Coding method and system for predicting channel state information

Abstract

The invention discloses a coding method for predicting channel state information. Wherein the encoder obtains first addition data according to the coherence time vector of the complex CFR data, the time mask of the input data of each batch and the input data of each batch. The first addition data serves as input of a first deep learning model, and the first deep learning model outputs predicted CFR data. According to the method, the time sequence characteristics of the channel state information are calculated, coding is added, and time and space information is incorporated into the input data of the deep learning model, so that the relative stability time of the channel at the future moment is considered, the problem of weak time sequence and space information capture capability in the model is solved, and the prediction precision of the channel state information is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to a channel state information (CSI) prediction technology, in particular to an encoding technology for channel state information prediction. BACKGROUND

[0002] Channel state information prediction (referred to as channel prediction) refers to predicting future channel state information using historical channel state information. Channel state information prediction is indispensable for link adaptation technology, because it can minimize the capacity loss caused by channel fluctuations and optimize the use efficiency of available bandwidth. In addition, under the condition of limited spectrum resources, accurate CSI prediction is particularly crucial for the optimization of spectrum resource allocation, which helps to improve spectrum efficiency and energy efficiency, thereby improving network performance and achieving efficient wireless resource management.

[0003] In recent years, researchers have attempted to use MLP (multilayer perceptron), CNN (convolutional neural networks), RNN (recurrent neural network), LSTM (long short-term memory), Transformer, Informer and other deep learning models to process the time series prediction task of channel state information. In the environment with limited computing resources, deep learning models based on CNN have become a popular solution because of their lightweight advantage and the fact that convolution kernels are suitable for processing channel state information, which has a spatial hierarchical structure.

[0004] However, deep learning models usually use discrete numbers or vectors to represent features, and the physical or semantic meaning of these features lacks interpretability. Moreover, the parameters (such as weight matrices and bias terms) of deep learning models are high-dimensional, and it is usually difficult to intuitively explain what they represent at the physical level, lacking physical meaning. Specifically, the "black box" nature of deep learning models makes it difficult to interpret the models physically. This limitation makes deep learning-based models face challenges when processing long time series data or signal processing.

[0005] In the Chinese patent application for invention with the application publication number CN117674926A and the application publication date of March 8, 2024, a channel state information processing method and device are disclosed. The channel state information is input into an encoder model to obtain extended enhanced channel state information output by the encoder model. The extended enhanced channel state information is sent to a network device. The encoder model includes an encoder and a channel state information extension enhancement module cascaded with the encoder. The document aims to multiplex AI (artificial intelligence) models under different network configuration parameters, reduce the complexity of AI model training, and save AI model deployment resources, but still cannot solve the above-mentioned lack of explanatory problem. SUMMARY

[0006] The technical problem to be solved by the present application is how to improve the physical interpretability of the deep learning model for channel state information prediction.

[0007] To solve the above technical problems, the application provides an encoding method for predicting channel state information, comprising the following steps. Step SA1: obtaining input data of each batch, input data time mask of each batch, fitting target of output data of each batch, output data time mask of each batch, and coherence time vector of complex CFR data according to the complex CFR data. Step SA2: an encoder obtains first added data according to the coherence time vector of the complex CFR data, the input data time mask of each batch, and the input data of each batch. Specifically, the encoder performs matrix calculation on the coherence time vector of the complex CFR data and the input data time mask of each batch to obtain a coherence time sequence of the complex CFR data of each batch. Then, the encoder maps the coherence time sequence of the complex CFR data of each batch to a high-dimensional space by an embedding function to obtain coherence time encoding of the complex CFR data of each batch, so that the data dimension of the coherence time encoding of the complex CFR data of each batch is equal to the data dimension of the input data of each batch. Then, the encoder adds the coherence time encoding of the complex CFR data of each batch to the input data of each batch to obtain the first added data. Step SA3: the first added data is input into a first deep learning model, and the first deep learning model outputs predicted CFR data. Step SA4: a decoder clips the predicted CFR data output by the first deep learning model to output predicted CFR data of each batch, and the clipping is to make the data dimension of the predicted CFR data of each batch equal to the data dimension of the fitting target of the output data of each batch. Then, a loss is calculated according to the predicted CFR data of each batch and the fitting target of the output data of each batch, and the first deep learning model is trained by back propagation, so that the predicted CFR data of each batch output by the decoder is as close as possible to the fitting target of the output data of the batch. Step SA5: first, input data of each batch, input data time mask of each batch, and coherence time vector of complex CFR data are obtained according to historical complex CFR data obtained in an actual scene. Then, an encoder obtains first added data according to the coherence time vector of the complex CFR data, the input data time mask of each batch, and the input data of each batch. Then, the first added data is input into a trained first deep learning model, and the trained first deep learning model outputs predicted CFR data. Then, a decoder clips the predicted CFR data output by the trained first deep learning model to obtain predicted CFR data of a prediction dimension.

[0008] Further, in the step SA1, the complex CFR data is used to obtain the input data of each batch, the input data time mask of each batch, the fitting target of the output data of each batch, and the output data time mask of each batch; specifically including the following steps. Step S21: Obtain the complex CFR data in the past period of time, split the complex CFR data into real and imaginary parts and longitudinally splice them to obtain real CFR data. Step S22: Obtain the input data set, the input data time mask set, the fitting target set of the output data, and the output data time mask set of the deep learning model from the real CFR data through a sliding time window. Step S23: Obtain the input data of each batch, the input data time mask of each batch, the fitting target of the output data of each batch, and the output data time mask of each batch from the input data set, the input data time mask set, the fitting target set of the output data, and the output data time mask set according to the data amount of each batch.

[0009] Further, in the step SA1, the complex CFR data is used to obtain the coherence time vector of the complex CFR data; including the following steps. Step S31: Calculate the first norm of the complex CFR data in the time dimension; the complex CFR data is a complex matrix of n rows and t columns, where n represents the number of subcarriers, and t represents the time length of the complex CFR data; first, difference the elements of two adjacent columns in the complex matrix of the complex CFR data, and then calculate the p-norm of the amplitude of the difference value of each adjacent two columns. CFRLtp j , p is a positive integer, and j takes an integer value from 0 to t-1; combine t CFRLtp j in the order from small to large to obtain the first norm of the complex CFR data in the time dimension; this step is to extract the difference feature of the complex CFR data with time. Step S32: Calculate the first change rate of the j position in time relative to the i position in time of the first norm of the complex CFR data in the time dimension, where i and j are integers, and 0≤i<j≤t-1. Step S33: Initialize tcVec_CFR0; calculate tcVec_CFR j using the first change rate and the first threshold; combine tcVec_CFR0and t-1 tcVec_CFR j in the order from small to large subscript to obtain the coherence time vector of the complex CFR data.

[0010] Further, in the step SA1, the coherence time vector of the complex CFR data is obtained according to the complex CFR data; comprising the following steps. Step S41: the complex CFR data is a complex matrix of n rows and t columns, wherein n represents the number of subcarriers, and t represents the time length of the complex CFR data; a difference matrix ΔCFR of the complex CFR data is calculated, which is a real matrix of t rows and t columns. Step S42: a second change rate of a j-th position in time after a i-th position in time in the difference matrix ΔCFR of the complex CFR data is calculated, wherein i and j are integers, and 0≤i<j≤t-1; wherein the second change rate comprises calculating a p-norm after differentiating the current time step and all previous time steps. Step S43: initializing tcVec_CFR0; using the second change rate and a second threshold to calculate tcVec_CFR j ; combining tcVec_CFR0and t-1 tcVec_CFR j in the order of the subscript from small to large to obtain the coherence time vector of the complex CFR data.

[0011] The application also proposes another encoding method for predicting channel state information, comprising the following steps. Step SB1: obtaining, according to complex channel frequency response (CFR) data, input data of each batch, input data time mask of each batch, fitting target of output data of each batch, output data time mask of each batch, and coherent spatial vector of the complex CFR data. Step SB2: an encoder obtains second added data according to the coherent spatial vector of the complex CFR data, the input data time mask of each batch, and the input data of each batch; specifically, the encoder performs matrix calculation on the coherent spatial vector of the complex CFR data and the input data time mask of each batch to obtain a coherent spatial sequence of the complex CFR data of each batch; then the encoder maps the coherent spatial sequence of the complex CFR data of each batch to a high-dimensional space by an embedding function to obtain coherent spatial encoding of the complex CFR data of each batch, so that the data dimension of the coherent spatial encoding of the complex CFR data of each batch is equal to the data dimension of the input data of each batch; then the encoder adds the coherent spatial encoding of the complex CFR data of each batch to the input data of each batch to obtain the second added data. Step SB3: the second added data is input into a second deep learning model, and the second deep learning model outputs predicted CFR data. Step SB4: a decoder performs clipping on the predicted CFR data output by the second deep learning model to output predicted CFR data of each batch, and the clipping is to make the data dimension of the predicted CFR data of each batch equal to the data dimension of the fitting target of the output data of each batch; then a loss is calculated according to the predicted CFR data of each batch and the fitting target of the output data of each batch, and the second deep learning model is trained by back propagation, so that the predicted CFR data of each batch output by the decoder is as close as possible to the fitting target of the output data of the batch. Step SB5: first, input data of each batch, input data time mask of each batch, and coherent spatial vector of the complex CFR data are obtained according to historical complex CFR data obtained in an actual scene; then the encoder obtains second added data according to the coherent spatial vector of the complex CFR data, the input data time mask of each batch, and the input data of each batch; then the second added data is input into a trained second deep learning model, and the trained second deep learning model outputs predicted CFR data; then the decoder performs clipping on the predicted CFR data output by the trained second deep learning model to obtain predicted CFR data of a prediction dimension.

[0012] Further, in step SB1, the coherence space vector of the complex CFR data is obtained based on the complex CFR data; this includes the following steps: Step S61: Calculate the third norm of the complex CFR data in the subcarrier dimension; the complex CFR data is an n-row × t-column complex matrix, where n represents the number of subcarriers and t represents the time length of the complex CFR data; first, differencing the elements of adjacent columns in the transpose matrix of the complex CFR data, and then calculating the p-norm CFRLsp of the magnitude of the difference value between each pair of adjacent columns. j p is a positive integer, and j is an integer ranging from 0 to n-1; n CFRLsp j The complex CFR data is combined in ascending order of j to obtain the third norm of the complex CFR data in the subcarrier dimension; this step is to extract the differential features of the complex CFR data as a function of the subcarrier. Step S62: Calculate the third rate of change of the third norm of the complex CFR data in the subcarrier dimension for the position j of the later subcarrier position relative to the position i of the earlier subcarrier position, where i and j are integers and 0 ≤ i < j ≤ n-1. Step S63: Initialize scVec_CFR0; calculate scVec_CFR using the third rate of change and the third threshold. j ; Set scVec_CFR0 and n-1 scVec_CFR j By combining the subscripts in ascending order, we obtain the coherent space vector of the complex CFR data.

[0013] Further, in step SB1, the coherence spatial vector of the complex CFR data is obtained based on the complex CFR data; this includes the following steps. Step S71: The complex CFR data is an n-row × t-column complex matrix, where n represents the number of subcarriers and t represents the time length of the complex CFR data; calculate the difference matrix ΔCFR of the transpose matrix of the complex CFR data. T This is an n x n real matrix. Step S72: Calculate the difference matrix ΔCFR of the transpose matrix of the complex CFR data. T The fourth rate of change of the subcarrier position j relative to the subcarrier position i, where i and j are integers and 0 ≤ i < j ≤ n-1; includes calculating the p-norm after differencing the current subcarrier with all previous subcarriers. Step S73: Initialize scVec_CFR0; calculate scVec_CFR using the fourth rate of change and the fourth threshold. j ; Set scVec_CFR0 and n-1 scVec_CFR j By combining the subscripts in ascending order, we obtain the coherent space vector of the complex CFR data.

