic information W. After projection, the subband dimension of V is transformed into a delay dimension, and the eigenvector (base station antenna port) dimension of V is transformed into a beam (angle) dimension, but the length of the dimension remains unchanged. Therefore, regardless of whether the downlink channel characteristic information W is obtained before or after projection, the dimension of the downlink channel characteristic information W is described as the number of subbands × the eigenvector length (the number of base station antenna ports).

[0153] Training Data Collection

[0154] In this scenario, both the training samples and the labels are the downlink channel response H or the downlink channel feature information W. The training data may include the downlink channel response H or the downlink channel feature information W in different dimensions. Specific dimensions are shown above.

[0155] Training data processing

[0156] The training data processing includes padding the first input training data, where the padded input training data is referred to as 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 based on the second input training data and the AI ​​model. The first output training data is truncated to determine second output training data, where the length of the second output training data is the same as the length of the labels.

[0157] In this embodiment, if the first input training data is a downlink channel response H, the downlink channel response H needs to be padded in the bandwidth dimension and / or the antenna port dimension. If the first input training data is downlink channel characteristic information W, the downlink channel characteristic information W needs to be padded in the subband dimension and / or the eigenvector dimension to obtain 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 based on the second input training data and the AI ​​model. If the label is a downlink channel response H, the first output training data needs to be truncated in the bandwidth dimension and / or the antenna port dimension. If the label is downlink channel characteristic information W, the first output training data needs to be truncated in the subband dimension and / or the eigenvector dimension to obtain second output training data, and the length of the second output training data is the same as the length of the label.

[0158] The specific padding method and the specific truncation method are the same as those in the above embodiment, and the details will not be described again.

[0159] Model training

[0160] In the model training process, the model training node may determine the structure of the AI ​​model, including 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 longer than or equal to the length of the longest training sample in the training data. The length of the output data in the output format of the AI ​​model is longer than or equal to the length of the longest label in the training data. The AI ​​model is trained by using training data obtained after processing the collected original training data containing multiple bandwidths and multiple different types of reference signals, or by using training data obtained after processing the collected original training data containing multiple subbands and / or multiple eigenvector lengths. In the AI ​​model training process, the training sample, i.e., the first input training data, is padded. The length of the padded input training data is the same as the length of the input data of the AI ​​model. The first output training data of the AI ​​model is truncated, and the length of the truncated output training data is the same as the length of the corresponding label. The model training node may determine the value of a loss function based on the second output training data and the label. If the value of the loss function is smaller than a threshold or satisfies a target requirement, the training of the AI ​​model is completed. Alternatively, if the value of the loss function is above a threshold or does not satisfy the target requirement, the parameters of the AI ​​model are adjusted and training of the AI ​​model continues.

[0161] Model deployment

[0162] In this application, the model training node needs to send the trained AI model to the model inference node and further needs to instruct the model inference node on a data processing method for the AI ​​model so that the model inference node can correctly use the same AI model to complete CSI feedback estimation tasks with different bandwidths and / or different numbers of antenna ports, or so that the model inference node can use the same AI model to complete CSI feedback tasks with different numbers of subbands and / or different eigenvector lengths. For example, the model deployment host may send instruction information to the model inference node, where the instruction information instructs a data processing method for the AI ​​model, and the instruction information may instruct an input data padding rule, an output data truncation rule, etc.

[0163] Model Inference

[0164] The model inference node performs model inference according to a data processing method instructed by the model training node for the AI ​​model to complete a CSI feedback task with a different bandwidth and / or a different number of antenna ports, or to complete a CSI feedback task with a different number of subbands and / or a different eigenvector length. For example, the input data is padded by using the same padding rule as that in the model training process, and the output data is truncated by using the same truncation rule as that in the model training process.

