Communication method and communication device

Abstract

The invention provides a communication method and a communication device, and relates to the field of communication. The method comprises the following steps: acquiring a first data set, wherein the first data set is used for verifying at least one first neural network model; and sending feedback information, wherein the feedback information is used for indicating the performance of the at least one first neural network model. The terminal and the network equipment can determine whether the at least one first neural network model is matched with the scene or determine the neural network model with the best performance in the at least one first neural network model according to the performance of the at least one first neural network model. Therefore, the problem that the performance may be reduced when the current neural network model is used for signal processing in a new scene is solved.

Description

Technical Field

[0001] The present application relates to the field of communications, and in particular to a communication method and a communication device. Background Art

[0002] The fifth generation of mobile communication systems is often referred to as new radio (NR) systems. With the development of NR, artificial intelligence models can be used to replace part or all of the signal processing modules of the physical layer of the wireless communication system to improve signal processing performance. As part of the physical layer signal processing module, the artificial intelligence model can be deployed on the base station side, the terminal side, or both the base station and the terminal to achieve different functions.

[0003] One of the most commonly used implementation methods of artificial intelligence models is to use a neural network model. A neural network model consists of two elements: structure and parameters. The structure and parameters of a neural network model generally need to be determined through training based on a data set. Model training is generally divided into two methods: offline training and online training / update. The current technology determines the structure and parameters of the air interface neural network model by offline training, where the data used for offline training is pre-defined or collected data. Specifically, if the neural network model needs to be deployed on the base station side, offline training is performed on the base station side; if the neural network model is to be deployed on the terminal side, offline training is performed on the terminal side; if the neural network model is to be deployed on both the base station and the terminal, assuming that the base station side model is model A and the terminal side model is model B, then model A and model B are simultaneously trained offline on the base station side, and model B is sent to the terminal side through the air interface for deployment, or model A and model B are simultaneously trained offline on the terminal side, and model A is sent to the base station side through the air interface for deployment.

[0004] Since offline training takes a long time, the update frequency of the parameters and structure of the neural network model is much lower than the change frequency of the air interface signal. When the scenario matching the training data set of the air interface neural network model (for example, the delay spread parameters of the channel, the interference intensity of the neighboring area, the terminal movement speed and other factors) changes, using the current neural network model for signal processing in the new scenario may result in performance degradation. Summary of the invention

[0005] The present application provides a communication method and a communication device, which help to solve the problem that performance may be degraded when using the current neural network model for signal processing in new scenarios.

[0006] In a first aspect, a communication method is provided, which can be executed by a terminal, or by a component of the terminal (such as a processor, a chip, or a chip system, etc.), or by a logic module or software that can implement all or part of the terminal functions.

[0007] The method includes: acquiring a first data set, the first data set is used to verify at least one first neural network model; sending feedback information, the feedback information is used to indicate the performance of the at least one first neural network model.

[0008] The first data set includes sample data and labels corresponding to the sample data (ie, sample labels). Acquiring the first data set may include: acquiring the sample data and / or sample labels locally, or receiving the sample data and / or sample labels from a network device.

[0009] According to the communication method provided in the present application, the terminal can verify at least one first neural network model based on the first data set, and then determine the performance of the at least one first neural network model, so as to determine whether the at least one first neural network model matches the scene or determine the neural network model with the best performance among the at least one first neural network model. Furthermore, the terminal can deploy the first neural network model with the best performance at the physical layer, perform signal processing based on the first neural network model, and ensure the signal processing performance as much as possible. On the other hand, the terminal can indicate the performance of the at least one first neural network model to the network device, so that the network device can also perform corresponding operations based on the feedback of the terminal, and avoid the performance degradation problem that may be caused by the use of a neural network model that does not match the current scene by itself or the terminal side as much as possible. For example, the network device uses the second neural network model corresponding to the first neural network model with the best performance for signal processing.

[0010] In a possible implementation, the at least one first neural network model is preconfigured or acquired through model information of the at least one first neural network model.

[0011] Exemplarily, the model information of the at least one first neural network model may be sent by the network device to the terminal. The model information may include structure parameters and weight parameters.

[0012] In a possible implementation, the first time unit in which the feedback information is sent is the same as the first time unit in which the first data set is acquired. In other words, the feedback information and the first data set are in the same first time unit.

[0013] That is, the signal sent in the same first time unit carries the feedback information and the first data set. Exemplarily, the first time unit may be a mini slot, a slot, a subframe, or a frame, or a time unit with a smaller or larger time length that has not yet been defined in the protocol.

[0014] Based on this solution, by transmitting feedback information and the first data set in the same first time unit, timely transmission of feedback information corresponding to the first data set can be achieved, thereby ensuring as much as possible that the feedback information on the terminal side can indicate the performance of each first neural network model in the current scenario.

[0015] In a possible implementation, the second time unit where the first data set is located is the same as the second time unit where the model information is located. In other words, the first data set and the model information are located in the same second time unit.

[0016] That is, the signal sent in the same second time unit carries the first data set and the model information. Exemplarily, the second time unit may be a mini-time slot, a time slot, a subframe or a frame, or a time unit with a smaller or larger time length that has not yet been defined in the protocol.

[0017] Based on this solution, the network device side can ensure as much as possible that the model verification result on the terminal side can indicate the performance of each first neural network model in the current scenario by sending the model information and the first data set of at least one first neural network model on the same second time unit.

[0018] In a possible implementation, the start time of the symbol carrying the model information does not lag behind the start time of the symbol carrying the first data set. For example, the start time of the symbol carrying the model information may be the same as the start time of the symbol carrying the first data set, or the start time of the symbol carrying the model information may be earlier than the start time of the symbol carrying the first data set.

[0019] Based on this solution, the terminal will first receive the model information and then receive the first data set, so that the at least one first neural network model can be verified based on the first data set without storing the first data set, thereby reducing cache overhead.

[0020] In a possible implementation, the third time unit in which the feedback information is sent is the same as the third time unit in which the model information is located and the third time unit in which the first data set is acquired. In other words, the feedback information, the model information, and the first data set are located in the same third time unit. Exemplarily, the third time unit may be a mini-time slot, a time slot, a subframe, or a frame, or a time unit that represents a smaller or larger time length that has not yet been defined in the protocol.

[0021] Based on this solution, by carrying the first data set, the model information of the at least one first neural network model and the feedback information on the same third time unit, it is possible to ensure as much as possible that the model verification result on the terminal side can indicate the performance of each first neural network model in the current scenario.

[0022] In a possible implementation, the first neural network model is trained based on the second neural network model and the second data set.

[0023] For example, the sample data in the second data set may include parameters related to the channel (or channel model) and / or amplifier nonlinearity.

[0024] In a possible implementation, before sending the feedback information, the method further includes: verifying the at least one first neural network model using the first data set and the at least one first neural network model.

[0025] For example, the sample data in the first data set may be input into the at least one first neural network model to obtain the output result of the at least one first neural network model. After obtaining the output result of the at least one first neural network model, the gap between the output result of the at least one first neural network model and the sample labels in the first data set may be calculated. For example, the gap may be the Euclidean distance between the output result of the first neural network model and the sample labels in the first data set.

[0026] Exemplarily, the feedback information may include: the gap between the output results of the at least one first neural network model and the sample labels in the first data set.

[0027] Exemplarily, the gap between the output result of the first neural network model and the sample label in the first data set can be used as an evaluation indicator of model performance. It should be understood that the smaller the gap, the better the performance of the corresponding first neural network model and the higher the degree of matching with the scene.

[0028] Based on this solution, the terminal can obtain the verification result of the at least one first neural network model by verifying the at least one first neural network model. The verification result can indicate the performance of the at least one first neural network model, and further determine whether the at least one first neural network model is applicable to the current scenario.

[0029] In a possible implementation, before sending the feedback information, the method further includes: receiving a control channel, where the control channel includes a sample label or indication information corresponding to the sample data in the first data set, where the indication information indicates the sample label.

[0030] Exemplarily, the fourth time unit in which the control channel is located is the same as the fourth time unit in which one or more of the following are located: sample data in the first data set, model information of the at least one first neural network model, or the feedback information. Exemplarily, the fourth time unit can be a mini-time slot, a time slot, a subframe or a frame, or a time unit representing a smaller or larger time length that has not yet been defined in the protocol.

[0031] It should be understood that the above-mentioned control channel may also be a channel that has not yet been defined in the existing protocol with control channel function. This channel is dedicated to carrying control information or indication information related to the AI ​​model, sample data or sample labels in the data set.

[0032] Based on this solution, by carrying the sample data in the first data set, the model information of the at least one first neural network model or one or more of the feedback information and the control channel on the same fourth time unit, it can be ensured as much as possible that the model verification result on the terminal side can indicate the performance of each first neural network model in the current scenario.

[0033] On the second aspect, a communication method is provided, which can be executed by a network device, or by a component of the network device (such as a processor, chip, or chip system, etc.), or by a logic module or software that can implement all or part of the network device functions.

[0034] The method includes: sending model information and / or a first data set of at least one first neural network model, the first data set is used to verify the at least one first neural network model; receiving feedback information, the feedback information is used to indicate the performance of the at least one first neural network model.

[0035] Exemplarily, the first data set includes sample data and labels corresponding to the sample data (ie, sample labels).

[0036] According to the communication method provided in the present application, the terminal can verify at least one first neural network model based on the first data set, and then determine the performance of the at least one first neural network model, so as to determine whether the at least one first neural network model matches the scene or determine the neural network model with the best performance among the at least one first neural network model. Furthermore, the terminal can deploy the first neural network model with the best performance at the physical layer, perform signal processing based on the first neural network model, and ensure the signal processing performance as much as possible. On the other hand, the terminal can indicate the performance of the at least one first neural network model to the network device, so that the network device can also perform corresponding operations based on the feedback of the terminal, and avoid the performance degradation problem that may be caused by the use of a neural network model that does not match the current scene by itself or the terminal side as much as possible. For example, the network device uses the second neural network model corresponding to the first neural network model with the best performance for signal processing.

[0037] In a possible implementation manner, the first time unit in which the feedback information is located is the same as the first time unit in which the first data set is located.

[0038] In a possible implementation manner, the second time unit where the first data set is located is the same as the second time unit where the model information is located.

[0039] In a possible implementation, the start time of the symbol carrying the model information does not lag behind the start time of the symbol carrying the first data set. For example, the start time of the symbol carrying the model information may be the same as the start time of the symbol carrying the first data set, or the start time of the symbol carrying the model information may be earlier than the start time of the symbol carrying the first data set.

[0040] In a possible implementation manner, the third time unit in which the feedback information is received is the same as the third time unit in which the model information is received and the third time unit in which the first data set is sent.

[0041] In a possible implementation, the first neural network model is trained based on the second neural network model and the second data set.

[0042] In a possible implementation manner, the first data set includes sample data, or the first data set includes the sample data and a sample label corresponding to the sample data.

[0043] In a possible implementation manner, a control channel is sent, where the control channel includes a sample label or indication information corresponding to the sample data in the first data set, where the indication information indicates the sample label.

[0044] Exemplarily, the fourth time unit in which the control channel is located is the same as the fourth time unit in which one or more of the following are located: sample data in the first data set, model information of the at least one first neural network model, or the feedback information. Exemplarily, the fourth time unit can be a subframe.

[0045] For possible implementation methods and beneficial effects of the second aspect, reference may be made to the description of the first aspect, and no further elaboration will be given in the second aspect.

[0046] On the third aspect, a communication method is provided, which can be executed by a terminal, or by a component of the terminal (such as a processor, chip, or chip system, etc.), or by a logic module or software that can implement all or part of the terminal functions.

[0047] The method includes: obtaining a first data set; obtaining at least one second neural network model based on the first data set and at least one first neural network model; obtaining a second data set, the second data set being used to verify the at least one first neural network model and / or the at least one second neural network model; and sending feedback information, the feedback information being used to indicate the performance of the at least one first neural network model and / or the performance of the at least one second neural network model.

[0048] In an embodiment of the present application, the at least one first neural network model can be updated using a first data set to obtain at least one updated first neural network model, and the at least one updated first neural network model is the at least one second neural network model.

[0049] It should be understood that if the at least one first neural network model is verified using the second data set, the feedback information may indicate the performance of the at least one first neural network model. If the at least one second neural network model is verified using the second data set, the feedback information may indicate the performance of the at least one second neural network model.

[0050] The first data set includes sample data and labels corresponding to the sample data (ie, sample labels). Acquiring the first data set may include: acquiring the sample data and / or sample labels locally, or receiving the sample data and / or sample labels from a network device.

[0051] The second data set includes sample data and labels corresponding to the sample data (ie, sample labels). Acquiring the second data set may include: acquiring the sample data and / or sample labels locally, or receiving the sample data and / or sample labels from a network device.

[0052] According to the communication method provided in the present application, the terminal can verify the performance of at least one neural network model (at least one first neural network model and / or at least one second neural network model) according to the second data set, and then determine the performance of the at least one neural network model, so as to determine whether the at least one neural network model matches the scene or determine the neural network model with the best performance among the at least one neural network model. Furthermore, the terminal can deploy the neural network model with the best performance at the physical layer, receive signals based on the neural network model, and ensure the signal processing performance as much as possible. On the other hand, the terminal can indicate the performance of the at least one neural network model to the network device, so that the network device can also perform corresponding operations based on the feedback of the terminal, and avoid the performance degradation problem that may be caused by the use of a neural network model that does not match the current scene on the terminal side as much as possible. For example, when the performance requirements of each neural network model are not met, the network device reconfigures the neural network model for the terminal or instructs the terminal to turn off the physical layer AI function.

[0053] In a possible implementation, the at least one first neural network model is preconfigured or acquired through model information of the at least one first neural network model.

[0054] Exemplarily, the model information of the at least one first neural network model may be sent by the network device to the terminal. The model information may include structure parameters and weight parameters.

[0055] In a possible implementation, the first time unit where the first data set and / or the second data set is located is the same as the first time unit where the feedback information is located. In other words, the first data set and / or the second data set and the feedback information are located in the same first time unit.

[0056] That is, the signal sent in the same first time unit carries one or both of the first data set and the second data set and feedback information. Exemplarily, the first time unit can be a mini-time slot, a time slot, a subframe or a frame, or a time unit with a smaller or larger time length that has not yet been defined in the protocol.

[0057] Based on this solution, by transmitting one or two of the first data set and the second data set and the feedback information in the same time unit, timely transmission of the feedback information corresponding to one or two of the first data set and the second data set can be achieved, thereby ensuring as much as possible that the feedback information on the terminal side can indicate the performance of each neural network model in the current scenario.

[0058] In a possible implementation, the second time unit where the first data set is located is the same as the second time unit where the second data set is located. In other words, the first data set and the second data set are located in the same second time unit.

[0059] That is, the signal sent in the same second time unit carries the first data set and the second data set. Exemplarily, the second time unit can be a mini-time slot, a time slot, a subframe or a frame, or a time unit with a smaller or larger time length that has not yet been defined in the protocol.

[0060] Based on this solution, the network device side sends the first data set and the second data set in the same second time unit, so as to ensure as much as possible that the model verification result on the terminal side can indicate the performance of the first neural network model and the second neural network model in the current scenario.

[0061] In a possible implementation, the symbol time domain start position of the first data set does not lag behind the symbol time domain start position of the second data set. In other words, the start time of the symbol carrying the first data set does not lag behind the start time of the symbol carrying the second data set. For example, the symbol time domain start position of the first data set may be the same as the symbol time domain start position of the second data set, or the symbol time domain start position of the first data set may be earlier than the symbol time domain start position of the second data set.

[0062] Based on this solution, the terminal will first receive the first data set and then the second data set, so that the at least one first neural network model can be verified based on the second data set and / or the at least one second neural network model obtained by updating the at least one first neural network model using the first data set can be verified without storing the second data set, thereby reducing cache overhead.

[0063] In one possible implementation, before sending the feedback information, the method further includes: verifying the at least one first neural network model through the second data set and the at least one first neural network model; or, verifying the at least one second neural network model through the second data set and the at least one second neural network model.

[0064] For example, the sample data in the second data set can be input into the at least one first neural network model and / or the at least one second neural network model to obtain the output results of the at least one first neural network model and / or the at least one second neural network model. After obtaining the output results of the at least one first neural network model and / or the at least one second neural network model, the gap between the output results of the at least one first neural network model and the sample labels in the second data set, and / or the gap between the output results of the at least one second neural network model and the sample labels in the second data set can be calculated. For example, the gap can be the Euclidean distance between the output results of the first neural network model / the second neural network model and the sample labels in the first data set.

[0065] Exemplarily, the feedback information may include: the gap between the output results of the at least one first neural network model and the sample labels in the second data set, and / or the gap between the output results of the at least one second neural network model and the sample labels in the second data set.

[0066] Exemplarily, the gap between the output result of the first neural network model / the second neural network model and the sample label in the second data set can be used as an evaluation index of model performance. It should be understood that the smaller the gap, the better the performance of the corresponding first neural network model / the second neural network model, and the higher the degree of matching with the scene.

[0067] Based on this scheme, the terminal can obtain the gap between the at least one first neural network model and / or the at least one second neural network model and the sample labels in the second data set by verifying the at least one first neural network model and / or the at least one second neural network model. The gap can indicate the performance of the at least one first neural network model and / or the at least one second neural network model, and further determine whether the at least one first neural network model and / or the at least one second neural network model is applicable to the current scenario.

[0068] In a possible implementation manner, both the first data set and the second data set include sample data, or both the first data set and the second data set include sample labels of the sample data.

[0069] In one possible implementation, before sending feedback information, the method also includes: receiving a control channel, the control channel including sample labels or first indication information corresponding to the sample data in the first data set, and / or, including sample labels or second indication information corresponding to the sample data in the second data set, the first indication information indicating the sample labels corresponding to the sample data in the first data set, and the second indication information indicating the sample labels corresponding to the sample data in the second data set.

[0070] Optionally, the third time unit where the control channel is located is the same as the third time unit where one or more of the following are located: sample data in the first data set, sample data in the second data set, or feedback information.

[0071] Exemplarily, the third time unit may be a mini-time slot, a time slot, a subframe or a frame, or a time unit representing a smaller or larger time length that has not yet been defined in the protocol.

[0072] Based on this solution, by carrying one or more of the sample data in the first data set, the sample data in the second data set or the feedback information and the control channel on the same third time unit, it can be ensured as much as possible that the model verification result on the terminal side can indicate the performance of each neural network model in the current scenario.

[0073] In a fourth aspect, a communication method is provided, which can be executed by a network device, or by a component of a network device (such as a processor, chip, or chip system, etc.), or by a logic module or software that can implement all or part of the functions of the network device.

[0074] The method includes: sending one or more of model information of at least one first neural network model, a first data set and a second data set, the first data set and the at least one first neural network model are used to obtain at least one second neural network model, and the second data set is used to verify the at least one first neural network model and / or at least one second neural network model; receiving feedback information, the feedback information is used to indicate the performance of the at least one first neural network model and / or the performance of the at least one second neural network model.

[0075] According to the communication method provided in the present application, the terminal can verify the performance of at least one neural network model (at least one first neural network model and / or at least one second neural network model) according to the second data set, and then determine the performance of the at least one neural network model, so as to determine whether the at least one neural network model matches the scene or determine the neural network model with the best performance among the at least one neural network model. Furthermore, the terminal can deploy the neural network model with the best performance at the physical layer, receive signals based on the neural network model, and ensure the signal processing performance as much as possible. On the other hand, the terminal can indicate the performance of the at least one first neural network model to the network device, so that the network device can also perform corresponding operations based on the feedback of the terminal, and avoid the performance degradation problem that may be caused by the use of a neural network model that does not match the current scene on the terminal side as much as possible. For example, when the performance requirements of each neural network model are not met, the network device reconfigures the neural network model for the terminal or instructs the terminal to turn off the physical layer AI function.

