Model configuration methods and apparatuses, devices and storage medium

Abstract

The present application belongs to the technical field of mobile communications. Disclosed are model configuration methods and apparatuses, devices and a storage medium. A method comprises: sending first information to a second device, the first information being used for indicating the normalization mode of input data of a first AI model, and the normalization mode comprising at least one of a power normalization mode and a phase normalization mode. The first information indicates the normalization mode of the input data of the first AI model, such that a device deploying an AI model can normalize input data of the AI model on the basis of the normalization mode used during training of the AI model, thereby ensuring that the normalization mode used for input data during model inference is the same as that used for the input data during model training, and ensuring the inference performance of the AI model.

Description

Model configuration method, device, apparatus, and storage medium TECHNICAL FIELD

[0001] The present application relates to the technical field of mobile communication, and in particular to a model configuration method, device, apparatus, and storage medium. BACKGROUND

[0002] In the process of communication between a terminal and a network device, using an artificial intelligence (AI) model to process information involved helps to improve the performance of communication between the terminal and the network device.

[0003] In training an AI model, input data in a training data set used in training is usually normalized to reduce the complexity of training. When using the trained AI model for inference, if the input data of the model is not normalized in the same way as the training data set (for example, no normalization is performed, or a different normalization method is used), the AI model will have difficulty in correctly processing the input data, resulting in incorrect output results. In particular, when the device training the model and the device deploying the model are different, how to ensure the consistency of the input data of training and inference, and thus ensure the inference performance of the AI model, is a problem to be solved.

[0004] SUMMARY

[0005] The present application provides a model configuration method, device, apparatus, and storage medium. The technical solution is as follows:

[0006] According to an aspect of the present application, a model configuration method is provided, which is performed by a first device, and the method comprises:

[0007] sending first information to a second device, the first information being used to indicate a normalization method of input data of a first AI model;

[0008] The normalization method comprises at least one of a power normalization method and a phase normalization method.

[0009] According to another aspect of the present application, a model configuration method is provided, which is performed by a second device, and the method comprises:

[0010] receiving first information sent by a first device, the first information being used to indicate a normalization method of input data of a first AI model;

[0011] The normalization method comprises at least one of a power normalization method and a phase normalization method.

[0012] According to another aspect of the present application, a configuration apparatus of a model is provided, the apparatus comprising:

[0013] a sending module configured to send first information to a second device, the first information being used to indicate a normalization manner of input data of the first AI model;

[0014] wherein the normalization manner comprises at least one of a power normalization manner and a phase normalization manner.

[0015] According to another aspect of the present application, a configuration apparatus of a model is provided, the apparatus comprising:

[0016] a receiving module configured to receive first information sent by a first device, the first information being used to indicate a normalization manner of input data of the first AI model;

[0017] wherein the normalization manner comprises at least one of a power normalization manner and a phase normalization manner.

[0018] According to another aspect of the present application, a first device is provided, the first device comprising:

[0019] a processor;

[0020] a transceiver connected to the processor;

[0021] a memory configured to store executable instructions of the processor;

[0022] wherein the processor is configured to load and execute the executable instructions to implement the configuration method of the model according to the above aspects.

[0023] According to another aspect of the present application, a second device is provided, the second device comprising:

[0024] a processor;

[0025] a transceiver connected to the processor;

[0026] a memory configured to store executable instructions of the processor;

[0027] wherein the processor is configured to load and execute the executable instructions to implement the configuration method of the model according to the above aspects.

[0028] According to another aspect of the present application, a computer readable storage medium is provided, the computer readable storage medium storing a computer program, the computer program being used to be executed by a processor to implement the configuration method of the model.

[0029] According to another aspect of the present application, a chip is provided, which includes programmable logic circuitry and / or program instructions, and when the chip is running on a communication device, is used to implement the configuration method of the above model.

[0030] According to another aspect of the present application, a computer program product is provided, which includes computer instructions stored in a computer readable storage medium; a processor of a communication device reads the computer instructions from the computer readable storage medium and executes the computer instructions, so that the communication device implements the configuration method of the above model.

[0031] According to another aspect of the present application, a computer program is provided, which is executed by a processor of a communication device to implement the configuration method of the above model.

[0032] The technical solutions provided by the embodiments of the present application have at least the following beneficial effects:

[0033] The first information indicates the normalization manner of the input data of the first AI model, so that the device deploying the AI model can normalize the input data of the AI model according to the normalization manner used when the AI model is trained, thereby ensuring that the normalization manner of the input data during model inference is the same as the normalization manner of the input data during model training, so as to ensure the performance of AI model inference. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0035] FIG. 1 is a schematic diagram of a neuron node provided by an example embodiment of the present application;

[0036] FIG. 2 is a schematic diagram of a neural network model provided by an example embodiment of the present application;

[0037] FIG. 3 is a structural schematic diagram of a neural network model for CSI feedback provided by an example embodiment of the present application;

[0038] FIG. 4 is a schematic diagram of the system architecture of a communication system provided by an example embodiment of the present application;

[0039] FIG. 5 is a flowchart of a model configuration method provided by an example embodiment of the present application;

[0040] FIG. 6 is a flow chart of a method for configuring a model according to an example embodiment of the present application;

[0041] FIG. 7 is a flow chart of a method for configuring a model according to an example embodiment of the present application;

[0042] FIG. 8 is a flow chart of a method for CSI prediction according to an example embodiment of the present application;

[0043] FIG. 9 is a block diagram of a device for configuring a model according to an example embodiment of the present application;

[0044] FIG. 10 is a block diagram of a device for configuring a model according to an example embodiment of the present application;

[0045] FIG. 11 is a structural diagram of a communication device according to an example embodiment of the present application. DETAILED DESCRIPTION

[0046] To make the objects, technical solutions and advantages of the present application clearer, the following will further describe the embodiments of the present application in detail with reference to the drawings. The example embodiments will be described in detail, and the examples are shown in the drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following example embodiments do not represent all the embodiments consistent with the present application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the present application as detailed in the appended claims. All other embodiments obtained by those of ordinary skill in the art without creative work on the basis of the embodiments in the present application are within the scope of protection of the present application. The terms used in the present disclosure are merely for the purpose of describing specific embodiments and are not intended to limit the present disclosure. The singular forms "a", "an" and "the" used in the present disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein means and includes any or all possible combinations of one or more associated listed items. It should be understood that although the terms first, second, third, etc. can be used in the present disclosure to describe various information, these information should not be limited to these terms. These terms are only used to distinguish one type of information from another. For example, without departing from the scope of the present disclosure, the first information can also be referred to as the second information, and similarly, the second information can also be referred to as the first information. Depending on the context, the word "if" as used herein can be interpreted as "when" or "upon determining" or "in response to determining".

[0047] The technical solutions described in some embodiments of the present application can be applied to various communication systems, for example: Global System for Mobile Communications (GSM), Code Division Multiple Access (CDMA) system, Wideband Code Division Multiple Access (WCDMA) system, General Packet Radio Service (GPRS) system, Long Term Evolution (LTE) system, Advanced Long Term Evolution (LTE-A) system, New Radio (NR) system, evolved system of NR system, LTE-based access to unlicensed spectrum (LTE-U) system, NR-based access to unlicensed spectrum (NR-U) system, Non-Terrestrial Networks (NTN) system, Universal Mobile Telecommunication System (UMTS), Wireless Local Area Networks (WLAN) system, Wireless Fidelity (WiFi) system, 5th Generation Mobile Communication Technology (5G) system, cellular Internet of Things system, cellular passive Internet of Things system, and can also be applied to the evolved system after 5G NR system, and can also be applied to 6th Generation Mobile Communication Technology (6G) system and the evolved system thereafter.

[0048] It should be understood that in some embodiments of the present application, "5G" can also be referred to as "5G NR" or "NR".

[0049] It should be understood that in the description of the embodiments of the present application, the term "corresponding" can represent a direct or indirect corresponding relationship between the two, can also represent an associated relationship between the two, or can indicate a relationship such as indicated, configured, and the like.

[0050] An AI model is introduced:

[0051] In some embodiments, the AI model comprises at least one of a neural network (NN) model, a deep learning (DL) model, a deep neural networks (DNN) model, a convolutional neural networks (CNN) model, a recurrent neural network (RNN) model, and a machine learning (ML) model. It should be noted that the AI model can also comprise a model other than the above models, which is not limited in the present application.

[0052] The neural network model is introduced as follows:

[0053] The neural network model is an operation model composed of multiple neuron nodes (neurons) connected to each other, wherein the connection between the nodes represents the weighted value from the input signal to the output signal, which is called weight; each node performs weighted summation on different input signals and outputs through a corresponding activation function. For example, FIG. 1 is a schematic diagram of a neuron node provided in an example embodiment of the present application. As shown in FIG. 1, the input data of the neuron node 101 includes a1-a n and 1, wherein a1-a n correspond to weights w1-w n , respectively, and 1 corresponds to the weight b. After inputting the input data into the neuron node 101, the neuron node 101 performs weighted summation on different input data according to the weights corresponding to the input data, and then calculates the result of the weighted summation through the activation function (f) to obtain the output (t) 102 of the neuron node 101.