[0014] The application also proposes another encoding method for predicting channel state information, comprising the following steps. Step SCI: obtaining, according to complex channel frequency response CFR data, input data of each batch, input data time mask of each batch, fitting target of output data of each batch, output data time mask of each batch, and coherence time vector of complex channel impulse response CIR data. Step SC2: an encoder obtains third added data according to the coherence time vector of complex CIR data, the input data time mask of each batch, and the input data of each batch; specifically, the encoder performs matrix calculation on the coherence time vector of complex CIR data and the input data time mask of each batch to obtain a coherence time sequence of complex CIR data of each batch; then the encoder maps the coherence time sequence of complex CIR data of each batch to a high-dimensional space by an embedding function to obtain coherence time encoding of complex CIR data of each batch after embedding, so that the data dimension of the coherence time encoding of complex CIR data of each batch is equal to the data dimension of the input data of each batch; then the encoder adds the coherence time encoding of complex CIR data of each batch to the input data of each batch to obtain the third added data. Step SC3: the third added data is input into a third deep learning model, and the third deep learning model outputs predicted CFR data. Step SC4: a decoder performs clipping on the predicted CFR data output by the third deep learning model to output predicted CFR data of each batch, and the clipping is to make the data dimension of the predicted CFR data of each batch equal to the data dimension of the fitting target of the output data of each batch; then a loss is calculated according to the predicted CFR data of each batch and the fitting target of the output data of each batch, and the third deep learning model is trained by back propagation, so that the predicted CFR data of each batch output by the decoder is as close as possible to the fitting target of the output data of the batch. Step SC5: first, input data of each batch, input data time mask of each batch, and coherence time vector of complex CIR data are obtained according to historical complex CFR data obtained in an actual scene; then the encoder obtains third added data according to the coherence time vector of complex CIR data, the input data time mask of each batch, and the input data of each batch; then the third added data is input into a trained third deep learning model, and the trained third deep learning model outputs predicted CFR data; then the decoder performs clipping on the predicted CFR data output by the trained third deep learning model to obtain predicted CFR data of a prediction dimension.

[0015] The application also proposes another encoding method for predicting channel state information, comprising the following steps. Step SD1: obtaining, according to complex channel frequency response CFR data, input data of each batch, input data time mask of each batch, fitting target of output data of each batch, output data time mask of each batch, and coherent spatial vector of complex channel impulse response CIR data. Step SD2: an encoder obtains fourth added data according to the coherent spatial vector of complex CIR data, the input data time mask of each batch, and the input data of each batch; specifically, the encoder performs matrix calculation on the coherent spatial vector of complex CIR data and the input data time mask of each batch to obtain a coherent spatial sequence of complex CIR data of each batch; then the encoder maps the coherent spatial sequence of complex CIR data of each batch to a high-dimensional space by an embedding function to obtain coherent spatial encoding of complex CIR data of each batch after embedding, so that the data dimension of the coherent spatial encoding of complex CIR data of each batch is equal to the data dimension of the input data of each batch; then the encoder adds the coherent spatial encoding of complex CIR data of each batch to the input data of each batch to obtain the fourth added data. Step SD3: the fourth added data is input into a fourth deep learning model, and the fourth deep learning model outputs predicted CFR data. Step SD4: a decoder clips the predicted CFR data output by the fourth deep learning model to output predicted CFR data of each batch, and the clipping is to make the data dimension of the predicted CFR data of each batch equal to the data dimension of the fitting target of the output data of each batch; then a loss is calculated according to the predicted CFR data of each batch and the fitting target of the output data of each batch, and the fourth deep learning model is trained by back propagation, so that the predicted CFR data of each batch output by the decoder is as close as possible to the fitting target of the output data of the batch. Step SD5: first, input data of each batch, input data time mask of each batch, and coherent spatial vector of complex CIR data are obtained according to historical complex CFR data obtained in an actual scene; then the encoder obtains fourth added data according to the coherent spatial vector of complex CIR data, the input data time mask of each batch, and the input data of each batch; then the fourth added data is input into a trained fourth deep learning model, and the trained fourth deep learning model outputs predicted CFR data; then the decoder clips the predicted CFR data output by the trained fourth deep learning model to obtain predicted CFR data of a prediction dimension.

[0016] The application also provides an encoding system for predicting channel state information, comprising a first preprocessor, a first encoder, a first deep learning model, and a first decoder. The first preprocessor is used to obtain input data of each batch, input data time mask of each batch, fitting target of output data of each batch, output data time mask of each batch, and coherence time vector of complex CFR data in a training stage according to complex channel frequency response (CFR) data. The first preprocessor is used to obtain input data of each batch, input data time mask of each batch, and coherence time vector of complex CFR data in an application stage according to historical complex CFR data obtained in an actual scene. The first encoder is used to obtain first added data in both the training stage and the application stage according to the coherence time vector of complex CFR data, the input data time mask of each batch, and the input data of each batch. Specifically, the first encoder performs matrix calculation on the coherence time vector of complex CFR data and the input data time mask of each batch to obtain a coherence time sequence of complex CFR data of each batch. Then, the first encoder maps the coherence time sequence of complex CFR data of each batch to a high-dimensional space embedding through an embedding function to obtain coherence time encoding of complex CFR data of each batch, so that the data dimension of the coherence time encoding of complex CFR data of each batch is equal to the data dimension of the input data of each batch. Then, the first encoder adds the coherence time encoding of complex CFR data of each batch to the input data of each batch to obtain the first added data. The first deep learning model is used to take the first added data as input and output predicted CFR data in both the training stage and the application stage. The first decoder is used to crop the predicted CFR data output by the first deep learning model in both the training stage and the application stage to output predicted CFR data of each batch. The cropping is to make the data dimension of the predicted CFR data of each batch equal to the data dimension of the fitting target of the output data of each batch or equal to a prediction dimension. In the training stage, the first decoder is also used to calculate a loss according to the predicted CFR data of each batch and the fitting target of the output data of each batch, and train the first deep learning model through back propagation, so that the predicted CFR data of each batch output by the first decoder is as close as possible to the fitting target of the output data of the batch.

[0017] The application also proposes another encoding system for predicting channel state information, comprising a second preprocessor, a second encoder, a second deep learning model, and a second decoder. The second preprocessor is used to obtain input data of each batch, input data time mask of each batch, fitting target of output data of each batch, output data time mask of each batch, and coherent spatial vector of complex CFR data in a training stage according to complex channel frequency response (CFR) data. The second preprocessor is used to obtain input data of each batch, input data time mask of each batch, and coherent spatial vector of complex CFR data in an application stage according to historical complex CFR data obtained in an actual scene. The second encoder is used to obtain second addition data in both the training stage and the application stage according to the coherent spatial vector of complex CFR data, the input data time mask of each batch, and the input data of each batch. Specifically, the second encoder performs matrix calculation on the coherent spatial vector of complex CFR data and the input data time mask of each batch to obtain a coherent spatial sequence of complex CFR data of each batch. Then, the second encoder maps the coherent spatial sequence of complex CFR data of each batch to a high-dimensional space embedding through an embedding function to obtain coherent spatial encoding of complex CFR data of each batch, so that the data dimension of the coherent spatial encoding of complex CFR data of each batch is equal to the data dimension of the input data of each batch. Then, the second encoder adds the coherent spatial encoding of complex CFR data of each batch to the input data of each batch to obtain the second addition data. The second deep learning model is used to take the second addition data as input and output predicted CFR data in both the training stage and the application stage. The second decoder is used to crop the predicted CFR data output by the second deep learning model in both the training stage and the application stage to output predicted CFR data of each batch. The cropping is to make the data dimension of the predicted CFR data of each batch equal to the data dimension of the fitting target of the output data of each batch or equal to a prediction dimension. In the training stage, the second decoder is also used to calculate a loss according to the predicted CFR data of each batch and the fitting target of the output data of each batch, and to train the second deep learning model through back propagation, so that the predicted CFR data of each batch output by the second decoder is as close as possible to the fitting target of the output data of the batch.

[0018] The application also proposes another encoding system for predicting channel state information, comprising a third preprocessor, a third encoder, a third deep learning model, and a third decoder. The third preprocessor is used to obtain input data of each batch, input data time mask of each batch, fitting target of output data of each batch, output data time mask of each batch, and coherence time vector of complex channel impulse response (CIR) data in a training stage according to complex channel frequency response (CFR) data. The third preprocessor is used to obtain input data of each batch, input data time mask of each batch, and coherence time vector of complex CIR data in an application stage according to historical complex CFR data obtained in an actual scene. The third encoder is used to obtain third addition data in the training stage and the application stage according to the coherence time vector of complex CIR data, the input data time mask of each batch, and the input data of each batch. Specifically, the third encoder performs matrix calculation on the coherence time vector of complex CIR data and the input data time mask of each batch to obtain a coherence time sequence of complex CIR data of each batch. Then, the third encoder maps the coherence time sequence of complex CIR data of each batch to a high-dimensional space by an embedding function to obtain coherence time encoding of complex CIR data of each batch, so that the data dimension of the coherence time encoding of complex CIR data of each batch is equal to the data dimension of the input data of each batch. Then, the third encoder adds the coherence time encoding of complex CIR data of each batch to the input data of each batch to obtain the third addition data. The third deep learning model is used to take the third addition data as input and output predicted CFR data in the training stage and the application stage. The third decoder is used to crop the predicted CFR data output by the third deep learning model in the training stage and the application stage to output predicted CFR data of each batch. The cropping is to make the data dimension of the predicted CFR data of each batch equal to the data dimension of the fitting target of the output data of each batch or equal to a prediction dimension. The third decoder is also used to calculate a loss according to the predicted CFR data of each batch and the fitting target of the output data of each batch in the training stage, and to train the third deep learning model by back propagation, so that the predicted CFR data of each batch output by the third decoder is as close as possible to the fitting target of the output data of the batch.

[0019] The application further provides another encoding system for predicting channel state information, comprising a fourth preprocessor, a fourth encoder, a fourth deep learning model and a fourth decoder. The fourth preprocessor is used to obtain input data of each batch, input data time mask of each batch, fitting target of output data of each batch and coherent spatial vector of complex channel impulse response (CIR) data in a training stage according to complex channel frequency response (CFR) data. In an application stage, the fourth preprocessor is used to obtain input data of each batch, input data time mask of each batch and coherent spatial vector of complex CIR data according to historical complex CFR data obtained in an actual scene. The fourth encoder is used to obtain fourth addition data according to the coherent spatial vector of complex CIR data, input data time mask of each batch and input data of each batch in the training stage and the application stage. Specifically, the fourth encoder performs matrix calculation on the coherent spatial vector of complex CIR data and the input data time mask of each batch to obtain a coherent spatial sequence of complex CIR data of each batch. Then, the fourth encoder maps the coherent spatial sequence of complex CIR data of each batch to a high-dimensional space by an embedding function to obtain coherent spatial encoding of complex CIR data of each batch, so that the data dimension of the coherent spatial encoding of complex CIR data of each batch is equal to the data dimension of the input data of each batch. Then, the fourth encoder adds the coherent spatial encoding of complex CIR data of each batch to the input data of each batch to obtain the fourth addition data. The fourth deep learning model is used to take the fourth addition data as input and output predicted CFR data in the training stage and the application stage. The fourth decoder is used to crop the predicted CFR data output by the fourth deep learning model to output predicted CFR data of each batch in the training stage and the application stage. The cropping is to make the data dimension of the predicted CFR data of each batch equal to the data dimension of the fitting target of the output data of each batch or equal to a prediction dimension. In the training stage, the fourth decoder is further used to calculate a loss according to the predicted CFR data of each batch and the fitting target of the output data of each batch, and train the fourth deep learning model by back propagation, so that the predicted CFR data of each batch output by the fourth decoder is as close as possible to the fitting target of the output data of the batch.