[0076] In a possible implementation manner, the first time unit in which the first data set and / or the second data set is located is the same as the first time unit in which the feedback information is located.

[0077] In a possible implementation, the second time unit where the first data set is located is the same as the second time unit where the second data set is located. Exemplarily, the first time unit may be a mini-time slot, a time slot, a subframe or a frame, or a time unit with a smaller or larger time length that has not yet been defined in the protocol.

[0078] In a possible implementation manner, the second time unit where the first data set is located is the same as the second time unit where the second data set is located.

[0079] In a possible implementation, the symbol time domain starting position of the first data set does not lag behind the symbol time domain starting position of the second data set. For example, the symbol time domain starting position of the first data set may be the same as the symbol time domain starting position of the second data set, or the symbol time domain starting position of the first data set may be earlier than the symbol time domain starting position of the second data set.

[0080] In a possible implementation manner, both the first data set and the second data set include sample data, or both the first data set and the second data set include sample labels of the sample data.

[0081] In one possible implementation, before receiving the feedback information, the method also includes: sending a control channel, the control channel including a sample label or first indication information corresponding to the sample data in the first data set, and / or including a sample label or second indication information corresponding to the sample data in the second data set, the first indication information indicating the sample label corresponding to the sample data in the first data set, and the second indication information indicating the sample label corresponding to the sample data in the second data set.

[0082] Optionally, the third time unit where the control channel is located is the same as the third time unit where one or more of the following are located: sample data in the first data set, sample data in the second data set, or feedback information.

[0083] Exemplarily, the third time unit may be a mini-time slot, a time slot, a subframe or a frame, or a time unit representing a smaller or larger time length that has not yet been defined in the protocol.

[0084] For possible implementation methods and beneficial effects of the fourth aspect, please refer to the description of the third aspect, and will not be repeated in the fourth aspect.

[0085] In a fifth aspect, a communication device is provided, comprising a module or unit for executing the method in the first aspect or any possible implementation of the first aspect, or comprising a module or unit for executing the method in the third aspect or any possible implementation of the third aspect.

[0086] In a sixth aspect, a communication device is provided, comprising a module or unit for executing the method in the second aspect or any possible implementation of the second aspect, or comprising a module or unit for executing the method in the fourth aspect or any possible implementation of the fourth aspect.

[0087] In the seventh aspect, a communication device is provided, comprising a processor, which, when executing a computer program (also referred to as code, or instruction) or instruction stored in a memory, causes the device to execute: the method in the first aspect or any possible implementation of the first aspect, or the method in the third aspect or any possible implementation of the third aspect.

[0088] In a possible implementation manner, the device also includes the memory.

[0089] In a possible implementation, there are one or more processors and / or one or more memories.

[0090] In a possible implementation, the memory may be integrated with the processor, or the memory may be separately provided with the processor.

[0091] In a possible implementation manner, the device further includes a communication interface, and the processor is coupled to the communication interface.

[0092] In one implementation, the device is a terminal. Exemplarily, the communication interface may be a transceiver, or an input / output interface.

[0093] In another implementation, the device is a chip in a terminal. Exemplarily, the communication interface may be an input / output interface.

[0094] In an eighth aspect, a communication device is provided, comprising a processor, wherein when the processor executes a computer program or instruction stored in a memory, the device executes: the method in the second aspect or any possible implementation of the second aspect, or the method in the fourth aspect or any possible implementation of the fourth aspect.

[0095] In a possible implementation manner, the device also includes the memory.

[0096] In a possible implementation, there are one or more processors and / or one or more memories.

[0097] In a possible implementation, the memory may be integrated with the processor, or the memory may be separately provided with the processor.

[0098] In a possible implementation manner, the device further includes a communication interface, and the processor is coupled to the communication interface.

[0099] In one implementation, the device is a network device. Exemplarily, the communication interface may be a transceiver, or an input / output interface.

[0100] In another implementation, the device is a chip in a network device. Exemplarily, the communication interface may be an input / output interface.

[0101] In a ninth aspect, a processor is provided, comprising: an input circuit, an output circuit, and a processing circuit. The processing circuit is used to receive a signal through the input circuit and transmit a signal through the output circuit, so that the processor executes the method in any one of the above aspects or any possible implementation of any one of the aspects.

[0102] In the specific implementation process, the processor can be a chip, the input circuit can be an input pin, the output circuit can be an output pin, and the processing circuit can be a transistor, a gate circuit, a trigger, and various logic circuits. The input signal received by the input circuit can be, for example, but not limited to, received and input by a receiver, and the signal output by the output circuit can be, for example, but not limited to, output to a transmitter and transmitted by the transmitter, and the input circuit and the output circuit can be the same circuit, which is used as an input circuit and an output circuit at different times. The embodiments of the present application do not limit the specific implementation methods of the processor and various circuits.

[0103] In the tenth aspect, a communication system is provided, comprising an apparatus for executing the method in the first aspect or any possible implementation of the first aspect, and / or an apparatus for executing the method in the second aspect or any possible implementation of the second aspect; or, comprising an apparatus for executing the method in the third aspect or any possible implementation of the third aspect, and / or an apparatus for executing the method in the fourth aspect or any possible implementation of the fourth aspect.

[0104] In an eleventh aspect, a computer program product is provided, the computer program product comprising: a computer program, when the computer program is executed, enables a computer to execute the method in any one of the above aspects or any possible implementation of any one of the aspects.

[0105] In the twelfth aspect, a computer-readable storage medium is provided, which stores a computer program (also referred to as code, or instruction). When the computer program runs on a computer, the computer executes a method in any one of the above aspects or any possible implementation of any one of the aspects.

[0106] In a thirteenth aspect, a chip is provided, comprising a processor for calling and running a computer program from a memory, so that a communication device equipped with the chip executes a method in any one of the above aspects or any possible implementation of any one of the aspects.

[0107] In a fourteenth aspect, a communication device is provided, comprising an interface and a processor, wherein the interface is used to send and / or receive signals, so that the processor executes a method in any one of the above aspects or any possible implementation of any one of the aspects. BRIEF DESCRIPTION OF THE DRAWINGS

[0108] Figure 1 It is a schematic diagram of the architecture of a mobile communication system provided by an embodiment of the present application;

[0109] Figure 2 is a schematic flow chart of a communication method provided in an embodiment of the present application;

[0110] Figure 3 It is a method of transmitting part or all of the first data set and feedback information provided by an embodiment of the present application;

[0111] Figure 4 It is a method of transmitting model information and part or all of the first data set provided by an embodiment of the present application;

[0112] Figure 5 It is a schematic diagram of the temporal position relationship between a part or all of the first data set and the model information provided in an embodiment of the present application;

[0113] Figure 6 It is a method of transmitting model information, part or all of the first data set and feedback information provided in an embodiment of the present application;

[0114] Figure 7 This is another method of transmitting model information, part or all of the first data set and feedback information provided in an embodiment of the present application;

[0115] Figure 8 It is a method of transmitting a control channel, model information, sample data in a first data set, and feedback information provided by an embodiment of the present application;

[0116] Fig. 9 This is another method of transmitting a control channel, model information, sample data in a first data set, and feedback information provided by an embodiment of the present application;

[0117] Fig.10 This is a schematic diagram of an application scenario provided by an embodiment of the present application;

[0118] Fig.11 is a schematic flow chart of a communication method provided in an embodiment of the present application;

[0119] Fig.12 This is a schematic diagram of an application scenario provided by an embodiment of the present application;

[0120] Fig.13 is a schematic flow chart of a communication method provided in an embodiment of the present application;

[0121] Fig.14 This is a schematic diagram of an application scenario provided by an embodiment of the present application;

[0122] Fig.15 is a schematic flow chart of a communication method provided in an embodiment of the present application;

[0123] Fig.16 is a schematic flow chart of a communication method provided in an embodiment of the present application;

[0124] Fig.17It is a method of transmitting part or all of the first data set and part or all of the second data set provided by an embodiment of the present application;

[0125] Fig.18 This is another method of transmitting part or all of the first data set and part or all of the second data set provided by an embodiment of the present application;

[0126] Fig.19 It is a schematic diagram of the temporal position relationship between part or all of the first data set and part or all of the second data set provided in an embodiment of the present application;

[0127] Fig. 20 It is a method of transmitting part or all of the first data set, part or all of the second data set and feedback information provided by an embodiment of the present application;

[0128] Fig.21 This is another method of transmitting part or all of the first data set, part or all of the second data set, and feedback information provided by an embodiment of the present application;

[0129] Fig. 22 A method for transmitting a control channel, sample data in a first data set, sample data in a second data set, and feedback information is provided in an embodiment of the present application;

[0130] Fig.23 This is another method of transmitting a control channel, sample data in a first data set, sample data in a second data set, and feedback information provided by an embodiment of the present application;

[0131] Fig.24 This is a schematic diagram of an application scenario provided by an embodiment of the present application;

[0132] Fig.25 is a schematic flow chart of a communication method provided in an embodiment of the present application;

[0133] Fig.26 is a schematic block diagram of a communication device provided in an embodiment of the present application;

[0134] Fig. 27 is a schematic block diagram of another communication device provided in an embodiment of the present application;

[0135] Fig.28 is a schematic structural diagram of a terminal provided in an embodiment of the present application;

[0136] Fig.29 It is a schematic structural diagram of a network device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0137] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0138] In the description of the present application, unless otherwise specified, " / " indicates that the objects associated before and after are in an "or" relationship, for example, A / B can represent A or B; "and / or" in the present application is only a kind of association relationship describing the associated objects, indicating that there can be three relationships, for example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone, where A and B can be singular or plural. In addition, in the description of the present application, unless otherwise specified, "multiple" refers to two or more than two. "At least one of the following" or its similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, c can be single or multiple. In addition, in order to facilitate the clear description of the technical solution of the embodiment of the present application, in the embodiment of the present application, the words "first" and "second" are used to distinguish the same or similar items with basically the same functions and effects. Those skilled in the art can understand that the words "first", "second", etc. do not limit the quantity and execution order, and the words "first", "second", etc. do not necessarily limit the differences.

[0139] In the various method embodiments of the present application, the size of the serial number does not mean the order of execution. The order of execution should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0140] It is understood that in this application, "under the circumstances of...", "if...", "when...", "if..." and similar descriptions can be used interchangeably. Moreover, these descriptions all refer to corresponding processing under certain objective circumstances, but do not limit the time, nor require judgment actions when implementing, nor do they mean the existence of other limitations.

[0141] It can be understood that in the present application, "greater than or equal to" can be replaced by "greater than", and correspondingly, "less than" can be replaced by "less than or equal to".

[0142] It can be understood that some optional features in the embodiments of the present application may be implemented independently in certain scenarios without relying on other features, such as the solution on which they are currently based, to solve corresponding technical problems and achieve corresponding effects, or may be combined with other features according to needs in certain scenarios. Accordingly, the devices provided in the embodiments of the present application may also realize these features or functions accordingly, which will not be elaborated here.

[0143] In this application, unless otherwise specified, the same or similar parts between the various embodiments can refer to each other. In the various embodiments in this application, and the various implementation methods / implementation methods / implementation methods in each embodiment, if there is no special explanation and logical conflict, the terms and / or descriptions between different embodiments and the various implementation methods / implementation methods / implementation methods in each embodiment are consistent and can be referenced to each other. The technical features in different embodiments and the various implementation methods / implementation methods / implementation methods in each embodiment can be combined to form new embodiments, implementation methods, implementation methods, or implementation methods according to their inherent logical relationships. The implementation methods of this application described below do not constitute a limitation on the scope of protection of this application.

[0144] The following is a brief introduction to some terms or concepts involved in this application.

[0145] 1. Artificial intelligence model: The artificial intelligence model is essentially a functional relationship, and this function does not have a corresponding analytical expression, but is expressed by a nested combination of multiple linear or nonlinear operations. Among them, the artificial intelligence model can be expressed as y=f(x), where x represents the model input, y represents the model output, and f represents the function composed of the model structure and parameters. For example, the expression of a simple artificial intelligence model - a single-layer fully connected neural network model is

[0146] y=σ(Wx+b)

[0147] Where σ is the activation function. For example, σ can be the hyperbolic tangent function, i.e., tanh. tanh(x)=[exp(x)-exp(-x)] / [exp(x)+exp(-x)]. If the input of tanh(x) is a vector, the tanh operation is performed element by element on each component of the vector. W and b are both trainable neural network parameters.

[0148] It should be noted that in the embodiments of the present application, the model structure may also be referred to as a structure, or may also be referred to as a model structure parameter or a structure parameter. The "parameter" in "model structure and parameters" refers to a weight parameter, i.e., the value of W and b in the above example.

[0149] 2. Training data set: refers to a set of data samples and labels corresponding to the data samples, and the samples in the set are used for model training. Let each sample in the set be (x, y'), where x represents the data sample and y' represents the label. For example, for image classification problems, x is a picture represented by a matrix or vector, and y' is the category corresponding to each picture.

[0150] It should be noted that in the embodiments of the present application, data samples are sometimes also referred to as sample data, and the labels corresponding to the data samples are also referred to as sample labels.

[0151] 3. Model Training: Given a training dataset, each sample (x, y’) in the set is used to optimize the parameters in the neural network. For example, the way to train a single-layer fully connected neural network y = tanh(Wx + b) is to optimize the trainable parameters W and b based on the training dataset using the gradient descent method. When the optimization algorithm converges, the training process is considered to be completed.

[0152] 4. Model Inference: Given the data x, it refers to the process of inputting the data x into the artificial intelligence model to obtain the output y = f(x).

[0153] 5. Validation Dataset: It can also be called a test / validation dataset. It has the same sample features as the training dataset but different uses. The samples in the validation dataset are not used to optimize the parameters in the artificial intelligence model, but to judge the performance of the current model. For example, let each sample in the validation dataset be (x, y’), and the model inference result be y = f(x). By comparing the difference between y and y’, the performance of the current model can be judged. For example, use ||.|| to represent the sum of the absolute values of each element of the vector (i.e., the L1 norm. Let (x1, x2,..., xn) be a vector, then

[0154] ||(x1, x2,..., xn)|| = |x1| + |x2| +... + |xn|, where |.| represents the absolute value symbol), and a judgment threshold th is given. If ||y - y’|| < th, it is considered that the inference result of the current model has a small difference from the true sample label and the performance meets the expectation; if ||y - y’|| > th, it is considered that the inference result of the current model has a large difference from the true sample label and the performance is lower than the expectation.

[0155] 6. Time Unit: It refers to the unit time required to transmit a signal of unit length, including but not limited to concepts such as micro-slots, time slots, sub-frames, or frames. The length of 1 time unit can be 100 microseconds, 1 millisecond, 1 second, 10 seconds, etc.

[0156] The embodiments of the present application can be applied to various communication systems, such as: Long Term Evolution (LTE) systems, 5th Generation (5G) mobile communication systems, New Radio (NR), and other future possible mobile communication systems, etc.

[0157] Figure 1 It is a schematic diagram of the architecture of a mobile communication system provided by the embodiments of the present application. As Figure 1 shown, the system 100 includes at least one terminal (for example, Figure 1 the terminal 110 shown) and at least one network device (for example, Figure 1The terminal 110 can communicate with the network device 120 via a wireless connection or an air interface. It should be understood that Figure 1 The system shown may also include more terminals and network devices.

[0158] The terminal in the embodiment of the present application, such as the terminal 110, is also referred to as user equipment (UE), terminal equipment, mobile station (MS), mobile terminal (MT), etc., which refers to a device that provides voice and / or data connectivity to a user. For example, the terminal can be a mobile phone, a tablet computer, a laptop computer, a PDA, a mobile internet device (MID), a wearable device, a virtual reality (VR) device, an augmented reality (AR) device, a wireless terminal in industrial control, a wireless terminal in self driving, a wireless terminal in remote medical surgery, a wireless terminal in smart grid, a wireless terminal in transportation safety, a wireless terminal in smart city, a wireless terminal in smart home, etc.

[0159] The network device in the embodiment of the present application, such as the network device 120, refers to a radio access network (RAN) node (or device) that connects the terminal to the wireless network, which can also be called a base station or access network device. For example, the network device can be an evolved NodeB (eNodeB), a transmission reception point (TRP), a next generation NodeB (gNB) in a 5G mobile communication system, a base station in a future mobile communication system or an access node (AP) in a WiFi system, a wireless controller in a cloud radio access network (CRAN) scenario, a relay station, an access point, a vehicle-mounted device, a wearable device, and network devices in other communication systems that will evolve in the future.

[0160] In one possible scenario, multiple RAN nodes collaborate to assist the terminal in achieving wireless access, and different RAN nodes respectively implement part of the functions of the base station. For example, the RAN node (i.e., the network device in this application) may be a centralized unit (CU), a distributed unit (DU), a CU-control plane (CP), a CU-user plane (UP), or a radio unit (RU), etc. The CU and DU may be set separately, or may be included in the same network element, such as a baseband unit (BBU). The RU may be included in a radio frequency device or a radio frequency unit, such as a remote radio unit (RRU), an active antenna unit (AAU), or a remote radio head (RRH). In different systems, CU (or CU-CP and CU-UP), DU or RU may also have different names, but those skilled in the art may understand their meanings. For example, in an open radio access network (ORAN) system, CU may also be referred to as O-CU (open CU), DU may also be referred to as O-DU, CU-CP may also be referred to as O-CU-CP, CU-UP may also be referred to as O-CU-UP, and RU may also be referred to as O-RU. Any unit in the CU (or CU-CP, CU-UP), DU and RU in this application may be implemented by a software module, a hardware module, or a combination of a software module and a hardware module. It should be understood that this application does not limit the specific technology and specific device form adopted by the network device.

[0161] In current technology, Figure 1 The structure and parameters of the air interface neural network model of the system 100 shown can be determined by offline training, wherein the data used for offline training is predefined or collected data. Specifically, if the neural network model needs to be deployed on the network device 120 side, offline training is performed on the network device 120 side; if the neural network model is to be deployed on the terminal 110 side, offline training is performed on the terminal 110 side; if the neural network model is to be deployed on both the network device 120 and the terminal 110 side, assuming that the model on the network device 120 side is model A and the model on the terminal 110 side is model B, then model A and model B are simultaneously trained offline on the network device 120 side, and model B is sent to the terminal 110 side through the air interface for deployment, or model A and model B are simultaneously trained offline on the terminal 110 side, and model A is sent to the network device 120 side through the air interface for deployment.

[0162] Since offline training takes a relatively long time, the update frequency of the parameters and structure of the neural network model is much lower than the change frequency of the air interface signal. When the scenario matched by the training data set of the air interface neural network model (for example, the delay spread parameters of the channel, the interference intensity of the neighboring area, the moving speed of the terminal 110, etc.) changes, using the current neural network model for signal processing in the new scenario may result in performance degradation.

[0163] In view of this, the present application provides a communication method, in which the terminal can verify at least one neural network model based on a verification data set, and then determine the performance of the at least one neural network model, so as to determine whether the at least one neural network model matches the current scene or determine the neural network model with the best performance among the at least one neural network model, so as to use the neural network model that matches the current scene or the neural network model with the best performance for signal processing. In addition, the terminal can return the verification result to the network device, so that the network device can also know the performance of the at least one neural network model or determine the performance of the at least one neural network model based on the verification result, so that the network device can perform corresponding processing according to the feedback of the terminal, such as using the neural network model corresponding to the neural network model with the best performance for signal processing. The method provided by the present application helps to solve the problem of performance degradation that may be caused by using a neural network model that does not match the current scene.

[0164] The method provided by the present application is described in detail below in conjunction with the relevant drawings. It can be understood that the flowchart provided by the present application mainly uses the terminal and the network device as the execution subject of the interactive diagram as an example to illustrate the method, but the present application does not limit the execution subject of the interactive diagram. For example, the terminal / network device in the flowchart can also be a chip, a chip system, or a processor that supports the terminal / network device to implement the method, or a logic module or software that can implement all or part of the terminal / network device functions. In addition, the meaning of the concepts introduced above will not be repeated below.