[0054] For example, FIG. 2 is a schematic diagram of a neural network model (neural network) provided in an example embodiment of the present application. As shown in FIG. 2, the neural network model comprises an input layer (Input Layer) 201, a hidden layer (Hidden Layer) 202, and an output layer (Output Layer) 203. By setting different connection modes for multiple neurons and setting corresponding weights and activation functions for the neurons, different outputs can be generated, thereby fitting the mapping relationship from the input to the output of the neural network model. In some embodiments, each upper-level node (neuron) in the neural network model is connected to all lower-level nodes (neurons). In this case, the neural network model is a fully connected model, which can also be called a deep neural network.

[0055] Introduction to AI model for communication

[0056] Through the processes of constructing a data set for an AI model, training the AI model, verifying the AI model, and testing the AI model, a neural network model (AI model) can be trained and obtained. In some embodiments, in the case where the neural network model is used for wireless communication between a terminal and a network device, the training of the neural network model can be divided into offline training and online training. The network device can obtain a static training result through offline training using a data set, which can be referred to as offline training here. During the use of the neural network model by the network device or the terminal, the network device can continue to collect more data for real-time online training to optimize the parameters of the neural network model and achieve better inference and prediction results as the terminal further measures and / or reports, which can be referred to as online training here.

[0057] After obtaining the trained neural network model, the corresponding model output can be inferred by inputting the current obtained information into the neural network model. When the neural network model is used for wireless communication, it can be divided into a single-end model and a double-end model. Among them, the single-end model can be used only after being deployed on one side of the terminal or the network device, and the training of the model can also be performed only on one side; the double-end model needs to be deployed in pairs on the terminal and the network device side, and the models on both sides need to be trained together, that is, the models deployed on both sides are corresponding and cannot be used or updated alone.

[0058] In order to realize different functions, different AI models can be introduced, and corresponding model inputs and model outputs can be defined. In some embodiments, when the AI model is used for channel state information (CSI) feedback, the channel information (such as eigenvectors, beam information, delay information, etc.) measured based on reference signals can be used as the input of the model on one side, so as to infer the corresponding CSI quantization bits. On the other side, there will be a corresponding AI model that takes the CSI quantization bits as input, and can infer the corresponding channel information. In some embodiments, when the AI model is used for data detection, that is, an AI receiver is used for detecting downlink signals, the terminal can use the received signal (and some additional information such as pilot sequences) as the input of the AI model for inference, so as to output the detected soft bit symbols (i.e., constellation points). In some embodiments, when the AI model is used for beam management, the obtained multiple reference signal receiving power (RSRP) values can be used as the input of the model, so as to output the optimal beam information and the optimal RSRP value. In addition, the AI model can also be used for positioning, channel coding, channel decoding, channel estimation, and other processes.

[0059] For example, FIG. 3 is a structural diagram of a neural network model for CSI feedback according to an example embodiment of the present application. As shown in FIG. 3, an encoder 301 for encoding is deployed at the terminal side, and a decoder 302 for decoding corresponding to the encoder 301 is deployed at the network device side. The terminal outputs CSI quantization bits, such as precoding matrix indicator (PMI) quantization bits, based on the encoder 301, and feeds back the CSI (PMI) quantization bits to the network device through an uplink channel (such as a physical uplink shared channel (PUSCH) or a physical uplink control channel (PUCCH)). The network device takes the CSI (PMI) quantization bits fed back by the terminal as the model input of the decoder 302, and thus outputs channel information corresponding to the input information of the terminal, such as a feature vector of each subband, for precoding of the network device for the downlink. In some embodiments, the terminal also performs performance monitoring on the AI model, and reports to the network device when the AI model performance is not good, so that the network device updates the AI model. Since the neural network models on both sides are matched, if the model is updated, the terminal needs to be notified through signaling.

[0060] During the training of the AI model, the input data in the data set is usually normalized (such as power normalization or phase normalization), so as to reduce the complexity of model training. When using the trained AI model for inference, if the input data of the AI model used for inference is not normalized in the same way as the input data in the data set used for training, the AI model will not be able to process the input data, resulting in that the inference result of the AI model is not available. In particular, when the device for training the model and the device for deploying the model are different, if the device for deploying the model does not know the data normalization method used when the model is trained, the device for deploying the model will not be able to correctly normalize the input data, resulting in that the correct model inference result cannot be obtained.

[0061] The method provided in the application can realize, in the case of model configuration, model reporting or model updating, carrying information indicating the normalization mode of input data of an AI model in information indicating model parameters of the AI model, enabling a device deploying the AI model to normalize input data of the AI model according to the normalization mode used during training of the AI model, thereby ensuring that the normalization mode of input data during model inference is the same as the normalization mode of input data in a data set during model training, so as to ensure the performance of AI model inference. Moreover, normalization of a data set during AI model training can make input data more regular (for example, all negative or all positive), thereby improving the reliability of the AI model.

[0062] FIG. 4 shows a schematic diagram of a system architecture of a communication system 400 provided in an embodiment of the application. The system architecture can include a terminal 10, an access network device 20 and a core network device 30.

[0063] The terminal 10 can refer to a UE (User Equipment), an access terminal, a user unit, a user station, a mobile station, a mobile station, a remote station, a remote terminal, a mobile device, a wireless communication device, a user agent or a user device. Alternatively, the terminal can also be a cellular phone, a cordless phone, a SIP (Session Initiation Protocol) phone, a WLL (Wireless Local Loop) station, a PDA (Personal Digital Assistant), a handheld device with wireless communication function, a computing device or other processing device connected to a wireless modem, a vehicle-mounted device, a wearable device, a terminal in a 5GS (5th Generation System) or a terminal in a future evolved PLMN (Public Land Mobile Network), etc. The embodiments of the application are not limited thereto. For convenience of description, the above-mentioned devices are collectively referred to as terminals.

[0064] It should be noted that the number of terminals 10 is usually multiple, and one or more terminals 10 can be distributed in the cell managed by each access network device 20. Moreover, one or more terminals 10 can also be distributed outside the cell managed by the access network device 20. Among them, different terminals 10 can communicate based on sidelink.

[0065] The access network device 20 is a device deployed in an access network to provide wireless communication functions for the terminal 10. The access network device 20 can include various forms of macro base stations, micro base stations, relay stations, access points, and the like. In systems using different wireless access technologies, the names of devices with access network device functions can be different, for example, in a 5G NR system, it is called gNodeB or gNB. As communication technology evolves, the name of the "access network device" can change. For ease of description, in the embodiments of the present application, the above-mentioned devices that provide wireless communication functions for the terminal 10 are collectively referred to as access network devices. Optionally, through the access network device 20, a communication relationship can be established between the terminal 10 and the core network device 30. Illustratively, in a long term evolution (Long Term Evolution, LTE) system, the access network device 20 can be an EUTRAN (Evolved Universal Terrestrial Radio Access Network) or one or more eNodeBs in the EUTRAN; in a 5G NR system, the access network device 20 can be a RAN or one or more gNBs in the RAN.

[0066] The main function of the core network device 30 is to provide user connection, management of users, and completion of bearer for services, and to provide an interface to an external network as a bearer network. For example, the core network device in a 5G NR system can include an AMF (Access and Mobility Management Function) entity, a UPF (User Plane Function) entity, and an SMF (Session Management Function) entity, and the like. The access network device 20 and the core network device 30 can be collectively referred to as network devices.

[0067] In one example, the access network device 20 and the core network device 30 communicate with each other through some air technology, such as an NG interface in a 5G NR system. The access network device 20 and the terminal 10 communicate with each other through some air technology, such as a Uu interface. The terminal 10 and the terminal 10 communicate with each other through some air technology, such as a PC5 interface.

[0068] FIG. 5 is a flowchart of a configuration method of a model according to an example embodiment of the present application. The method can be performed by a first device. The method includes:

[0069] Step 502: sending first information to a second device, the first information being used to indicate a normalization manner of input data of the first AI model.

[0070] In some embodiments, the first AI model comprises a model implemented based on AI-related technologies. In some embodiments, the AI model in the present application comprises at least one of a neural network (NN) model, a deep learning (DL) model, a deep neural network (DNN) model, a convolutional neural network (CNN) model, a recurrent neural network (RNN) model, and a machine learning (ML) model.

[0071] In some embodiments, the first AI model is deployed in the second device. In some embodiments, the output information of the first AI model is used for communication between the first device and the second device. In some embodiments, the output information of the first AI model is related to the communication between the first device and the second device. In some embodiments, the first AI model is used for the communication between the first device and the second device. In some embodiments, the input information of the first AI model comprises information involved in the communication process between the first device and the second device. In some embodiments, the first AI model affects the communication process between the first device and the second device. In some embodiments, the deployment of the first AI model in the second device comprises at least one of the following: the first AI model is a model deployed in the second device alone; the first AI model is deployed in the second device, and a second AI model corresponding to the first AI model deployed in the second device is deployed in the first device. In some embodiments, in the case that a second AI model corresponding to the first AI model is deployed in the first device, the output information of the first AI model deployed in the second device is used as the input information of the second AI model deployed in the first device, or the output information of the second AI model deployed in the first device is used as the input information of the first AI model deployed in the second device.