[0020] The application has the beneficial effects that the time and space information is included in the input data of the deep learning model by calculating the time sequence characteristics of the channel state information and adding the encoding, so that the relative stability time of the channel in the future time is considered, the problem of weak time sequence and space information capturing ability in the model is solved, and the interpretability and accuracy of the channel state prediction are improved. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1This is a flowchart illustrating an embodiment of the coding method for predicting channel state information proposed in this application.

[0022] Figure 2 This is a flowchart illustrating the process of obtaining data for each batch in step SA1.

[0023] Figure 3 This is a schematic diagram illustrating the specific process of the first implementation of calculating the coherent time vector in step SA1.

[0024] Figure 4 This is a schematic diagram illustrating the specific process of the second implementation of calculating the coherent time vector in step SA1.

[0025] Figure 5 This is a flowchart illustrating a second embodiment of the coding method for predicting channel state information proposed in this application.

[0026] Figure 6 This is a schematic diagram illustrating the specific process of the first implementation of calculating the coherent spatial vector in step SB1.

[0027] Figure 7 This is a schematic diagram illustrating the specific process of the second implementation of calculating the coherent spatial vector in step SB1.

[0028] Figure 8 This is a flowchart illustrating Embodiment 3 of the coding method for predicting channel state information proposed in this application.

[0029] Figure 9 This is a flowchart illustrating Embodiment 4 of the coding method for predicting channel state information proposed in this application.

[0030] Figure 10 This is a schematic diagram of an embodiment of the coding system for predicting channel state information proposed in this application.

[0031] Figure 11 This is a schematic diagram of a second embodiment of the coding system for predicting channel state information proposed in this application.

[0032] Figure 12 This is a schematic diagram of the structure of Embodiment 3 of the coding system for predicting channel state information proposed in this application.

[0033] Figure 13 This is a schematic diagram of the fourth embodiment of the coding system for predicting channel state information proposed in this application.

[0034] Figure legend: first preprocessor A1, first encoder A2, first deep learning model A3, first decoder A4, second preprocessor B1, second encoder B2, second deep learning model B3, second decoder B4, third preprocessor C1, third encoder C2, third deep learning model C3, third decoder C4, fourth preprocessor D1, fourth encoder D2, fourth deep learning model D3, fourth decoder D4. DETAILED DESCRIPTION

[0035] Please refer to Figure 1 The embodiment one of the encoding method for predicting channel state information proposed in the present application includes the following steps.

[0036] Step SA1: obtaining, according to complex CFR (channel frequency response) data, input data of each batch batch_x, input data time mask of each batch batch_x_mark, fitting target of output data of each batch batch_y, output data time mask of each batch batch_y_mark, and coherence time vector tcVec_CFR of the complex CFR data. Preferably, the complex CFR data in this step is data after data preprocessing such as deduplication and interpolation. The coherence time vector tcVec_CFR of the complex CFR data embodies the characteristics of the CFR data in the time dimension.

[0037] Step SA2: the encoder obtains first addition data according to the coherence time vector tcVec_CFR of the complex CFR data, the input data time mask of each batch batch_x_mark, and the input data of each batch batch_x.

[0038] Specifically, the encoder performs matrix calculation on the coherence time vector tcVec_CFR of the complex CFR data and the input data time mask batch_x_mark of each batch to obtain the coherence time sequence batch_tcSequence_CFR of the complex CFR data of each batch. Then the encoder maps the coherence time sequence batch_tcSequence_CFR of the complex CFR data of each batch to a high-dimensional space by an embedding function to obtain the coherence time encoding batch_tcEmbedding_CFR of the complex CFR data of each batch. The embedding function is used to map discrete data to a continuous high-dimensional vector space, and the data dimension of the coherence time encoding batch_tcEmbedding_CFR of the complex CFR data of each batch is equal to the data dimension of the input data batch_x of each batch. Then the encoder adds the coherence time encoding batch_tcEmbedding_CFR of the complex CFR data of each batch to the input data batch_x of each batch, for example, directly adding the numerical values of each element of the two, to obtain first addition data.

[0039] Step SA3: The first addition data is input into the first deep learning model, and the first deep learning model outputs predicted CFR data. The first deep learning model is used to predict future CFR data from past CFR data, and focuses on the features of historical CFR data in the time dimension. For example, the first deep learning model sequentially includes a one-dimensional convolution layer, a batch normalization layer, an activation function layer, and a pooling layer to extract features and learn weights.

[0040] Step SA4: The decoder crops the predicted CFR data output by the first deep learning model to output the predicted CFR data of each batch. The cropping is to perform matrix calculation on the predicted CFR data and the output data time mask batch_y_mark of each batch to obtain the predicted CFR data of each batch, and the data dimension of the predicted CFR data of each batch is equal to the data dimension of the fitting target batch_y of the output data of each batch. For example, the decoder contains a fully connected layer, and this step crops the output of the fully connected layer. Then the loss is calculated according to the predicted CFR data of each batch and the fitting target batch_y of the output data of each batch, so that the predicted CFR data of each batch output by the decoder is as close as possible to the fitting target batch_y of the output data of the batch.

[0041] The steps SA2 to SA4 are the processes of training the first deep learning model joint encoder and decoder. In the training process, for example, using the mean squared error (MSE) function as the loss function, the error between the prediction result and the fitting target is calculated and the first deep learning model is trained by back propagation. In the training process, for example, using the adaptive moment estimation (Adam) as the optimization algorithm, the calculation efficiency is improved.

[0042] Step SA5: using the trained first deep learning model joint encoder and decoder, input the historical complex CFR data obtained in the actual scene to output the predicted future CFR data. The trained first deep learning model, encoder and decoder are deployed in the chip of the mobile terminal for example. The historical CFR data obtained in the actual scene is collected by the sensor for example.

[0043] Specifically, first, the input data batch_x of each batch, the input data time mask batch_x_mark of each batch, and the coherence time vector tcVec_CFR of the complex CFR data are obtained according to the historical complex CFR data obtained in the actual scene (refer to step SA1); then the first added data is obtained by the encoder according to the coherence time vector tcVec_CFR of the complex CFR data, the input data time mask batch_x_mark of each batch, and the input data batch_x of each batch (refer to step SA2); then the first added data is input into the first deep learning model, and the first deep learning model outputs the predicted CFR data (refer to step SA3); then the decoder crops the predicted CFR data output by the first deep learning model to obtain the predicted CFR data of the prediction dimension (refer to step SA4).

[0044] Please refer to Figure 2 In step SA1, the input data batch_x of each batch, the input data time mask batch_x_mark of each batch, the fitting target batch_y of the output data of each batch, and the output data time mask batch_y_mark of each batch are obtained according to the complex CFR data; specifically including the following steps.

[0045] Step S21: Obtain complex CFR data in a past period of time, split the complex CFR data into real and imaginary parts and longitudinally splice them to obtain real CFR data. The complex CFR data is a complex matrix of n rows and t columns, where n represents the number of subcarriers, and t represents the time length of the complex CFR data. Each matrix element in the complex CFR data is a complex number in the form of a+j×b, where a represents the real part of the complex number, j represents the imaginary symbol, and b represents the imaginary part of the complex number. The real part of each matrix element in the complex CFR data forms a real matrix of n rows and t columns, and the imaginary part of each matrix element in the complex CFR data also forms a real matrix of n rows and t columns. The two real matrices are longitudinally spliced (vertically spliced) to obtain a real matrix of 2n rows and t columns, which becomes a real matrix of t rows and 2n columns after transposition, which is the real CFR data; n represents the number of subcarriers, and t represents the time length of the real CFR data.

[0046] Step S22: Obtain the input data set X, the input data time mask set X_mark, the fitting target set Y of the output data, and the output data time mask set Y_mark of the deep learning model from the real CFR data through a sliding time window.

[0047] Specifically, a time window with a length of seq_len+pred_len is used to slide on the real CFR data at a specified step length step until the time window covers all time steps of the real CFR data. In a time series prediction model, a time step refers to the number of consecutive observation points. For example, data in one or more time steps in the past (i.e., one or more consecutive observation points in the past) is used to predict data in one or more time steps in the future (i.e., one or more consecutive prediction points in the future). The length of the time window covered each time is seq_len+pred_len in the real CFR data, the data with a time length of seq_len before is referred to as x, and the time window with a time length of seq_len covered by the data x is used as the input data time mask x_mark; the data with a time length of pred_len after is referred to as y, and the time window with a time length of pred_len covered by the data y is used as the output data time mask y_mark. Splicing the data x of each time window obtains a sequence X as the input data set of the deep learning model. Splicing the input data time mask x_mark of each time window obtains a sequence X_mark as the input data time mask set of the deep learning model. Splicing the data y of each time window obtains a sequence Y as the fitting target set of the output data of the deep learning model. Splicing the output data time mask y_mark of each time window obtains a sequence Y_mark as the output data time mask set of the deep learning model.

[0048] Step S23: The artificial neural network trains the data in batches, and respectively obtains, according to the data amount batch_size of each batch, the input data batch_x, the input data time mask batch_x_mark, the fitting target batch_y of the output data, and the output data time mask batch_y_mark of each batch from the input data set, the input data time mask set, the fitting target set of the output data, and the output data time mask set.

[0049] Please refer to Figure 3 In the step SA1, a coherence time vector tcVec_CFR of the complex CFR data is obtained according to the complex CFR data. The first implementation manner includes the following steps.

[0050] Step S31: A first norm of the complex CFR data in the time dimension is calculated.

[0051] The complex CFR data is a complex matrix of n rows and t columns, where n represents the number of subcarriers, and t represents the time length of the complex CFR data. The elements of adjacent two columns (for example, the jth column and the (j-1)th column) in the complex matrix of the complex CFR data are first differentiated, the adjacent two columns represent adjacent two time steps, and then the p-norm CFRLtp of the amplitude of the difference value of each adjacent two column elements is calculated. j p is a positive integer, and j is an integer between 0 and t-1.

[0052] In general, the p-norm of the amplitude of the difference value of each adjacent two column elements in the complex matrix of the complex CFR data is , where i represents a subcarrier index, i is an integer between 0 and n-1, j represents a time index, and j is an integer between 0 and t-1. CFR i,j represents the element in the (i+1)th row and the (j+1)th column of the complex matrix of the complex CFR data. CFR i,j-1 represents the element in the (i+1)th row and the jth column of the complex matrix of the complex CFR data. When j is 0, CFR i,-1 has no actual physical meaning, for example, all values are 0. The abs() function represents calculating the amplitude of a complex number.

[0053] For example, when p is 1, the L1-norm of the amplitude of the difference value of each adjacent two column elements in the complex matrix of the complex CFR data is .

[0054] For another example, when p is 2, the L2-norm of the amplitude of the difference value of each adjacent two column elements in the complex matrix of the complex CFR data is .

[0055] p-norm of the amplitude of the difference value of each pair of adjacent elements in the complex matrix of the complex CFR data CFR j There are t, and the value range of j is an integer from 0 to t-1. The t CFRs are combined in the order of j from small to large to obtain the first norm of the complex CFR data in the time dimension CFR j The first norm of the complex CFR data in the time dimension is a real one-dimensional matrix (i.e., a row vector) with a length of t. This step is to extract the differential characteristics of the complex CFR data over time. The first norm of the complex CFR data in the time dimension is a real one-dimensional matrix (i.e., a row vector) with a length of t. This step is to extract the differential characteristics of the complex CFR data over time.

[0056] Step S32: Calculate the first change rate of the j-th position in the time dimension of the first norm of the complex CFR data CFR where i and j are integers, and 0≤i<j≤t-1. .

[0057] Step S33: Use the first change rate and the first threshold to calculate tcVec_CFR j The initialized tcVec_CFR0=0. When , for all i that satisfy the condition, take the minimum value min(i) of i, tcVec_CFR j =1-min(i). When , tcVec_CFR j =0. Since 1≤j≤t-1, tcVec_CFR j There are t-1 in total.