[0165] Figure 2 2 is a schematic flow chart of a communication method provided in an embodiment of the present application. In the method 200, a second neural network model is deployed on the network device side. The method 200 may include one or more steps from S210 to S230, and each step of the method 200 is described below.

[0166] S210: The terminal obtains a first data set.

[0167] The first data set may include sample data and labels corresponding to the sample data (i.e., sample labels). The sample labels are known to the terminal and the network device. Exemplarily, the sample labels may be reference signals, such as a demodulation reference signal (DMRS), a channel sounding reference signal (SRS), a channel state information reference signal (CSI-RS), a positioning reference signal (PRS), or a reference signal that has not yet been defined in the current protocol.

[0168] The terminal can obtain the sample data and / or sample labels in the first data set locally. Exemplarily, the sample data and the sample label can be historical data received by the terminal side from the network side; or, the sample data can be historical data received by the terminal side from the network side, and the sample label is a sample label calculated by the terminal according to a predefined rule, such as a sample label for a certain type of reference signal defined by the standard. Exemplarily, the sample data can also be predefined, and the sample data can also be used as a sample label. For example, the sample data is a channel sample generated according to a channel model of a typical 3GPP scenario (for example, urban macrocell Uma, urban microcell Umi, etc.), and the channel sample is also a sample label.

[0169] The terminal may also obtain the sample data and / or sample label in the first data set from other communication devices (e.g., another terminal or network device). For example, the network device may send the sample data and the sample label to the terminal; or, the network device may send the sample data and indication information to the terminal, the indication information indicating the sample label, for example, the indication information indicating that the sample label is a certain type of reference signal defined by the standard.

[0170] Exemplarily, the network device may send the sample tag or the indication information through a control channel, or may send the sample tag or the indication information through a channel dedicated to carrying the sample tag.

[0171] Exemplarily, the network device may send the sample data through a data channel, or may send the sample data through a channel dedicated to carrying the sample data.

[0172] The first data set is used to verify at least one first neural network model.

[0173] Exemplarily, the at least one first neural network model is pre-configured on the terminal side. Alternatively, the model information of the at least one first neural network model is pre-configured on the terminal side. Alternatively, the model information of the at least one first neural network model is sent by the network device to the terminal. The terminal can obtain each first neural network model based on the model information of each first neural network model. Among them, the model information includes model structure and parameters. Exemplarily, for a fully connected neural network model including multiple hidden layers, the model structure refers to hyperparameters such as the number of hidden layers, the dimensions of the input and output nodes of each hidden layer, and the model parameters refer to the weight parameters of each neural network node in each hidden layer.

[0174] Exemplarily, the first neural network model is trained based on the second neural network model and the second data set.

[0175] In some embodiments, the at least one first neural network model corresponds to the same second neural network model. That is, by using the same second neural network model and at least one second data set for model training, at least one first neural network model corresponding to at least one second data set can be obtained. For example, at least one second data set can be a channel (or channel model) that obeys different statistical distributions, the second neural network model can be used for signal modulation, and the first neural network model can be used for signal demodulation.

[0176] In some other embodiments, the second neural network model corresponding to the at least one first neural network model is different. That is, by using at least one second neural network model and at least one second data set for model training, a first neural network model corresponding to at least one second neural network model and at least one second data set can be obtained, wherein one second neural network model and one second data set correspond to only one first neural network model. For example, at least one second data set may be a channel subject to different statistical distributions, the second neural network model may be used for precoding vector compression, and the first neural network model may be used for precoding vector decompression.

[0177] Furthermore, the at least one first neural network model may be obtained by training based on at least one second data set that obeys a different statistical distribution, or may be obtained by training based on at least one different second data set that obeys the same statistical distribution.

[0178] In one example, the first neural network model is assumed to be a demodulator model, that is, the first neural network model is used for channel demodulation, the at least one first neural network model includes two models g1 and g2, the at least one second data set includes two data sets H1 and H2, the samples in the data sets are channel vectors or matrices, the channels in H1 obey the statistical distribution p1, and the channels in H2 obey the statistical distribution p2, then H1 can be used to train g1, and H2 can be used to train g2. Based on this solution, when the channel statistical distribution changes, the neural network model parameters can be matched with the current channel statistical distribution.

[0179] In another example, the first neural network model is assumed to be a demodulator model, that is, the first neural network model is used for channel demodulation, the at least one first neural network model includes two models g1 and g2, the second data set includes two data sets H1 and H2, the samples in the data sets are all channels, at least one channel sample in H1 and H2 has a different specific value, but the channels in the data sets all obey the statistical distribution p, then H1 can be used to train g1, and H2 can be used to train g2. Based on this scheme, the number of samples of H1 and H2 may be different, and the scale of the training data set can be adaptively adjusted to avoid overfitting or underfitting of the neural network model and improve the model reasoning performance.

[0180] For example, different statistical distributions can be distinguished by different distribution density functions. When the distribution density functions are the same, they can be distinguished by different distribution parameters. For example, suppose the data in the first data set obeys a uniform distribution, and the distribution density function is p1(x)=1 / 2, and the data in the second data set obeys a Gaussian distribution, and the distribution density function is Where exp represents an exponential function, then when p1 and p2 are not equal, or the distance D = ∫|p1(x)-p2(x)|dx between p1 and p2 in the function space is greater than a given threshold (e.g., 0.1), the two data sets are considered to have different distributions, where |x| represents the absolute value of x. For another example, suppose that the data in the first data set and the second data set both obey a Gaussian distribution, with distribution densities of and The means are m1 and m2, and the variances are v1 and v2 respectively. When the difference between the two distributions is large enough, for example, |m1-m2|>th or |v1-v2|>th or the distance D = ∫ |p1(x)-p2(x)|dx between p1 and p2 in the function space is greater than a given threshold th (for example, th = 0.1), it can be considered that the two data sets obey different distributions.

[0181] In a possible implementation, the method may further include:

[0182] S220, the terminal verifies the at least one first neural network model through the first data set and the at least one first neural network model to obtain a verification result.

[0183] Alternatively, S220 can be described as: verifying the at least one first neural network model through the first data set to obtain a verification result.

[0184] Specifically, the first data set can be input into the at least one first neural network model to obtain the output result of the first neural network model. This operation is actually to use the at least one first neural network model to perform model inference on the sample data in the first data set.

[0185] In one example, after obtaining the output result of the at least one first neural network model, the terminal may further calculate the gap between the output result of the at least one first neural network model and the sample labels in the first data set to obtain at least one gap. For example, the gap may be the Euclidean distance between the output result of the first neural network model and the sample labels in the first data set.

[0186] In one example, after obtaining the at least one gap, the at least one gap may be compared with a preset threshold to obtain a comparison result.

[0187] In one example, after obtaining the comparison result, the terminal may also determine a first neural network model whose corresponding gap in the at least one first neural network model is smaller than a preset threshold. For ease of description, any first neural network model whose corresponding gap in the at least one first neural network model is smaller than a preset threshold is hereinafter referred to as a target neural network model.

[0188] In one example, the terminal may also determine a first neural network model or a target neural network model that minimizes the corresponding gap.

[0189] In one example, the terminal may also determine the average (may also be a calculation method such as total number, maximum, minimum, etc.) quantization bit number of some or all first neural network model parameters or target neural network model parameters in the at least one first neural network model. The first neural network model parameters / target neural network model parameters include the structure and parameters of the first neural network model / target neural network model.

[0190] In one example, the terminal may also determine the neural network model with the least number of average quantization bits (it may also be calculated by total number, maximum, minimum, etc.) of parameters in the at least one first neural network model or multiple target neural network models.

[0191] In one example, the terminal may also determine the inference delay of some or all of the first neural network models or some or all of the target neural network models in the at least one first neural network model.

[0192] In one example, the terminal may also determine a first neural network model or a target neural network model with the smallest inference delay.

[0193] In some embodiments, the operations described in the examples described in S220 after obtaining the gaps between the output results of the at least one first neural network model and the sample labels in the first data set can also be considered as part of the verification of the at least one first neural network model.

[0194] In some embodiments, the verification result may indicate one or more of the following described above: part or all of the at least one gap (for example, it may be the gap corresponding to the target neural network model), a comparison result (which may be the comparison result corresponding to part or all of the at least one first neural network model), part or all of the target neural network models, the first neural network model or target neural network model that minimizes the corresponding gap, the average quantization bit number of the first neural network model parameters or the target neural network model parameters, the neural network model with the least average quantization bit number of the parameters in the at least one first neural network model or multiple target neural network models (it may also be calculated by the total number, maximum, minimum, etc.), the inference delay of part or all of the first neural network models or part or all of the target neural network models in the at least one first neural network model, and the first neural network model or target neural network model with the least inference delay.

[0195] Exemplarily, the preset threshold in the embodiment of the present application can be configured by the network device to the terminal.

[0196] S230, the terminal sends feedback information to the network device. Correspondingly, the network device receives the feedback information from the terminal.

[0197] Exemplarily, the feedback information can be sent from the terminal to the network device via an uplink control channel, can be sent from the terminal to the network device via an uplink data channel, or can be sent from the terminal to the network device via a channel dedicated to sending neural network model performance indication information.

[0198] The feedback information is used to indicate or determine the performance of the at least one first neural network model.

[0199] In one example, the feedback information may be the verification result obtained in S220. For example, the feedback information may indicate one or more of the following: part or all of the gaps in the at least one gap (for example, it may be the gap corresponding to the target neural network model), a comparison result (it may be the comparison result corresponding to part or all of the first neural network models in the at least one first neural network model), part or all of the target neural network models, the first neural network model or target neural network model that minimizes the corresponding gap, the average quantization bit number of the first neural network model parameter or the target neural network model parameter, the neural network model with the least average quantization bit number of the parameters in the at least one first neural network model or multiple target neural network models (it may also be calculated by the total number, maximum, minimum, etc.), the inference delay of part or all of the first neural network models or part or all of the target neural network models in the at least one first neural network model, and the first neural network model or target neural network model with the least inference delay.

[0200] For example, the at least one first neural network model includes model 1 and model 2, and the feedback information indicates that the gap corresponding to model 1 (i.e., the gap between the output result obtained by model inference on model 1 and the sample label in the first data set) is smaller than a preset threshold (i.e., model 1 is the target first neural network model), the inference delay of model 1 is smaller than a given threshold, and the gap corresponding to model 2 is greater than the preset threshold.

[0201] For example, the at least one first neural network model includes model 1 and model 2, and the feedback information indicates that the gaps corresponding to model 1 and model 2 are both smaller than a preset threshold.

[0202] For example, the at least one first neural network model includes model 1 and model 2, and the feedback information indicates that the gap corresponding to model 1 is smaller than a preset threshold, and the gap corresponding to model 2 is larger than the preset threshold.

[0203] In one example, the feedback information may be part or all of the at least one gap obtained in S220. Furthermore, the operations performed by the terminal described above after obtaining the at least one gap may also be performed by the network device. For example, the network device may determine one or more of the following based on part or all of the at least one gap: a comparison result (which may be a comparison result corresponding to part or all of the first neural network models in the at least one first neural network model), part or all of the target neural network models, a first neural network model or a target neural network model that minimizes the corresponding gap, an average quantization bit number of a first neural network model parameter or a target neural network model parameter, a neural network model with the least average quantization bit number of parameters in the at least one first neural network model or multiple target neural network models (which may also be calculated by the total number, maximum, minimum, etc.), an inference delay of part or all of the first neural network models or part or all of the target neural network models in the at least one first neural network model, and a first neural network model or a target neural network model with the least inference delay.

[0204] In one example, the feedback information is the output result of some or all of the first neural network models in the at least one first neural network model obtained in S220. Based on the output result, the network device can determine one or more of the following: the at least one gap, the comparison result (which can be the comparison result corresponding to some or all of the first neural network models in the at least one first neural network model), some or all of the target neural network models, the first neural network model or target neural network model that minimizes the corresponding gap, the average quantization bit number of the first neural network model parameter or the target neural network model parameter, the neural network model with the least average quantization bit number of the parameters in the at least one first neural network model or multiple target neural network models (which can also be calculated by the total number, maximum, minimum, etc.), the reasoning delay of some or all of the first neural network models or some or all of the target neural network models in the at least one first neural network model, and the first neural network model or target neural network model with the smallest reasoning delay.

[0205] In one example, the gap between the output result of the first neural network model and the sample label in the first data set can be used as an evaluation indicator of model performance. For example, the gap is the Euclidean distance described above, then the smaller the gap is, the better the performance of the corresponding first neural network model is, and the higher the degree of matching with the scene is.

[0206] In one example, the function value of the function of the independent variable, which is the gap between the output result of the first neural network model and the sample label in the first data set, can be used as an evaluation index of model performance. Taking the gap as the Euclidean distance described above as an example, for example, the function can be the reciprocal of the gap, or the function can be the opposite of the gap. Then, the larger the function value, the better the performance of the corresponding first neural network model, and the higher the degree of matching with the scene.

[0207] In one example, the average (or the total number, maximum, minimum, etc.) quantization bit number of the neural network model parameters can be used as an evaluation index of the model performance. It should be understood that the smaller the bit number, the lower the overhead required to store the first neural network model parameters, and the better the performance of the first neural network model.

[0208] In one example, the inference delay of the neural network model can be used as an evaluation index of model performance. It should be understood that the smaller the inference delay, the better the performance of the first neural network model and the higher the degree of matching with the scene.

[0209] It should be understood that some or all of the multiple evaluation indicators described above may also be used to comprehensively evaluate the performance of the first neural network model.

[0210] According to the communication method provided in the present application, the terminal can verify at least one first neural network model based on the first data set, and then determine the performance of the at least one first neural network model, so as to determine whether the at least one first neural network model matches the scene or determine the neural network model with the best performance among the at least one first neural network model. Furthermore, the terminal can deploy the first neural network model with the best performance at the physical layer, perform signal processing based on the first neural network model, and ensure the signal processing performance as much as possible. On the other hand, the terminal can indicate the performance of the at least one first neural network model to the network device, so that the network device can also perform corresponding operations based on the feedback of the terminal, and avoid the performance degradation problem that may be caused by the use of a neural network model that does not match the current scene by itself or the terminal side as much as possible. For example, the network device uses the second neural network model corresponding to the first neural network model with the best performance for signal processing.

[0211] In a possible implementation, the first time unit in which the terminal sends the feedback information is the same as the first time unit in which the terminal obtains part or all of the first data set from the network device. That is, the feedback information and part or all of the first data set are in the same first time unit.

[0212] For example, see Figure 3, the time unit in which the network device sends part or all of the first data set to the terminal and the time unit in which the terminal sends feedback information are the same time unit. It should be understood that Figure 3 The time unit shown in is the first time unit.

[0213] Among them, part of the first data set refers to the sample data in the first data set. In the scenario where the terminal only obtains the sample data from the network device but does not obtain the sample label, it is considered that the terminal side can confirm the sample label by itself, such as the terminal side pre-stores the sample label or can calculate the sample label according to a pre-defined rule. The entire first data set refers to the sample data and sample labels in the first data set.

[0214] By transmitting feedback information and part or all of the first data set in the same first time unit, timely transmission of feedback information corresponding to the first data set can be achieved, thereby ensuring as much as possible that the feedback information on the terminal side can indicate the performance of each first neural network model in the current scenario.

[0215] Exemplarily, the first time unit in the embodiment of the present application may be a mini-time slot, a time slot, a subframe or a frame, or a time unit representing a smaller or larger time length that has not yet been defined in the protocol.

[0216] In a possible implementation, the second time unit in which part or all of the first data set is located is the same as the second time unit in which the model information of the at least one first neural network model is located. That is, the network device will send the model information of the at least one first neural network model and part or all of the first data set to the terminal in the same second time unit.

[0217] For example, see Figure 4 , the time unit in which the network device sends part or all of the first data set to the terminal and the time unit in which the model information is sent are the same time unit. It should be understood that Figure 4 The time unit shown in is the second time unit.

[0218] By sending the model information of at least one first neural network model and part or all of the first data set on the same second time unit, the network device side can ensure as much as possible that the model verification result on the terminal side can indicate the performance of each first neural network model in the current scenario.

[0219] Exemplarily, the second time unit in the embodiment of the present application may be a micro-time slot, a time slot, a subframe or a frame, or a time unit representing a smaller or larger time length that has not yet been defined in the protocol.

[0220] In one possible implementation, the start time of the symbol carrying the model information does not lag behind the start time of the symbol carrying part or all of the first data set, that is, the time domain start position of the symbol of the model information does not lag behind the time domain start position of the symbol of the first data set.

[0221] For example, the symbol time domain starting position of the model information may be before the symbol time domain starting position of the first data set, or the symbol time domain starting position of the model information may be the same as the symbol time domain starting position of the first data set.

[0222] For example, the temporal position relationship between part or all of the first data set and the model information may be Figure 5 Any of the positional relationships shown.

[0223] Since the symbol time domain starting position of the model information does not lag behind the symbol time domain starting position of the first data set, the terminal will receive the model information first and then the first data set. In this way, the at least one first neural network model can be verified based on the first data set without storing the first data set, thereby reducing cache overhead.

[0224] In one possible implementation, the third time unit where part or all of the first data set is located, the third time unit where the model information of the at least one first neural network model is located, and the third time unit where the feedback information is located are the same.

[0225] For example, see Figure 6 , the third time unit where part or all of the first data set is located, the third time unit where the model information of the at least one first neural network model is located, and the feedback information are located in the same time unit. It should be understood that Figure 6 The time unit shown in is the second time unit.

[0226] By carrying the model information of at least one first neural network model, part or all of the first data set and feedback information on the same third time unit, it is possible to ensure as much as possible that the model verification result on the terminal side can indicate the performance of each first neural network model in the current scenario.

[0227] Exemplarily, the third time unit in the embodiment of the present application may be a micro-time slot, a time slot, a subframe or a frame, or a time unit representing a smaller or larger time length that has not yet been defined in the protocol.

[0228] It should be noted that part or all of the first data set, the model information of the at least one first neural network model, and the feedback information may not be in the same third time unit. For example, part or all of the first data set, the model information of the at least one first neural network model, and the feedback information may be located in different third time units. Alternatively, at least one of part or all of the first data set, the model information of the at least one first neural network model, and the feedback information may occupy several symbols in at least two adjacent time units. For example, Figure 7 The example shows that part or all of the first data set, the model information of the at least one first neural network model, and the feedback information are not in the same time unit. It should be understood that Figure 7 The time unit shown in is the third time unit.

[0229] As mentioned above, the sample label in the first data set or the indication information indicating the sample label can be sent through the control channel. In one possible implementation, the fourth time unit where the control channel is located is the same as the fourth time unit where one or more of the following are located: the sample data in the first data set, the model information of at least one first neural network model, or the feedback information.

[0230] For example, the control channel includes indication information, and the control channel can be located in (or carried on) the same fourth time unit as the sample data in the first data set, the model information, and the feedback information. For example, the control channel includes sample labels, the sample data in the first data set can be sent through the data channel, and the control channel can be located in the same fourth time unit as the data channel and the model information of at least one first neural network model. For example, the control channel can be located in the same fourth time unit as the feedback information.

[0231] For example, Figure 8 In the example shown, the control channel may be located in the same time unit as the sample data in the first data set, the model information and the feedback information, wherein the control channel includes indication information. It should be understood that Figure 8 The time unit shown in is the fourth time unit.

[0232] Exemplarily, the fourth time unit in the embodiment of the present application may be a micro-time slot, a time slot, a subframe or a frame, or a time unit representing a smaller or larger time length that has not yet been defined in the protocol.