[0165] In a CSI feedback scenario, the input data processing method for the AI ​​model may further indicate whether the input data of the AI ​​model is a downlink channel or feature information of the downlink channel, whether the feature information of the downlink channel is a feature matrix of the downlink channel or a feature matrix obtained after sparse projection, etc. The output data processing method for the AI ​​model may further indicate whether the output data of the AI ​​model is a downlink channel or feature information of the downlink channel, whether the feature information of the downlink channel is a feature matrix of the downlink channel or a feature matrix obtained after sparse projection, etc.

[0166] As described above, the data processing method for the AI ​​model includes at least one of the following: an input data processing method for the AI ​​model or an output data processing method for the AI ​​model. In this design, it is mainly described that the input data processing method for the AI ​​model is an input data truncation method, and the output data processing method for the AI ​​model is an output data padding method. Alternatively, it may be described as follows: the data processing method for the AI ​​model includes at least one of the following: an input data truncation method for the AI ​​model or an output data padding method for the AI ​​model.

[0167] In this design, the model training node performing model training according to the data processing method for the AI ​​model includes truncating first input training data according to an input data truncation method to obtain second input training data, determining first output training data based on the second input training data and the AI ​​model, padding the first output training data according to an output data padding method to obtain second output training data, and performing parameter adjustment on the AI ​​model based on the second output training data.

[0168] In this design, the model inference node performing model inference according to the data processing method for the AI ​​model includes truncating first input data to obtain second input data according to the input data truncation method, determining first output data based on the second input data and the AI ​​model, and padding the first output data according to the output data padding method. The padded 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.

[0169] In this design, the input data truncation method may include at least one of the following: a length obtained after truncation, or a data truncation rule. The output data padding method may include at least one of the following: a length obtained after padding, a padding data type, a data padding rule, etc. For specific truncation and padding descriptions, please refer to the above design.

[0170] For example, if some data in the input data is valid and the remaining data is invalid, or if the energy of the input data is mainly concentrated in some data, the input data may be truncated in the model inference process or the model training process. Input data of different lengths is truncated to have a uniform length. The length of the truncated input data is the same as the length of the input data of the AI ​​model. Output data is determined based on the truncated input data and the AI ​​model. Then, the output data of the AI ​​model is padded.

[0171] A CSI feedback scenario is used as an example. When the downlink channel feature information W is a feature matrix obtained after sparse projection, W is sparse in the angle domain and / or delay domain, i.e., the dimension of W is the number of subbands × the eigenvector length, but the values ​​of most elements in the matrix are small, and the total energy of the matrix is ​​mainly concentrated in some angle and delay paths. In this case, the downlink channel feature information W may be truncated, and only some elements with large values ​​are reserved to obtain the second input data. The first output data is determined based on the second input data and the AI, and then the first output data is converted into the second output data, i.e., the downlink channel feature information restored by using the AI ​​model.

number

[0172] In the above design, the input data is truncated and the output data is padded, so that task inference in different scenarios can be completed by using one AI model.

[0173] In another design, the AI ​​model includes a first sub-model and a second sub-model. The data processing method for the AI ​​model includes at least one of the following: performing data truncation on an output of the first sub-model or performing data padding on an input of the second sub-model. Alternatively, it may be written as: the data processing method for the AI ​​model includes at least one of the following: an output data truncation method for the AI ​​model (e.g., the first sub-model) or an input data padding method for the AI ​​model (e.g., the second sub-model).

[0174] In this design, the model training node performing model training according to the data processing method for the AI ​​model includes determining first output training data of the first submodel based on the input training data of the first submodel and the first submodel; performing data truncation on the first output training data according to the output data truncation method to obtain second output training data of the first submodel and to obtain first input training data of the second submodel based on the second output training data of the first submodel; performing data padding on the first input training data of the second submodel according to the input data padding method to obtain second input training data; determining output training data of the second submodel based on the second input training data and the second submodel; and adjusting at least one of the following model parameters of the first submodel or model parameters of the second submodel based on the input training data of the first submodel and the output training data of the second submodel.