[0058] Combine tcVec_CFR0and the t-1 tcVec_CFR j in the order of the subscript from small to large to obtain the coherence time vector tcVec_CFR of the complex CFR data, which is a one-dimensional matrix (i.e., a row vector) with a length of t.

[0059] Please refer to Figure 4 , in the step SA1, the coherence time vector tcVec_CFR of the complex CFR data is obtained according to the complex CFR data; the second implementation manner includes the following steps.

[0060] Step S41: The complex CFR data is a complex matrix with n rows and t columns, where n represents the number of subcarriers, and t represents the time length of the complex CFR data. Calculate the difference matrix ΔCFR of the complex CFR data, which is a real matrix with t rows and t columns, where each element ΔCFR i,j is calculated as follows.

[0061] When i < j, where k denotes the subcarrier index, and k takes integer values from 0 to n-1. i and j both denote the time index, and i and j take integer values from 0 to t-1. CFR k,j denotes the element in the (k+1)th row and the (j+1)th column of the complex matrix of complex CFR data. CFR k,i denotes the element in the (k+1)th row and the (i+1)th column of the complex matrix of complex CFR data.

[0062] When i = j, .

[0063] When i > j, ΔCFR i,j = 0.

[0064] ΔCFR is an upper triangular matrix, and the lower triangular is all 0 elements.

[0065] Step S42: Calculate the second rate of change of the j-th position in time relative to the i-th position in time in the difference matrix ΔCFR of complex CFR data where i and j are integers, and 0≤i . Where, denotes the p-norm of a vector (called the second norm in the time dimension), and p is a positive integer. ΔCFR i,j denotes the element in the (i+1)th row and the (j+1)th column of the difference matrix ΔCFR. ΔCFR j,j denotes the element in the (j+1)th row and the (j+1)th column of the difference matrix ΔCFR. The elements in the difference matrix ΔCFR are all complex numbers, and any complex number can be represented as a two-dimensional vector. The first norm in step S31 is to calculate the p-norm after differencing adjacent time steps, and the second norm in this step is to calculate the p-norm after differencing the current time step and all previous time steps.

[0066] Step S43: Calculate tcVec_CFR using the second rate of change and a second threshold j . Initialize tcVec_CFR0=0. When , for all i that satisfy the condition, take the maximum value max(i) of i, . When , tcVec_CFR j = 0. Since 1≤j≤t-1, there are t-1 tcVec_CFR j .

[0067] tcVec_CFR0and the t-1 tcVec_CFRj The coherence time vector tcVec_CFR of the complex CFR data is obtained by combining the subscripts in ascending order. The tcVec_CFR is a one-dimensional matrix (i.e., a row vector) with a length of t.

[0068] Referring to Figure 5 Embodiment two of the encoding method for predicting channel state information proposed in the present application includes the following steps.

[0069] Step SB1: Obtain the input data batch_x, the input data time mask batch_x_mark, the fitting target batch_y of the output data, the output data time mask batch_y_mark, and the coherence spatial vector scVec_CFR of each batch according to the complex CFR data. Preferably, the complex CFR data in this step is data after data preprocessing such as deduplication and interpolation. The coherence spatial vector scVec_CFR of the complex CFR data reflects the characteristics of the CFR data in the spatial dimension.

[0070] Wherein, the input data batch_x, the input data time mask batch_x_mark, the fitting target batch_y of the output data, and the output data time mask batch_y_mark of each batch are obtained according to the complex CFR data; for example, the method shown in Figure 2 is adopted.

[0071] Step SB2: The encoder obtains the second addition data according to the coherence spatial vector scVec_CFR of the complex CFR data, the input data time mask batch_x_mark, and the input data batch_x of each batch.

[0072] Specifically, the encoder performs matrix calculation on the coherent spatial vector scVec_CFR of the complex CFR data and the input data time mask batch_x_mark of each batch to obtain the coherent spatial sequence batch_scSequence_CFR of the complex CFR data of each batch. Then the encoder maps the coherent spatial sequence batch_scSequence_CFR of the complex CFR data of each batch to a high-dimensional space embedding through an embedding function to obtain the coherent spatial encoding batch_scEmbedding_CFR of the complex CFR data of each batch. The embedding function is used to map discrete data to a continuous high-dimensional vector space, and the data dimension of the coherent spatial encoding batch_scEmbedding_CFR of the complex CFR data of each batch is equal to the data dimension of the input data batch_x of each batch. Then the encoder adds the coherent spatial encoding batch_scEmbedding_CFR of the complex CFR data of each batch to the input data batch_x of each batch, for example, directly adding the numerical value of each element of the two, to obtain second addition data.

[0073] Step SB3: The second addition data is input into a second deep learning model, and the second deep learning model outputs predicted CFR data. The second deep learning model is used to predict future CFR data from past CFR data, and focuses on the features of historical CFR data in the spatial dimension. For example, the second deep learning model sequentially includes a one-dimensional convolution layer, a batch normalization layer, an activation function layer, and a pooling layer to extract features and learn weights.

[0074] Step SB4: The decoder crops the predicted CFR data output by the second deep learning model to output the predicted CFR data of each batch. The cropping is to perform matrix calculation on the predicted CFR data and the output data time mask batch_y_mark of each batch to obtain the predicted CFR data of each batch, and the data dimension of the predicted CFR data of each batch is equal to the data dimension of the fitting target batch_y of the output data of each batch. For example, the decoder contains a fully connected layer, and this step crops the output of the fully connected layer. Then the loss is calculated according to the predicted CFR data of each batch and the fitting target batch_y of the output data of each batch, so that the predicted CFR data of each batch output by the decoder is as close as possible to the fitting target batch_y of the output data of the batch.

[0075] The steps SB2 to SB4 are the processes of training the second deep learning model joint encoder and decoder. In the training process, for example, using the mean square error function as the loss function, the error between the predicted result and the fitting target is calculated and the second deep learning model is trained by back propagation. In the training process, for example, using the adaptive moment estimation optimizer as the optimization algorithm, the calculation efficiency is improved.

[0076] Step SB5: Using the trained second deep learning model joint encoder and decoder, the historical complex CFR data obtained in the actual scene is taken as input, and the predicted future CFR data is output. The trained second deep learning model, encoder and decoder are deployed in the chip of the mobile terminal, for example. The historical CFR data obtained in the actual scene is collected by a sensor, for example.

[0077] Specifically, first, the input data batch_x of each batch, the input data time mask batch_x_mark of each batch, and the coherent spatial vector scVec_CFR of the complex CFR data are obtained according to the historical complex CFR data obtained in the actual scene (refer to step SB1); then the second addition data is obtained by the encoder according to the coherent spatial vector scVec_CFR of the complex CFR data, the input data time mask batch_x_mark of each batch, and the input data batch_x of each batch (refer to step SB2); then the second addition data is taken as the input of the second deep learning model, and the second deep learning model outputs the predicted CFR data (refer to step SB3); then the decoder crops the predicted CFR data output by the second deep learning model to obtain the predicted CFR data of the prediction dimension (refer to step SB4).

[0078] Please refer to Figure 6 In step SB1, the coherent spatial vector scVec_CFR of the complex CFR data is obtained according to the complex CFR data. The first implementation manner includes the following steps.

[0079] Step S61: Calculate the third norm of the complex CFR data in the subcarrier dimension.

[0080] The complex CFR data is an n-row × t-column complex matrix, where n represents the number of subcarriers, and t represents the time length of the complex CFR data. The transpose matrix of the complex CFR data is represented as , which is a t-row × n-column complex matrix. The difference between the elements of the adjacent two columns (for example, the jth column and the j-1th column) in the transpose matrix of the complex CFR data is calculated first, and the adjacent two columns represent two adjacent subcarriers. Then the p-norm of the amplitude of the difference value of each adjacent two columns is calculated , p is a positive integer, and j is an integer ranging from 0 to n-1.

[0081] The p-norm of the amplitude of the difference value of each adjacent two columns of elements in the transpose matrix of the complex CFR data , where i represents a time index, i is an integer ranging from 0 to t-1. j represents a subcarrier index, j is an integer ranging from 0 to n-1. represents the element in the i+1th row and the j+1th column of the transpose matrix of the complex CFR data. represents the element in the i+1th row and the jth column of the transpose matrix of the complex CFR data. When j takes 0, has no actual physical meaning, for example, all values are 0. The abs() function represents calculating the amplitude of a complex number.

[0082] The p-norm of the amplitude of the difference value of each adjacent two columns of elements in the transpose matrix of the complex CFR data j There are n, corresponding to j, which is an integer ranging from 0 to n-1. The n CFRLsp j are combined in the order of j from small to large, and the third norm of the complex CFR data in the subcarrier dimension CFRLsp is obtained. The third norm of the complex CFR data in the subcarrier dimension CFRLsp is a real one-dimensional matrix (i.e. a row vector) with a length of n. This step is to extract the differential characteristics of the complex CFR data with respect to the subcarrier.

[0083] Step S62: Calculate the third change rate of the subcarrier position j behind the subcarrier position i in front of the third norm of the complex CFR data in the subcarrier dimension CFRLsp , where i and j are integers, and 0≤i<j≤n-1. .

[0084] Step S63: Calculate scVec_CFR using the third change rate and the third threshold j . The initialized scVec_CFR0=0. When , for all i that meet the conditions, take the minimum value min(i) of i, scVec_CFR j =1-min(i). When , scVec_CFR j =0. Since 1≤j≤n-1, scVec_CFR j There are n-1.

[0085] scVec_CFR = scVec_CFR + n-1 j The coherent spatial vector scVec_CFR of complex CFR data is obtained by combining the phases in the order of the subscript from small to large, which is a one-dimensional matrix of length n (i.e., a row vector).

[0086] Please refer to Figure 7 , the step SB1, the coherent spatial vector scVec_CFR of complex CFR data is obtained according to the complex CFR data; the second implementation manner comprises the following steps.

[0087] Step S71: The complex CFR data is a complex matrix of n rows × t columns, where n represents the number of subcarriers, and t represents the time length of the complex CFR data. The transposed matrix of the complex CFR data is represented as , which is a complex matrix of t rows × n columns. The difference matrix ΔCFR of the transposed matrix of the complex CFR data T is an n row × n column real matrix, where each element is calculated according to the following formula.

[0088] When i < j, . Wherein, k represents the time index, and k takes an integer value between 0 and t-1. i and j both represent the subcarrier index, and i and j both take an integer value between 0 and n-1. represents the element in the k+1th row and the j+1th column of the transposed matrix of the complex CFR data. represents the element in the k+1th row and the i+1th column of the transposed matrix of the complex CFR data.

[0089] When i = j, .

[0090] When i > j, = 0.

[0091] ΔCFR T is an upper triangular matrix, and the lower triangular is all 0 elements.

[0092] Step S72: Calculate the difference matrix ΔCFR of the transposed matrix of the complex CFR data T The fourth change rate of the subcarrier position after j relative to the subcarrier position before i in the difference matrix ΔCFR , wherein i and j are both integers, and 0≤i < j≤n-1. . Wherein, The p-norm of a vector (called the fourth norm in the subcarrier dimension) is denoted as, p is a positive integer. The third norm in step S61 is the p-norm calculated after the difference between adjacent subcarriers, and the fourth norm in this step is the p-norm calculated after the difference between the current subcarrier and all previous subcarriers.

[0093] Step S73: using the fourth change rate and the fourth threshold to calculate scVec_CFR j . The initialized scVec_CFR0=0. When , take the maximum value max(i) of i for all i that meet the condition, When , scVec_CFR j =0. Since 1≤j≤n-1, scVec_CFR j has n-1 in total.