[0233] It should be noted that the control channel may not be in the same fourth time unit as one or more of the following: sample data in the first data set, model information or feedback information of at least one first neural network model. For example, the control channel may be in a different fourth time unit than one or more of the following: sample data in the first data set, model information or feedback information of at least one first neural network model. For example, see Fig. 9 , the control channel, the sample data in the first data set, the model information and the feedback information are located in different time units. Alternatively, the control channel is in the first time unit, the model information is in the second time unit, the sample data in the first data set occupies several symbols in the second time unit and several symbols in the third time unit, and the feedback information is in the fourth time unit. It should be understood that Fig. 9 The time unit shown in is the fourth time unit.

[0234] In a possible implementation, any information involved in method 200 (eg, part or all of the first data set, model information, or feedback information, etc.), such as Figures 3 to 9 Any information shown may occupy one or more consecutive symbols in one time unit, or one or more consecutive symbols in multiple time units, which is not limited in the present application. For example, the symbol in the embodiment of the present application may be an OFDM symbol.

[0235] It should be understood that Figures 3 to 9 The positional relationship of each information in the time domain and the frequency domain shown is only some examples, and these examples should not constitute any limitation to the present application. The positional relationship of each information in the time domain and the frequency domain involved in the method 200 of the present application can also be Figures 3 to 9 The positional relationships shown are different.

[0236] Figure 2 The method 200 shown can be applied in a variety of scenarios, for example, Fig.10 , Fig.12 as well as Fig.14 The scene shown.

[0237] Fig.10 is a schematic diagram of an application scenario provided by an embodiment of the present application. Fig.10, neural network models can be deployed on both the network equipment and the terminal side, and the traditional modulation and demodulation can be replaced by neural network models to realize modulation and demodulation based on neural network models. Among them, when the modulation order is N, the input of the modulation module is a string of bits b0b1b2b3…bN-1, bi={0,1}, and the output is the modulation symbol x, where x is a complex number, so the modulation module can be expressed as x=f(b0b1b2b3…bN-1). For example, when N=4, x may be any one of the complex number set {1+1j,1-1j,-1+1j,-1-1j} (where j represents a complex unit, which will not be repeated later), f(0000)=-1-1j. The input of the demodulation module is a complex number z, and the output is a string of bits b0b1b2b3…bN-1, so the input-output relationship of the demodulation module can be expressed as b0b1b2b3…bN-1=g(z). Since the input of the demodulation module is generally the result of the output of the modulation module after passing through the channel, that is, z = h*x + n, n is noise, it can be known that b0b1b2b3 ... bN-1 = g(h*f(b0b1b2b3 ... bN-1) + n). For example, when N = 4, b0b1b2b3 = 0000, x = f(0000) = -1-1j, h = 2, n = 0.1 + 0.1j, then z = -1.9-1.9j, g(-1.9-1.9j) = 0000. The structure or parameters of f and g are related to the statistical characteristics of the current channel h and n.

[0238] Fig.11 is a schematic flow chart of a communication method provided by the present application. The method 600 shows Fig.10 The scenario shown uses Figure 2 An example of the method shown.

[0239] S601, the network device determines at least one first neural network model according to at least one second data set and a second neural network model f. For ease of understanding, the embodiment of the present application is described by taking at least one first neural network model as g1 and g2 as an example.

[0240] Among them, f is used for signal modulation, and its structure and parameters can be determined by the network equipment. g1 and g2 can be used for signal demodulation.

[0241] In one example, the second data set is (b0b1b2b3…bN-1, (h, b0b1b2b3…bN-1), where b0b1b2b3…bN-1 is a sample label, and (h, b0b1b2b3…bN-1) is sample data. The data samples in at least one second data set, i.e., h and / or b0b1b2b3…bN-1, have different distributions. Since b0b1b2b3…bN-1=g(h*f(b0b1b2b3…bN-1)+n), when f, h and b0b1b2b3…bN-1 are known, the parameters in g can be trained. For example, when N=4, for a certain sample data, b0b1b2b3=0 01, h=1-1j, f(0001)=-1-1j, n=0.1+0.1j, then g(-1.9+0.1j)=0001. There are multiple sample data as shown above in each data set, and h in different sample data is different, so a statistical distribution can be used to describe h in different sample data in the same data set. Among them, the structure and parameters of g1 are obtained by training together with f when the distribution of h obeys the first statistical model (for example, the delay spread is t1), and the structure and parameters of g2 are obtained by training when the distribution of h obeys the second statistical model (for example, the delay spread is t2), wherein all or part of the distribution parameters of the first statistical model and the second statistical model are different.

[0242] In another example, the second data set is (b0b1b2b3…bN-1, (n, b0b1b2b3…bN-1)), where b0b1b2b3…bN-1 is a sample label and (n, b0b1b2b3…bN-1) is sample data). The distribution of data samples in at least one of the second data sets, i.e., n and / or b0b1b2b3…bN-1, is different. Since b0b1b2b3…bN-1=g(h*f(b0b1b2b3…bN-1)+n), when f, n, h and b0b1b2b3…bN-1 are known, the parameters in g can be trained. For example, when N=4, for a certain sample data, b0b1b2b3=1001, h=1-1j, f(1001)=-1-1j, n=0.1+0.1j, then g(-1.9+0.1j)=1001. There are multiple sample data as shown above in each data set, and n in different sample data is different, so a statistical distribution can be used to describe n in different sample data in the same data set. Among them, the structure and parameters of g1 are obtained by training together with f when the distribution of n obeys the first statistical model (for example, n obeys a complex Gaussian distribution with a mean of 0 and a variance of 1), and the structure and parameters of g2 are obtained by training when the distribution of n obeys the second statistical model (for example, n obeys a complex Gaussian distribution with a mean of 0 and a variance of 2), wherein all or part of the distribution parameters of the first statistical model and the second statistical model are different.

[0243] S602: The network device sends a first data set D and model information of g1 and g2 for model verification to the terminal.

[0244] Wherein, each sample in D is (b0b1b2b3…bN-1, f(b0b1b2b3…bN-1)), b0b1b2b3…bN-1 represents the sample label, and f(b0b1b2b3…bN-1) represents the sample data. In addition, if all sample labels are known on the terminal and the network device side, each sample in D only needs to include f(b0b1b2b3…bN-1); otherwise, the sample in D is (b0b1b2b3…bN-1, f(b0b1b2b3…bN-1)). For example, N=6, b0b1b2b3b4b5=001110, f(001110)=1+1j, then 1+1j represents the sample data, and 001110 represents the sample label.

[0245] Exemplarily, the sample data may be sent via a data channel or a channel not yet defined in an existing protocol for carrying sample data; the sample tag may be sent via a control channel or a channel not yet defined in an existing protocol for carrying sample tags.

[0246] Exemplarily, D and the model information of g1 and g2 correspond to the same time unit. That is, the signal sent in the same time unit includes D and the model information of g1 and g2. For example, the time unit can be a mini-time slot, a time slot, a subframe or a frame, or a time unit with a smaller or larger time length that has not yet been defined in the protocol.

[0247] Further, the signal includes multiple OFDM symbols, and the starting position of the OFDM symbol carrying D in the multiple OFDM symbols does not lag behind the starting position of the OFDM symbol carrying the model information of g1 and g2.

[0248] The model information of g1 and g2 includes the structure and parameters of g1 and g2.

[0249] S603, the terminal uses g1 and g2 to perform model inference process on the data in D respectively, and obtains g1(h*f(b0b1b2b3…bN-1)+n) and g2(h*f(b0b1b2b3…bN-1)+n).

[0250] That is, the terminal inputs the sample data f(b0b1b2b3…bN-1) in D into g1 and g2 respectively, and obtains the output g1(h*f(b0b1b2b3…bN-1)+n) of g1 and the output g2(h*f(b0b1b2b3…bN-1)+n) of g2. For example, N=4, b0b1b2b3=0011, f(b0b1b2b3…bN-1)=1+1j, h=2, n=0.1-0.1j, then g1(2.1+1.9j)=0011, g2(2.1+1.9j)=1011.

[0251] S604, the terminal compares the difference between g1(h*f(b0b1b2b3…bN-1)+n) and g2(h*f(b0b1b2b3…bN-1)+n) and the sample label, that is, the actual bit sequence b0b1b2b3…bN-1.

[0252] For example, the terminal calculates Li = ||gi(h*f(b0b1b2b3…bN-1)+n)-b0b1b2b3…bN-1||, and can obtain L1 and L2, where ||.|| represents the sum of the absolute values ​​of each element of the vector, and i = 1, 2. For example, N = 4, b0b1b2b3 = 0011, f(b0b1b2b3…bN-1) = 1+1j, h = 2, n = 0.1-0.1j, then g1(2.1+1.9j) = 0011, g2(2.1+1.9j) = 1011, and L1 = 0, L2 = 1.

[0253] For another example, the terminal calculates Li=1 / ||gi(h*f(b0b1b2b3…bN-1)+n)-b0b1b2b3…bN-1||, or the terminal calculates Li=-||gi(h*f(b0b1b2b3…bN-1)+n)-b0b1b2b3…bN-1||, and L1 and L2 can be obtained, where ||.|| represents the sum of the absolute values ​​of each element of the vector, and i=1, 2. For example, N=4, b0b1b2b3=0011, f(b0b1b2b3…bN-1)=1+1j, h=2, n=0.1-0.1j, then g1(2.1+1.9j)=0011, g2(2.1+1.9j)=1011, and L1=0, L2=-1 can be obtained.

[0254] S605: The terminal sends feedback information to the network device. Correspondingly, the network device receives the feedback information.

[0255] The feedback information includes L1 and L2 calculated by the terminal.

[0256] Exemplarily, the time unit in which the terminal sends the feedback information is the same as the time unit in which the network device sends the model information of D and / or g1, g2. For example, the time unit may be a mini-time slot, a time slot, a subframe or a frame, or a time unit with a smaller or larger time length that has not yet been defined in the protocol.

[0257] S606, the terminal selects the model gi that minimizes / maximizes Li as the currently used demodulation neural network model, and deploys it at its physical layer for signal demodulation.

[0258] Assume that the smaller Li is, the better the performance of the neural network model is. For example, Li = ||gi(h*f(b0b1b2b3…bN-1)+n)-b0b1b2b3…bN-1||, N = 4, b0b1b2b3 = 0011, f(b0b1b2b3…bN-1) = 1+1j, h = 2, n = 0.1-0.1j, then g1(2.1+1.9j) = 0011, g2(2.1+1.9j) = 1011, and L1 = 0, L2 = 1. Then, in S606, the terminal can select the model gi that minimizes Li (that is, the model g1 corresponding to L1) as the currently used demodulation neural network model deployed in its physical layer.

[0259] In a possible implementation, S606 is executed only when the minimum Li is less than a preset threshold. For example, N=4, b0b1b2b3=0011, f(b0b1b2b3…bN-1)=1+1j, h=2, n=0.1-0.1j, then g1(2.1+1.9j)=0011, g2(2.1+1.9j)=1011, and L1=0, L2=1 can be obtained, the minimum Li is L1=0, the preset threshold is 0.1, and the minimum Li is less than the preset threshold, and S606 can be executed.

[0260] In one possible implementation, if the minimum Li is greater than or equal to the preset threshold, the terminal can also indicate to the network device in the feedback information in S605 that all Li are greater than or equal to the preset threshold. For example, N = 4, b0b1b2b3 = 0011, f(b0b1b2b3…bN-1) = 1+1j, h = 2, n = 0.1-0.1j, then g1(2.1+1.9j) = 1111, g2(2.1+1.9j) = 1011, and L1 = 2, L2 = 1, the preset threshold is 0.1, and the reasoning results of all models are greater than or equal to the preset threshold. At this time, the terminal can send feedback information to notify the network device that all reasoning results are greater than or equal to the preset threshold. All network devices can resend a new neural network model to the terminal based on the received feedback information, restart the process of this solution, or instruct the terminal to turn off the physical layer AI function and fall back to non-AI mode.

[0261] Assuming that the larger Li is, the better the performance of the neural network model is, for example, Li = -||gi(h*f(b0b1b2b3…bN-1)+n)-b0b1b2b3…bN-1||, then, in S606, the terminal selects the model gi that makes Li the largest as the currently used demodulation neural network model deployed in its physical layer. For example, N = 4, b0b1b2b3 = 0011, f(b0b1b2b3…bN-1) = 1+1j, h = 2, n = 0.1-0.1j, then g1(2.1+1.9j) = 0011, g2(2.1+1.9j) = 1011, and L1 = 0, L2 = -1 can be obtained. At this time, S606 is executed and g1 is selected to be deployed in the physical layer.

[0262] In a possible implementation, S606 is executed only when the maximum Li is greater than a preset threshold. For example, N=4, b0b1b2b3=0011, f(b0b1b2b3…bN-1)=1+1j, h=2, n=0.1-0.1j, then g1(2.1+1.9j)=0011, g2(2.1+1.9j)=1011, L1=0, L2=-1, the preset threshold is -0.1, then S606 is executed and g1 is selected to be deployed at the physical layer.

[0263] In a possible implementation, if the maximum Li is less than or equal to the preset threshold, the terminal may also indicate to the network device in the feedback information in S605 that all Li are less than or equal to the preset threshold. Based on the received feedback information, the network device may resend a new neural network model to the terminal, restart the process of this solution, or instruct the terminal to turn off the physical layer AI function and fall back to the non-AI mode.

[0264] It should be understood that the present application does not limit the execution order between S605 and S606. S605 may be executed first and then S606, S606 may be executed first and then S605, or S605 and S606 may be executed simultaneously.

[0265] According to the communication method provided in the present application, the network device can send multiple demodulation neural network models (i.e., g1 and g2) and a data set D for model performance verification to the terminal. The terminal can use the data set D to select the model that best matches the statistical distribution parameters of the current channel (e.g., the parameter representing h or the parameter representing n) from multiple demodulation neural network models, thereby improving the reasoning performance of the neural network model. Furthermore, the terminal can also feedback the performance of each demodulation neural network model to the network device, and the network device can determine the performance of each demodulation neural network model based on the feedback information, so that when each demodulation neural network model does not meet the performance requirements, the demodulation neural network model can be reconfigured for the terminal or the terminal can be instructed to turn off the physical layer AI function.

[0266] Fig.12 is another application scenario diagram provided by the embodiment of the present application. Fig.12 , when there is power amplifier nonlinearity, a neural network model can be deployed on both the network equipment and the terminal side, and the neural network model can be used to replace the traditional modulation and demodulation to learn the nonlinear model of the power amplifier and improve the demodulation performance of the signal. Among them, the nonlinear model of the power amplifier can be expressed as x'=J(x), that is, when the ideal modulation symbol is x, the transmitted modulation symbol is x', then z=h*x'+n, and b0b1b2b3…bN-1=g(h*J(f(b0b1b2b3…bN-1))+n), at this time, the structure or parameters of f and g obtained by training are also related to the nonlinear model J(x). For the meaning of each parameter in this scenario, please refer to the above description. Fig.10 For example, N=4, b0b1b2b3=0001, x=f(0001)=1+1j, J(x)=x^2, x'=-2j, h=1, n=0.1+0.1j, then z=0.1-1.9j, g(0.1-1.9j)=0001.

[0267] Fig.13 is a schematic flow chart of a communication method provided by the present application. The method 800 shows Fig.12 The scenario shown uses Figure 2 An example of the method shown.

[0268] S801, the network device determines at least one first neural network model according to at least one second data set and a second neural network model f. For ease of understanding, the embodiment of the present application is described by taking at least one first neural network model as g1 and g2 as an example.

[0269] Among them, f is used for signal modulation, and its structure and parameters can be determined by the network equipment. g1 and g2 can be used for signal demodulation.

[0270] The second data set is (b0b1b2b3…bN-1, (J(f(b0b1b2b3…bN-1)), b0b1b2b3…bN-1)), where b0b1b2b3…bN-1 is a sample label, and ((J(f(b0b1b2b3…bN-1)), b0b1b2b3…bN-1) is sample data. The data samples in at least one of the second data sets, i.e., (J(f(b0b1b2b3…bN-1)) and / or b0b1b2b3…bN-1, have different distributions. Since b0b1b2b3…bN-1=g(h*J(f(b0b1b2b3…bN-1))+n), when J, h and b0b1b2b3…bN-1 are known, the parameters in g can be trained. For example, when N=4, for For a certain sample data, b0b1b2b3=1001, J(x)=x^2, h=2, f(1001)=-1-1j, n=0.1+0.1j, then g(0.1-1.9j)=1001. There are multiple sample data as shown above in each data set, and J and / or h in different sample data may be different, so a statistical distribution can be used to describe J and / or in different sample data in the same data set. Among them, the structure and parameters of g1 are obtained by training together with f when J obeys the first nonlinear model (for example, J(x)=x^2), and the structure and parameters of g2 are obtained by training when J obeys the second nonlinear model (for example, J(x)=2x^2), and all or part of the distribution parameters of the first nonlinear model and the second nonlinear model are different.

[0271] S802: The network device sends a first data set D and model information of g1 and g2 for model verification to the terminal.

[0272] Among them, each sample in D is (b0b1b2b3…bN-1, J(f(b0b1b2b3…bN-1)), b0b1b2b3…bN-1 represents the sample label, and J(f(b0b1b2b3…bN-1) represents the sample data. In addition, if all sample labels are known on the terminal and network device sides, each sample in D only needs to include J(f(b0b1b2b3…bN-1); otherwise, the sample in D is (b0b1b2b3…bN-1, J(f(b0b1b2b3…bN-1)). For example, when N=4, for a certain sample data, b0b1b2b3=1000, f(1000)=1-1j, J(x)=x^2, then a sample in D is (1000, -2j).

[0273] Exemplarily, the sample data may be sent via a data channel or a channel not yet defined in an existing protocol for carrying sample data; the sample tag may be sent via a control channel or a channel not yet defined in an existing protocol for carrying sample tags.

[0274] Exemplarily, D and the model information of g1 and g2 correspond to the same time unit. That is, the signal sent in the same time unit includes D and the model information of g1 and g2.

[0275] Further, the signal includes multiple OFDM symbols, and the starting position of the OFDM symbol carrying D in the multiple OFDM symbols does not lag behind the starting position of the OFDM symbol carrying the model information of g1 and g2.

[0276] The model information of g1 and g2 includes the structure and parameters of g1 and g2.

[0277] S803, the terminal uses g1 and g2 to perform model inference process on the data in D respectively, and obtains g1(h*J(f(b0b1b2b3…bN-1))+n) and g2(h*J(f(b0b1b2b3…bN-1))+n).

[0278] That is, the terminal inputs the sample data f(b0b1b2b3…bN-1) in D into g1 and g2 respectively, and obtains the output g1(h*J(f(b0b1b2b3…bN-1))+n) of g1 and the output g2(h*J(f(b0b1b2b3…bN-1))+n) of g2. For example, N=4, b0b1b2b3=1000, f(1000)=1-1j, J(x)=x^2, h=1, n=0.1-0.1j, then g1(0.1-2.1j)=1000, g2(0.1-2.1j)=1001.

[0279] S804, the terminal compares the difference between g1(h*J(f(b0b1b2b3…bN-1))+n) and g2(h*J(f(b0b1b2b3…bN-1))+n) and the sample label, that is, the actual bit sequence b0b1b2b3…bN-1.

[0280] For example, the terminal calculates Li = ||gi(h*J(f(b0b1b2b3…bN-1))+n)-b0b1b2b3…bN-1||, and L1 and L2 can be obtained, where ||.|| represents the absolute value summation sign of the vector, i = 1, 2. For example, when N = 4, b0b1b2b3 = 1000, f(1000) = 1-1j, J(x) = x^2, h = 1, n = 0.1-0.1j, then g1(0.1-2.1j) = 1000, g2(0.1-2.1j) = 1001, and L1 = 0, L2 = 1 can be obtained.