[0072] In some embodiments, the first device comprises a terminal, and the second device comprises a network device. In some embodiments, the first device comprises a network device, and the second device comprises a terminal.

[0073] In some embodiments, the use of the first AI model comprises at least one of: CSI prediction; CSI compression; beam management; beam prediction; data detection. In some embodiments, in the case that the first AI model is used for CSI prediction (CSI feedback), the input information of the first AI model comprises at least one of: channel information obtained based on reference signal measurement, such as channel eigenvectors, beam information, delay information, and the output information of the first AI model comprises corresponding CSI quantization bits, such as PMI quantization bits; or the input information of the first AI model comprises CSI quantization bits, such as PMI quantization bits, and the output information of the first AI model comprises corresponding channel information, such as eigenvectors of each subband of the channel, for precoding of the downlink. In some embodiments, in the case that the first AI model is used for CSI compression, the input information of the first AI model comprises channel eigenvectors, which are obtained by measurement, and the output information of the first AI model comprises PMI quantization bits, for CSI compression. In some embodiments, in the case that the first AI model is used for beam management and / or beam prediction, the input information of the first AI model comprises a plurality of RSRP values (a RSRP set) measured on reference signals in a first reference signal set, and the output information of the first AI model comprises reference signal indication information corresponding to a second reference signal set and / or RSRP values corresponding to the second reference signal set, the second reference signal set being an optimal reference signal set corresponding to the first reference signal set, and the reference signal indication information being used to indicate reference signals in the second reference signal set. In some embodiments, in the case that the first AI model is used for data detection, the input data of the first AI model comprises received signals and / or information related to the received signals, such as pilot sequences and some additional information, and the output information of the first AI model comprises soft bit symbols, such as constellation points. It should be noted that, in addition to the above uses, the first AI model in the present application can also have other uses, which are not limited in the present application.

[0074] In some embodiments, the first information is used to indicate a normalization manner adopted on input data when inference is performed by the first AI model, and the normalization manner indicated by the first information is the same as a normalization manner adopted on input data when the first AI model is trained. In some embodiments, the first information is used to indicate a normalization manner adopted on input data when the first AI model is trained. In some embodiments, the first information is used to indicate a normalization manner adopted on input data when a model corresponding to the first AI model is trained, and the output information of the first AI model is used as input information of the model corresponding to the first AI model, or the output information of the model corresponding to the first AI model is used as input information of the first AI model.

[0075] In some embodiments, the first information is carried in second information, and the second information includes / carrys the first information. The second information is used to indicate the parameters of the first AI model. In some embodiments, the second information includes the parameters of the first AI model, and the parameters of the first AI model include the first information. In some embodiments, the first information is equivalent to / replacable by normalization information, normalization manner information, normalization manner indication information, input data normalization manner information. In some embodiments, the second information is equivalent to / replacable by the parameters of the first AI model, model parameters, parameter indication information, model parameter indication information.

[0076] In some embodiments, the first information is sent in at least one of the following situations: model configuration is performed; model reporting is performed; model updating is performed.

[0077] In some embodiments, the normalization manner indicated by the first information includes at least one of a power normalization manner and a phase normalization manner. In some embodiments, the power normalization manner refers to a manner of normalizing input data based on the power value of the input data. For example, different input data is scaled in proportion to the power value. In some embodiments, the phase normalization manner refers to a manner of normalizing input data based on the phase of the input data. For example, different input data is offset by the same phase. It should be noted that in addition to the above-mentioned power normalization manner and phase normalization manner, the first information can also indicate other normalization manners, which are not limited in the present application. For example, other normalization manners include the min-max normalization manner, which is represented as x2=(x1-x min ) / (x max -x min ), where x2 represents the normalized input data, x1 represents the input data before normalization, x min represents the input data with the minimum value, and x max represents the input data with the maximum value.

[0078] In some embodiments, the power normalization manner comprises at least one of the following: power normalization with the first input data as the reference; power normalization with the last input data as the reference; power normalization with the input data having the highest power value as the reference; power normalization with the input data having the lowest power value as the reference; and no power normalization. In some embodiments, power normalization with a certain input data as the reference means that the power value of the input data as the reference is processed according to a preset manner, and other input data is processed in the same manner as the processing of the input data as the reference. No power normalization means that no power normalization processing is performed on the input data. In some embodiments, the first AI model corresponds to an input data set, the input data set comprises data for inputting the first AI model, the first input data is the first data in the input data set, the last input data is the last data in the input data set, the input data having the highest power value is the input data having the highest power value in the input data set, and the input data having the lowest power value is the input data having the lowest power value in the input data set.

[0079] In some embodiments, the power value of the input data as the reference after power normalization comprises at least one of the following: 1; 0 dB; and 0 dBm. In some embodiments, power normalization comprises scaling all input data by the power value in a proportional manner. For example, all input data is scaled by power reduction or amplification (equivalent to subtracting or adding the same power dB value), and the scaling ratio of power reduction or amplification depends on the power normalization process of the input data as the reference. For example, the input data set comprises k input data, and in the process of power normalization with the first input data in the input data set as the reference, the first input data is scaled by power to make the power of the first input data 1, and other input data in the input data set is scaled by power in the same manner, so as to complete the power normalization of all input data. In some embodiments, the input data subjected to normalization processing based on the power normalization manner comprises at least one of the following: an amplitude value; a transmission power value; an RSRP value; an RSRP difference value; and a signal to interference plus noise ratio (SINR) value.

[0080] In some embodiments, the phase normalization manner comprises at least one of the following: phase normalization with the first input data as the reference; phase normalization with the last input data as the reference; phase normalization with the input data having the highest amplitude value as the reference; phase normalization with the input data having the lowest amplitude value as the reference; and no phase normalization. In some embodiments, phase normalization with a certain input data as the reference means that the phase of the input data as the reference is processed according to a preset manner, and other input data is processed in the same manner as the processing of the input data as the reference. No phase normalization means that no phase normalization processing is performed on the input data. In some embodiments, the first AI model corresponds to an input data set, the input data set comprises data for inputting the first AI model, the first input data is the first data in the input data set, the last input data is the last data in the input data set, the input data having the highest amplitude value is the data having the highest amplitude value in the input data set, and the input data having the lowest amplitude value is the data having the lowest amplitude value in the input data set.

[0081] In some embodiments, the phase of the input data as the reference after phase normalization is 0, or the imaginary part of the input data as the reference after phase normalization is 0. In some embodiments, phase normalization comprises the same phase offset of the input data. For example, all input data is offset by the same phase, and the same phase offset depends on the phase normalization process of the input data as the reference. For example, the input data set comprises k input data, and in the process of phase normalization with the first input data in the input data set as the reference, the first input data is offset by a target phase to make the phase of the first input data 0, and other input data in the input data set is also offset by the target phase, thereby completing the phase normalization of all input data. In some embodiments, the input data subjected to normalization processing based on the phase normalization manner comprises at least one of the following: a feature vector; a channel matrix; and a channel covariance matrix.

[0082] In some embodiments, after receiving the first information, the second device performs the inference process of the first AI model according to the normalization manner indicated by the first information. In some embodiments, the second device performs normalization processing on the input data according to the normalization manner indicated by the first information, for example, performs power normalization processing or phase normalization processing, and inputs the processed input data into the first AI model to perform the inference process of the first AI model, thereby obtaining the corresponding output information of the first AI model. In some embodiments, the second device inputs the first information and the input data into the first AI model as inputs to perform the inference process of the first AI model, thereby obtaining the corresponding output information of the first AI model. The input data comprises data for the inference of the first AI model.

[0083] In some embodiments, for the case of taking the first information as the input of the first AI model, the first information can be a bit, through which different power normalization manners and / or phase normalization manners can be indicated. For example, the first information can be 2-bit information, through which one of 4 different power normalization manners (corresponding to 4 different reference input data) or one of 4 different phase normalization manners (corresponding to 4 different reference input data) can be indicated, and if the first information is not input, it can be understood as implicitly indicating that no power normalization or no phase normalization is performed. In some embodiments, for the case of taking the first information as the input of the first AI model, the first information can be understood as a feature of the input data for inference of the first AI model, or a normalization layer for normalization processing is provided in the first AI model, and through input of the first information, the manner of normalization processing of the input information by the normalization layer can be indicated.

[0084] In some embodiments, when collecting the data set for model training and inputting the data set for inference to the model for the same AI model, the unified data normalization manner is adopted by different devices for normalization processing of the data set, so as to ensure that all training data and inference data of the AI model applied to the same function / purpose have the same data normalization manner. That is, the data normalization manner is agreed by each collection device and model deployment device of the data set in advance, and is kept consistent in the data collection and model inference process. The training data includes the input data for AI model training, and the inference data includes the input data for AI model inference.