[0094] Combine scVec_CFR0and the n-1 scVec_CFR j in order of decreasing subscript to obtain the coherence spatial vector scVec_CFR of complex CFR data. scVec_CFR is a one-dimensional matrix of length n (i.e., a row vector).

[0095] Please refer to Figure 8 , the third embodiment of the encoding method for predicting channel state information proposed in this application includes the following steps.

[0096] Step SC1: obtaining, from the complex CFR data, input data batch_x of each batch, input data time mask batch_x_mark of each batch, fitting target batch_y of output data of each batch, output data time mask batch_y_mark of each batch, and coherence time vector tcVec_CIR of complex CIR (channel impulse response) data. Preferably, the complex CFR data in this step is data after data preprocessing such as de-duplication and interpolation. The coherence time vector tcVec_CIR of complex CIR data reflects the characteristics of the CFR data converted into CIR data in the time dimension.

[0097] wherein the input data batch_x of each batch, the input data time mask batch_x_mark of each batch, the fitting target batch_y of output data of each batch, and the output data time mask batch_y_mark of each batch are obtained from the complex CFR data; for example, the method shown in Figure 2 is used.

[0098] The coherence time vector tcVec_CIR of the complex CIR data is obtained according to the complex CFR data in the following manner. First, the complex CIR data is obtained by performing inverse Fourier transform on the complex CFR data, and then the complex CIR data is obtained according to the complex CIR data in any one of the implementation manners shown in Figure 3 Or Figure 4 The complex CIR data is obtained according to the complex CIR data in any one of the implementation manners shown in Figure 3 And Figure 4 The complex CFR data originally described in the complex CIR data is changed to the complex CIR data.

[0099] Step SC2: The encoder obtains third addition data according to the coherence time vector tcVec_CIR of the complex CIR data, the input data time mask batch_x_mark of each batch, and the input data batch_x of each batch.

[0100] Specifically, the encoder performs matrix calculation on the coherence time vector tcVec_CIR of the complex CIR data and the input data time mask batch_x_mark of each batch to obtain the coherence time sequence batch_tcSequence_CIR of the complex CIR data of each batch. Then the encoder maps the coherence time sequence batch_tcSequence_CIR of the complex CIR data of each batch to a high-dimensional space by an embedding function to obtain the coherence time encoding batch_tcEmbedding_CIR of the complex CIR data of each batch. The embedding function is used to map discrete data to a continuous high-dimensional vector space, and the data dimension of the coherence time encoding batch_tcEmbedding_CIR of the complex CIR data of each batch is equal to the data dimension of the input data batch_x of each batch. Then the encoder adds the coherence time encoding batch_tcEmbedding_CIR of the complex CIR data of each batch to the input data batch_x of each batch, for example, directly adds the numerical value of each element of the two, to obtain the third addition data.

[0101] Step SC3: The third addition data is input into the third deep learning model, and the third deep learning model outputs the predicted CFR data. The third deep learning model is used to predict the future CFR data according to the past CFR data, and focuses on the features of the historical CFR data in the time dimension after the historical CFR data is converted into historical CIR data. For example, the third deep learning model sequentially includes a one-dimensional convolution layer, a batch normalization layer, an activation function layer, and a pooling layer to extract features and learn weights.

[0102] Step SC4: the decoder crops the predicted CFR data output by the third deep learning model, and outputs the predicted CFR data of each batch. The cropping is to perform matrix calculation on the predicted CFR data and the output data time mask batch_y_mark of each batch, so that the data dimension of the predicted CFR data of each batch is equal to the data dimension of the fitting target batch_y of the output data of each batch. For example, the decoder includes a fully connected layer, and this step crops the output of the fully connected layer. Then, the loss is calculated according to the predicted CFR data of each batch and the fitting target batch_y of the output data of each batch, so that the predicted CFR data of each batch output by the decoder is as close as possible to the fitting target batch_y of the output data of the batch.

[0103] Steps SC2 to SC4 are the process of training the third deep learning model combined with the encoder and the decoder. In the training process, for example, using the mean square error function as the loss function, the error between the prediction result and the fitting target is calculated and the third deep learning model is trained by back propagation. In the training process, for example, using the adaptive moment estimation optimizer as the optimization algorithm, the calculation efficiency is improved.

[0104] Step SC5: using the trained third deep learning model combined with the encoder and the decoder, the historical complex CFR data obtained in the actual scene is taken as input, and the predicted future CFR data is output. The trained third deep learning model, the encoder and the decoder are deployed in the chip of a mobile terminal, for example. The historical CFR data obtained in the actual scene is obtained by a sensor, for example.

[0105] Specifically, first, the input data batch_x of each batch, the input data time mask batch_x_mark of each batch, and the coherence time vector tcVec_CIR of the complex CIR data are obtained according to the historical complex CFR data obtained in the actual scene (refer to step SC1); then the third added data is obtained by the encoder according to the coherence time vector tcVec_CIR of the complex CIR data, the input data time mask batch_x_mark of each batch, and the input data batch_x of each batch (refer to step SC2); then the third deep learning model takes the third added data as input, and outputs the predicted CFR data (refer to step SC3); then the decoder crops the predicted CFR data output by the third deep learning model, and obtains the predicted CFR data of the prediction dimension (refer to step SC4).

[0106] Please refer to Figure 9 Embodiment four of the encoding method for predicting channel state information provided in the present application includes the following steps.

[0107] Step SD1: obtaining, according to the complex CFR data, input data of each batch batch_x, an input data time mask of each batch batch_x_mark, a fitting target of output data of each batch batch_y, an output data time mask of each batch batch_y_mark, and a coherent spatial vector scVec_CIR of the complex CIR data. Preferably, the complex CFR data in this step is data after data preprocessing such as deduplication and interpolation. The coherent spatial vector scVec_CIR of the complex CIR data reflects the characteristics of the CFR data converted into the CIR data in the spatial dimension.

[0108] wherein the input data of each batch batch_x, the input data time mask of each batch batch_x_mark, the fitting target of output data of each batch batch_y, and the output data time mask of each batch batch_y_mark are obtained according to the complex CFR data; for example, the method shown in FIG. 8 is adopted. Figure 2

[0109] wherein the coherent spatial vector scVec_CIR of the complex CIR data is obtained according to the complex CFR data; the following method is adopted. First, inverse Fourier transform is performed on the complex CFR data to obtain the complex CIR data, and then any one of the implementation manners shown in FIGS. 9 to 11 is referred to. Figure 6 Figure 7 wherein the coherent spatial vector scVec_CIR of the complex CIR data is obtained according to the complex CIR data, Figure 6 Figure 7 and the complex CFR data originally recorded in steps SD1 to SD3 is replaced by the complex CIR data.

[0110] Step SD2: obtaining, by the encoder, fourth addition data according to the coherent spatial vector scVec_CIR of the complex CIR data, the input data time mask of each batch batch_x_mark, and the input data of each batch batch_x.

[0111] ​​​Specifically, the encoder performs matrix calculation on the coherence space vector scVec_CIR of the complex CIR data and the input data time mask batch_x_mark of each batch to obtain the coherence space sequence batch_scSequence_CIR of the complex CIR data of each batch. Then the encoder maps the coherence space sequence batch_scSequence_CIR of the complex CIR data of each batch to a high-dimensional space embedding through an embedding function to obtain the coherence space encoding batch_scEmbedding_CIR of the complex CIR data of each batch. The embedding function is used to map discrete data to a continuous high-dimensional vector space, and the data dimension of the coherence space encoding batch_scEmbedding_CIR of the complex CIR data of each batch is equal to the data dimension of the input data batch_x of each batch. Then the encoder adds the coherence space encoding batch_scEmbedding_CIR of the complex CIR data of each batch to the input data batch_x of each batch, for example, directly adding the numerical values of each element of the two, to obtain fourth addition data.

[0112] Step SD3: The fourth addition data is input into the fourth deep learning model, and the fourth deep learning model outputs predicted CFR data. The fourth deep learning model is used to predict future CFR data from past CFR data, and focuses on the features of the historical CFR data in the subcarrier dimension after the historical CFR data is converted into historical CIR data. For example, the fourth deep learning model sequentially includes a one-dimensional convolution layer, a batch normalization layer, an activation function layer, and a pooling layer to extract features and learn weights.

[0113] Step SD4: The decoder crops the predicted CFR data output by the fourth deep learning model to output predicted CFR data of each batch. The cropping is to perform matrix calculation on the predicted CFR data and the output data time mask batch_y_mark of each batch to obtain predicted CFR data of each batch, and the data dimension of the predicted CFR data of each batch is equal to the data dimension of the fitting target batch_y of the output data of each batch. For example, the decoder includes a fully connected layer, and this step crops the output of the fully connected layer. Then the loss is calculated according to the predicted CFR data of each batch and the fitting target batch_y of the output data of each batch, so that the predicted CFR data of each batch output by the decoder is as close as possible to the fitting target batch_y of the output data of the batch.

[0114] The steps SD2 to SD4 are the process of training the fourth deep learning model joint encoder and decoder. In the training process, for example, using the mean square error function as the loss function, the error between the predicted result and the fitting target is calculated and back propagation is used to train the fourth deep learning model. In the training process, for example, using the adaptive moment estimation optimizer as the optimization algorithm, the calculation efficiency is improved.

[0115] Step SD5: Using the trained fourth deep learning model joint encoder and decoder, the historical complex CFR data obtained in the actual scene is taken as input, and the predicted future CFR data is output. The trained fourth deep learning model, encoder and decoder are deployed in the chip of the mobile terminal, for example. The historical CFR data obtained in the actual scene is collected by a sensor, for example.

[0116] Specifically, first, the input data batch_x of each batch, the input data time mask batch_x_mark of each batch, and the coherent spatial vector scVec_CIR of the complex CIR data are obtained according to the historical complex CFR data obtained in the actual scene (refer to step SD1); then the fourth addition data is obtained by the encoder according to the coherent spatial vector scVec_CIR of the complex CIR data, the input data time mask batch_x_mark of each batch, and the input data batch_x of each batch (refer to step SD2); then the fourth addition data is taken as the input of the fourth deep learning model, and the fourth deep learning model outputs the predicted CFR data (refer to step SD3); then the decoder crops the predicted CFR data output by the fourth deep learning model to obtain the predicted CFR data of the predicted dimension (refer to step SD4).

[0117] The basic principle of the encoding method for predicting channel state information proposed in the present application is that in channel state information prediction, CFR data is taken as the prediction target. CFR data can describe the influence of the channel on signal propagation from the amplitude-frequency characteristic and the phase-frequency characteristic, respectively. With the channel frequency response of a specific signal, the propagation characteristics of different multipath paths can be calculated, from which useful information for positioning and perception can be analyzed. The amplitude-frequency characteristic represents the characteristic of the ratio of the amplitude of the output signal to the amplitude of the input signal varying with frequency, and the phase-frequency characteristic represents the characteristic of the phase difference of the output signal to the phase of the input signal varying with frequency. At the same time, the encoding method proposed in the present application uses the rate of change of historical CFR data to describe the change of relative coherence time (Embodiment One), or uses the rate of change of historical CFR data to describe the change of relative coherence space (Embodiment Two), or uses the rate of change of historical CIR data to describe the change of relative coherence time (Embodiment Three), or uses the rate of change of historical CIR data to describe the change of relative coherence space (Embodiment Four), so as to be able to fuse more rich time information or more rich space information. The relative coherence time refers to the time required for the CFR data or CIR data to change under a certain threshold. The relative coherence space refers to the number of subcarriers required for the CFR data or CIR data to change under a certain threshold. The method proposed in the present application is suitable for deep learning models that lack physical meaning and are difficult to explain, such as CNN models, temporal convolutional networks (TCN), etc. The present application adds additional time sequence information encoding after data preprocessing (steps SA2, SB2, SC2, SD2) to help the deep learning model obtain more channel state information.