[0175] The model inference node performing model inference according to the data processing method for the AI ​​model includes determining first output data of a first sub-model and performing data truncation on the first output data according to an output data truncation method. The model inference node performing model inference according to the data processing method for the AI ​​model includes padding the first input data to obtain second input data according to an input data padding method, and determining first output data based on the second input data and the second sub-model. For specific input data padding methods and specific output data truncation methods, please refer to the design description above.

[0176] In the present application, the model inference node of the first sub-model and the model inference node of the second sub-model may be the same node or different nodes. This is not limited. For example, in a downlink channel measurement scenario, the inference node of the first sub-model may be a UE, and the inference node of the second sub-model may be a base station. Alternatively, in an uplink channel measurement scenario, the inference node of the first sub-model may be a base station, and the inference node of the second sub-model may be a UE, etc.

[0177] A CSI feedback scenario of a downlink channel is used as an example to describe CSI feedback in the case of different lengths of feedback bits by using the same AI model.

[0178] In a wireless network, the length of the feedback bit B is determined by the base station and is usually not fixed. For example, if the base station expects to obtain higher CSI recovery accuracy, the UE may feed back more feedback bits. In this case, the length of the feedback bit B is long. Alternatively, if the base station expects to reduce CSI feedback overhead, the UE may feed back fewer feedback bits. In this case, the length of the feedback bit B is short. In design, since the length of the feedback bit B is not fixed, for each feedback bit B, a pair of corresponding sub-models f and g needs to be trained. As a result, the overhead for training and maintaining the AI ​​model is high.

[0179] In this application, an AI model may be designed, and the AI ​​model may be applicable to cases of different lengths of the feedback bit B. Subsequently, this application is described in terms of training data collection, training data processing, model training, model deployment, model inference, etc.

[0180] Training Data Collection

[0181] In the CSI feedback scenario, the feedback bits are intermediate results of the entire CSI feedback AI model, and can also be considered as output data of submodel f and input data of submodel g. In this application, the feedback bits B can also be considered as training data.

[0182] The training data set is determined based on an application scenario of the AI ​​model. The training data set includes training data of all possible lengths of the feedback bit B in the application scenario so that the AI ​​model can achieve good performance with feedback bit B of various lengths. In this application, each of the training data includes a training sample, a feedback bit B, and a label. The training sample and the label are the same and are both the downlink channel or the downlink channel characteristic information H. For example, if an AI model that can be applied to CSI feedback with feedback bit B lengths of 20, 30, and 40 is expected to be trained, the training data includes feedback bit B with lengths of 20, 40, and 60. Indeed, the training data also needs to include training samples and labels corresponding to feedback bit B of all lengths. For example, for feedback bit B with a length of 20, the training data needs to include a training sample, feedback bit B(20), and a label. The training sample may be the downlink channel or the downlink channel characteristic information H acquired by the UE, and the label is the downlink channel or the downlink channel characteristic information accurately restored by the base station.

number

[0183] Training data processing

[0184] In this application, as shown in FIG. 13, the training data processing includes truncating the output data of sub-model f so that the length of the output data of sub-model f is the same as the length of feedback bit B, and padding feedback bit B so that the length of feedback bit B is the same as the length of input data of sub-model g.

[0185] In this application, the truncation rule may be one or more of, but is not limited to, truncation at front, truncation at back, truncation at middle, truncation at both sides, truncation at equal intervals, truncation at unequal intervals, etc.

[0186] For example, the output data of submodel f is A=[A1,A2,...,A 60 ].

[0187] If the length of the feedback bits B is 20, then 20 elements may be truncated from the output data A of the sub-model f. For example, the first 20 elements may be truncated from A, and the truncated output data is the feedback bits B=[A1, A2, . . . , A 20 Alternatively, one element is truncated from A at intervals of three elements, and the feedback bits are B = [A1, A4, , A 58 ].