[0281] For another example, the terminal calculates Li = -||gi(h*J(f(b0b1b2b3…bN-1))+n)-b0b1b2b3…bN-1||, or the terminal calculates Li = -||gi(h*J(f(b0b1b2b3…bN-1))+n)-b0b1b2b3…bN-1||, and L1 and L2 can be obtained, where ||.|| represents It shows the sum of the absolute values ​​of each element of the vector i=1,2. For example, when N=4, b0b1b2b3=1000, f(1000)=1-1j, J(x)=x^2, h=1, n=0.1-0.1j, then g1(0.1-2.1j)=1000, g2(0.1-2.1j)=1001, and we can get L1=0, L2=-1.

[0282] S805: The terminal sends feedback information to the network device. Correspondingly, the network device receives the feedback information.

[0283] The feedback information includes L1 and L2 calculated by the terminal.

[0284] Exemplarily, the time unit in which the terminal sends the feedback information is the same as the time unit in which the network device sends the model information of D and / or g1, g2. For example, the time unit may be a mini-time slot, a time slot, a subframe or a frame, or a time unit with a smaller or larger time length that has not yet been defined in the protocol.

[0285] S806, the terminal selects the model gi that minimizes / maximizes Li as the currently used demodulation neural network model, and deploys it at its physical layer for signal demodulation.

[0286] Assuming that the smaller Li is, the better the performance of the neural network model is, for example, Li = ||gi(h*J(f(b0b1b2b3…bN-1))+n)-b0b1b2b3…bN-1||, then, in S806, the terminal selects the model gi that minimizes Li as the currently used demodulation neural network model deployed in its physical layer. For example, when N = 4, b0b1b2b3 = 1000, f(1000) = 1-1j, J(x) = x^2, h = 1, n = 0.1-0.1j, then g1(0.1-2.1j) = 1000, g2(0.1-2.1j) = 1001, at this time it can be obtained that L1 = 0, L2 = 1, and in S806, the model g1 corresponding to L1 is selected and deployed in the physical layer.

[0287] In a possible implementation, S806 is executed only when the minimum Li is less than a preset threshold. For example, when N=4, b0b1b2b3=1000, f(1000)=1-1j, J(x)=x^2, h=1, n=0.1-0.1j, then g1(0.1-2.1j)=1000, g2(0.1-2.1j)=1001, at this time, L1=0, L2=1, the preset threshold is 0.1, so in S806, the model g1 corresponding to L1 is selected and deployed at the physical layer.

[0288] In one possible implementation, if the minimum Li is greater than or equal to the preset threshold, the terminal can also indicate to the network device in the feedback information in S805 that all Li are greater than or equal to the preset threshold. For example, N = 4, b0b1b2b3 = 0011, f(b0b1b2b3…bN-1) = 1+1j, h = 2, n = 0.1-0.1j, then g1(2.1+1.9j) = 1111, g2(2.1+1.9j) = 1011, and L1 = 2, L2 = 1, the preset threshold is 0.1, and the reasoning results of all models are greater than or equal to the preset threshold. At this time, the terminal can send feedback information to notify the network device that all reasoning results are greater than or equal to the preset threshold. The network device can resend a new neural network model to the terminal based on the received feedback information, restart the process of this solution, or instruct the terminal to turn off the physical layer AI function and fall back to non-AI mode.

[0289] Assuming that the larger Li is, the better the performance of the neural network model is, for example, Li = -||gi(h*J(f(b0b1b2b3…bN-1))+n)-b0b1b2b3…bN-1||, then, in S806, the terminal selects the model gi that makes Li the largest as the currently used demodulation neural network model deployed in its physical layer. For example, when N = 4, b0b1b2b3 = 1000, f(1000) = 1-1j, J(x) = x^2, h = 1, n = 0.1-0.1j, then g1(0.1-2.1j) = 1000, g2(0.1-2.1j) = 1001, at this time it can be obtained that L1 = 0, L2 = -1, and in S806, the model g1 corresponding to L1 is selected and deployed in the physical layer.

[0290] In a possible implementation, S806 is executed only when the maximum Li is greater than a preset threshold. For example, when N=4, b0b1b2b3=1000, f(1000)=1-1j, J(x)=x^2, h=1, n=0.1-0.1j, then g1(0.1-2.1j)=1000, g2(0.1-2.1j)=1001, at this time, L1=0, L2=-1, the preset threshold is -0.1, so in S806, the model g1 corresponding to L1 is selected and deployed at the physical layer.

[0291] In a possible implementation, if the maximum Li is less than or equal to the preset threshold, the terminal can also indicate to the network device in the feedback information in S805 that all Li are less than or equal to the preset threshold. Based on the received feedback information, the network device can resend a new neural network model to the terminal, restart the process of this solution, or instruct the terminal to turn off the physical layer AI function and fall back to the non-AI mode.

[0292] It should be understood that the present application does not limit the execution order between S805 and S806. S805 may be executed first and then S806, S806 may be executed first and then S805, or S805 and S806 may be executed simultaneously.

[0293] According to the communication method provided in the present application, the network device can send multiple demodulation neural network models (i.e., g1 and g2) and a data set D for model performance verification to the terminal. The terminal can use the data set D to select the model that best matches the current power amplifier nonlinear model from multiple demodulation neural network models, thereby improving the reasoning performance of the neural network model. Furthermore, the terminal can also feedback the performance of each demodulation neural network model to the network device, and the network device can determine the performance of each demodulation neural network model based on the feedback information, so that when each demodulation neural network model does not meet the performance requirements, the demodulation neural network model can be reconfigured for the terminal or the terminal can be instructed to turn off the physical layer AI function.

[0294] Fig.14 is another application scenario diagram provided by the embodiment of the present application. Fig.14 , the neural network model can be deployed on both the network device and the terminal side to realize the scenario of precoding vector indication compression, that is, the precoding vector is compressed by the neural network model on the network device side and transmitted to the terminal side, which can achieve the effect of reducing the indication overhead. Among them, the precoding vector to be compressed by the neural network model on the network device side is Vin, the compressed result is Vq, and the precoding vector to be decompressed by the neural network model on the terminal side is Vout. The neural network model on the network device side is f, and the neural network model on the terminal side is g, then Vq = f(Vin), Vout = g(Vq).

[0295] Fig.15 is a schematic flow chart of a communication method provided by the present application. The method 1000 shows Fig.14 The scenario shown uses Figure 2 An example of the method shown.

[0296] S1001, the network device determines at least one first neural network model based on at least one second data set and at least one second neural network model.

[0297] For ease of understanding, in the embodiments of the present application, at least one second neural network model is f1 and f2, and at least one first neural network model is g1 and g2.

[0298] Among them, f1 and f2 can be used for precoding vector compression, and g1 and g2 can be used for precoding vector decompression. The structure and parameters of g1 are obtained by training together with f1 when the distribution of h obeys the first statistical model (for example, the LOS channel model), and the structure and parameters of g2 are obtained by training together with f2 when the distribution of h obeys the second statistical model (for example, the NLOS channel model). All or part of the distribution parameters of the first statistical model and the second statistical model are different.

[0299] An example of the above training process is that, assuming h is given, Vin=Vout=V is obtained based on h (for example, V can be obtained by performing SVD decomposition on h and calculating a feature vector based on the decomposition result). Ideally, V=g(f(V)+n), where n is the random noise introduced during signal processing or signal quantization. When V and n are known, the parameters of f and g can be obtained through the training process. Wherein, V is the sample data and sample label in the second data set. The above training process requires multiple second data sets, and the data samples in the multiple second data sets, that is, the distribution of V, are different. For example, assuming V=(0.5,0.5,-0.5,0.5), f((0.5,0.5,-0.5,0.5))=1-1j, n=0.1+0.1j, then g(1.1-0.9j)=(0.5,0.5,-0.5,0.5).

[0300] S1002: The network device sends a first data set D and model information of g1 and g2 for model verification to the terminal.

[0301] Wherein, each sample in D is (Vin, Vq), where Vin represents the true value of the precoding vector, i.e., the sample label, and Vq represents the sample data. In addition, if all sample labels are known on both the terminal and the network device side (for example, a reference signal sequence known to both the transceiver and the receiver is used as a sample label), then each sample in D only needs to include Vq; otherwise, the sample in D is (Vin, Vq). Wherein, the reference signal may be DMRS, SRS, CSI-RS, PRS, or other reference signals that have not yet been defined in the current protocol.

[0302] Exemplarily, the sample data may be sent via a data channel or a channel not yet defined in an existing protocol for carrying sample data; the sample tag may be sent via a control channel or a channel not yet defined in an existing protocol for carrying sample tags.

[0303] Exemplarily, D and the model information of g1 and g2 correspond to the same time unit. That is, the signal sent in the same time unit includes D and the model information of g1 and g2. For example, the time unit can be a mini-time slot, a time slot, a subframe or a frame, or a time unit with a smaller or larger time length that has not yet been defined in the protocol.

[0304] Further, the signal includes multiple OFDM symbols, and the starting position of the OFDM symbol carrying D in the multiple OFDM symbols does not lag behind the starting position of the OFDM symbol carrying the model information of g1 and g2.

[0305] The model information of g1 and g2 includes the structure and parameters of g1 and g2.

[0306] S1003, the terminal uses g1 and g2 to perform a model inference process on the data in D, and obtains Vout1 = g1 (Vq) and Vout2 = g2 (Vq).

[0307] That is, the terminal inputs the sample data Vq in D into g1 and g2 respectively, and obtains the output Vout1=g1(Vq) of g1 and the output Vout2=g2(Vq) of g2.

[0308] S1004, the terminal compares the difference between Vout1 and Vout2 and the sample label Vin.

[0309] For example, the terminal calculates Li = ||Vouti-Vin||2, and can obtain L1 and L2, where ||.||2 represents the 2-norm of the orientation quantity. i=1, 2. For example, Vin=(0.5, 0.5, -0.5, 0.5), Vout1=(0.5, 0.5, -0.6, 0.5), Vout2=(0.5, 0.5, 0.5, 0.5), then L1=0.1, L2=1.

[0310] For another example, the terminal calculates Li = 1 / ||Vouti-Vin||2, or the terminal calculates Li = -||Vouti-Vin||2, and L1 and L2 can be obtained, where ||.||2 represents the 2-norm of the orientation quantity, and i = 1, 2. For example, Vin = (0.5, 0.5, -0.5, 0.5), Vout1 = (0.5, 0.5, -0.6, 0.5), Vout2 = (0.5, 0.5, 0.5, 0.5), if Li = 1 / ||Vouti-Vin||2, then L1 = 10, L2 = 1; if Li = -||Vouti-Vin||2, then L1 = -0.1, L2 = -1. S1005, the terminal sends feedback information to the network device. Correspondingly, the network device receives the feedback information.

[0311] The feedback information includes L1 and L2 calculated by the terminal.

[0312] Exemplarily, the time unit in which the terminal sends the feedback information is the same as the time unit in which the network device sends the model information of D and / or g1, g2. For example, the time unit is a mini-time slot, a time slot, a subframe or a frame, or a time unit with a smaller or larger time length that has not yet been defined in the protocol.

[0313] S1006, the terminal selects the model gi that minimizes / maximizes Li as the currently used decompression neural network model, and deploys it at its physical layer for the decompression operation of the precoded vector.

[0314] Assume that the smaller Li is, the better the performance of the neural network model is. For example, Li = ||Vouti-Vin||2, Vin = (0.5, 0.5, -0.5, 0.5), Vout1 = (0.5, 0.5, -0.6, 0.5), Vout2 = (0.5, 0.5, 0.5, 0.5), then L1 = 0.1, L2 = 1, then, in S1006, the terminal selects the model g1 corresponding to L1 as the currently used decompression neural network model deployed in its physical layer.

[0315] In a possible implementation, S806 is executed only when the minimum Li is less than a preset threshold. For example, Li = || Vouti - Vin || 2, Vin = (0.5, 0.5, -0.5, 0.5), Vout1 = (0.5, 0.5, -0.6, 0.5), Vout2 = (0.5, 0.5, 0.5, 0.5), then L1 = 0.1, L2 = 1, the preset threshold is 0.5, then S806 is executed.

[0316] In one possible implementation, if the minimum Li is greater than or equal to the preset threshold, the terminal can also indicate to the network device in the feedback information in S1005 that all Li are greater than or equal to the preset threshold. For example, Li = || Vouti-Vin || 2, Vin = (0.5, 0.5, -0.5, 0.5), Vout1 = (0.5, 0.5, -0.6, 0.5), Vout2 = (0.5, 0.5, 0.5, 0.5), then L1 = 0.1, L2 = 1, the preset threshold is 0.05, then the terminal needs to send feedback information to the network device, indicating that all Li are greater than or equal to the preset threshold. Based on the received feedback information, the network device can resend a new neural network model to the terminal, restart the process of this solution, or instruct the terminal to turn off the physical layer AI function and fall back to non-AI mode.

[0317] Assume that the larger Li is, the better the performance of the neural network model is. For example, Li = 1 / ||Vouti-Vin||2, Vin = (0.5, 0.5, -0.5, 0.5), Vout1 = (0.5, 0.5, -0.6, 0.5), Vout2 = (0.5, 0.5, 0.5, 0.5), then L1 = 10, L2 = 1. At this time, the neural network model corresponding to L1 has better performance than the neural network model corresponding to L2. Then, in S1006, the terminal selects the model gi that makes Li the largest (that is, the model g1 corresponding to L1) as the currently used decompression neural network model deployed at its physical layer.

[0318] In a possible implementation, S1006 is executed only when the maximum Li is greater than a preset threshold. For example, Vin = (0.5, 0.5, -0.5, 0.5), Vout1 = (0.5, 0.5, -0.6, 0.5), Vout2 = (0.5, 0.5, 0.5, 0.5), if Li = 1 / ||Vouti-Vin||2, then L1 = 10, L2 = 1, when the preset threshold is 5, S1006 should be executed.

[0319] In one possible implementation, if the maximum Li is less than or equal to the preset threshold, the terminal can also indicate to the network device in the feedback information in S1005 that all Li are less than or equal to the preset threshold. For example, Vin = (0.5, 0.5, -0.5, 0.5), Vout1 = (0.5, 0.5, -0.6, 0.5), Vout2 = (0.5, 0.5, 0.5, 0.5), if Li = 1 / ||Vouti-Vin||2, then L1 = 10, L2 = 1, the preset threshold is 11, then the terminal should send feedback information to the network device, indicating that all neural network models of the network device are less than or equal to the preset threshold. Based on the received feedback information, the network device can resend a new neural network model to the terminal, restart the process of this solution, or instruct the terminal to turn off the physical layer AI function and fall back to non-AI mode.

[0320] It should be understood that the present application does not limit the execution order between S1005 and S1006. S1005 may be executed first and then S1006, S1006 may be executed first and then S1005, or S1005 and S1006 may be executed simultaneously.

[0321] S1007, the network device selects fi corresponding to the model gi that makes Li minimum / maximum as the currently used compression neural network model, deploys it at its physical layer, and uses it for compression operations of precoding vectors.

[0322] If the model gi selected in S1006 minimizes Li, then fi corresponding to the model gi that minimizes Li will be selected in S1007. If the model gi selected in S1006 maximizes Li, then fi corresponding to the model gi that maximizes Li will be selected in S1007. In other words, fi corresponding to the gi selected in S1006 will be selected in S1007. For example, Li=||Vouti-Vin||2, Vin=(0.5,0.5,-0.5,0.5), Vout1=(0.5,0.5,-0.6,0.5), Vout2=(0.5,0.5,0.5,0.5), then L1=0.1, L2=1, at this time the neural network models f1 and g1 corresponding to L1 are selected to be deployed at the physical layer; for another example, Li=1 / ||Vouti-Vin||2, Vin=(0.5,0.5,-0.5,0.5), Vout1=(0.5,0.5,-0.6,0.5), Vout2=(0.5,0.5,0.5,0.5), then L1=10, L2=1, at this time the neural network models f1 and g1 corresponding to L1 are selected to be deployed at the physical layer.

[0323] Assuming that the smaller Li is, the smaller the gap described in S1004 is, in a possible implementation, S1007 is executed only when the minimum Li is less than the preset threshold. For example, Li = || Vouti - Vin || 2, Vin = (0.5, 0.5, -0.5, 0.5), Vout1 = (0.5, 0.5, -0.6, 0.5), Vout2 = (0.5, 0.5, 0.5, 0.5), then L1 = 0.1, L2 = 1, when the preset threshold is 0.5, S1007 is executed.

[0324] Assuming that the larger Li is, the smaller the gap described in S1004 is, in a possible implementation, S1007 is executed only when the smallest Li is greater than or equal to the preset threshold. For example, Vin=(0.5,0.5,-0.5,0.5), Vout1=(0.5,0.5,-0.6,0.5), Vout2=(0.5,0.5,0.5,0.5), if Li=1 / ||Vouti-Vin||2, then L1=10, L2=1, when the preset threshold is 5, S1007 should be executed.

[0325] According to the communication method provided in the present application, the network device can send multiple precoding vector decompression neural network models (i.e., g1 and g2) and a data set D for model performance verification to the terminal. The terminal can use the data set D to select the model that best matches the statistical distribution parameters of the current channel from multiple precoding vector decompression neural network models, thereby improving the reasoning performance of the neural network model. Furthermore, the terminal can also feedback the performance of each precoding vector decompression neural network model to the network device, so that the network device can use the precoding vector compression model that matches the terminal side to perform precoding vector compression according to the feedback of the terminal, or the network device can reconfigure the precoding vector decompression neural network model for the terminal or instruct the terminal to turn off the physical layer AI function when all precoding vector decompression neural network models do not meet the performance requirements.

[0326] Fig.16 11 is a schematic flow chart of another communication method provided by the present application. The method 1100 may include one or more steps from S1110 to S1160. Each step of the method 1100 is described below.

[0327] S1110: The terminal obtains a first data set.

[0328] The first data set may include sample data and a label corresponding to the sample data (i.e., a sample label). The sample label is known to the terminal and the network device. Exemplarily, the sample label may be a reference signal, such as DMRS, SRS, CSI-RS, PRS, or a reference signal not yet defined in the current protocol.

[0329] The terminal can obtain the sample data and / or sample labels in the first data set locally. Exemplarily, the sample data and the sample label can be historical data received by the terminal side from the network side; or, the sample data can be historical data received by the terminal side from the network side, and the sample label is a sample label calculated by the terminal according to a predefined rule, such as a sample label for a certain type of reference signal defined by the standard. Exemplarily, the sample data can also be predefined, and the sample data can also be used as a sample label. For example, the sample data is a channel sample generated according to a channel model of a typical 3GPP scenario (for example, urban macrocell Uma, urban microcell Umi, etc.), and the channel sample is also a sample label.

[0330] The terminal may also obtain the sample data and / or sample labels in the first data set from other communication devices (for example, another terminal or a network device). For example, the network device may send sample data and sample labels to the terminal; or, the network device may send sample data and indication information to the terminal, where the indication information indicates the sample label, for example, the indication information indicates that the sample label is a certain type of reference signal defined by the standard. Exemplarily, the network device may send the sample label or the indication information through a control channel, or send the sample label or the indication information through a channel that is not yet defined in an existing protocol dedicated to carrying sample labels. Exemplarily, the network device may send the sample data through a data channel, or send the sample data through a channel that is not yet defined in an existing protocol dedicated to carrying sample data.

[0331] S1120, the terminal obtains at least one second neural network model based on the first data set and at least one first neural network model.

[0332] The terminal can use the first data set to update part or all of the parameters (i.e., part or all of the weights) of at least one first neural network model, and / or update / adjust the connection relationship (i.e., part or all of the structure / structural parameters) of part or all of the neural network nodes to obtain at least one second neural network model.

[0333] Exemplarily, the at least one first neural network model may be pre-deployed on the terminal side.

[0334] It should be understood that the at least one second neural network model is an updated at least one first neural network model.

[0335] S1130: The terminal obtains a second data set.