[0085] In some embodiments, when the AI model is a neural network model, the first m (several) input layers (for example, the first layer) of the neural network model are used to implement the function of data normalization, through which the same phase normalization or power normalization processing of the input data of the neural network model in inference and the input data in training can be implemented, so as to ensure that the normalization manners used in the training process and the inference process are consistent. In this way, no matter whether the input data is normalized or not or what normalization manner is adopted, the AI model containing the normalization processing layer can obtain correct inference results.

[0086] In summary, the method provided by the embodiment can indicate the normalization manner of the input data of the first AI model through the first information, so that the device deploying the AI model can normalize the input data of the AI model according to the normalization manner used in the training of the AI model, thereby ensuring that the normalization manner of the input data in the model inference is the same as that of the input data in the model training, so as to ensure the performance of the AI model inference.

[0087] The method provided in the embodiment further guarantees that the power normalization manner of the input data during model inference is the same as the power normalization manner of the input data during model training by providing different power normalization manners. By providing the power value of the reference input data after power normalization and the processing manner of power normalization on different input data, the power normalization on the input data is implemented, which helps to improve the inference efficiency of the model. Different input data for power normalization is provided, and the power normalization on different types of input data is implemented. Different phase normalization manners are provided, so as to guarantee that the phase normalization manner of the input data during model inference is the same as the phase normalization manner of the input data during model training. By providing the phase or imaginary part of the reference input data after phase normalization and the processing manner of phase normalization on different input data, the phase normalization on the input data is implemented, which helps to improve the inference efficiency of the model. Different input data for phase normalization is provided, and the phase normalization on different types of input data is implemented. The first information is carried in the second information indicating the parameters of the model, so that the normalization manner of the input data is indicated synchronously in the process of indicating the parameters of the model, and the signaling overhead is reduced. The first information is sent in different cases, and the normalization manner of the input data of the AI model is indicated flexibly.

[0088] FIG. 6 is a flowchart of a configuration method of a model according to an example embodiment of the present application. The method can be performed by a second device. The method includes:

[0089] Step 602: receiving first information sent by a first device, the first information being used to indicate a normalization manner of input data of a first AI model.

[0090] In some embodiments, the first AI model includes a model implemented based on AI-related technologies. In some embodiments, the AI model in the present application includes at least one of a NN model, a DL model, a DNN model, a CNN model, a RNN model, and a ML model.

[0091] In some embodiments, the first AI model is deployed in the second device. In some embodiments, the output information of the first AI model is used for communication between the first device and the second device. In some embodiments, the output information of the first AI model is related to the communication between the first device and the second device. In some embodiments, the first AI model is used for the communication between the first device and the second device. In some embodiments, the input information of the first AI model includes information involved in the communication process between the first device and the second device. In some embodiments, the first AI model affects the communication process between the first device and the second device.

[0092] In some embodiments, the first device comprises a terminal, and the second device comprises a network device. In some embodiments, the first device comprises a network device, and the second device comprises a terminal.

[0093] In some embodiments, the use of the first AI model comprises at least one of: CSI prediction; CSI compression; beam management; beam prediction; data detection. It should be noted that in addition to the above uses, the first AI model in the present application can also have other uses, which are not limited in the present application.

[0094] In some embodiments, the first information is used to indicate a normalization manner adopted for input data when inference is performed by the first AI model, and the normalization manner indicated by the first information is the same as a normalization manner adopted for input data when the first AI model is trained. In some embodiments, the first information is used to indicate a normalization manner adopted for input data when the first AI model is trained. In some embodiments, the first information is used to indicate a normalization manner adopted for input data when a model corresponding to the first AI model is trained, and output information of the first AI model is used as input information of the model corresponding to the first AI model, or output information of the model corresponding to the first AI model is used as input information of the first AI model.

[0095] In some embodiments, the first information is carried in the second information, and the second information comprises / carrys the first information. The second information is used to indicate parameters of the first AI model. In some embodiments, the second information comprises the parameters of the first AI model, and the parameters of the first AI model comprise the first information. In some embodiments, the first information is equivalent to / replacable by normalization information, normalization manner information, normalization manner indication information, input data normalization manner information. In some embodiments, the second information is equivalent to / replacable by parameters of the first AI model, model parameters, parameter indication information, model parameter indication information.

[0096] In some embodiments, the first information is sent in at least one of the following cases: model configuration is performed; model reporting is performed; model updating is performed.

[0097] In some embodiments, the normalization manner indicated by the first information comprises at least one of a power normalization manner and a phase normalization manner. In some embodiments, the power normalization manner refers to a manner of normalizing input data based on a power value of the input data. In some embodiments, the phase normalization manner refers to a manner of normalizing input data based on a phase of the input data. It should be noted that in addition to the above power normalization manner and phase normalization manner, the first information can also indicate other normalization manners, which are not limited in the present application.

[0098] In some embodiments, the power normalization manner comprises at least one of: power normalization with the first input data as reference; power normalization with the last input data as reference; power normalization with the input data having the highest power value as reference; power normalization with the input data having the lowest power value as reference; and no power normalization. In some embodiments, power normalization with a certain input data as reference means that the power value of the input data as reference is processed according to a preset manner, and other input data is processed in the same manner as the input data as reference. No power normalization means that no power normalization processing is performed on the input data.

[0099] In some embodiments, the power value of the input data as reference after power normalization comprises at least one of: 1; 0 dB; and 0 dBm. In some embodiments, power normalization comprises equal proportion scaling of the input data according to the power value. In some embodiments, the input data subjected to normalization processing based on the power normalization manner comprises at least one of: an amplitude value; a transmission power value; an RSRP value; an RSRP difference value; and an SINR value.

[0100] In some embodiments, the phase normalization manner comprises at least one of: phase normalization with the first input data as reference; phase normalization with the last input data as reference; phase normalization with the input data having the highest amplitude value as reference; phase normalization with the input data having the lowest amplitude value as reference; and no phase normalization. In some embodiments, phase normalization with a certain input data as reference means that the phase of the input data as reference is processed according to a preset manner, and other input data is processed in the same manner as the input data as reference. No phase normalization means that no phase normalization processing is performed on the input data.

[0101] In some embodiments, the phase of the input data as reference after phase normalization is 0, or the imaginary part of the input data as reference after phase normalization is 0. In some embodiments, phase normalization comprises the same phase offset of the input data. In some embodiments, the input data subjected to normalization processing based on the phase normalization manner comprises at least one of: a feature vector; a channel matrix; and a channel covariance matrix.

[0102] In some embodiments, after receiving the first information, the second device performs the inference process of the first AI model according to the normalization manner indicated by the first information. In some embodiments, the second device normalizes the input data according to the normalization manner indicated by the first information, for example, performs power normalization processing or phase normalization processing, and inputs the processed input data into the first AI model to perform the inference process of the first AI model, thereby obtaining the corresponding output information of the first AI model. In some embodiments, the second device inputs the first information and the input data into the first AI model as inputs to perform the inference process of the first AI model, thereby obtaining the corresponding output information of the first AI model. The input data includes data used for inference of the first AI model.

[0103] To sum up, the method provided in this embodiment can realize that the device deploying the AI model can normalize the input data of the AI model according to the normalization manner used during training of the AI model, thereby ensuring that the normalization manner of the input data during inference of the model is the same as the normalization manner of the input data during training of the model, so as to ensure the performance of the AI model inference.

[0104] The method provided in this embodiment also provides different power normalization manners, thereby ensuring that the power normalization manner of the input data during inference of the model is the same as the power normalization manner of the input data during training of the model. By providing the power value of the reference input data after power normalization and the processing manner of power normalization of different input data, the power normalization of the input data is realized, which helps to improve the inference efficiency of the model. By providing different input data used for power normalization, the power normalization of different types of input data is realized. Different phase normalization manners are provided, thereby ensuring that the phase normalization manner of the input data during inference of the model is the same as the phase normalization manner of the input data during training of the model. By providing the phase or imaginary part of the reference input data after phase normalization and the processing manner of phase normalization of different input data, the phase normalization of the input data is realized, which helps to improve the inference efficiency of the model. By providing different input data used for phase normalization, the phase normalization of different types of input data is realized. By carrying the first information in the second information indicating the parameters of the model, the normalization manner of the input data is indicated synchronously in the process of indicating the parameters of the model, thereby reducing the signaling overhead. By sending the first information in different situations, the normalization manner of the input data of the AI model is indicated flexibly.

[0105] FIG. 7 is a flowchart of a method for configuring a model according to an example embodiment of the present application. The method can be used in the system shown in FIG. 4. The method comprises:

[0106] Step 702: The first device sends first information to the second device.

[0107] The first information is used to indicate a normalization manner of input data of the first AI model. In some embodiments, the first AI model comprises a model implemented based on AI-related technologies. In some embodiments, the AI model in the present application comprises at least one of a NN model, a DL model, a DNN model, a CNN model, a RNN model, and a ML model. It should be noted that the AI model can also comprise a model other than the above-mentioned models, which is not limited in the present application.