[0188] If the length of the feedback bits B is 40, then 40 elements can be truncated from A. For example, the first 40 elements are truncated from A, and the feedback bits are B=[A1,A2,...,A 40 Alternatively, two elements are truncated from A at intervals of three elements, and the feedback bits are B = [A1, A2, A4, A5, , A 58 ,A 60 ].

[0189] If the length of the feedback bits is 60, data truncation may not be performed on A, and the feedback bits are B=[A1,A2, ,A 60 ].

[0190] After truncation, the training data contains feedback bits of different lengths, so the training data needs to be padded. The padded data is used as the input of sub-model g. The padding rule may be one or more of: padding at the front, padding at the back, padding in the middle, padding at both sides, padding with equal intervals, or padding with unequal intervals. This is not limited. The padding value may be 0, or a very large positive number, or a very small negative number, etc. This is not limited.

[0191] For example, if the length of the feedback bit B is 20 and the first input data of the submodel g is the feedback bit B=[A1,A2,...,A 20 ] and the input length of submodel f is 60, then the padded input data of submodel g is C=[A1,A2, ,A 20 ,0,0,···,0].

[0192] The length of the feedback bit B is 40, and the first input data of the submodel g is the feedback bit B=[A1,A2,...,A 40 ] and the input length of submodel f is 60, then the padded input data of submodel g is C=[A1,A2, ,A 40 ,0,0,···,0].

[0193] If the length of the feedback bit B is 60 and the length of the input data of the submodel f is 60, the feedback bit B is no longer padded and the input of the submodel g is C=B=[A1,A2, ,A 60 ].

[0194] The padding rule may be independent of the truncation rule or may coincide with the truncation rule, e.g., if elements at the front of the output data of submodel f are truncated, then feedback bits may be padded after the feedback bits, etc.

[0195] It should be noted that although the feedback bits B are regarded as training data in this application, in reality, the feedback bits B are intermediate results generated by the AI ​​model, i.e., the feedback bits B are generated only in the training or inference process, and the above truncation and padding operations on the feedback bits B are performed in the training or inference process.

[0196] Model training

[0197] The lengths of the input data and output data of the AI ​​model are determined, where the length of the output data of sub-model f is longer than or equal to the length of the longest feedback bit B in the training data, and the length of the input data of sub-model g is longer than or equal to the length of the longest feedback bit B in the training data.

[0198] The sub-models f and g are trained by using training data obtained after the above training data including the feedback bits B of multiple lengths is processed. As described above, the training data includes training samples, labels, and feedback bits B. The model training process includes the following steps: inputting input training data (e.g., training samples) into submodel f to obtain first output training data of submodel f; truncating the first output training data of submodel f to obtain second output data of submodel f, where the length of the first output training data is truncated to a target length, the target length being all possible lengths of feedback bits in the application scenario of the AI ​​model; obtaining first input training data of submodel g based on the second output data of submodel f; padding the first input training data, where the length of the padded first input training data is the same as the input length of submodel g, where the padded first input training data is referred to as second input training data; obtaining output training data of submodel g based on the second input training data and submodel g; and adjusting at least one of the following: parameters of submodel f or parameters of submodel g based on the input training data of submodel f and the output training data of submodel g. For example, one input training data (i.e., training sample) and corresponding label for a first submodel f is determined, and the value of a loss function is determined based on the label and the output training data of submodel g. If the value of the loss function is less than a threshold or satisfies the target requirement, the training of submodel f and submodel g is completed; otherwise, if the value of the loss function is greater than or equal to the threshold or does not satisfy the target requirement, the parameters of submodel f and / or the parameters of submodel g are adjusted, and the training of submodel f and submodel g continues.

[0199] Model deployment

[0200] When the node that performs model training is different from the node that performs model inference, model deployment needs to be performed. Specifically, the model training node needs to send the trained AI model to the model inference node. In this application, the model training node needs to send the AI ​​model to the model inference node and further needs to instruct the model inference node on the data processing method for the AI ​​model so that the model inference node can correctly use the same AI model to complete the CSI feedback task with feedback bits B of different lengths. The data processing method for the AI ​​model includes data padding rules, data truncation rules, etc.