[0336] The first data set may include sample data and a label corresponding to the sample data (i.e., a sample label). The sample label is known to the terminal and the network device. Exemplarily, the sample label may be a reference signal, such as DMRS, SRS, CSI-RS, PRS, or a reference signal not yet defined in the current protocol.

[0337] The method for obtaining the second data set is similar to that for the first data set, and will not be described in detail here.

[0338] The second data set is used to verify at least one first neural network model and / or at least one second neural network model.

[0339] In a possible implementation manner, the method may further include S1140 and / or S1150.

[0340] S1140, verifying the at least one first neural network model using the second data set and the at least one first neural network model to obtain a first verification result.

[0341] Alternatively, S1140 can be described as: verifying the at least one first neural network model through the second data set to obtain a first verification result.

[0342] Specifically, the second data set can be input into the at least one first neural network model to obtain the output result of the first neural network model. This operation is actually to use the at least one first neural network model to perform model inference on the sample data in the second data set.

[0343] In one example, after obtaining the output result of the at least one first neural network model, the terminal may further calculate the gap between the output result of the at least one first neural network model and the sample labels in the second data set to obtain at least one gap. For example, the gap may be the Euclidean distance between the output result of the first neural network model and the sample labels in the second data set.

[0344] In one example, after obtaining the at least one gap, the at least one gap may be compared with a preset threshold to obtain a comparison result.

[0345] In one example, after obtaining the comparison result, the terminal may also determine a first neural network model whose corresponding gap in the at least one first neural network model is smaller than a preset threshold. For ease of description, any first neural network model whose corresponding gap in the at least one first neural network model is smaller than a preset threshold is hereinafter referred to as: a first target neural network model.

[0346] In one example, the terminal may also determine a first neural network model or a first target neural network model that minimizes the corresponding gap.

[0347] In one example, the terminal may also determine the average (may also be a calculation method such as total number, maximum, minimum, etc.) quantization bit number of some or all first neural network model parameters or first target neural network model parameters in the at least one first neural network model. The first neural network model parameters / first target neural network model parameters include the structure and parameters of the first neural network model / first target neural network model.

[0348] In one example, the terminal may also determine the neural network model with the least number of average quantization bits (or calculation methods such as total number, maximum, minimum, etc.) of parameters in the at least one first neural network model or multiple first target neural network models.

[0349] In one example, the terminal may also determine the inference delay of some or all of the first neural network models or some or all of the first target neural network models in the at least one first neural network model.

[0350] In one example, the terminal may also determine a first neural network model or a first target neural network model with the smallest inference delay.

[0351] In some embodiments, the operations described in the examples, performed after obtaining the gaps between the output results of the at least one first neural network model and the sample labels in the second data set as described in S220, can also be considered as part of the verification of the at least one first neural network model.

[0352] In some embodiments, the verification result may include an indication of one or more of the following described above: part or all of the gaps in the at least one gap (for example, it may be the gap corresponding to the first target neural network model), a comparison result (which may be the comparison result corresponding to part or all of the first neural network models in the at least one first neural network model), part or all of the first target neural network models, the first neural network model or the first target neural network model that minimizes the corresponding gap, the first neural network model parameter or the average quantization bit number of the first target neural network model parameter, the neural network model with the least average quantization bit number of the parameters in the at least one first neural network model or multiple first target neural network models (which may also be calculated by the total number, maximum, minimum, etc.), the inference delay of part or all of the first neural network models or part or all of the first target neural network models in the at least one first neural network model, and the first neural network model or the first target neural network model with the least inference delay.

[0353] Exemplarily, the preset threshold in the embodiment of the present application can be configured by the network device to the terminal.

[0354] S1150, verifying the at least one second neural network model using the second data set and the at least one second neural network model to obtain a second verification result.

[0355] Alternatively, S1150 can be described as: verifying the at least one second neural network model through a second data set to obtain a second verification result.

[0356] Specifically, the second data set can be input into the at least one second neural network model to obtain the output result of the second neural network model. This operation is actually to use the at least one second neural network model to perform model inference on the sample data in the second data set.

[0357] In one example, after obtaining the output result of the at least one second neural network model, the terminal may also calculate the gap between the output result of the at least one second neural network model and the sample labels in the second data set to obtain at least one gap. For example, the gap may be the Euclidean distance between the output result of the second neural network model and the sample labels in the second data set.

[0358] In one example, after obtaining the at least one gap, the at least one gap may be compared with a preset threshold to obtain a comparison result.

[0359] In one example, after obtaining the comparison result, the terminal may also determine a second neural network model whose corresponding gap in the at least one second neural network model is smaller than a preset threshold. For ease of description, any second neural network model whose corresponding gap in the at least one second neural network model is smaller than a preset threshold is hereinafter referred to as: a second target neural network model.

[0360] In one example, the terminal may also determine a second neural network model or a second target neural network model that minimizes the corresponding gap.

[0361] In one example, the terminal may also determine the average (may also be a calculation method such as total number, maximum, minimum, etc.) quantization bit number of some or all second neural network model parameters or second target neural network model parameters in the at least one second neural network model. The second neural network model parameters / second target neural network model parameters include the structure and parameters of the second neural network model / second target neural network model.

[0362] In one example, the terminal may also determine the neural network model with the least average quantization bit number (it may also be calculated by total number, maximum, minimum, etc.) of parameters in the at least one second neural network model or multiple second target neural network models.

[0363] In one example, the terminal may also determine the inference delay of part or all of the second neural network models in the at least one second neural network model or part or all of the second target neural network models.

[0364] In one example, the terminal may also determine a second neural network model or a second target neural network model with the smallest inference delay.

[0365] In some embodiments, the operations described in the examples described in S220 after obtaining the gaps between the output results of the at least one second neural network model and the sample labels in the second data set can also be considered as part of the verification of the at least one second neural network model.

[0366] In some embodiments, the verification result may indicate one or more of the following described in S1150: part or all of the gaps in the at least one gap (for example, it may be the gap corresponding to the second target neural network model), a comparison result (which may be the comparison result corresponding to part or all of the second neural network models in the at least one second neural network model), part or all of the second target neural network models, the second neural network model or the second target neural network model that minimizes the corresponding gap, the second neural network model parameter or the average quantization bit number of the second target neural network model parameter, the neural network model with the least average quantization bit number of the parameters in the at least one second neural network model or multiple second target neural network models (it may also be calculated by the total number, maximum, minimum, etc.), the inference delay of part or all of the second neural network models or part or all of the second target neural network models in the at least one second neural network model, and the second neural network model or the second target neural network model with the least inference delay.

[0367] Exemplarily, the preset threshold in the embodiment of the present application can be configured by the network device to the terminal. In addition, the preset thresholds involved in S1150 and S1140 can be the same or different, and the present application does not limit this.

[0368] S1160: The terminal sends feedback information to the network device. Correspondingly, the network device receives the feedback information from the terminal.

[0369] The feedback information is used to indicate or determine the performance of the at least one first neural network model and / or the performance of the at least one second neural network model.

[0370] Exemplarily, the feedback information may be sent from the terminal to the access network device via an uplink control channel, may be sent from the terminal to the access network device via an uplink data channel, or may be sent from the terminal to the access network device via a channel that has not yet been defined in an existing protocol specifically for sending neural network model performance indication information.

[0371] In one example, the feedback information may include the first verification result obtained in S1140 and / or the second verification result obtained in S1150. For example, the feedback information may indicate one or more of the following described in S1140 above: part or all of the gaps in the at least one gap (for example, it may be the gap corresponding to the first target neural network model), a comparison result (it may be the comparison result corresponding to part or all of the first neural network models in the at least one first neural network model), part or all of the first target neural network models, the first neural network model or the first target neural network model that minimizes the corresponding gap, the first neural network model parameter or the average quantization bit number of the first target neural network model parameter, the neural network model with the least average quantization bit number (which may also be calculated by the total number, maximum, minimum, etc.) of the parameters in the at least one first neural network model or multiple first target neural network models, the inference delay of part or all of the first neural network models or part or all of the first target neural network models in the at least one first neural network model, and the first neural network model or the first target neural network model with the smallest inference delay. For example, the feedback information may indicate one or more of the following described above in S1150: part or all of the gaps in the at least one gap (for example, it may be the gap corresponding to the second target neural network model), a comparison result (which may be the comparison result corresponding to part or all of the second neural network models in the at least one second neural network model), part or all of the second target neural network models, the second neural network model or the second target neural network model that minimizes the corresponding gap, the second neural network model parameters or the average quantization bit number of the second target neural network model parameters, the neural network model with the least average quantization bit number (which may also be calculated by the total number, maximum, minimum, etc.) of the parameters in the at least one second neural network model or multiple second target neural network models, the inference delay of part or all of the second neural network models or part or all of the second target neural network models in the at least one second neural network model, and the second neural network model or the second target neural network model with the least inference delay.

[0372] For example, the at least one first neural network model includes model 1 and model 2, and the at least one first neural network model includes model 3 and model 4. The feedback information indicates that the gap corresponding to model 1 (i.e., the gap between the output result obtained by model inference of model 1 and the sample label in the first data set) is smaller than a preset threshold (i.e., model 1 is the target first neural network model), the inference delay of model 1 is smaller than a given threshold, the gap corresponding to model 2 is larger than the preset threshold, the gap corresponding to model 3 is smaller than the preset threshold, the inference delay of model 3 is smaller than the given threshold, and the gap corresponding to model 4 is larger than the preset threshold.

[0373] For example, the at least one first neural network model includes model 1 and model 2, and the at least one first neural network model includes model 3 and model 4, and the feedback information indicates that the gaps corresponding to model 3 and model 4 are both smaller than a preset threshold.

[0374] For example, the at least one first neural network model includes model 1 and model 2, and the at least one first neural network model includes model 3 and model 4. The feedback information indicates that the gap corresponding to model 1 is smaller than a preset threshold, the gap corresponding to model 2 is larger than the preset threshold, the gap corresponding to model 3 is larger than the preset threshold, and the gap corresponding to model 4 is smaller than the preset threshold.

[0375] In one example, the feedback information may include the at least one gap obtained in S1140 and / or the at least one gap obtained in S1150. Furthermore, the operation performed by the terminal after obtaining the at least one gap described in S1140 may also be performed by the network device. And / or, the operation performed by the terminal after obtaining the at least one gap described in S1150 may also be performed by the network device. For example, the network device can determine one or more of the following based on part or all of the gap in the at least one described in S1140: a comparison result (which may be a comparison result corresponding to part or all of the first neural network models in the at least one first neural network model), part or all of the first target neural network models, the first neural network model or the first target neural network model that minimizes the corresponding gap, the first neural network model parameter or the average quantization bit number of the first target neural network model parameter, the neural network model with the least average quantization bit number of the parameters in the at least one first neural network model or multiple first target neural network models (which may also be calculated by the total number, maximum, minimum, etc.), the inference delay of part or all of the first neural network models or part or all of the first target neural network models in the at least one first neural network model, and the first neural network model or the first target neural network model with the least inference delay. For example, the network device can determine one or more of the following based on part or all of the gap in the at least one described in S1140: a comparison result (which may be a comparison result corresponding to part or all of the second neural network models in the at least one second neural network model), part or all of the second target neural network models, a second neural network model or a second target neural network model that minimizes the corresponding gap, a second neural network model parameter or an average quantization bit number of second target neural network model parameters, a neural network model with the least average quantization bit number of parameters in the at least one second neural network model or multiple second target neural network models (which may also be calculated by the total number, maximum, minimum, etc.), an inference delay of part or all of the second neural network models or part or all of the second target neural network models in the at least one second neural network model, and a second neural network model or a second target neural network model with the least inference delay.

[0376] In one example, the feedback information includes the output results of some or all of the first neural network models in the at least one first neural network model obtained in S1140 and / or the output results of some or all of the second neural network models in the at least one second neural network model obtained in S1150. The network device can determine one or more of the following based on the output results of some or all of the first neural network models in the at least one first neural network model: the at least one gap, the comparison result (which can be the comparison result corresponding to some or all of the first neural network models in the at least one first neural network model), some or all of the first target neural network models, the first neural network model or the first target neural network model that minimizes the corresponding gap, the first neural network model parameter or the average quantization bit number of the first target neural network model parameter, the neural network model with the least average quantization bit number of the parameters in the at least one first neural network model or multiple first target neural network models (which can also be calculated by the total number, maximum, minimum, etc.), the reasoning delay of some or all of the first neural network models or some or all of the first target neural network models in the at least one first neural network model, and the first neural network model or the first target neural network model with the smallest reasoning delay. The network device can also determine one or more of the following based on the output results of some or all of the second neural network models in the at least one second neural network model: the at least one gap, the comparison result (which can be the comparison results corresponding to some or all of the second neural network models in the at least one second neural network model), some or all of the second target neural network models, the second neural network model or the second target neural network model that minimizes the corresponding gap, the second neural network model parameters or the average quantization bit number of the second target neural network model parameters, the neural network model with the least average quantization bit number of the parameters in the at least one second neural network model or multiple second target neural network models (which can also be calculated by the total number, maximum, minimum, etc.), the inference delay of some or all of the second neural network models or some or all of the second target neural network models in the at least one second neural network model, and the second neural network model or the second target neural network model with the least inference delay.

[0377] In one example, the gap between the output result of the neural network model (e.g., the first neural network model or the second neural network model) and the sample label in the second data set can be used as an evaluation indicator of model performance. For example, the gap is the Euclidean distance described above, then the smaller the gap is, the better the performance of the corresponding first neural network model or the second neural network model is, and the higher the degree of matching with the scene is.

[0378] In one example, the function value of the function of the independent variable, which is the gap between the output result of the neural network model (e.g., the first neural network model or the second neural network model) and the sample label in the second data set, can be used as an evaluation index of model performance. Taking the gap as the Euclidean distance described above as an example, for example, the function can be the reciprocal of the gap, or the function can be the opposite of the gap. Then, the larger the function value, the better the performance of the corresponding first neural network model or the second neural network model, and the higher the degree of matching with the scene.

[0379] In one example, the average (or the total number, maximum, minimum, etc.) quantization bit number of the neural network model parameters can be used as an evaluation index of model performance. It should be understood that the smaller the bit number, the lower the overhead required to store the first neural network model parameters or the second neural network model parameters, and the better the performance of the first neural network model or the second neural network model.

[0380] In one example, the inference delay of the neural network model can be used as an evaluation indicator of model performance. It should be understood that the smaller the inference delay, the better the performance of the first neural network model or the second neural network model, and the higher the degree of matching with the scene.

[0381] It should be understood that some or all of the multiple evaluation indicators described above may also be used to comprehensively evaluate the performance of the first neural network model and / or the second neural network model.

[0382] According to the communication method provided in the present application, the terminal can verify the performance of at least one neural network model (at least one first neural network model and / or at least one second neural network model) according to the second data set, and then determine the performance of the at least one neural network model, so as to determine whether the at least one neural network model matches the scene or determine the neural network model with the best performance among the at least one neural network model. Furthermore, the terminal can deploy the neural network model with the best performance at the physical layer, receive signals based on the neural network model, and ensure the signal processing performance as much as possible. On the other hand, the terminal can indicate the performance of the at least one first neural network model to the network device, so that the network device can also perform corresponding operations based on the feedback of the terminal, and avoid the performance degradation problem that may be caused by the use of a neural network model that does not match the current scene on the terminal side as much as possible. For example, when the performance requirements of each neural network model are not met, the network device reconfigures the neural network model for the terminal or instructs the terminal to turn off the physical layer AI function.

[0383] In a possible implementation, the second time unit in which part or all of the first data set is located is the same as the second time unit in which part or all of the second data set is located. That is, the network device will send part or all of the first data set and part or all of the second data set to the terminal in the same second time unit.

[0384] For example, see Fig.17 , the time unit in which the network device sends part or all of the first data set and part or all of the second data set to the terminal is the same. It should be understood that Fig.17 The time unit shown in is the second time unit.

[0385] By sending part or all of the first data set and part or all of the first data set on the same second time unit, the network device side can ensure that the model verification result on the terminal side can indicate the performance of each neural network model in the current scenario.

[0386] Exemplarily, the second time unit in the embodiment of the present application may be a micro-time slot, a time slot, a subframe or a frame, or a time unit representing a smaller or larger time length that has not yet been defined in the protocol.

[0387] It should be noted that part or all of the first data set and part or all of the second data set may not be in the same second time unit. For example, part or all of the first data set and part or all of the second data set may be in different second time units, or at least one of part or all of the first data set and part or all of the second data set may occupy several symbols in at least two adjacent time units. For example, see Fig.18 , part or all of the first data set and part or all of the second data set may not be in the same time unit. It should be understood that Fig.18 The time unit shown in is the second time unit.

[0388] In a possible implementation, the start time of the symbol carrying part or all of the first data set does not lag behind the start time of the symbol carrying part or all of the second data set, that is, the time domain start position of the symbol of part or all of the first data set does not lag behind the time domain start position of the symbol of part or all of the second data set. For example, the time domain start position of the symbol of part or all of the first data set may be before the time domain start position of the symbol of part or all of the second data set, or the time domain start position of the symbol of part or all of the first data set may be the same as the time domain start position of the symbol of part or all of the second data set.

[0389] For example, the temporal position relationship between part or all of the first data set and part or all of the second data set may be: Fig.19 Any of the positional relationships shown.

[0390] Based on this solution, the terminal will first receive the first data set and then the second data set, so that the at least one first neural network model can be verified based on the second data set and / or the at least one second neural network model obtained by updating the at least one first neural network model using the first data set can be verified without storing the second data set, thereby reducing cache overhead.

[0391] In a possible implementation manner, the first time unit in which the terminal obtains part or all of the first data set or part and / or all of the second data set from the network device is the same as the first time unit in which the terminal sends feedback information.

[0392] Among them, part of the first data set refers to the sample data in the first data set. In the scenario where the terminal only obtains the sample data from the network device but does not obtain the sample label, it is considered that the terminal side can confirm the sample label by itself, such as the terminal side pre-stores the sample label or can calculate the sample label according to a pre-defined rule. The entire first data set refers to the sample data and sample labels in the first data set.

[0393] Similarly, part of the second data set refers to the sample data in the second data set. In the scenario where the terminal only obtains the sample data from the network device but does not obtain the sample label, it is considered that the terminal side can confirm the sample label by itself, such as the terminal side pre-stores the sample label or can calculate the sample label according to a pre-defined rule. The entire second data set refers to the sample data and sample labels in the second data set.

[0394] For example, see Fig. 20 , part or all of the first data set, part or all of the second data set and the feedback information are located in the same time unit. It should be understood that Fig. 20 The time unit shown in is the second time unit.

[0395] By transmitting part or all of the first data set or part or all of the second data set and feedback information in the same first time unit, timely transmission of the feedback information corresponding to the first data set or the second data set can be achieved, thereby ensuring as much as possible that the feedback information on the terminal side can indicate the performance of each neural network model in the current scenario.

[0396] Exemplarily, the first time unit in the embodiment of the present application may be a mini-time slot, a time slot, a subframe or a frame, or a time unit representing a smaller or larger time length that has not yet been defined in the protocol.

[0397] It should be noted that part or all of the first data set, part or all of the second data set, and the feedback information may not be in the same first time unit. For example, part or all of the first data set, part or all of the second data set, and the feedback information may be located in different first time units, or at least one of part or all of the first data set, part or all of the second data set, and the feedback information may occupy several symbols in at least two adjacent time units. For example, Fig.21 It shows the situation that part or all of the first data set, part or all of the second data set and the feedback information are not in the same time unit. It should be understood that Fig.21 The time unit shown in is the second time unit.

[0398] As described above, the sample label in the first data set or the indication information indicating the sample label may be sent through the control channel. Similarly, the sample label in the second data set or the indication information indicating the sample label may be sent through the control channel. The third time unit where the control channel is located is the same as the third time unit where one or more of the following are located: the sample data in the first data set, the sample data in the second data set, or the feedback information.