[0108] In some embodiments, the first AI model is deployed in the second device. In some embodiments, the output information of the first AI model is used for communication between the first device and the second device. In some embodiments, the output information of the first AI model is related to the communication between the first device and the second device. In some embodiments, the first AI model is used for the communication between the first device and the second device. In some embodiments, the input information of the first AI model comprises information involved in the communication process between the first device and the second device. In some embodiments, the first AI model affects the communication process between the first device and the second device.

[0109] In some embodiments, the first device comprises a terminal, and the second device comprises a network device. In some embodiments, the first device comprises a network device, and the second device comprises a terminal.

[0110] In some embodiments, the use of the first AI model comprises at least one of the following: CSI prediction; CSI compression; beam management; beam prediction; data detection. It should be noted that in addition to the above-mentioned uses, the first AI model in the present application can also have other uses, which are not limited in the present application.

[0111] In some embodiments, the first information is used to indicate a normalization manner adopted for input data when inference is performed by the first AI model, and the normalization manner indicated by the first information is the same as a normalization manner adopted for input data when the first AI model is trained. In some embodiments, the first information is used to indicate a normalization manner adopted for input data when the first AI model is trained. In some embodiments, the first information is used to indicate a normalization manner adopted for input data when a model corresponding to the first AI model is trained, and the output information of the first AI model is used as input information of the model corresponding to the first AI model, or the output information of the model corresponding to the first AI model is used as input information of the first AI model.

[0112] In some embodiments, the first information is carried in second information, and the second information includes / carrys the first information. The second information is used to indicate the parameters of the first AI model. In some embodiments, the second information includes the parameters of the first AI model, and the parameters of the first AI model include the first information. In some embodiments, the first information is equivalent to / replacable by normalization information, normalization manner information, normalization manner indication information, input data normalization manner information. In some embodiments, the second information is equivalent to / replacable by the parameters of the first AI model, model parameters, parameter indication information, model parameter indication information.

[0113] In some embodiments, the first information is sent in at least one of the following cases: model configuration is performed; model reporting is performed; model updating is performed.

[0114] In some embodiments, the normalization manner indicated by the first information includes at least one of a power normalization manner and a phase normalization manner. In some embodiments, the power normalization manner refers to a manner of normalizing input data based on the power value of the input data. In some embodiments, the phase normalization manner refers to a manner of normalizing input data based on the phase of the input data. It should be noted that in addition to the above-mentioned power normalization manner and phase normalization manner, the first information can also indicate other normalization manners, which are not limited in the present application.

[0115] In some embodiments, the power normalization manner includes at least one of the following: power normalization is performed with the first input data as a reference; power normalization is performed with the last input data as a reference; power normalization is performed with the input data having the highest power value as a reference; power normalization is performed with the input data having the lowest power value as a reference; no power normalization is performed. In some embodiments, performing power normalization with a certain input data as a reference means that the power value of the input data as a reference is processed in a predetermined manner, and other input data is processed in the same manner as the input data as a reference. No power normalization means that no power normalization processing is performed on the input data.

[0116] In some embodiments, the power value of the input data as a reference after power normalization includes at least one of the following: 1; 0dB; 0dBm. In some embodiments, power normalization includes scaling the input data in proportion to the power value. In some embodiments, the input data subjected to normalization processing based on the power normalization manner includes at least one of the following: amplitude value; transmission power value; RSRP value; RSRP difference value; SINR value.

[0117] In some embodiments, the phase normalization manner comprises at least one of the following: phase normalization with the first input data as reference; phase normalization with the last input data as reference; phase normalization with the input data having the highest amplitude value as reference; phase normalization with the input data having the lowest amplitude value as reference; and no phase normalization. In some embodiments, phase normalization with a certain input data as reference means that the phase of the input data as reference is processed according to a preset manner, and other input data is processed in the same manner as the processing of the input data as reference. No phase normalization means that no phase normalization processing is performed on the input data.

[0118] In some embodiments, the phase of the input data as reference after phase normalization is 0, or the imaginary part of the input data as reference after phase normalization is 0. In some embodiments, phase normalization comprises the same phase offset of the input data. In some embodiments, the input data subjected to normalization processing based on the phase normalization manner comprises at least one of the following: a feature vector; a channel matrix; and a channel covariance matrix.

[0119] Step 704: The second device performs normalization processing on the input data according to the normalization manner indicated by the first information.

[0120] In some embodiments, the second device performs power normalization processing or phase normalization processing on the input data according to the normalization manner indicated by the first information, and inputs the processed input data into the first AI model to perform an inference process of the first AI model, thereby obtaining corresponding output information of the first AI model. The input data comprises data used for inference of the first AI model.

[0121] Step 706: The second device inputs the first information into the first AI model together with the input data.

[0122] In some embodiments, the second device inputs the first information into the first AI model together with the input data to perform an inference process of the first AI model, thereby obtaining corresponding output information of the first AI model. The input data comprises data used for inference of the first AI model.

[0123] In some embodiments, for the case that the first information is input as the input of the first AI model, the first information can be a bit, by which different power normalization manners and / or phase normalization manners can be indicated. For example, the first information can be 2-bit information, by which one of 4 different power normalization manners (corresponding to 4 different reference input data) or one of 4 different phase normalization manners (corresponding to 4 different reference input data) can be indicated, and if the first information is not input, it can be understood as implicitly indicating that no power normalization or no phase normalization is performed.

[0124] It should be noted that the above steps 704 and 706 are parallel steps, and one of them can be selected for execution. For example, in the case that the first AI model does not support input of the first information, steps 702 and 704 are executed; in the case that the first AI model supports input of the first information, steps 702 and 706 are executed.

[0125] In this embodiment, steps 702, 704 and 706 are optional, and in different embodiments, one or more of these steps can be omitted or replaced.

[0126] Step 702 can be implemented as an independent embodiment, such as a configuration method of a model on the first device side or a configuration method of a model on the second device side. Step 704 can be implemented as an independent embodiment, such as a model input method on the second device side. Step 706 can be implemented as an independent embodiment, such as a model input method on the second device side.

[0127] In summary, the method provided in this embodiment can indicate the normalization manner of the input data of the first AI model by the first information, so that the device deploying the AI model can normalize the input data of the AI model according to the normalization manner used when the AI model is trained, thereby ensuring that the normalization manner of the input data when the model is inferred is the same as the normalization manner of the input data when the model is trained, so as to ensure the performance of the AI model inference.

[0128] The method provided in the embodiment also ensures that the power normalization manner of the input data during model inference is the same as the power normalization manner of the input data during model training by providing different power normalization manners. By providing the power value of the reference input data after power normalization and the processing manner of power normalization of different input data, the power normalization of the input data is implemented, which helps to improve the inference efficiency of the model. Different input data for power normalization is provided, and the power normalization of different types of input data is implemented. Different phase normalization manners are provided, so as to ensure that the phase normalization manner of the input data during model inference is the same as the phase normalization manner of the input data during model training. By providing the phase or imaginary part of the reference input data after phase normalization and the processing manner of phase normalization of different input data, the phase normalization of the input data is implemented, which helps to improve the inference efficiency of the model. Different input data for phase normalization is provided, and the phase normalization of different types of input data is implemented. The first information is carried in the second information indicating the parameters of the model, so that the normalization manner of the input data is indicated synchronously in the process of indicating the parameters of the model, and the signaling overhead is reduced. The first information is sent in different cases, and the normalization manner of the input data of the AI model is indicated flexibly.

[0129] Taking the normalization of the feature vector by the method provided in the application as an example:

[0130] In step A1, the network device indicates the parameters of the AI model to the terminal, and the indicated parameters include the first information for indicating the normalization manner of the input data of the AI model. The first information is used to indicate the manner of phase normalization of the input data of the AI model.

[0131] In some embodiments, the AI model is a model for CSI compression. The input of the AI model is a channel feature vector measured by the terminal, and the output is PMI quantization bits. In some embodiments, when the network device configures the model information (such as the input and output of the model) of the AI model used for CSI compression for the terminal, the normalization manner of the input data is also indicated simultaneously. This process can occur in the model deployment, model update, etc.

[0132] In some embodiments, the first information is used to indicate the manner of phase normalization of the feature vector input into the AI model. In some embodiments, the manner of phase normalization of the feature vector includes one of the following manners:

[0133] · Phase normalization with the first element of the feature vector as a reference;

[0134] · Phase normalization with the last element of the feature vector as a reference;

[0135] • phase normalization with reference to the element with the highest amplitude value in the feature vector;

[0136] • phase normalization with reference to the element with the lowest amplitude value in the feature vector;

[0137] • no phase normalization.

[0138] In some embodiments, the first information indicates a phase normalization manner for the feature vector input to the AI model from the above-mentioned multiple phase normalization manners, and the first information can also indicate no phase normalization.

[0139] In some embodiments, the phase of the element taken as reference in the normalized feature vector is 0 or the imaginary part is 0. The phase normalization includes the same phase offset for all elements in the feature vector, so that the phase of the element taken as reference is 0 or the imaginary part is 0. When it is indicated that no phase normalization is performed, the feature vector can be the original feature vector obtained by channel decomposition, without normalization processing.