[0201] Model Inference

[0202] The model inference node can correctly use the AI ​​model to complete the CSI feedback task with different lengths of feedback bits B according to the data processing method for the AI ​​model instructed by the model training node. In other words, the model inference node processes data during model inference by using the same rules as those of the training data processing method. For example, when the UE uses submodel f, the UE obtains downlink channel response H and inputs the downlink channel response H into submodel f to obtain output data of submodel f. The UE truncates the output data of submodel f during inference based on the length of feedback bits B set by the base station and by using the same truncation rule as used for the output data of submodel f during training to obtain feedback bits B for inference. The UE feeds back feedback bits B to the base station. The base station pads feedback bits during inference with the same padding scheme as used to pad feedback bits B during training, and then restores downlink channel response H by using submodel B.

[0203] In this design, the output of submodel f is truncated to feedback bits B of different lengths, and the feedback bits B of different lengths are padded to input data of the same length of submodel g, so that an AI model that can be used for CSI feedback in the case of feedback bits B of multiple lengths can be trained.

[0204] In this application, an example in which the first output data of the first submodel is trained and the input data of the second submodel is padded is used for explanation purposes and is not intended to limit the application. For example, in the design, the output data of the first submodel may be padded, and the input data of the second submodel may be truncated. Furthermore, in this application, a process of processing the output data of the first submodel and the input data of the second submodel is mainly described. In this application, whether the input data of the first submodel and the output data of the second submodel are processed is not limited. For example, in the design, the first input data of the first submodel may be processed to obtain second input data of the first submodel, the first output data of the first submodel is determined based on the second input data of the first submodel and the first submodel, and the first output data of the first submodel is truncated (or padded) to obtain the second output data of the first submodel. The first input data of the second submodel is obtained based on the second output data of the first submodel. The first input data of the second submodel is padded (or truncated) to obtain the second input data of the second submodel. The first output data of the second sub-model is obtained 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 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. For example, the present application may be combined with the above-mentioned processes for processing input data and output data. For example, the first input data of the first sub-model is padded or truncated, and the first output data of the second sub-model is truncated or padded.

[0205] In another design, the data processing method for the AI ​​model includes at least one of the following: an input data partitioning method or an output data reassembly method.

[0206] In this design, the model training node performing model training according to the data processing method for the AI ​​model includes performing data partitioning on first input training data according to an input data partitioning method to obtain second input training data; determining first output training data based on the second input training data and the AI ​​model; performing data reassembly on the first output training data according to an output data reassembly method to obtain second output training data; and performing parameter tuning on the AI ​​model based on the second output training data.

[0207] The model inference node performing model inference according to the data processing method for the AI ​​model includes performing data division on first input data according to an input data division method to obtain second input data, obtaining first output data based on the first input data and the AI ​​model, and performing data reassembly on the first output data according to an output data reassembly method to obtain second output data.

[0208] For example, an AI model such as a Transformer, in which input data and / or output data needs to be processed, is used as an illustrative example.

[0209] The Transformer is a sequence-to-sequence AI model and is the oldest to be applied in natural language processing, such as translation. Sentences vary in length, and one of the Transformer's features is that it supports inputs of any length. Therefore, in theory, the Transformer can be applied to signal processing for different wireless resource configurations. As shown in Figure 14, both the input and output of the Transformer are vector sets. In wireless signal processing, the objects that need to be processed are typically wireless signals, wireless channels, etc. The dimensions of wireless signals or wireless channels typically include the time domain, frequency domain, and spatial domain. Therefore, when the Transformer is applied to wireless signal processing, the original data must be processed and converted into a vector set and input to the Transformer. The vector set output by the Transformer is then converted into target data.