[0118] Please refer to Figure 10 Embodiment One of the encoding system for predicting channel state information proposed in the present application includes a first preprocessor A1, a first encoder A2, a first deep learning model A3, and a first decoder A4. Figure 10 Embodiment One of the system shown corresponds to Figure 1 Embodiment One of the method shown.

[0119] The first preprocessor A1 is configured to obtain, in the training stage, the input data batch_x, the input data time mask batch_x_mark, the fitting target batch_y of the output data, the output data time mask batch_y_mark of each batch, and the coherence time vector tcVec_CFR of the complex CFR data according to the complex CFR data. The first preprocessor A1 is configured to obtain, in the application stage, the input data batch_x, the input data time mask batch_x_mark, and the coherence time vector tcVec_CFR of the complex CFR data according to the historical complex CFR data obtained in the actual scene.

[0120] The first encoder A2 is configured to obtain, in the training stage and the application stage, the first added data according to the coherence time vector tcVec_CFR of the complex CFR data, the input data time mask batch_x_mark of each batch, and the input data batch_x of each batch. Specifically, the first encoder A2 performs matrix calculation on the coherence time vector tcVec_CFR of the complex CFR data and the input data time mask batch_x_mark of each batch to obtain the coherence time sequence batch_tcSequence_CFR of the complex CFR data of each batch. Then, the first encoder A2 maps the coherence time sequence batch_tcSequence_CFR of the complex CFR data of each batch to a high-dimensional space by an embedding function to obtain the coherence time encoding batch_tcEmbedding_CFR of the complex CFR data of each batch, so that the data dimension of the coherence time encoding batch_tcEmbedding_CFR of the complex CFR data of each batch is equal to the data dimension of the input data batch_x of each batch. Then, the first encoder A2 adds the coherence time encoding batch_tcEmbedding_CFR of the complex CFR data of each batch to the input data batch_x of each batch to obtain the first added data.

[0121] The first deep learning model A3 is configured to take the first added data as input and output the predicted CFR data in the training stage and the application stage.

[0122] The first decoder A4 is configured to crop the predicted CFR data output by the first deep learning model A3 in both the training phase and the application phase, and output the predicted CFR data of each batch, the purpose of the cropping being to make the data dimension of the predicted CFR data of each batch equal to the data dimension of the fitting target batch_y of the output data of each batch (in the training phase) or equal to the prediction dimension (in the application phase). In the training phase, the first decoder A4 is further configured to calculate a loss according to the predicted CFR data of each batch and the fitting target batch_y of the output data of each batch, and to train the first deep learning model A3 by back propagation, so that the predicted CFR data of each batch output by the first decoder A4 is as close as possible to the fitting target batch_y of the output data of the batch.

[0123] For example, the first deep learning model A3 can be a convolutional neural network (CNN) or a recurrent neural network (RNN). Figure 11 The second embodiment of the encoding system for predicting channel state information provided in the present application includes a second preprocessor B1, a second encoder B2, a second deep learning model B3, and a second decoder B4. Figure 11 The second embodiment of the system shown corresponds to Figure 5 The second embodiment of the method shown.

[0124] In the training phase, the second preprocessor B1 is configured to obtain, according to the complex CFR data, the input data batch_x of each batch, the input data time mask batch_x_mark of each batch, the fitting target batch_y of the output data of each batch, the output data time mask batch_y_mark of each batch, and the coherent spatial vector scVec_CFR of the complex CFR data.

[0125] The second encoder B2 is configured to obtain second addition data according to the coherent spatial vector scVec_CFR of the complex CFR data, the input data time mask batch_x_mark of each batch, and the input data batch_x of each batch, in both the training phase and the application phase. Specifically, the second encoder B2 performs matrix calculation on the coherent spatial vector scVec_CFR of the complex CFR data and the input data time mask batch_x_mark of each batch to obtain the coherent spatial sequence batch_scSequence_CFR of the complex CFR data of each batch. Then, the second encoder B2 maps the coherent spatial sequence batch_scSequence_CFR of the complex CFR data of each batch to a high-dimensional space embedding through an embedding function to obtain the coherent spatial encoding batch_scEmbedding_CFR of the complex CFR data of each batch, so that the data dimension of the coherent spatial encoding batch_scEmbedding_CFR of the complex CFR data of each batch is equal to the data dimension of the input data batch_x of each batch. Then, the second encoder B2 adds the coherent spatial encoding batch_scEmbedding_CFR of the complex CFR data of each batch to the input data batch_x of each batch to obtain the second addition data.

[0126] The second deep learning model B3 is configured to take the second addition data as input and output predicted CFR data, in both the training phase and the application phase.

[0127] The second decoder B4 is configured to crop the predicted CFR data output by the second deep learning model, and output the predicted CFR data of each batch, in both the training phase and the application phase. The purpose of the cropping is to make the data dimension of the predicted CFR data of each batch equal to the data dimension of the fitting target batch_y of the output data of each batch (in the training phase) or equal to the prediction dimension (in the application phase). In the training phase, the second decoder B4 is further configured to calculate a loss according to the predicted CFR data of each batch and the fitting target batch_y of the output data of each batch, and train the second deep learning model B3 through back propagation, so that the predicted CFR data of each batch output by the second decoder B4 is as close as possible to the fitting target batch_y of the output data of the batch.

[0128] For example, the method for predicting channel state information shown in the embodiment three of the encoding system includes the following steps: Figure 12 The embodiment three of the encoding system for predicting channel state information includes a third preprocessor C1, a third encoder C2, a third deep learning model C3, and a third decoder C4. Figure 12 The embodiment three of the system shown corresponds to the method shown in the embodiment three. Figure 8 The embodiment three of the method shown.

[0129] The third preprocessor C3 is configured to obtain, in the training stage, the input data batch x, the input data time mask batch x mark, the fitting target of the output data batch y, the output data time mask batch y mark, and the coherence time vector tcVec CIR of the complex CIR data according to the complex CFR data. The third preprocessor C3 is configured to obtain, in the application stage, the input data batch x, the input data time mask batch x mark, and the coherence time vector tcVec CIR of the complex CIR data according to the historical complex CFR data obtained in the actual scene.

[0130] The third encoder C2 is configured to obtain, in the training stage and the application stage, the third added data according to the coherence time vector tcVec CIR of the complex CIR data, the input data time mask batch x mark, and the input data batch x. Specifically, the third encoder C2 performs matrix calculation on the coherence time vector tcVec CIR of the complex CIR data and the input data time mask batch x mark to obtain the coherence time sequence batch tcSequence CIR of the complex CIR data. Then, the third encoder C2 maps the coherence time sequence batch tcSequence CIR of the complex CIR data to a high-dimensional space by using an embedding function to obtain the coherence time encoding batch tcEmbedding CIR of the complex CIR data, so that the data dimension of the coherence time encoding batch tcEmbedding CIR of the complex CIR data is equal to the data dimension of the input data batch x. Then, the third encoder C2 adds the coherence time encoding batch tcEmbedding CIR of the complex CIR data to the input data batch x to obtain the third added data.

[0131] The third deep learning model C3 is configured to take the third added data as input and output the predicted CFR data in the training stage and the application stage.

[0132] The third decoder C4 is configured to crop the predicted CFR data output by the third deep learning model in both the training stage and the application stage, and output the predicted CFR data of each batch, so that the data dimension of the predicted CFR data of each batch is equal to the data dimension of the fitting target batch_y of the output data of each batch (in the training stage) or equal to the prediction dimension (in the application stage). In the training stage, the third decoder C4 is further configured to calculate a loss according to the predicted CFR data of each batch and the fitting target batch_y of the output data of each batch, and train the third deep learning model C3 through back propagation, so that the predicted CFR data of each batch output by the third decoder C4 is as close as possible to the fitting target batch_y of the output data of the batch.

[0133] Referring to Figure 13 Embodiment four of the encoding system for predicting channel state information provided in the present application includes a fourth preprocessor D1, a fourth encoder D2, a fourth deep learning model D3, and a fourth decoder D4. Figure 13 Embodiment four of the system shown corresponds to Figure 9 Embodiment four of the method shown.

[0134] The fourth preprocessor D1 is configured to obtain, in the training stage, the input data batch_x, the input data time mask batch_x_mark, the fitting target batch_y of the output data of each batch, the output data time mask batch_y_mark, and the coherent spatial vector scVec_CIR of the complex CIR data according to the complex CFR data. In the application stage, the fourth preprocessor D1 is configured to obtain the input data batch_x, the input data time mask batch_x_mark, and the coherent spatial vector scVec_CIR of the complex CIR data according to the historical complex CFR data obtained in the actual scene.

[0135] The fourth encoder D2 is configured to obtain fourth added data from the coherent spatial vector scVec_CIR of the complex CIR data, the input data time mask batch_x_mark of each batch, and the input data batch_x of each batch in both the training stage and the application stage. Specifically, the fourth encoder D2 performs matrix calculation on the coherent spatial vector scVec_CIR of the complex CIR data and the input data time mask batch_x_mark of each batch to obtain the coherent spatial sequence batch_scSequence_CIR of the complex CIR data of each batch. Then, the fourth encoder D2 maps the coherent spatial sequence batch_scSequence_CIR of the complex CIR data of each batch to a high-dimensional space embedding by an embedding function to obtain the coherent spatial encoding batch_scEmbedding_CIR of the complex CIR data of each batch, so that the data dimension of the coherent spatial encoding batch_scEmbedding_CIR of the complex CIR data of each batch is equal to the data dimension of the input data batch_x of each batch. Then, the fourth encoder D2 adds the coherent spatial encoding batch_scEmbedding_CIR of the complex CIR data of each batch to the input data batch_x of each batch to obtain the fourth added data.

[0136] The fourth deep learning model D3 is configured to take the fourth added data as input and output predicted CFR data in both the training stage and the application stage.

[0137] The fourth decoder D4 is configured to crop the predicted CFR data output by the fourth deep learning model in both the training stage and the application stage to output the predicted CFR data of each batch. The purpose of the cropping is to make the data dimension of the predicted CFR data of each batch equal to the data dimension of the fitting target batch_y of the output data of each batch (in the training stage) or equal to the prediction dimension (in the application stage). In the training stage, the fourth decoder D4 is further configured to calculate a loss from the predicted CFR data of each batch and the fitting target batch_y of the output data of each batch, and train the fourth deep learning model D3 by back propagation, so that the predicted CFR data of each batch output by the fourth decoder D4 is as close as possible to the fitting target batch_y of the output data of the batch.

[0138] Compared with the prior art, the present application has the following beneficial technical effects. First, the present application introduces coherence information coding calculated from CFR data into the input data of the deep learning model, so that the deep learning model obtains more channel stability information and improves the accuracy of the model in predicting future channel states. Second, the present application uses CFR data for time series coding. This method uses easily accessible data, has an intuitive principle and is simple to calculate. It can be widely used in various deep learning-based models without changing the existing model architecture, only requiring a small amount of additional computing resources, while covering a wealth of physical information. In this way, the performance of the signal state prediction task can be significantly improved, showing its wide application prospects and high practicality. Third, the present application calculates the channel state information features and adds coding, which incorporates the coherence time or coherence space in the physical sense into the model, thereby considering the relative stability of the channel at future time or space, and solving the problem of weak physical meaning representation in the deep learning model, thereby improving the accuracy of channel state prediction.

[0139] In wireless communication, coherence time and coherence bandwidth are core concepts for describing the statistical characteristics of the channel over time and frequency. Coherence time represents the time interval within which the statistical characteristics of the channel remain basically stable, i.e., the multipath distribution, phase relationship and power spectrum, etc. do not change dramatically. Coherence bandwidth describes the high correlation of the channel within a certain frequency range. These coherence intervals reflect the physical laws of wireless channels: within a small range of time and space, the channel exhibits high consistency. Therefore, in a deep learning CSI prediction model, if this physical prior can be explicitly introduced, the model will be more consistent with the physical laws of the actual communication environment.