[0140] For example, assume that the feature vector is a Kx1 vector, where the i-th element is When the phase normalization manner is to take the first element of the feature vector as reference, each element is subjected to the same phase offset (i.e., multiplied by ) so that the phase of the first element is 0, and each element of the feature vector obtained after phase normalization is Similarly, when the normalization manner is to take the k-th element of the feature vector as reference, each element is subjected to the same phase offset (i.e., multiplied by ) so that the phase of the k-th element is 0, and each element of the feature vector obtained after phase normalization is

[0141] In some embodiments, the input data of the AI model includes a feature vector and / or a channel matrix. For example, the feature vector is a vector obtained by eigenvalue decomposition of a channel covariance matrix, or a vector obtained by singular value decomposition of a channel matrix. For the case where the input data is a channel matrix, the process of phase normalization can refer to the process of phase normalization of the feature vector described above, which will not be described here.

[0142] In step A2, the terminal receives first information indicating the normalization manner of the input data, and performs CSI compression based on the AI model according to the first information.

[0143] In some embodiments, the terminal performs corresponding phase normalization processing on the input data (i.e., the feature vector) of the AI model according to the first information, and inputs the processed data into the AI model to obtain the corresponding PMI bit output. In some embodiments, the terminal takes the normalization mode (the first information) as the input of the AI model, and then performs inference on the AI model. For example, the input of the AI model includes information bits corresponding to the normalization mode, and 2 bits of information are used to correspond to 4 different phase normalization modes. In some embodiments, when the first information indicates that no phase normalization is performed, no normalization processing is required, and the original feature vector is directly taken as the input of the AI model.

[0144] The method provided in this embodiment can also be used for the terminal to report the first information indicating the normalization mode of the input data to the network device. At this time, the terminal reports the model information to the network device while providing the first information indicating the normalization mode of the input data, so that the network device can perform phase normalization on the input data in a corresponding mode.

[0145] For example, the method provided in the embodiments of the present application is applied to the normalization of RSRP:

[0146] In step B1, the network device indicates the parameters of the AI model to the terminal, and the parameters include the first information indicating the normalization mode of the input data. The first information is used to indicate the normalization mode of the input data of the AI model.

[0147] In some embodiments, the AI model is a model for beam management. The input of the AI model is a plurality of RSRP values (a set of RSRP values) corresponding to a first set of reference signals measured by the terminal, and the output is optimal reference signal indication information and a corresponding RSRP value corresponding to a second set of reference signals.

[0148] In some embodiments, the network device indicates the normalization mode of the input data when configuring the model information (e.g., the input and output of the model) of the AI model for beam management for the terminal. This process can occur during model deployment, model updating, etc.

[0149] In some embodiments, the first information is used to indicate the normalization mode of the input RSRP set. In some embodiments, the normalization mode of the RSRP set includes one of the following modes:

[0150] · power normalization with the first RSRP value as the reference;

[0151] · power normalization with the last RSRP value as the reference;

[0152] · power normalization with the highest RSRP value as the reference;

[0153] • power normalization is performed with the lowest RSRP value in the RSRP set as reference;

[0154] • no power normalization is performed.

[0155] In some embodiments, the first information indicates a power normalization manner to be used for the RSRP set as input of the AI model from the above-mentioned multiple power normalization manners, and the first information can also indicate that no power normalization is performed.

[0156] In some embodiments, the value of the RSRP value taken as reference in the normalized RSRP set is 0 (dBm). The power normalization includes scaling the received power values in the RSRP set by the same dB (i.e., adding the same dB) so that the value of the RSRP value taken as reference is 0 (dBm). When the first information indicates that no power normalization is performed, the RSRP set is the original RSRP set measured by the terminal and no normalization processing is needed.

[0157] For example, assume that the RSRP set contains K RSRP values, where the i-th RSRP value is N i (dBm). When the power normalization manner is to perform power normalization with the first RSRP value as reference, each RSRP value is scaled by the same N1dB (i.e., subtracting the same N1dB) so that the value of the first RSRP is 0 (dBm), and the elements in the RSRP set obtained after power normalization are {N i -N1}(dBm). Similarly, when the power normalization manner is to perform power normalization with the k-th RSRP value in the RSRP set as reference, each RSRP value is scaled by the same N k dB) so that the value of the k-th RSRP is 0 (dBm), and the i-th element in the RSRP set obtained after power normalization is {N i -N k}.

[0158] In some embodiments, the input data of the AI model can also be amplitude values, transmission power values, or RSRP difference values. When the input data is amplitude values, the value of the amplitude value taken as a reference (which can be the first, last, maximum, or minimum amplitude value) in the set of normalized amplitude values is 1; when the input data is transmission power values, the value of the transmission power value taken as a reference (which can be the first, last, maximum, or minimum transmission power value) in the set of normalized transmission power values is 0 (dBm); and when the input data is RSRP difference values, the value of the RSRP difference value taken as a reference (which can be the first, last, maximum, or minimum RSRP difference value) in the set of normalized RSRP difference values is 0 (dB).

[0159] In step B2, the terminal receives first information indicating the normalization method of the input data, and performs beam prediction based on the AI model according to the first information.

[0160] In some embodiments, the terminal performs corresponding power normalization processing on the input data (i.e., the RSRP set) of the AI model according to the first information, and inputs the processed data into the AI model to obtain the corresponding beam prediction output. In some embodiments, the terminal takes the normalization method (i.e., the first information) as the input of the AI model, and thus performs inference of the AI model. For example, the input of the AI model includes information bits corresponding to the normalization method, and 2 bits of information are used to correspond to 4 different power normalization methods. In some embodiments, when the first information indicates that no power normalization is performed, no normalization processing is needed, and the original RSRP set is directly taken as the input of the model.

[0161] The method of this embodiment can also be used for the terminal to report the first information indicating the normalization method of the input data to the network device. At this time, when the terminal reports the model information for beam prediction to the network device, the first information indicating the normalization method of the input data is also provided, so that the network device can perform power normalization on the input data (i.e., the RSRP set reported by the terminal) in a corresponding manner.

[0162] FIG. 8 is a flowchart of a CSI prediction method according to an example embodiment of the present application. The method can be used in the system shown in FIG. 4. The method includes the following steps:

[0163] Step 802: The network device sends first information to the terminal.

[0164] The first information is used to indicate a normalization manner of input data of the encoder for CSI prediction. In some embodiments, the encoder is deployed in the terminal. The encoder for CSI prediction corresponds to a decoder for CSI prediction, which is deployed in the network device. The output information of the encoder is used as the input information of the decoder. In some embodiments, the encoder and the decoder for CSI prediction are constructed based on a neural network. For the structure of the encoder and the decoder, reference can be made to the description of FIG. 3, which will not be repeated here.

[0165] In some embodiments, the first information is carried in the second information, and the second information includes / carryes the first information. The second information is used to indicate the parameters of the encoder. In some embodiments, the normalization manner indicated by the first information includes a phase normalization manner.

[0166] In some embodiments, the phase normalization manner includes at least one of the following: performing phase normalization with the first input data as a reference; performing phase normalization with the last input data as a reference; performing phase normalization with the input data having the highest amplitude value as a reference; performing phase normalization with the input data having the lowest amplitude value as a reference; and not performing phase normalization. In some embodiments, the phase of the input data after phase normalization as a reference is 0, or the imaginary part of the input data after phase normalization as a reference is 0. In some embodiments, the phase normalization includes the same phase offset to the input data. The phase normalization manner indicated by the first information is consistent with the normalization manner used to the input data in the process of training the encoder.

[0167] Step 804: The terminal performs normalization processing on the first channel information according to the normalization manner indicated by the first information.

[0168] In some embodiments, the input information of the encoder includes the first channel information. The first channel information includes channel information measured by the terminal based on a reference signal, such as at least one of a channel feature vector, beam information, and time delay information. The embodiments mainly take the channel information as the channel feature vector as an example for description.

[0169] Before inputting the first channel information into the encoder, the terminal performs phase normalization processing on the first channel information according to the phase normalization manner indicated by the first information, so as to obtain processed first channel information. The process of performing phase normalization processing on the first channel information can refer to the related content in the foregoing description, which will not be repeated here.

[0170] Step 806: The terminal inputs the processed first channel information into the encoder for CSI prediction, and obtains CSI quantization bits.

[0171] In some embodiments, the CSI quantization bits, such as PMI quantization bits, can be predicted by the encoder by inputting the normalized first channel information into the encoder.

[0172] At step 808, the terminal sends the CSI quantization bits to the network device.

[0173] After the CSI quantization bits are predicted by the encoder, the terminal sends the CSI quantization bits to the network device. For example, the CSI quantization bits can be sent to the network device through PUSCH and / or PUCCH.

[0174] At step 810, the network device inputs the CSI quantization bits into the decoder for CSI prediction to obtain second channel information.