[0399] For example, the control channel may be located in (or carried on) the same third time unit as the sample data in the first data set, the sample data in the second data set, and the feedback information, or the control channel may be located in the same third time unit as the sample data in the first data set and the sample data in the second data set. For example, the control channel may be located in the same third time unit as the feedback information.

[0400] For example, Fig. 22 In the example shown, the control channel may be located in the same time unit as the sample data in the first data set, the sample data in the second data set, and the feedback information, wherein the control channel includes first indication information and second indication information, the first indication information indicates the sample labels in the first data set, and the second data set indicates the sample labels in the second data set. It should be understood that Fig. 22 The time unit shown in is the third time unit.

[0401] Exemplarily, the third time unit in the embodiment of the present application may be a micro-time slot, a time slot, a subframe or a frame, or a time unit representing a smaller or larger time length that has not yet been defined in the protocol.

[0402] It should be noted that the control channel may also be in a different fourth time unit from one or more of the following: sample data in the first data set, sample data in the second data set, or feedback information. For example, the control channel may be in a different fourth time unit from one or more of the following: sample data in the first data set, sample data in the second data set, or feedback information. For example, see Fig.23 , the control channel, the sample data in the first data set, the sample data in the second data set, and the feedback information are located in different time units. Alternatively, the control channel is in the first time unit, the sample data in the first data set is in the second time unit, the sample data in the second data set occupies several symbols in the second time unit and several symbols in the third time unit, and the feedback information is in the fourth time unit. It should be understood that Fig.23 The time unit shown in is the third time unit.

[0403] In a possible implementation, any information involved in method 1100 (eg, part or all of the first data set, part or all of the second data set, model information, or feedback information, etc.), such as Figures 17 to 23 Any information shown may occupy one or more consecutive symbols in one time unit, or one or more consecutive symbols in multiple time units, which is not limited in the present application. For example, the symbol in the embodiment of the present application may be an OFDM symbol.

[0404] It should be understood that Figure 17 to Figure 13 The positional relationship of each information in the time domain and the frequency domain shown is only some examples, and these examples should not constitute any limitation to the present application. The positional relationship of each information involved in the method 1100 of the present application in the time domain and the frequency domain can also be Figures 17 to 23 The positional relationships shown are different.

[0405] Fig.16 The method 1100 shown can be applied in a variety of scenarios, for example, Fig.24 The scene shown.

[0406] Fig.24 is a schematic diagram of an application scenario provided by an embodiment of the present application. Fig.24, a neural network model can be deployed on the terminal side, and the traditional demodulation can be replaced by the neural network model to realize demodulation based on the neural network model. When the modulation order is N, the input of the modulation module is a string of bits b0b1b2b3…bN-1, bi={0,1}, and the output is the modulation symbol x, where x is a complex number, so the modulation module can be expressed as x=f(b0b1b2b3…bN-1). For example, when N=4, x may be any one of the complex number set {1+1j,1-1j,-1+1j,-1-1j}, f(0000)=-1-1j. The modulation module f can be implemented based on the neural network model or not. The input of the demodulation module is the complex number z, and the output is a string of bits b0b1b2b3…bN-1, so the input-output relationship of the demodulation module can be expressed as b0b1b2b3…bN-1=g(z). For example, when N=4, z may be any one of the complex number set {1+1j, 1-1j, -1+1j, -1-1j}, and g(-1-1j)=0000. Since the input of the demodulation module is generally the result of the output of the modulation module after passing through the channel, that is, z=h*x+n, n is the noise, it can be known that b0b1b2b3…bN-1=g(h*f(b0b1b2b3…bN-1)+n). The value of z is not limited to a complex number set with a finite element, but may be any complex number. The structure or parameters of f and g are related to the statistical characteristics of the current channel h, and may also be related to the intensity of the neighboring cell interference. The greater the intensity of the neighboring cell interference, the greater the variance of the noise n. For example, when N=4, b0b1b2b3=0000, x=f(0000)=-1-1j, h=2, n=0.1+0.1j, then z=-1.9-1.9j, g(-1.9-1.9j)=0000.

[0407] Fig.25 is a schematic flow chart of a communication method provided by the present application. The method 1500 shows Fig.24 The scenario shown uses Fig.16 An example of the method shown. In the method 1500, the terminal side currently has at least one first neural network model gi for signal demodulation.

[0408] S1501, the network device generates a first data set D1 for online model training / updating and a second data set D2 for model verification.

[0409] For example, if the channel model of the cell where the current terminal is located is the Uma model, the network device can generate a channel data sample based on the Uma channel model specified by the NR protocol, and the channel data sample is D1 or D2.

[0410] Among them, each sample in D1 is (b0b1b2b3…bN-1, f(b0b1b2b3…bN-1)), b0b1b2b3…bN-1 represents the sample label, and f(b0b1b2b3…bN-1) represents the sample data. In addition, if all sample labels are known on the terminal and the network device side, each sample in D1 only needs to include f(b0b1b2b3…bN-1), for example, N=4, b0b1b2b3=0001, f(0001)=1+1j represents a sample; otherwise, the sample in D1 is (b0b1b2b3…bN-1, f(b0b1b2b3…bN-1)). For example, N=4, b0b1b2b3=0011, (0011, f(0011)=1-1j) represents a sample.

[0411] Similarly, each sample in D2 is (b0b1b2b3…bN-1, f(b0b1b2b3…bN-1)), b0b1b2b3…bN-1 represents the sample label, and f(b0b1b2b3…bN-1) represents the sample data. In addition, if all sample labels are known on the terminal and network device side, each sample in D2 only needs to include f(b0b1b2b3…bN-1); otherwise, the sample in D2 is (b0b1b2b3…bN-1, f(b0b1b2b3…bN-1)).

[0412] Exemplarily, the sample data may be sent via a data channel, and the sample tag may be sent via a control channel.

[0413] Exemplarily, D1 and D2 correspond to the same time unit. That is, the signal sent in the same time unit includes D1 and D2. For example, the time unit can be a mini-time slot, a time slot, a subframe or a frame, or a time unit with a smaller or larger time length that has not yet been defined in the protocol.

[0414] Further, the signal includes multiple OFDM symbols, and a starting position of an OFDM symbol carrying D1 in the multiple OFDM symbols does not lag behind a starting position of an OFDM symbol carrying D2.

[0415] S1502: The network device sends D1 and D2 to the terminal. Correspondingly, the terminal receives D1 and D2.

[0416] S1503, the terminal uses D1 to update all or part of the parameters in at least one first neural network model gi, and / or adjusts the connection relationship of all or part of the neural network nodes in gi, to obtain at least one second neural network model ki. It should be understood that the second neural network model ki is the updated first neural network model gi. i=1, 2, ...

[0417] S1504, the terminal uses gi and ki to perform model inference operations on the sample data in D2, and obtains gi(h*f(b0b1b2b3…bN-1)+n) and ki(h*f(b0b1b2b3…bN-1)+n). That is, the terminal inputs the sample data f(b0b1b2b3…bN-1) in D2 into gi and ki, and obtains the output of gi gi(h*f(b0b1b2b3…bN-1)+n) and the output of ki ki(h*f(b0b1b2b3…bN-1)+n). For example, N=4, b0b1b2b3=0001, f(0001)=1+1j, h=2, n=0.1+0.1j, gi(2.1+2.1j)=1001, ki(2.1+2.1j)=0001.

[0418] S1505, the terminal compares the difference between gi(h*f(b0b1b2b3…bN-1)+n) and ki(h*f(b0b1b2b3…bN-1)+n) and the sample label, that is, the actual bit sequence b0b1b2b3…bN-1.

[0419] For example, the terminal calculates Lgi=||gi(h*f(b0b1b2b3…bN-1)+n)-b0b1b2b3…bN-1|| and Lki=||ki(h*f(b0b1b2b3…bN-1)+n)-b0b1b2b3…bN-1||, where ||.|| represents the sum of the absolute values ​​of each element of the vector, for example, N=4, b0b1b2b3=0001, f(0001)=1+1j, h=2, n=0.1+0.1j, gi(2.1+2.1j)=1001, ki(2.1+2.1j)=0001, Lgi=1, Lki=0.

[0420] For another example, the terminal calculates Li=1 / ||gi(h*f(b0b1b2b3…bN-1)+n)-b0b1b2b3…bN-1|| and Lki=1 / ||ki(h*f(b0b1b2b3…bN-1)+n)-b0b1b2b3…bN-1||, or, the terminal calculates Li=-||gi(h*f(b0b1b2b3…bN-1)+n)-b0b1b2b3…bN-1|| and Lki=-||ki(h*f(b0b1b2b3…bN-1)+n)-b0b1b2b3…bN-1||, where ||.|| means summing the absolute values ​​of each element of the vector. For example, the terminal calculates Lgi=-||gi(h*f(b0b1b2b3…bN-1)+n)-b0b1b2b3…bN-1|| and Lki=-||ki(h*f(b0b1b2b3…bN-1)+n)-b0b1b2b3…bN-1||, where ||.|| represents the sum of the absolute values ​​of each element of the vector, for example, N=4, b0b1b2b3=0001, f(0001)=1+1j, h=2, n=0.1+0.1j, gi(2.1+2.1j)=1001, ki(2.1+2.1j)=0001, Lgi=-1, Lki=0.

[0421] S1506: The terminal sends feedback information to the network device. Correspondingly, the network device receives the feedback information.

[0422] The feedback information includes Lgi and Lki calculated by the terminal.

[0423] Exemplarily, the time unit in which the terminal sends the feedback information is the same as the time unit in which the network device sends D1 and / or D2. For example, the time unit is a mini-slot, a time slot, a subframe or a frame, or a time unit with a smaller or larger time length that has not yet been defined in the protocol.

[0424] S1507, the terminal selects the model that minimizes / maximizes Li as the currently used demodulation neural network model, deploys it in its physical layer, and uses it for signal demodulation.

[0425] It should be understood that the minimum (maximum) Li refers to the minimum (maximum) value of Lgi and Lki. Accordingly, the model that minimizes (maximums) Li is the neural network model corresponding to the minimum (maximum) value of Lgi and Lki.

[0426] Assume that the smaller Li is, the better the performance of the neural network model is. For example, Lgi=||gi(h*f(b0b1b2b3…bN-1)+n)-b0b1b2b3…bN-1||, Lki=||ki(h*f(b0b1b2b3…bN-1)+n)-b0b1b2b3…bN-1||, N=4, b0b1b2b3=0001, f(0001)=1+1j, h=2, n=0.1+0.1j, gi(2.1+2.1j)=1001, ki(2.1+2.1j)=0001, Lgi=1, Lki=0, then, in S1507, the terminal should select the model ki that minimizes Li as the currently used demodulation neural network model deployed at its physical layer.

[0427] In a possible implementation, when the minimum Li is less than a preset threshold, S1507 is executed. For example, N=4, b0b1b2b3=0001, f(0001)=1+1j, h=2, n=0.1+0.1j, gi(2.1+2.1j)=1001, ki(2.1+2.1j)=0001, Lgi=1, Lki=0, when the preset threshold is 0.1, S1507 is executed.

[0428] In one possible implementation, if the minimum Li is greater than or equal to the preset threshold, the terminal can also indicate to the network device in the feedback information in S1506 that all Li are greater than or equal to the preset threshold. For example, N = 4, b0b1b2b3 = 0001, f (0001) = 1 + 1j, h = 2, n = 0.1 + 0.1j, gi (2.1 + 2.1j) = 1101, ki (2.1 + 2.1j) = 1001, Lgi = 2, Lki = 1, when the preset threshold is 0.1, all Li are greater than or equal to the preset threshold, and the terminal should send feedback information to the network device, indicating that all Li are greater than or equal to the preset threshold. Based on the received feedback information, the network device can resend a new neural network model to the terminal, restart the process of this solution, or instruct the terminal to turn off the physical layer AI function and fall back to non-AI mode.

[0429] Assuming that the larger Li is, the better the performance of the neural network model is, for example, Li = -||gi(h*f(b0b1b2b3…bN-1)+n)-b0b1b2b3…bN-1||, Lki = -||ki(h*f(b0b1b2b3…bN-1)+n)-b0b1b2b3…bN-1||, then, in S1507, the terminal selects the model gi or ki that maximizes Li as the currently used demodulation neural network model deployed in its physical layer. For example, N = 4, b0b1b2b3 = 1001, f(0001) = 1+1j, h = 2, n = 0.1+0.1j, gi(2.1+2.1j) = 1101, ki(2.1+2.1j) = 1001, Lgi = -1, Lki = 0, the model ki that maximizes Li should be selected and deployed in the physical layer.

[0430] In a possible implementation, when the maximum Li is greater than a preset threshold, S1507 is executed. For example, N=4, b0b1b2b3=1001, f(0001)=1+1j, h=2, n=0.1+0.1j, gi(2.1+2.1j)=1101, ki(2.1+2.1j)=1001, Lgi=-1, Lki=0, when the preset threshold is -0.1, S1507 is executed.

[0431] In one possible implementation, if the maximum Li is less than or equal to the preset threshold, the terminal can also indicate to the network device in the feedback information in S1506 that all Li are less than or equal to the preset threshold. For example, N = 4, b0b1b2b3 = 0001, f (0001) = 1 + 1j, h = 2, n = 0.1 + 0.1j, gi (2.1 + 2.1j) = 1101, ki (2.1 + 2.1j) = 1001, Lgi = -2, Lki = -1, when the preset threshold is -0.1, the terminal sends feedback information to the network device, indicating that the maximum Li of the network device is less than or equal to the preset threshold. Based on the received feedback information, the network device can resend a new neural network model to the terminal, restart the process of this solution, or instruct the terminal to turn off the physical layer AI function and fall back to non-AI mode.

[0432] It should be understood that the present application does not limit the execution order between S1506 and S1507. S1506 may be executed first and then S1507, S1507 may be executed first and then S1506, or S1506 and S1507 may be executed simultaneously.

[0433] According to the communication method provided in the present application, the network device can send a training data set D1 for updating model parameters and a data set D2 for model performance verification to the terminal. The terminal can use the data set D1 to update at least one demodulation neural network model, and use the data set D2 to select the model that best matches the statistical distribution parameters of the current channel or the interference intensity of the neighboring area from at least one demodulation neural network model, thereby improving the reasoning performance of the neural network model. Furthermore, the terminal can also feedback the performance of each demodulation neural network model to the network device, and the network device can determine the performance of each demodulation neural network model based on the feedback information, so that when each demodulation neural network model does not meet the performance requirements, the demodulation neural network model can be reconfigured for the terminal or the terminal can be instructed to turn off the physical layer AI function.

[0434] Fig.26 is a schematic block diagram of a communication device provided in an embodiment of the present application. Fig.26 As shown, the communication device 2000 may include at least one of a communication unit 2100 and a processing unit 2200. The communication unit 2100 may implement a corresponding communication function, and the communication may be internal communication of the communication device 2000 or communication between the communication device 2000 and other devices; the processing unit 2200 may implement a corresponding processing function. The communication unit 2100 may also be referred to as a communication interface or a transceiver unit. Optionally, the communication device 2000 may also include a storage unit, which may be used to store instructions and / or data, and the processing unit 2200 may read the instructions and / or data in the storage unit, so that the communication device 2000 implements the aforementioned method embodiment.

[0435] In one possible design, the communication device 2000 may be a terminal in the above method 200, method 600, method 800 or method 1000, or may be a module or chip applied to the terminal in the method 200, method 600, method 800 or method 1000. The communication device 2000 may be used to execute the steps or processes executed by the terminal in the above method 200, method 600, method 800 or method 1000.

[0436] Specifically, the communication unit 2100 or the processing unit 2200 is used to obtain a first data set, which is used to verify at least one first neural network model; the communication unit 2100 is used to send feedback information, which is used to indicate the performance of the at least one first neural network model.

[0437] Optionally, the at least one first neural network model is preconfigured or acquired through model information of the at least one first neural network model.

[0438] Optionally, the model information of the at least one first neural network model is acquired from a network device.

[0439] Optionally, the first time unit in which the feedback information is sent is the same as the first time unit in which the first data set is acquired.

[0440] Optionally, the second time unit where the first data set is located is the same as the second time unit where the model information is located.

[0441] Optionally, a start time of a symbol carrying the model information does not lag behind a start time of a symbol carrying the first data set.

[0442] Optionally, the third time unit in which the feedback information is sent is the same as the third time unit in which the model information is sent and the third time unit in which the first data set is acquired.

[0443] Optionally, the at least one first neural network model is trained based on a second neural network model and a second data set.

[0444] Optionally, the communication unit 2100 is also used to verify the at least one first neural network model through the first data set and the at least one first neural network model.

[0445] Optionally, the first data set includes sample data, or the first data set includes sample data and sample labels corresponding to the sample data.

[0446] Optionally, the communication unit 2100 is further used to receive a control channel, where the control channel includes a sample label or indication information corresponding to the sample data in the first data set, where the indication information indicates the sample label.

[0447] Optionally, the fourth time unit where the control channel is located is the same as the fourth time unit where one or more of the following are located: sample data in the first data set, model information of the at least one first neural network model, or the feedback information.

[0448] In one possible design, the communication device 2000 may be the network device in the above method 200, method 600, method 800 or method 1000, or may be a module or chip applied to the network device in the method 200, method 600, method 800 or method 1000. The communication device 2000 may be used to execute the steps or processes executed by the network device in the above method 200, method 600, method 800 or method 1000.

[0449] Specifically, the communication unit 2100 is used to: send model information and / or a first data set of at least one first neural network model, wherein the first data set is used to verify the at least one first neural network model; and receive feedback information, wherein the feedback information is used to indicate the performance of the at least one first neural network model.

[0450] Optionally, the first time unit in which the feedback information is sent is the same as the first time unit in which the first data set is acquired.

[0451] Optionally, the second time unit where the first data set is located is the same as the second time unit where the model information is located.

[0452] Optionally, a start time of a symbol carrying the model information does not lag behind a start time of a symbol carrying the first data set.

[0453] Optionally, the third time unit in which the feedback information is sent is the same as the third time unit in which the model information is sent and the third time unit in which the first data set is acquired.

[0454] Optionally, the at least one first neural network model is trained based on a second neural network model and a second data set.

[0455] Optionally, the first data set includes sample data, or the first data set includes the sample data and a sample label corresponding to the sample data.

[0456] Optionally, the communication unit 2100 is further used to send a control channel, where the control channel includes a sample label or indication information corresponding to the sample data in the first data set, where the indication information indicates the sample label.

[0457] Optionally, the fourth time unit where the control channel is located is the same as the fourth time unit where one or more of the following are located: sample data in the first data set, model information of the at least one first neural network model, or the feedback information.

[0458] In one possible design, the communication device 2000 may be the terminal in the above method 1100 or method 1500, or may be a module or chip applied to the terminal in the method 1100 or method 1500. The communication device 2000 may be used to execute the steps or processes executed by the terminal in the above method 1100 or method 1500.

[0459] Specifically, the communication unit 2100 or the processing unit 2200 is used to obtain a first data set; the processing unit 2200 is used to obtain at least one second neural network model based on the first data set and at least one first neural network model; the communication unit 2100 or the processing unit 2200 is also used to obtain a second data set, and the second data set is used to verify the at least one first neural network model and / or the at least one second neural network model; the communication unit 2100 is also used to send feedback information, and the feedback information is used to indicate the performance of the at least one first neural network model and / or the performance of the at least one second neural network model.

[0460] Optionally, the at least one first neural network model is preconfigured or acquired through model information of the at least one first neural network model.

[0461] Optionally, the first time unit where the first data set and / or the second data set is located is the same as the first time unit where the feedback information is located.

[0462] Optionally, the second time unit where the first data set is located is the same as the second time unit where the second data set is located.

[0463] Optionally, a time-domain starting position of symbols of the first data set does not lag behind a time-domain starting position of symbols of the second data set.