[0140] The inspiration for the coherence embedding designed in the present application comes from the segment embedding in the BERT (bidirectional encoder representations from transformers) model in the field of natural language processing. In BERT, segment embedding can clearly distinguish different semantic segments in the input sequence, letting the model know which tokens belong to which part, thereby providing context structure information for downstream tasks. Similarly, in the CSI prediction task, all time steps within the same coherence time or coherence space can be assigned to the same group, and the same embedding vector is assigned to the group. This embedding vector is superimposed or spliced with the original CSI input representation in the model to form a new input representation. The idea behind this is to incorporate the slowly changing ( ) Statistical properties are modeled separately from fast-varying (time-step) small-scale fading. Specifically, the encoder focuses on capturing fast-varying instantaneous multipath information, while the coherence embedding serves as a group-level context vector, carrying the scene background and slow-varying statistical features. Such a design is highly consistent with the actual physical process, equivalent to decomposing the channel prediction into the sum of slow-varying statistical components and fast-varying residual components, fully utilizing the channel coherence in wireless communication, and directly embedding the physical meaning into the model structure.

[0141] In terms of interpretability, since the same embedding is used for the same coherence time or coherence space in the present application, the embedding vector itself can be regarded as a "scene label" or "large-scale environment code". Therefore, the introduction of the coherence embedding directly encodes the coherence characteristics of the wireless channel as the input prior of the deep learning model, achieving the separate modeling of the slow-varying statistical properties and fast-varying instantaneous properties of the channel. This not only conforms to the physical laws of classical channel models, but also makes the model have a clear physical explanation in structure. Through the analysis and visualization of the embedding, researchers can intuitively understand the internal representation of the model and its correspondence with the real physical world, thereby significantly improving the interpretability and generalization ability of the CSI prediction model.

[0142] The above is only a preferred embodiment of the present application and is not intended to limit the present application. Those skilled in the art can make various changes and modifications to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A coding method for predicting channel state information, characterized in that, Includes the following steps; Step SA1: Obtain the input data for each batch, the input data time mask for each batch, the fitting target for the output data for each batch, the output data time mask for each batch, and the coherent time vector of the complex CFR data based on the complex channel frequency response (CFR) data. Step SA2: The encoder obtains the first summed data based on the coherent time vector of the complex CFR data, the input data time mask for each batch, and the input data for each batch; specifically, the encoder performs matrix calculation on the coherent time vector of the complex CFR data and the input data time mask for each batch to obtain the coherent time sequence of the complex CFR data for each batch. The encoder then uses an embedding function to map the coherent time series of complex CFR data in each batch to a high-dimensional space. After embedding, the coherent time code of complex CFR data in each batch is obtained, so that the data dimension of the coherent time code of complex CFR data in each batch is equal to the data dimension of the input data in each batch. The encoder then adds the coherent time encoding of the complex CFR data of each batch to the input data of each batch to obtain the first summed data; Step SA3: The first summed data is used as the input to the first deep learning model, and the first deep learning model outputs the predicted CFR data; Step SA4: The decoder prunes the predicted CFR data output by the first deep learning model to output the predicted CFR data for each batch. The pruning is to make the data dimension of the predicted CFR data for each batch equal to the data dimension of the fitting target of the output data for each batch. Then, the loss is calculated based on the fitting target of the predicted CFR data of each batch and the output data of each batch. Backpropagation is then performed to train the first deep learning model so that the predicted CFR data of each batch output by the decoder is as close as possible to the fitting target of the output data of that batch. Step SA5: First, based on the historical complex CFR data obtained from the actual scenario, obtain the input data for each batch, the time mask for each batch of input data, and the coherent time vector of the complex CFR data; then, the encoder obtains the first summed data based on the coherent time vector of the complex CFR data, the time mask for each batch of input data, and the input data for each batch; then, the first summed data is used as the input to the trained first deep learning model, and the trained first deep learning model outputs the predicted CFR data; then, the decoder prunes the predicted CFR data output by the trained first deep learning model to obtain the predicted CFR data for the predicted dimension.

2. The coding method for predicting channel state information according to claim 1, characterized in that, In step SA1, the input data for each batch, the time mask for the input data for each batch, the fitting target for the output data for each batch, and the time mask for the output data for each batch are obtained based on the complex CFR data; specifically, the following steps are included. Step S21: Obtain complex CFR data over a past period, split the complex CFR data into real and imaginary parts and concatenate them vertically to obtain real CFR data; Step S22: Obtain the input dataset, input data time mask set, output data fitting target set, and output data time mask set of the deep learning model from real CFR data through a sliding time window; Step S23: According to the amount of data in each batch, obtain the input data, the input data time mask, the fitting target, and the output data time mask for each batch from the input dataset, the input data time mask set, the fitting target set, and the output data time mask set respectively.

3. The coding method for predicting channel state information according to claim 1, characterized in that, In step SA1, the coherent time vector of the complex CFR data is obtained based on the complex CFR data; this includes the following steps: Step S31: Calculate the first norm of the complex CFR data in the time dimension; the complex CFR data is an n-row × t-column complex matrix, where n represents the number of subcarriers and t represents the time length of the complex CFR data; For complex CFR data, first differencing the elements of adjacent columns in the complex matrix, then calculating the p-norm of the magnitude of the difference between each pair of adjacent columns (CFRLtp). j p is a positive integer, and j takes values ​​from 0 to t-1. t CFRLtp j By combining the numbers j in ascending order, we obtain the first norm of the complex CFR data in the time dimension; this step is to extract the difference features of the complex CFR data over time. Step S32: Calculate the first rate of change of the complex CFR data at the first norm in the time dimension, which is the j position later in time relative to the i position earlier in time, where i and j are integers and 0 ≤ i < j ≤ t-1; Step S33: Initialize tcVec_CFR0; calculate tcVec_CFR using the first rate of change and the first threshold. j ; Set tcVec_CFR0 and t-1 tcVec_CFR j By combining the indices in ascending order, we obtain the coherent time vector of complex CFR data.

4. The coding method for predicting channel state information according to claim 1, characterized in that, In step SA1, the coherent time vector of the complex CFR data is obtained based on the complex CFR data; this includes the following steps: Step S41: Complex CFR data is an n-row × t-column complex matrix, where n represents the number of subcarriers and t represents the time length of the complex CFR data; Calculate the difference matrix ΔCFR of the complex CFR data, which is a real matrix with t rows and t columns; Step S42: Calculate the second rate of change in the difference matrix ΔCFR of the complex CFR data, where the time position j is later in time relative to the time position i is earlier in time, where i and j are integers and 0 ≤ i < j ≤ t-1; this includes calculating the p-norm after differencing the current time step with all previous time steps. Step S43: Initialize tcVec_CFR0; calculate tcVec_CFR using the second rate of change and the second threshold. j ; Set tcVec_CFR0 and t-1 tcVec_CFR j By combining the indices in ascending order, we obtain the coherent time vector of complex CFR data.

5. A coding method for predicting channel state information, characterized in that, Includes the following steps; Step SB1: Obtain the input data for each batch, the input data time mask for each batch, the fitting target for the output data for each batch, the output data time mask for each batch, and the coherence space vector of the complex CFR data based on the complex channel frequency response (CFR) data. Step SB2: The encoder obtains the second summed data based on the coherence spatial vector of the complex CFR data, the time mask of the input data for each batch, and the input data for each batch; specifically, the encoder performs matrix calculation on the coherence spatial vector of the complex CFR data and the time mask of the input data for each batch to obtain the coherence spatial sequence of the complex CFR data for each batch. Subsequently, the encoder maps the coherent spatial sequence of the complex CFR data of each batch to a high-dimensional space through an embedding function to obtain the coherent spatial encoding of the complex CFR data of each batch, so that the data dimension of the coherent spatial encoding of the complex CFR data of each batch is equal to the data dimension of the input data of each batch. The encoder then adds the coherent spatial encoding of the complex CFR data of each batch to the input data of each batch, resulting in the second summed data; Step SB3: The second summed data is used as input to the second deep learning model, and the second deep learning model outputs the predicted CFR data; Step SB4: The decoder prunes the predicted CFR data output by the second deep learning model, outputting the predicted CFR data for each batch. Pruning makes the data dimension of the predicted CFR data for each batch equal to the data dimension of the fitting target of the output data for each batch. Then, the loss is calculated based on the fitting target of the predicted CFR data of each batch and the output data of each batch. Backpropagation is then performed to train the second deep learning model so that the predicted CFR data of each batch output by the decoder is as close as possible to the fitting target of the output data of that batch. Step SB5: First, based on the historical complex CFR data obtained from the actual scenario, obtain the input data for each batch, the time mask for each batch of input data, and the coherence space vector of the complex CFR data; then, the encoder obtains the second summed data based on the coherence space vector of the complex CFR data, the time mask for each batch of input data, and the input data for each batch; then, the second summed data is used as the input to the trained second deep learning model, and the trained second deep learning model outputs the predicted CFR data; then, the decoder prunes the predicted CFR data output by the trained second deep learning model to obtain the predicted CFR data for the predicted dimension.

6. The coding method for predicting channel state information according to claim 5, characterized in that, In step SB1, the coherence space vector of the complex CFR data is obtained based on the complex CFR data; this includes the following steps: Step S61: Calculate the third norm of the complex CFR data in the subcarrier dimension; the complex CFR data is an n x t complex matrix, where n represents the number of subcarriers and t represents the time length of the complex CFR data; first, differ between adjacent columns of the transpose matrix of the complex CFR data, then calculate the p-norm (CFRLsp) of the magnitude of the difference between each pair of adjacent columns. j p is a positive integer, and j is an integer ranging from 0 to n-1; n CFRLsp j By combining the numbers in ascending order of j, we obtain the third norm of the complex CFR data in the subcarrier dimension; this step is to extract the differential features of the complex CFR data as the subcarrier changes. Step S62: Calculate the third rate of change of the third norm of the complex CFR data in the subcarrier dimension at position j, which is later than position i, where i and j are integers and 0 ≤ i < j ≤ n-1. Step S63: Initialize scVec_CFR0; calculate scVec_CFR using the third rate of change and the third threshold. j ; Set scVec_CFR0 and n-1 scVec_CFR j By combining the subscripts in ascending order, we obtain the coherent space vector of the complex CFR data.

7. The coding method for predicting channel state information according to claim 5, characterized in that, In step SB1, the coherence space vector of the complex CFR data is obtained based on the complex CFR data; this includes the following steps: Step S71: The complex CFR data is an n-row × t-column complex matrix, where n represents the number of subcarriers and t represents the time length of the complex CFR data; calculate the difference matrix ΔCFR of the transpose matrix of the complex CFR data. T This is an n-row × n-column real matrix; Step S72: Calculate the difference matrix ΔCFR of the transpose matrix of the complex CFR data. T The fourth rate of change of the subcarrier position j relative to the subcarrier position i, where i and j are integers and 0 ≤ i < j ≤ n-1; this includes calculating the p-norm by differentiating the current subcarrier from all previous subcarriers. Step S73: Initialize scVec_CFR0; calculate scVec_CFR using the fourth rate of change and the fourth threshold. j ; Set scVec_CFR0 and n-1 scVec_CFR j By combining the subscripts in ascending order, we obtain the coherent space vector of the complex CFR data.