[0175] After the CSI quantization bits are obtained, the network device inputs the CSI quantization bits into the decoder, and the second channel information is predicted by the decoder. The second channel information includes information of the channel corresponding to the CSI quantization bits, such as eigenvectors of each subband of the channel, which is used for precoding of the downlink. The second channel information predicted by using the decoder can improve the downlink transmission performance of the network device.

[0176] In summary, the method provided in the embodiment can realize that the terminal deploying the encoder can normalize the input data of the encoder according to the normalization manner used when the encoder is trained, so as to ensure that the normalization manner of the input data when the encoder is inferred is the same as the normalization manner of the input data when the encoder is trained, so as to ensure the performance of the encoder inference.

[0177] It should be noted that the order of the method steps provided in the embodiments of the present application can be appropriately adjusted, the steps can be appropriately increased or decreased according to the situation, and the steps can be freely combined to form new embodiments. Any person skilled in the art can easily think of changes within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application, and thus will not be described here. In addition, the order of the above-mentioned different situations does not have a preferred meaning, but is only for convenient description.

[0178] FIG. 9 is a block diagram of a configuration device of a model provided in an example embodiment of the present application, which can be implemented as a first device or a part of the first device through software or hardware or a combination of both. The device includes a sending module 901.

[0179] The sending module 901 is configured to send first information to the second device, the first information being used to indicate a normalization manner of input data of the first AI model.

[0180] In some embodiments, the first AI model is deployed in the second device. In some embodiments, the apparatus comprises a terminal and the second device comprises a network device. In some embodiments, the apparatus comprises a network device and the second device comprises a terminal. In some embodiments, the use of the first AI model comprises at least one of the following: CSI prediction; CSI compression; beam management; beam prediction; data detection.

[0181] In some embodiments, the first information is used to indicate a normalization manner adopted for the input data when inference is performed by the first AI model, and the normalization manner indicated by the first information is the same as a normalization manner adopted for the input data when the first AI model is trained. In some embodiments, the first information is used to indicate a normalization manner adopted for the input data when the first AI model is trained. In some embodiments, the first information is used to indicate a normalization manner adopted for the input data when a model corresponding to the first AI model is trained, the output information of the first AI model being used as input information of the model corresponding to the first AI model, or the output information of the model corresponding to the first AI model being used as input information of the first AI model.

[0182] In some embodiments, the first information is carried in second information, and the second information comprises / carrying the first information. The second information is used to indicate parameters of the first AI model. In some embodiments, the first information is sent in at least one of the following cases: model configuration; model reporting; model updating.

[0183] In some embodiments, the normalization manner indicated by the first information comprises at least one of a power normalization manner and a phase normalization manner. In some embodiments, the power normalization manner refers to a manner of normalizing the input data based on a power value of the input data. In some embodiments, the phase normalization manner refers to a manner of normalizing the input data based on a phase of the input data.

[0184] In some embodiments, the power normalization manner comprises at least one of the following: power normalization with the first input data as reference; power normalization with the last input data as reference; power normalization with the input data having the highest power value as reference; power normalization with the input data having the lowest power value as reference; and no power normalization. In some embodiments, power normalization with a certain input data as reference means that the power value of the input data as reference is processed according to a preset manner, and other input data is processed in the same manner as the input data as reference. No power normalization means that no power normalization processing is performed on the input data. In some embodiments, the power value of the input data as reference after power normalization comprises at least one of the following: 1; 0 dB; and 0 dBm. In some embodiments, power normalization comprises equal proportion scaling of the input data according to the power value. In some embodiments, the input data subjected to normalization processing based on the power normalization manner comprises at least one of the following: amplitude value; transmission power value; RSRP value; RSRP difference value; and SINR value.

[0185] In some embodiments, the phase normalization manner comprises at least one of the following: phase normalization with the first input data as reference; phase normalization with the last input data as reference; phase normalization with the input data having the highest amplitude value as reference; phase normalization with the input data having the lowest amplitude value as reference; and no phase normalization. In some embodiments, phase normalization with a certain input data as reference means that the phase of the input data as reference is processed according to a preset manner, and other input data is processed in the same manner as the input data as reference. No phase normalization means that no phase normalization processing is performed on the input data. In some embodiments, the phase of the input data as reference after phase normalization is 0, or the imaginary part of the input data as reference after phase normalization is 0. In some embodiments, phase normalization comprises the same phase offset of the input data. In some embodiments, the input data subjected to normalization processing based on the phase normalization manner comprises at least one of the following: feature vector; channel matrix; and channel covariance matrix.

[0186] FIG. 10 is a block diagram of a configuration apparatus of a model according to an example embodiment of the present application. The apparatus can be implemented as a second device or a part of the second device through software or hardware or a combination of both. The apparatus comprises a receiving module 1001 and a processing module 1002.

[0187] The receiving module 1001 is configured to receive first information sent by a first device, the first information being used to indicate a normalization manner of input data of a first AI model.

[0188] In some embodiments, the first AI model is deployed in the apparatus. In some embodiments, the first device comprises a terminal and the apparatus comprises a network device. In some embodiments, the first device comprises a network device and the apparatus comprises a terminal. In some embodiments, the use of the first AI model comprises at least one of: CSI prediction; CSI compression; beam management; beam prediction; data detection.

[0189] In some embodiments, the first information is used to indicate a normalization manner adopted for input data when inference is performed by the first AI model, and the normalization manner indicated by the first information is the same as a normalization manner adopted for input data when the first AI model is trained. In some embodiments, the first information is used to indicate a normalization manner adopted for input data when the first AI model is trained. In some embodiments, the first information is used to indicate a normalization manner adopted for input data when a model corresponding to the first AI model is trained, and output information of the first AI model is used as input information of the model corresponding to the first AI model, or output information of the model corresponding to the first AI model is used as input information of the first AI model.

[0190] In some embodiments, the first information is carried in second information, and the second information comprises / carryes the first information. The second information is used to indicate parameters of the first AI model. In some embodiments, the first information is sent in at least one of the following cases: model configuration; model reporting; model updating.

[0191] In some embodiments, the normalization manner indicated by the first information comprises at least one of a power normalization manner and a phase normalization manner. In some embodiments, the power normalization manner refers to a manner of normalizing input data based on a power value of the input data. In some embodiments, the phase normalization manner refers to a manner of normalizing input data based on a phase of the input data.

[0192] In some embodiments, the power normalization manner comprises at least one of: power normalization with the first input data as reference; power normalization with the last input data as reference; power normalization with the input data having the highest power value as reference; power normalization with the input data having the lowest power value as reference; and no power normalization. In some embodiments, power normalization with a certain input data as reference means that the power value of the input data as reference is processed according to a preset manner, and other input data is processed in the same manner as the input data as reference. No power normalization means that no power normalization processing is performed on the input data. In some embodiments, the power value of the input data as reference after power normalization comprises at least one of: 1; 0 dB; and 0 dBm. In some embodiments, power normalization comprises equal proportion scaling of the input data according to the power value. In some embodiments, the input data subjected to normalization processing based on the power normalization manner comprises at least one of: amplitude value; transmission power value; RSRP value; RSRP difference value; and SINR value.

[0193] In some embodiments, the phase normalization manner comprises at least one of: phase normalization with the first input data as reference; phase normalization with the last input data as reference; phase normalization with the input data having the highest amplitude value as reference; phase normalization with the input data having the lowest amplitude value as reference; and no phase normalization. In some embodiments, phase normalization with a certain input data as reference means that the phase of the input data as reference is processed according to a preset manner, and other input data is processed in the same manner as the input data as reference. No phase normalization means that no phase normalization processing is performed on the input data. In some embodiments, the phase of the input data as reference after phase normalization is 0, or the imaginary part of the input data as reference after phase normalization is 0. In some embodiments, phase normalization comprises the same phase offset of the input data. In some embodiments, the input data subjected to normalization processing based on the phase normalization manner comprises at least one of: feature vector; channel matrix; and channel covariance matrix.

[0194] In some embodiments, after receiving the first information, the apparatus performs the inference process of the first AI model according to the normalization manner indicated by the first information. In some embodiments, the processing module 1002 is configured to normalize the input data according to the normalization manner indicated by the first information, for example, perform power normalization processing or phase normalization processing, and input the processed input data into the first AI model to perform the inference process of the first AI model, so as to obtain the corresponding output information of the first AI model. In some embodiments, the processing module 1002 is configured to input the first information and the input data together as the input of the first AI model to perform the inference process of the first AI model, so as to obtain the corresponding output information of the first AI model. The input data includes data used for the inference of the first AI model.

[0195] It should be noted that the apparatus provided by the above embodiments is only exemplified by the above division of various functional modules when realizing its functions. In actual applications, the above functions can be completed by different functional modules according to actual needs, that is, the content structure of the device is divided into different functional modules to complete all or part of the above described functions.

[0196] As for the apparatus in the above embodiments, the specific manner in which each module performs operations has been described in detail in the embodiments of the method, and will not be described in detail here.

[0197] FIG. 11 is a structural schematic diagram of a communication device (network device or terminal) provided by an embodiment of the present application. The communication device can include a processor 1101, a receiver 1102, a transmitter 1103, a memory 1104, and a bus 1105.