[0464] Optionally, the processing unit 2200 is further used to verify the at least one first neural network model through the second data set and the at least one first neural network model; or, to verify the at least one second neural network model through the second data set and the at least one second neural network model.

[0465] Optionally, both the first data set and the second data set include sample data, or both the first data set and the second data set include sample labels of the sample data.

[0466] Optionally, the communication unit 2100 is also used to receive a control channel, which includes sample labels or first indication information corresponding to the sample data in the first data set, and / or includes sample labels or second indication information corresponding to the sample data in the second data set, the first indication information indicating the sample labels corresponding to the sample data in the first data set, and the second indication information indicating the sample labels corresponding to the sample data in the second data set.

[0467] Optionally, the third time unit where the control channel is located is the same as the third time unit where one or more of the following are located: sample data in the first data set, sample data in the second data set, or feedback information.

[0468] In one possible design, the communication device 2000 may be the network device in the above method 1100 or method 1500, or may be a module or chip applied to the network device in the method 1100 or method 1500. The communication device 2000 may be used to execute the steps or processes executed by the network device in the above method 1100 or method 1500.

[0469] Specifically, the communication unit 2100 is used to: send one or more of model information of at least one first neural network model, a first data set, and a second data set, the first data set and the at least one first neural network model are used to obtain at least one second neural network model, and the second data set is used to verify the at least one first neural network model and / or at least one second neural network model; receive feedback information, the feedback information is used to indicate the performance of the at least one first neural network model and / or the performance of the at least one second neural network model.

[0470] Optionally, the first time unit where the first data set and / or the second data set is located is the same as the first time unit where the feedback information is located.

[0471] Optionally, the second time unit where the first data set is located is the same as the second time unit where the second data set is located.

[0472] Optionally, a time-domain starting position of symbols of the first data set does not lag behind a time-domain starting position of symbols of the second data set.

[0473] Optionally, both the first data set and the second data set include sample data, or both the first data set and the second data set include sample labels of the sample data.

[0474] Optionally, the communication unit 2100 is also used to send a control channel, which includes sample labels or first indication information corresponding to the sample data in the first data set, and / or includes sample labels or second indication information corresponding to the sample data in the second data set, the first indication information indicating the sample labels corresponding to the sample data in the first data set, and the second indication information indicating the sample labels corresponding to the sample data in the second data set.

[0475] Optionally, the third time unit where the control channel is located is the same as the third time unit where one or more of the following are located: sample data in the first data set, sample data in the second data set, or feedback information.

[0476] Regarding the steps or processes executed by each unit in the communication device 2000, specific reference may be made to the corresponding method embodiments above, which will not be described in detail here.

[0477] It should be understood that the "unit" in the communication device 2000 can be implemented by hardware, can be implemented by software, and can also be implemented by hardware executing the corresponding software. For example, the "unit" can refer to an application specific integrated circuit (ASIC), an electronic circuit, a processor (such as a shared processor, a dedicated processor or a group processor, etc.) and a memory for executing one or more software or firmware programs, a combined logic circuit and / or other suitable components that support the described functions. For another example, the communication unit 2100 can be replaced by a transceiver transceiver circuit (for example, it can include a receiving circuit and a transmitting circuit), and the processing unit 2200 can be replaced by a processor or a processing circuit.

[0478] Fig. 27 A schematic block diagram of another communication device 3000 provided in an embodiment of the present application is shown. The communication device 3000 may be a terminal or a network device, or may be a chip, a chip system, or a processor that supports the terminal or the network device to implement the above method. The communication device 3000 may be used to implement the method described in the above method embodiment, and the details may refer to the description in the above method embodiment.

[0479] The communication device 3000 may include one or more processors 3100, which may also be referred to as a processing unit, and may implement certain control functions. The processor 3100 may be a general-purpose processor or a dedicated processor, etc. For example, it may be a baseband processor or a central processing unit. The baseband processor may be used to process the communication protocol and the communication data, and the central processing unit may be used to control the communication device (e.g., a base station, a baseband chip, a user chip, a DU or a CU, etc.), execute the software program, and process the data of the software program.

[0480] In an optional design, the processor 3100 may also store instructions and / or data, and the instructions and / or data can be executed by the processor 3100 so that the communication device 3000 executes the method described in the above method embodiment.

[0481] In another optional design, the communication device 3000 may include a communication interface 3200 for implementing the receiving and sending functions. For example, the communication interface 3200 may be a transceiver circuit, an interface, an interface circuit, or a transceiver. The transceiver circuit, interface, interface circuit, or transceiver for implementing the receiving and sending functions may be separate or integrated. The above-mentioned transceiver circuit, interface, interface circuit, or transceiver may be used for reading and writing code / data, or the above-mentioned transceiver circuit, interface, interface circuit, or transceiver may be used for transmitting or delivering signals.

[0482] Optionally, the communication device 3000 may include one or more memories 3300, on which instructions may be stored, and the instructions may be executed on the processor 3100, so that the communication device 3000 performs the method described in the above method embodiment. Optionally, data may also be stored in the memory 3300. Optionally, instructions and / or data may also be stored in the processor 3100. The processor 3100 and the memory 3300 may be provided separately or integrated together.

[0483] Fig.28 The present application provides a schematic diagram of the structure of a terminal 4000. The above communication device 2000 or the communication device 3000 can be configured in the terminal 4000. Alternatively, the communication device 2000 or the communication device 3000 itself can be the terminal 4000. In other words, the terminal 4000 can perform the actions performed by the terminal in the above method embodiment. For the convenience of explanation, Fig.28 Only the main components of the terminal are shown. Fig.28 As shown, terminal 4000 includes a processor, a memory, a control circuit, an antenna, and an input and output device.

[0484] The processor is mainly used to process the communication protocol and communication data, as well as to control the entire terminal, execute the software program, and process the data of the software program, for example, to support the terminal to perform the actions described in the above method embodiment. The memory is mainly used to store software programs and data. The control circuit is mainly used for the conversion of baseband signals and radio frequency signals and the processing of radio frequency signals. The control circuit and the antenna can also be called a transceiver, which is mainly used to send and receive radio frequency signals in the form of electromagnetic waves. Input and output devices, such as touch screens, display screens, keyboards, etc. are mainly used to receive data input by users and output data to users.

[0485] When the terminal is turned on, the processor can read the software program in the storage unit, interpret and execute the instructions of the software program, and process the data of the software program. When data needs to be sent wirelessly, the processor performs baseband processing on the data to be sent, and outputs the baseband signal to the RF circuit. The RF circuit performs RF processing on the baseband signal and then sends the RF signal outward in the form of electromagnetic waves through the antenna. When data is sent to the terminal, the RF circuit receives the RF signal through the antenna, converts the RF signal into a baseband signal, and outputs the baseband signal to the processor. The processor converts the baseband signal into data and processes the data.

[0486] Those skilled in the art will appreciate that for ease of description, Fig.28 Only one memory and processor are shown. In an actual terminal, there may be multiple processors and memories. The memory may also be referred to as a storage medium or a storage device, etc., which is not limited in the embodiments of the present application.

[0487] For example, the processor may include a baseband processor and a central processing unit. The baseband processor is mainly used to process the communication protocol and communication data, and the central processing unit is mainly used to control the entire terminal, execute software programs, and process data of the software programs. Fig.12 The processor in the integrated baseband processor and the central processor functions, those skilled in the art will appreciate that the baseband processor and the central processor may also be independent processors, interconnected by bus and other technologies. Those skilled in the art will appreciate that the terminal may include multiple baseband processors to adapt to different network formats, the terminal may include multiple central processors to enhance its processing capabilities, and the various components of the terminal may be connected through various buses. The baseband processor may also be described as a baseband processing circuit or a baseband processing chip. The central processor may also be described as a central processing circuit or a central processing chip. The function of processing the communication protocol and the communication data may be built into the processor, or may be stored in the storage unit in the form of a software program, and the processor executes the software program to implement the baseband processing function.

[0488] For example, in the embodiment of the present application, the antenna and the control circuit having the transceiver function may be regarded as the transceiver unit 4100 of the terminal 4000, and the processor having the processing function may be regarded as the processing unit 4200 of the terminal 4000. Fig.28 As shown, the terminal 4000 includes a transceiver unit 4100 and a processing unit 4200. The transceiver unit may also be referred to as a transceiver, a transceiver, a transceiver device, etc. Optionally, the device used to implement the receiving function in the transceiver unit 4100 may be regarded as a receiving unit, and the device used to implement the sending function in the transceiver unit 4100 may be regarded as a sending unit, that is, the transceiver unit 4100 includes a receiving unit and a sending unit. Exemplarily, the receiving unit may also be referred to as a receiver, a receiver, a receiving circuit, etc., and the sending unit may be referred to as a transmitter, a transmitter, or a transmitting circuit, etc.

[0489] Fig.29 The present invention provides a schematic diagram of the structure of a network device 5000 provided in an embodiment of the present application. The above-mentioned communication device 2000 or communication device 3000 can be configured in the network device 5000. Alternatively, the communication device 2000 or communication device 3000 itself can be the network device 5000. Alternatively, the network device 5000 can perform the actions performed by the network device in the above-mentioned method embodiment.

[0490] like Fig.29As shown, the network device 5000 may include one or more DUs 5010 and one or more CUs 5020. The CU 5020 may communicate with the NG core (Next Generation Core Network, NC). The DU 5010 may include at least one antenna 5011, at least one radio frequency unit 5012, at least one processor 5013 and at least one memory 5014. The DU 5010 part is mainly used for receiving and transmitting radio frequency signals, converting radio frequency signals to baseband signals, and part of baseband processing. The CU5020 may include at least one processor 5022 and at least one memory 5021. The CU 5020 and the DU 5010 may communicate through an interface, wherein the control plane (CP) interface may be Fs-C, such as F1-C, and the user plane (UP) interface may be Fs-U, such as F1-U.

[0491] The CU 5020 part is mainly used for baseband processing, controlling the network device 5000, etc. The DU 5010 and the CU 5020 can be physically set together or physically separated, that is, a distributed base station. The CU 5020 is the control center of the network device 5000, which can also be called a processing unit, and is mainly used to complete the baseband processing function. For example, the CU 5020 can be used to control the network device 5000 to execute the operation process of the network device in the above method embodiment.

[0492] Specifically, the baseband processing on the CU and DU can be divided according to the protocol layer of the wireless network, for example, the functions of the PDCP layer and above protocol layers are set in the CU, and the functions of the protocol layers below the PDCP, such as the RLC layer and the MAC layer, are set in the DU. For another example, the CU implements the functions of the RRC layer and the PDCP layer, and the DU implements the functions of the RLC layer, the MAC layer, and the PHY layer.

[0493] In addition, optionally, the network device 5000 may include one or more radio frequency units (RUs), one or more DUs, and one or more CUs. The DU may include at least one processor 5013 and at least one memory 5014, the RU may include at least one antenna 5011 and at least one radio frequency unit 5012, and the CU may include at least one processor 5022 and at least one memory 5021.

[0494] In one example, the CU 5020 may be composed of one or more single boards, and multiple single boards may jointly support a wireless access network with a single access indication (such as a 5G network), or may respectively support wireless access networks with different access standards (such as an LTE network, a 5G network, or other networks). The memory 5021 and the processor 5022 may serve one or more single boards. In other words, a memory and a processor may be separately set on each single board. It may also be that multiple single boards share the same memory and processor. In addition, necessary circuits may be set on each single board. The DU 5010 may be composed of one or more single boards, and multiple single boards may jointly support a wireless access network with a single access indication (such as a 5G network), or may respectively support wireless access networks with different access standards (such as an LTE network, a 5G network, or other networks). The memory 5014 and the processor 5013 may serve one or more single boards. In other words, a memory and a processor may be separately set on each single board. It may also be that multiple single boards share the same memory and processor. In addition, necessary circuits may also be set on each single board.

[0495] It should be understood that Fig.29 The network device 5000 shown can implement each process of the actions performed by the network device in the above method embodiment. The operations and / or functions of each module in the network device 5000 are respectively to implement the corresponding processes in the above method embodiment. For details, please refer to the description in the above method embodiment. To avoid repetition, the detailed description is appropriately omitted here.

[0496] It should be understood that Fig.29 The network device 5000 shown is only a possible architecture of the network device and should not constitute any limitation to the present application. The method provided in the present application can be applied to network devices of other architectures. For example, a network device including a CU, a DU, and an AAU, or a network device that does not adopt a CU-DU separation architecture. The present application does not limit the specific architecture of the network device.

[0497] It should be understood that in a possible design, each step in the method embodiment provided by the present application can be completed by an integrated logic circuit of hardware in a processor or an instruction in the form of software. The steps of the method disclosed in conjunction with the embodiment of the present application can be directly embodied as a hardware processor for execution, or a combination of hardware and software modules in a processor for execution. The software module can be located in a mature storage medium in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in a memory, and the processor reads the information in the memory and completes the steps of the above method in conjunction with its hardware. To avoid repetition, it is not described in detail here.

[0498] It should be noted that the processor in the embodiment of the present application can be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method embodiment can be completed by the hardware integrated logic circuit or software instructions in the processor. The above processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in the embodiment of the present application can be directly embodied as a hardware decoding processor to perform, or the hardware and software modules in the decoding processor are combined and performed. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, and other mature storage media in the art. The storage medium is located in a memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.

[0499] It can be understood that the memory in the embodiments of the present application can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory can be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), and direct RAM (DR RAM). It should be noted that the memory of the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0500] The present application also provides a computer program product, which includes: computer program code, when the computer program code is run on a computer, the computer executes each step or process executed by the terminal or network device in any of the above method embodiments.

[0501] The present application also provides a computer-readable storage medium, which stores program code. When the program code runs on a computer, the computer executes each step or process executed by the terminal or network device in any of the above method embodiments.

[0502] The present application also provides a communication device, including a processor and an interface, wherein the interface is used to send and / or receive signals, so that the processor executes each step or process executed by the terminal or network device in any of the above method embodiments.

[0503] The present application also provides a communication system, which includes at least one of a terminal and a network device.

[0504] The above-mentioned various device embodiments and method embodiments completely correspond to each other, and the corresponding steps are executed by the corresponding modules or units. For example, the communication unit or communication interface executes the receiving or sending steps in the method embodiment, and the other steps except sending and receiving can be executed by the processing unit or processor.

[0505] In the embodiments of the present application, each term and English abbreviation is provided for the convenience of description and shall not constitute any limitation to the present application. The present application does not exclude the possibility of defining other terms that can achieve the same or similar functions in existing or future protocols.

[0506] The terms "component", "module", "system", etc. used in this specification are used to represent computer-related entities, hardware, firmware, a combination of hardware and software, software, or software in execution. For example, a component can be, but is not limited to, a process running on a processor, a processor, an object, an executable file, an execution thread, a program and / or a computer. By way of illustration, both applications running on a computing device and a computing device can be components. One or more components may reside in a process and / or an execution thread, and a component may be located on a computer and / or distributed between two or more computers. In addition, these components may be executed from various computer-readable storage media having various data structures stored thereon. Components may, for example, communicate through local and / or remote processes according to signals having one or more data packets (e.g., data from two components interacting with another component between a local system, a distributed system and / or a network, such as the Internet interacting with other systems through signals).

[0507] Those of ordinary skill in the art will appreciate that the various illustrative logical blocks and steps described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or in a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0508] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can be based on the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0509] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0510] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0511] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0512] In the above embodiments, the functions of each functional unit can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions (programs). When the computer program instructions (programs) are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transmitted from a website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server, data center, etc. that contains one or more available media integrated. The available medium may be a magnetic medium (eg, a floppy disk, a hard disk, a magnetic tape), an optical medium (eg, a DVD), or a semiconductor medium (eg, a solid state disk (SSD)).

[0513] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0514] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A communication method, characterized in that: The method comprises: Acquire a first data set, where the first data set is used to verify at least one first neural network model; Sending feedback information, wherein the feedback information is used to indicate the performance of the at least one first neural network model.

2. The method according to claim 1, characterized in that: The at least one first neural network model is preconfigured or acquired through model information of the at least one first neural network model.

3. The method according to claim 1 or 2, characterized in that The first time unit during which the feedback information is sent is the same as the first time unit during which the first data set is acquired.

4. The method according to claim 2, characterized in that The second time unit where the first data set is located is the same as the second time unit where the model information is located.

5. The method according to claim 2 or 4, characterized in that The third time unit in which the feedback information is sent is the same as the third time unit in which the model information is sent and the third time unit in which the first data set is acquired.

6. The method according to claim 2, 4 or 5, characterized in that The start time of the symbol carrying the model information does not lag behind the start time of the symbol carrying the first data set.

7. The method according to any one of claims 1 to 6, characterized in that The at least one first neural network model is trained based on the second neural network model and the second data set.

8. The method according to any one of claims 1 to 7, characterized in that Before sending the feedback information, the method further includes: The at least one first neural network model is verified using the first data set and the at least one first neural network model.

9. The method according to any one of claims 1 to 8, characterized in that The first data set includes sample data, or the first data set includes sample data and sample labels corresponding to the sample data.

10. The method according to any one of claims 1 to 9, characterized in that Before sending the feedback information, the method further includes: A control channel is received, where the control channel includes a sample label or indication information corresponding to the sample data in the first data set, where the indication information indicates the sample label.

11. The method according to any one of claims 1 to 10, characterized in that The fourth time unit where the control channel is located is the same as the fourth time unit where one or more of the following are located: sample data in the first data set, model information of the at least one first neural network model, or the feedback information.

12. A communication method, characterized in that: The method comprises: Sending model information and / or a first data set of at least one first neural network model, wherein the first data set is used to verify the at least one first neural network model; Feedback information is received, where the feedback information is used to indicate the performance of the at least one first neural network model.

13. The method according to claim 12, characterized in that The first time unit where the feedback information is located is the same as the first time unit where the first data set is located.

14. The method according to claim 12 or 13, characterized in that The second time unit where the first data set is located is the same as the second time unit where the model information is located.

15. The method according to any one of claims 12 to 14, characterized in that The third time unit where the feedback information is located is the same as the third time unit where the model information is located and the third time unit where the first data set is located.

16. The method according to any one of claims 12 to 15, characterized in that The start time of the symbol carrying the model information does not lag behind the start time of the symbol carrying the first data set.

17. The method according to any one of claims 12 to 16, characterized in that The at least one first neural network model is trained based on at least one second neural network model and a second data set.

18. The method according to any one of claims 12 to 17, characterized in that The first data set includes sample data, or the first data set includes sample data and sample labels corresponding to the sample data.

19. The method according to any one of claims 12 to 18, characterized in that The method further comprises: A control channel is sent, where the control channel includes a sample label or indication information corresponding to the sample data in the first data set, where the indication information indicates the sample label.

20. The method according to any one of claims 12 to 19, characterized in that The fourth time unit where the control channel is located is the same as the fourth time unit where one or more of the following are located: sample data in the first data set, the model information, or the feedback information.

21. A communication device, characterized in that: The method comprises means for executing the steps of the method according to any one of claims 1 to 11.

22. A communication device, characterized in that: The method comprises means for executing the steps of the method according to any one of claims 12 to 20.

23. A communication device, characterized in that: The device comprises a processor, and when the processor executes a program or instruction stored in a memory, the device executes a method as described in any one of claims 1-11 or any one of claims 12-20.

24. A communication device, characterized in that: The method comprises a processor and an interface, wherein the interface is used to send and / or receive signals, so that the processor executes the method according to any one of claims 1 to 11 or any one of claims 12 to 20.

25. A readable storage medium having a computer program or instruction stored thereon, characterized in that: When the computer program or instruction is executed, the computer is caused to perform the method according to any one of claims 1 to 11 or any one of claims 12 to 20.

26. A computer program product, characterized in that The method comprises computer program instructions, wherein the computer program instructions cause the computer to execute the method according to any one of claims 1 to 11 or any one of claims 12 to 20.

Images