[0210] There may be multiple ways to transform original data into a vector set and transform a vector set into target data. For example, a wireless channel whose original data is 12 subcarriers and 14 OFDM symbols is used as an example. As shown in FIG. 15, in the time domain dimension, the 12×14 original data can be directly divided into 14 vectors with a dimension of 12. Alternatively, as shown in FIG. 16, the 12×14 original data may first be divided into 12 2×7 matrices, and then each 2×7 matrix is ​​converted into a vector with a dimension of 14. Certainly, other division methods may also be used. Similarly, there are multiple ways to reassemble the target data into multiple vectors output by the Transformer.

[0211] There are multiple methods for splitting raw data and multiple methods for reassembling target data. Therefore, the model training node needs to send the trained AI model to the model inference node, and further needs to inform the model inference node of the method for splitting raw data and the method for reassembling target data during training to ensure that the model inference node can correctly use the AI ​​model to complete model inference. The model inference node splits the raw data and reassembles the output data of the AI ​​model by using the same method used during training.

[0212] In this application, for an AI model such as a Transformer whose inputs and outputs follow specific requirements, the model training node instructs the model inference node on how to split the raw data and how to reassemble the AI's output data so that the AI ​​model can accurately process signals within the wireless network.

[0213] To implement the functionality in the above method, a model training node and a model inference It can be understood that a node includes a corresponding hardware structure and / or software module for performing a function. Those skilled in the art will easily realize that the present application can be implemented by hardware or a combination of hardware and computer software with reference to the units and method steps in the examples described herein. Whether the function is performed by hardware or hardware driven by computer software depends on the specific application scenario and design constraints of the technical solution.

[0214] 17 and 18 are diagrams of possible structures of communication devices according to the present application, respectively, which may be configured to implement the functions of the model training node or the model inference node in the above method, and thus can also realize the advantageous effects of the above method.

[0215] As shown in Figure 17, the communication device 1700 includes a processing unit 1710 and a transceiver unit 1720. The communication device 1700 is configured to implement the function of the first node in the method shown in Figure 8. The first node may be a model training node or a model inference node. When the communication device 1700 is configured to implement the function of the first node in the method shown in Figure 8, it is as follows.

[0216] In design, the processing unit 1710 is configured to determine a data processing method for the model, and perform at least one of the following according to the data processing method for the model: perform model training or perform model inference, wherein: The data processing method for the model includes at least one of the following: an input data padding method for the model or an output data truncation method for the model, or alternatively, the data processing method for the model includes at least one of the following: an input data truncation method for the model or an output data padding method for the model.

[0217] In another design, the transceiver unit 1720 is configured to send instructional information to the second node, the instructional information instructing how to process the data for the model, wherein: The data processing method for the model includes at least one of the following: an input data padding method for the model or an output data truncation method for the model, or alternatively, the data processing method for the model includes at least one of the following: an input data truncation method for the model or an output data padding method for the model.

[0218] For a more detailed description of the processing unit 1710 and the transceiver unit 1720, please directly refer to the relevant description of the method shown in Figure 8. Details will not be described here.

[0219] 18, the 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. It may be understood that the interface circuit 1820 may be a transceiver or an input / output interface. Optionally, the communication device 1800 may further include a memory 1830 configured to store instructions to be executed by the processor 1810, or to store input data required for execution of the instructions by the processor 1810, or to store data generated after the processor 1810 executes the instructions.

[0220] When the communications device 1800 is configured to implement the above method, the processor 1810 is configured to implement the functionality of the processing unit 1710 and the interface circuit 1820 is configured to implement the functionality of the transceiver unit 1720.

[0221] When the communication device is a module applied to the first node, the module of the first node implements the function of the first node in the above method. The module of the first node receives information from another module (e.g., a radio frequency module or an antenna) of the first node, and then the information is sent to the first node by the second node. Alternatively, the module of the first node sends information to another module (e.g., a radio frequency module or an antenna) of the first node, and then the information is sent to the second node by the first node. Here, the module of the first node may be a baseband chip of the first node or another module.