[0198] The processor 1101 includes one or more processing cores. The processor 1101 performs various functional applications and information processing by running software programs and modules.

[0199] The receiver 1102 and the transmitter 1103 can be implemented as a transceiver 1106, which can be a communication chip.

[0200] The memory 1104 is connected to the processor 1101 through the bus 1105. The memory 1104 can be used to store computer programs, and the processor 1101 is configured to execute the computer programs to implement each step performed by the network device or the terminal in the above method embodiments.

[0201] Moreover, the memory 1104 can be implemented by any type of volatile or nonvolatile storage devices or a combination thereof, including but not limited to: RAM (Random-Access Memory) and ROM (Read-Only Memory), EPROM (Erasable Programmable Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory or other solid state storage technology, CD-ROM (Compact Disc Read-Only Memory), DVD (Digital Video Disc), or other optical storage, magnetic cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices.

[0202] The embodiment of the present application further provides a computer readable storage medium, wherein the storage medium stores a computer program, and the computer program is used for being executed by a processor of a terminal or a network device to implement each step in the configuration method of the model.

[0203] In some embodiments, the computer readable storage medium can include a ROM (Read-Only Memory), a RAM (Random-Access Memory), a SSD (Solid State Drives) or an optical disc, etc. Wherein, the random access memory can include a ReRAM (Resistance Random Access Memory) and a DRAM (Dynamic Random Access Memory).

[0204] The embodiment of the present application further provides a chip, wherein the chip includes a programmable logic circuit and / or program instructions, and when the chip is running on a terminal or a network device, is used for implementing each step in the configuration method of the model.

[0205] The embodiment of the present application further provides a computer program product or a computer program, wherein the computer program product or the computer program includes computer instructions, the computer instructions are stored in a computer readable storage medium, a processor of a communication device reads and executes the computer instructions from the computer readable storage medium, and each step in the configuration method of the model is implemented.

[0206] Those skilled in the art should be aware that, in the above one or more examples, the functions described in the embodiments of the present application can be implemented in hardware, software, firmware or any combination thereof. When implemented in software, the functions can be stored in a computer readable medium or transmitted as one or more instructions or code on the computer readable medium. The computer readable medium includes computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. The storage media can be any available media that can be accessed by a general purpose or special purpose computer.

[0207] The above merely provides exemplary embodiments of the present application, but is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method of configuring a model, the method comprising: The method is performed by a first device, and the method comprises: sending, to a second device, first information used to indicate a normalization manner of input data of a first artificial intelligence (AI) model; wherein the normalization manner comprises at least one of a power normalization manner and a phase normalization manner.

2. The method of claim 1, wherein, The power normalization manner comprises at least one of: performing power normalization with a first input data as a reference; performing the power normalization with a last input data as the reference; and performing the power normalization with an input data having a highest power value as the reference. performing the power normalization with an input data having a lowest power value as the reference. not performing the power normalization.

3. The method of claim 2, wherein, The power value of the input data taken as the reference after the power normalization comprises at least one of: 1; 0 dB; and 0 dBm.

4. The method according to claim 2 or 3, characterized in that, The power normalization comprises performing equal proportion scaling on the input data according to the power value.

5. The method according to any one of claims 2 to 4, characterized in that, The input data comprises at least one of: an amplitude value; a transmission power value; a reference signal receiving power (RSRP) value; an RSRP difference value; and a signal to interference plus noise ratio (SINR) value.

6. The method according to any one of claims 1 to 5, characterized in that, The phase normalization manner comprises at least one of: performing phase normalization with a first input data as a reference; performing the phase normalization with a last input data as the reference; and performing the phase normalization with an input data having a highest amplitude value as the reference. performing the phase normalization with an input data having a lowest amplitude value as the reference. not performing the phase normalization.

7. The method of claim 6, wherein, The phase of the input data taken as the reference after the phase normalization is 0 or the imaginary part is 0.

8. The method according to claim 6 or 7, characterized in that, The phase normalization comprises performing the same phase offset on the input data.

9. The method according to any one of claims 6 to 8, characterized in that, The input data comprises at least one of: a feature vector; a channel matrix; and a channel covariance matrix.

10. The method according to any one of claims 1 to 9, characterized in that, The first information is carried in second information used to indicate a parameter of the first AI model.

11. The method according to any one of claims 1 to 10, characterized in that, The first information is sent in at least one of the following cases: model configuration; model reporting; and model updating.

12. The method according to any one of claims 1 to 11, characterized in that, The first AI model is used for at least one of: channel state information (CSI) prediction; CSI compression; beam management; beam prediction; and data detection.

13. The method according to any one of claims 1 to 12, characterized in that, The first device comprises a terminal, and the second device comprises a network device; or the first device comprises a network device, and the second device comprises a terminal.

14. A method of configuring a model, the method comprising: The method is performed by a second device, and the method comprises: receiving first information sent by a first device, the first information being used to indicate a normalization manner of input data of a first AI model; wherein the normalization manner comprises at least one of a power normalization manner and a phase normalization manner.

15. The method of claim 14, wherein, The power normalization manner comprises at least one of: performing power normalization with a first input data as a reference; performing the power normalization with a last input data as the reference; and performing the power normalization with an input data having a highest power value as the reference. performing the power normalization with an input data having a lowest power value as the reference. not performing the power normalization.

16. The method of claim 15, wherein, The power value of the input data taken as reference after the power normalization comprises at least one of the following: 1; 0 dB; 0 dBm.

17. The method according to claim 15 or 16, characterized in that, The power normalization comprises scaling the input data in proportion to the power value.

18. The method of any one of claims 15 to 17, wherein, The input data comprises at least one of the following: amplitude value; transmission power value; RSRP value; RSRP difference value; SINR value.

19. The method according to any one of claims 14 to 18, characterized in that, The phase normalization mode comprises at least one of the following: performing phase normalization taking the first input data as reference; performing phase normalization taking the last input data as reference; performing phase normalization taking the input data with the highest amplitude value as reference; performing phase normalization taking the input data with the lowest amplitude value as reference; not performing the phase normalization.

20. The method of claim 19, wherein, The phase of the input data taken as reference after the phase normalization is 0 or the imaginary part is 0.

21. The method of claim 19 or 20, wherein, The phase normalization comprises the same phase offset to the input data.

22. The method of any one of claims 19 to 21, wherein, The input data comprises at least one of the following: feature vector; channel matrix; channel covariance matrix.

23. The method of any one of claims 14 to 22, wherein, The method further comprises: normalizing the input data according to the normalization mode indicated by the first information, and inputting the processed input data into the first AI model.

24. The method of any one of claims 14 to 22, wherein, The method further comprises: inputting the first information together with the input data as the input of the first AI model.

25. The method of any one of claims 14 to 24, wherein, The first information is carried in second information, and the second information is used to indicate the parameters of the first AI model.

26. The method of any one of claims 14 to 25, wherein, The first information is sent in at least one of the following cases: model configuration; model reporting; model updating.

27. The method of any one of claims 14 to 26, wherein, The use of the first AI model comprises at least one of the following: CSI prediction; CSI compression; beam management; beam prediction; data detection.

28. The method of any one of claims 14 to 27, wherein, The first device comprises a terminal, and the second device comprises a network device; or, the first device comprises a network device, and the second device comprises a terminal.

29. An apparatus for configuring a model, the apparatus comprising: The apparatus comprises: a sending module, configured to send first information to a second device, the first information being used to indicate the normalization mode of input data of a first AI model; wherein the normalization mode comprises at least one of the following: power normalization mode; phase normalization mode.

30. An apparatus for configuring a model, the apparatus comprising: The apparatus comprises: a receiving module, configured to receive first information sent by a first device, the first information being used to indicate the normalization mode of input data of a first AI model; wherein the normalization mode comprises at least one of the following: power normalization mode; phase normalization mode.

31. A first device, comprising: The first device comprises: a processor; a transceiver connected to the processor; a memory for storing executable instructions of the processor; wherein the processor is configured to load and execute the executable instructions to implement the model configuration method according to any one of claims 1 to 13.

32. A second device, comprising: The second device comprises: a processor; a transceiver connected to the processor; a memory for storing executable instructions of the processor; wherein the processor is configured to load and execute the executable instructions to implement the model configuration method according to any one of claims 14 to 28.

33. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is used by a processor to implement the model configuration method in any one of claims 1 to 28.

34. A chip, characterized by The chip comprises programmable logic circuit and / or program instructions, and when the chip is running on a communication device, the programmable logic circuit and / or program instructions are used to implement the model configuration method in any one of claims 1 to 28.

35. A computer program product, characterised in that, The computer program product comprises computer instructions stored in a computer readable storage medium; a processor of a communication device reads the computer instructions from the computer readable storage medium and executes the computer instructions, so that the communication device implements the model configuration method in any one of claims 1 to 28.

36. A computer program, characterized in that, The computer program is executed by a processor of a communication device to implement the model configuration method in any one of claims 1 to 28.

Images