8. A coding method for predicting channel state information, characterized in that, Includes the following steps; Step SC1: Obtain the input data for each batch, the input data time mask for each batch, the fitting target for the output data for each batch, the output data time mask for each batch, and the coherent time vector of the complex channel impulse response (CIR) data based on the complex channel frequency response (CFR) data. Step SC2: The encoder obtains the third summed data based on the coherent time vector of the complex CIR data, the time mask of the input data for each batch, and the input data for each batch; specifically, the encoder performs matrix calculation on the coherent time vector of the complex CIR data and the time mask of the input data for each batch to obtain the coherent time sequence of the complex CIR data for each batch. The encoder then uses an embedding function to map the coherent time series of each batch of complex CIR data into a high-dimensional space. After embedding, the coherent time code of each batch of complex CIR data is obtained, so that the data dimension of the coherent time code of each batch of complex CIR data is equal to the data dimension of the input data of each batch. The encoder then adds the coherent time encoding of the complex CIR data of each batch to the input data of each batch, resulting in the third summed data; Step SC3: The third summed data is used as input to the third deep learning model, and the third deep learning model outputs the predicted CFR data; Step SC4: The decoder prunes the predicted CFR data output by the third deep learning model, outputting the predicted CFR data for each batch. Pruning makes the data dimension of the predicted CFR data for each batch equal to the data dimension of the fitting target of the output data for each batch. Then, the loss is calculated based on the fitting target of the predicted CFR data of each batch and the output data of each batch. Backpropagation is then performed to train the third deep learning model so that the predicted CFR data of each batch output by the decoder is as close as possible to the fitting target of the output data of that batch. Step SC5: First, based on the historical complex CFR data obtained from the actual scenario, the input data for each batch, the time mask for each batch of input data, and the coherent time vector of the complex CIR data are obtained. Then, the encoder obtains the third summed data based on the coherent time vector of the complex CIR data, the time mask for each batch of input data, and the input data for each batch. The third summed data is then used as the input to the trained third deep learning model, and the trained third deep learning model outputs the predicted CFR data. Then, the decoder prunes the predicted CFR data output by the trained third deep learning model to obtain the predicted CFR data for the predicted dimension.

9. A coding method for predicting channel state information, characterized in that, Includes the following steps; Step SD1: Obtain the input data for each batch, the input data time mask for each batch, the fitting target for the output data for each batch, the output data time mask for each batch, and the coherence space vector of the complex channel impulse response (CIR) data based on the complex channel frequency response (CFR) data. Step SD2: The encoder obtains the fourth summed data based on the coherence spatial vector of the complex CIR data, the time mask of the input data for each batch, and the input data for each batch; specifically, the encoder performs matrix calculation on the coherence spatial vector of the complex CIR data and the time mask of the input data for each batch to obtain the coherence spatial sequence of the complex CIR data for each batch. The encoder then uses an embedding function to map the coherent spatial sequence of the complex CIR data of each batch to a high-dimensional space. After embedding, the coherent spatial encoding of the complex CIR data of each batch is obtained, so that the data dimension of the coherent spatial encoding of the complex CIR data of each batch is equal to the data dimension of the input data of each batch. The encoder then adds the coherent spatial encoding of the complex CIR data of each batch to the input data of each batch, resulting in the fourth summed data; Step SD3: The fourth summed data is used as input to the fourth deep learning model, and the fourth deep learning model outputs the predicted CFR data; Step SD4: The decoder prunes the predicted CFR data output by the fourth deep learning model to output the predicted CFR data for each batch. Pruning makes the data dimension of the predicted CFR data for each batch equal to the data dimension of the fitting target of the output data for each batch. Then, the loss is calculated based on the fitting target of the predicted CFR data of each batch and the output data of each batch. Backpropagation is then performed to train the fourth deep learning model so that the predicted CFR data of each batch output by the decoder is as close as possible to the fitting target of the output data of that batch. Step SD5: First, based on the historical complex CFR data obtained in the actual scenario, obtain the input data for each batch, the time mask for each batch of input data, and the coherence space vector of the complex CIR data; then, the encoder obtains the fourth summed data based on the coherence space vector of the complex CIR data, the time mask for each batch of input data, and the input data for each batch; then, the fourth summed data is used as the input to the trained fourth deep learning model, and the trained fourth deep learning model outputs the predicted CFR data; then, the decoder prunes the predicted CFR data output by the trained fourth deep learning model to obtain the predicted CFR data for the predicted dimension.

10. A coding system for predicting channel state information, characterized in that, Includes a first preprocessor, a first encoder, a first deep learning model, and a first decoder; During the training phase, the first preprocessor is used to obtain the input data for each batch, the input data time mask for each batch, the fitting target for the output data for each batch, the output data time mask for each batch, and the coherent time vector of the complex CFR data based on the complex channel frequency response (CFR) data. During the application phase, the first preprocessor is used to obtain the input data for each batch, the input data time mask for each batch, and the coherent time vector of the complex CFR data based on the historical complex CFR data obtained in the actual scenario. In both the training and application phases, the first encoder is used to obtain the first summed data based on the coherent time vector of the complex CFR data, the time mask of the input data for each batch, and the input data for each batch. Specifically, the first encoder performs matrix calculations on the coherent time vector of the complex CFR data and the time mask of the input data for each batch to obtain the coherent time series of the complex CFR data for each batch. Subsequently, the first encoder maps the coherent time series of the complex CFR data for each batch to a high-dimensional space and embeds it using an embedding function to obtain the coherent time encoding of the complex CFR data for each batch, so that the data dimension of the coherent time encoding of the complex CFR data for each batch is equal to the data dimension of the input data for each batch. The first encoder then adds the coherent time encoding of the complex CFR data of each batch to the input data of each batch, to obtain the first summed data; The first deep learning model is used as input to the first summed data and outputs the predicted CFR data in both the training and application phases. The first decoder is used in both the training and application phases to prune the predicted CFR data output by the first deep learning model, outputting predicted CFR data for each batch. Pruning ensures that the data dimension of the predicted CFR data for each batch is equal to the data dimension of the fitting target of the output data for each batch or equal to the predicted dimension. In the training phase, the first decoder is also used to calculate the loss based on the predicted CFR data for each batch and the fitting target of the output data for each batch, and to perform backpropagation to train the first deep learning model, so that the predicted CFR data output by the first decoder for each batch is as close as possible to the fitting target of the output data for that batch.

11. A coding system for predicting channel state information, characterized in that, This includes a second preprocessor, a second encoder, a second deep learning model, and a second decoder; During the training phase, the second preprocessor is used to obtain the input data for each batch, the input data time mask for each batch, the fitting target for the output data for each batch, the output data time mask for each batch, and the coherence space vector of the complex CFR data based on the complex channel frequency response (CFR) data. During the application phase, the second preprocessor is used to obtain the input data for each batch, the input data time mask for each batch, and the coherence space vector of the complex CFR data based on the historical complex CFR data obtained in the actual scenario. The second encoder is used in both the training and application phases to obtain the second summed data based on the coherence space vector of the complex CFR data, the time mask of the input data for each batch, and the input data for each batch. Specifically, the second encoder performs matrix calculations on the coherence space vector of the complex CFR data and the time mask of the input data for each batch to obtain the coherence space sequence of the complex CFR data for each batch. Subsequently, the second encoder maps the coherent spatial sequence of the complex CFR data of each batch to a high-dimensional space through an embedding function to obtain the coherent spatial encoding of the complex CFR data of each batch, so that the data dimension of the coherent spatial encoding of the complex CFR data of each batch is equal to the data dimension of the input data of each batch. The second encoder then adds the coherent spatial encoding of the complex CFR data of each batch to the input data of each batch, resulting in the second summed data; The second deep learning model is used in both the training and application phases to take the second summed data as input and output the predicted CFR data. The second decoder is used in both the training and application phases to prune the predicted CFR data output by the second deep learning model, outputting predicted CFR data for each batch. Pruning ensures that the data dimension of the predicted CFR data for each batch is equal to the data dimension of the fitting target of the output data for each batch or equal to the predicted dimension. In the training phase, the second decoder is also used to calculate the loss based on the predicted CFR data for each batch and the fitting target of the output data for each batch, and to perform backpropagation to train the second deep learning model, so that the predicted CFR data output by the second decoder for each batch is as close as possible to the fitting target of the output data for that batch.

12. A coding system for predicting channel state information, characterized in that, This includes a third preprocessor, a third encoder, a third deep learning model, and a third decoder; The third preprocessor is used during the training phase to obtain the input data for each batch, the input data time mask for each batch, the fitting target for the output data for each batch, the output data time mask for each batch, and the coherent time vector of the complex channel frequency response (CFR) data based on the complex channel frequency response (CFR) data. The third preprocessor is used in the application phase to obtain the input data for each batch, the time mask of the input data for each batch, and the coherent time vector of the complex CIR data based on the historical complex CFR data obtained in the actual scenario. The third encoder is used in both the training and application phases to obtain the third summed data based on the coherent time vector of the complex CIR data, the time mask of the input data for each batch, and the input data for each batch. Specifically, the third encoder performs matrix calculations on the coherent time vector of the complex CIR data and the time mask of the input data for each batch to obtain the coherent time sequence of the complex CIR data for each batch. Subsequently, the third encoder maps the coherent time series of the complex CIR data of each batch to a high-dimensional space through an embedding function to obtain the coherent time code of the complex CIR data of each batch, so that the data dimension of the coherent time code of the complex CIR data of each batch is equal to the data dimension of the input data of each batch; then the third encoder adds the coherent time code of the complex CIR data of each batch to the input data of each batch, to obtain the third summed data; The third deep learning model is used as input to the third summed data and outputs the predicted CFR data in both the training and application phases. The third decoder is used in both the training and application phases to prune the predicted CFR data output by the third deep learning model, outputting predicted CFR data for each batch. Pruning ensures that the data dimension of the predicted CFR data for each batch is equal to the data dimension of the fitting target of the output data for each batch or equal to the predicted dimension. In the training phase, the third decoder is also used to calculate the loss based on the predicted CFR data for each batch and the fitting target of the output data for each batch, and to perform backpropagation to train the third deep learning model, so that the predicted CFR data output by the third decoder for each batch is as close as possible to the fitting target of the output data for that batch.

13. A coding system for predicting channel state information, characterized in that, This includes a fourth preprocessor, a fourth encoder, a fourth deep learning model, and a fourth decoder; The fourth preprocessor is used in the training phase to obtain the input data for each batch, the input data time mask for each batch, the fitting target for the output data for each batch, the output data time mask for each batch, and the coherence space vector of the complex channel impulse response (CIR) data based on the complex channel frequency response (CFR) data; the fourth preprocessor is used in the application phase to obtain the input data for each batch, the input data time mask for each batch, and the coherence space vector of the complex CIR data based on the historical complex CFR data obtained in the actual scenario. The fourth encoder is used in both the training and application phases to obtain the fourth summed data based on the coherence space vector of the complex CIR data, the time mask of the input data for each batch, and the input data for each batch. Specifically, the fourth encoder performs matrix calculations on the coherence space vector of the complex CIR data and the time mask of the input data for each batch to obtain the coherence space sequence of the complex CIR data for each batch. Subsequently, the fourth encoder maps the coherent spatial sequence of the complex CIR data of each batch to a high-dimensional space through an embedding function to obtain the coherent spatial encoding of the complex CIR data of each batch, so that the data dimension of the coherent spatial encoding of the complex CIR data of each batch is equal to the data dimension of the input data of each batch. The fourth encoder then adds the coherent spatial encoding of the complex CIR data of each batch to the input data of each batch, resulting in the fourth summed data; The fourth deep learning model is used in both the training and application phases to take the fourth summed data as input and output the predicted CFR data. The fourth decoder is used in both the training and application phases to prune the predicted CFR data output by the fourth deep learning model, outputting predicted CFR data for each batch. Pruning ensures that the data dimension of the predicted CFR data for each batch is equal to the data dimension of the fitting target of the output data for each batch or equal to the predicted dimension. In the training phase, the fourth decoder is also used to calculate the loss based on the predicted CFR data for each batch and the fitting target of the output data for each batch, and to perform backpropagation to train the fourth deep learning model, so that the predicted CFR data output by the fourth decoder for each batch is as close as possible to the fitting target of the output data for that batch.

Images