[0222] The processor herein may be a central processing unit (CPU), or may be another general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic device, transistor logic device, hardware component, or any combination thereof. The general-purpose processor may be a microprocessor, or may be any conventional processor.

[0223] The memory of this application may be random access memory, flash memory, read-only memory, programmable read-only memory, erasable programmable read-only memory, electrically erasable programmable read-only memory, registers, hard disk, removable hard disk, CD-ROM, or any other form of storage medium well known in the art.

[0224] For example, a storage medium may be coupled to the processor such that the processor can read information from, and write information to, the storage medium. The storage medium may alternatively be components of the processor. The processor and the storage medium may reside in an ASIC. Further, the ASIC may reside in a base station or terminal. Indeed, the processor and the storage medium may reside as discrete components in a base station or terminal.

[0225] Some or all of the methods of the present application may be implemented by software, hardware, firmware, or any combination thereof. When software is used to implement the methods, some or all of the methods may be implemented in the form of a computer program product. The computer program product includes one or more computer programs or instructions. When the computer programs or instructions are loaded into a computer and executed, all or some of the procedures or functions according to the present application are performed. The computer may be a general-purpose computer, a special-purpose computer, a computer network, a network device, a user device, a core network device, an OAM, or other programmable device. The computer program or instructions may be stored in a computer-readable storage medium or transmitted from a computer-readable storage medium to another computer-readable storage medium. For example, the computer program or instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired or wireless method. The computer-readable storage medium may be any useful medium or data storage device accessible by a computer, such as a server or data center incorporating one or more useful media. Useful media may be magnetic media such as floppy disks, hard disks, or magnetic tape, or optical media such as digital video disks, or semiconductor media such as solid state drives. The computer readable storage medium may be a volatile or non-volatile storage medium, or may include two types of storage media: volatile and non-volatile storage media.

[0226] In this application, unless otherwise stated or there is no logical contradiction, the terms and / or descriptions in different embodiments may be consistent and cross-referenced, and the technical features in different embodiments may be combined based on their internal logical relationships to form new embodiments.

[0227] In this application, "at least one" means one or more, and "plurality" means two or more. "And / or" describes an association relationship between related objects and indicates that three relationships may exist. For example, A and / or B can indicate that A exists alone, that both A and B exist, and that B exists alone, and A and B may be singular or plural. In text descriptions in this application, the " / " character typically indicates an "or" relationship between related objects. In formulas in this application, the " / " character indicates a "divide by" relationship between related objects. "Comprising at least one of A, B, or C" can indicate including A, including B, including C, including A and B, including A and C, including B and C, or including A, B, and C.

[0228] It can be understood that the various numbers used in this application are merely distinguished for ease of description and are not intended to limit the scope of this application. The sequential numbers of the above processes do not imply an execution order, and the execution order of the processes should be determined based on the functions and internal logic of the processes.

Claims

1. 1. A data processing method for a model, comprising: determining the data processing method for the model; and and performing at least one of the following in accordance with the data processing method for the model: performing model training; or performing model inference; The data processing method for the model includes an input data processing method selected based on the length of sample data input to the model and the input format of the model, and an output data processing method selected based on the output format of the model and the length of a processing result corresponding to the sample data, and when an input data padding method is selected as the input data processing method, an output data truncation method is selected as the output data processing method, or when the input data truncation method is selected as the input data processing method, an output data padding method is selected as the output data processing method. method.

2. Determining the data processing method for the model comprises: receiving, from a first node, instructional information instructing the data processing method for the model; or determining the data processing method for the model as defined in a protocol; The method of claim 1.

3. The input data padding method or the output data padding method has at least one of the following padded data length, padding data type, or data padding rule: The method of claim 1.

4. The output data truncation method or the input data truncation method has at least one of the following data length after truncation or data truncation rule:

4. The method according to any one of claims 1 to 3.

5. performing the model inference according to the data processing method for the model, performing data padding on the first input data according to the input data padding method to obtain second input data; determining first output data based on the second input data and the model; performing data truncation on the first output data according to the output data truncation method.

4. The method according to any one of claims 1 to 3.

6. performing the model training according to the data processing method for the model, performing data padding on the first input training data according to the input data padding method to obtain second input training data; determining first output training data based on the second input training data and the model; performing data truncation on the first output training data according to the output data truncation method to obtain second output training data; and performing parameter adjustment on the model based on the second output training data.

4. The method according to any one of claims 1 to 3.

7. performing the model inference according to the data processing method for the model, performing data truncation on the first input data according to the input data truncation method to obtain second input data; determining first output data based on the second input data and the model; performing data padding on the first output data according to the output data padding method.

4. The method according to any one of claims 1 to 3.

8. performing the model training according to the data processing method for the model, performing data truncation on the first input training data according to the input data truncation method to obtain second input training data; determining first output training data based on the second input training data and the model; performing data padding on the first output training data according to the output data padding method to obtain second output training data; and performing parameter adjustment on the model based on the second output training data.

4. The method according to any one of claims 1 to 3.

9. the model has a first sub-model, and performing the model inference according to the data processing method for the model includes: determining first output data for the first sub-model; performing data truncation on the first output data according to the output data truncation method.

4. The method according to any one of claims 1 to 3.

10. the model has a second sub-model, and performing the model inference according to the data processing method for the model includes: performing data padding on the first input data according to the input data padding method to obtain second input data; determining first output data based on the second input data and the second sub-model; 4. The method according to any one of claims 1 to 3.

11. the model has a first sub-model and a second sub-model, and performing the model training according to the data processing method for the model includes: determining first output training data for the first sub-model based on the input training data for the first sub-model and the first sub-model; performing data truncation on the first output training data according to the output data truncation method to obtain first input training data for the second sub-model; performing data padding on the first input training data of the second sub-model according to the input data padding method to obtain second input training data; determining output training data for the second sub-model based on the second input training data and the second sub-model; and adjusting at least one of model parameters of the first sub-model or model parameters of the second sub-model based on the input training data of the first sub-model and the output training data of the second sub-model.

4. The method according to any one of claims 1 to 3.

12. 1. A data processing method for a model, comprising: sending instruction information to a second node instructing the data processing method for the model; the data processing method for the model includes an input data processing method selected based on the length of sample data input to the model and the input format of the model, and an output data processing method selected based on the output format of the model and the length of a processing result corresponding to the sample data, and when the input data padding method is selected as the input data processing method, an output data truncation method is selected as the output data processing method, or when the input data truncation method is selected as the input data processing method, an output data padding method is selected as the output data processing method. method.

13. The input data padding method or the output data padding method has at least one of the following padded data length, padding data type, or data padding rule: The method of claim 12.

14. The output data truncation method or the input data truncation method has at least one of the following data length after truncation or data truncation rule:

14. The method according to claim 12 or 13.

15. A communication device comprising a unit adapted to perform the method of claim 1.

16. a processor and a memory; the processor is coupled to the memory; The processor is configured to perform the method of claim 1. Communication equipment.

17. A communication device comprising a unit adapted to perform the method according to claim 12.

18. a processor and a memory; the processor is coupled to the memory; The processor is configured to perform the method of claim 12. Communication equipment.

19. a communication device according to claim 15 or 16; The communication device according to claim 17 or 18. A communication system having:

20. A computer-readable storage medium storing instructions, comprising: The instructions, when executed on a computer, enable the computer to perform the method of claim 1. A computer-readable storage medium.

21. A computer-readable storage medium storing instructions, comprising: The instructions, when executed on a computer, enable the computer to perform the method of claim 12. A computer-readable storage medium.

22. Have commands, The instructions, when executed on a computer, enable the computer to perform the method of claim 1. Computer program.

23. Have commands, The instructions, when executed on a computer, enable the computer to perform the method of claim 12. Computer program.

Images