A channel state information compression and decompression method for 5G positioning

Abstract

The present invention discloses a channel state information compression and decompression method for 5G positioning, which is characterized in that the method uses collected CSI samples to generate a data set for deep neural network training, verification and testing, trains the deep neural network, obtains a CSI compression and decompression model including a CSI encoder and a decoder, deploys the CSI encoder part in the base station, and deploys the CSI decoder part in the positioning module that needs to extract the CSI in the base station. The CSI processed by the encoder will be compressed into a codeword with a small number of bytes, extracted through an interface to the CSI decoder in the positioning module for decompression, and restored to the CSI information before compression. Compared with the prior art, the present invention has higher recovery accuracy and better generalization characteristics for CSI data in different scenarios. It has great advantages and commercial prospects in 5G positioning and other 5G application environments that require rapid extraction of CSI.

Description

Technical Field

[0001] The present invention relates to the field of channel state information technology, and specifically to a 5G large-scale antenna array channel state information (Channel State Information, CSI) compression and decompression method based on a transformer deep neural network applied in 5G positioning. Background Art

[0002] With the popularization and development of 5G technology, research on using 5G technology for indoor positioning has been gradually carried out. In 5G positioning technology, very accurate signal arrival time (TOA) and arrival time difference (TDOA) estimation, namely super-resolution TOA\TDOA estimation, is required. The channel state information (CSI) used for super-resolution TOA\TDOA estimation can be obtained in the 5G physical layer measurement module in the base station. Based on the 5G physical layer uplink signal processing flow, the positioning module design can be integrated into the communication system, and the 5G positioning module obtains CSI from the 5G physical layer measurement module. Among them, CSI is transmitted to the positioning module through the inter-module interface. The positioning module uses the acquired CSI to obtain TOA\TDOA measurement using a super-resolution TOA estimation algorithm, and then uses the source positioning algorithm to solve the user equipment position, and finally uses the error source elimination algorithm to optimize the positioning result.

[0003] In the above-mentioned 5G positioning process, researchers found that it is very difficult to extract CSI from base stations. This is mainly because with the application of large-scale MIMO technology, the amount and complexity of CSI information are gradually increasing. Due to design limitations, insufficient resources and other issues, the interface for transmitting CSI cannot quickly transmit a large amount of such complex CSI. In order to optimize this step, the demand and research on CSI compression and decompression technology have been triggered. In the 5G communication system, there are also CSI compression methods and feedback steps.

[0004] The CSI compression methods in the prior art are usually codebook-based and compressed sensing-based CSI compression recovery methods. Due to the popularization of large-scale MIMO technology in 5G systems, CSI information has become more complex, and the size of the channel matrix has also increased with the increase in the number of antennas. The codebook-based and compressed sensing-based CSI compression methods can no longer meet system requirements, and the recovery accuracy is not enough. There is a problem that it is difficult for the positioning module in the 5G indoor positioning system to extract 5G channel state information (CSI) from the 5G physical layer measurement module of the base station. Summary of the invention

[0005] The purpose of the present invention is to provide a channel state information compression and decompression method for 5G positioning in view of the deficiencies of the prior art. The deep learning method based on the transformer structure neural network is used to compress and decompress the 5G channel state information. The auxiliary positioning module extracts the required CSI from the base station at high speed and high precision, and compresses the CSI in the 5G base station. The recovery accuracy of the decompressed CSI meets the requirements of super-resolution TOA estimation in 5G positioning, which greatly reduces the complexity of calculation. Without increasing the interface load, the extraction speed of CSI by the 5G positioning module can be greatly improved. It has good adaptability to solving the problem that it is difficult for the positioning module to extract CSI data from the 5G measurement module of the base station in the current 5G indoor positioning. It is also suitable for other 5G application scenarios that require fast CSI extraction. It has good generalization characteristics for CSI data in different scenarios. The extraction of CSI inside the 5G base station plays an important role in 5G positioning technology. It has great advantages and commercial prospects in 5G positioning and other 5G application environments that require fast CSI extraction.

[0006] The purpose of the present invention is to achieve: a channel state information compression and decompression method for 5G positioning, which is characterized by using the actually collected CSI data to create a data set, using a transformer-based deep neural network for training, to obtain a CSI compression and decompression model, which is deployed in a base station and a positioning module where CSI needs to be extracted, using the encoder of the model to extract and compress the features of the channel matrix, and using the decoder of the model to restore the channel matrix before compression, specifically including the following steps:

[0007] Step 1: Sample collection

[0008] The user equipment to be located is placed in a range near the base station and moves randomly. CSI samples are losslessly collected from the base station using an uncompressed and uncompressed method as the data set for model training. Generally, no less than 100,000 CSI samples need to be collected to meet the needs of neural network model training.

[0009] Step 2: Sample pretreatment

[0010] The collected samples are subjected to discrete Fourier transform, cropped, saved as sample matrices, and other operations, so that the collected original CSI samples conform to the input format for neural network model training.

[0011] Step 3: Dataset division.

[0012] The samples after data preprocessing are randomly divided into training set, validation set and test set in proportion. The number of samples in the training set accounts for the largest proportion. The number of samples in the validation set and the test set is the same, and the number of samples in the two sets does not exceed 1 / 2 of the number of samples in the training set.

[0013] Step 4: Network model training

[0014] First, configure the model hyperparameters such as compression rate and number of module layers required for network model initialization, as well as parameters such as the number of training cycles, number of batch samples, learning rate optimization method, and random seed during the training process. Then, load the data set obtained in step 1 to the neural network model and run the training program. During the training of the neural network, use visualization tools to record the training data log to facilitate the analysis of the training process after each training. After each training, adjust the model hyperparameters according to the changes in the CSI recovery accuracy until a hyperparameter configuration with a recovery accuracy that meets the system requirements is obtained. Fix this hyperparameter configuration and save the model file to obtain the CSI compression and decompression model for system deployment.

[0015] Step 5: Deploy the CSI compression and decompression model

[0016] The CSI compression and decompression model for system deployment mentioned above includes two parts: encoder and decoder (which can be deployed and configured separately). The encoder part is deployed in the base station, the pre-trained parameter file of the model is configured, and the CSI output interface in the base station is connected to the input interface of the model, and the output of the model is connected to the communication interface between the base station and the positioning module. Then, in a similar way, the decoder part of the model is deployed in the server that needs to extract the CSI in the base station, the parameter configuration is completed, and the CSI output interface of the server is connected to the input interface of the model decoder in the server to complete the deployment of the CSI compression and decompression model.

[0017] Step 6: Application of 5G positioning

[0018] When the server needs to read the CSI data of the base station, it first sends a request to the base station through the communication interface. After receiving the request, the base station starts to read the CSI. The CSI data will first be output to the model encoder through the output module inside the base station. After compression by the encoder, a codeword that occupies a much smaller number of bytes than the original CSI information is formed. This codeword is transmitted to the model decoder in the server through the base station-server communication interface, and then decompressed by the decoder to restore the CSI information with high precision for the server to call. This process can speed up the server to read CSI data from the base station and reduce the occupation of interface resources.

[0019] The sample preprocessing specifically includes the following steps:

[0020] 1) Assuming that the number of collected samples is M, for each collected CSI sample, that is, each channel matrix Using two-dimensional discrete Fourier transform, the channel matrix The following formula (a) is used to transform from the space-frequency domain to the angle-delay domain to obtain a sparse matrix H:

[0021]

[0022] Where: For one Row, N t A complex-valued matrix of columns; F a 、F d is the discrete Fourier transform matrix; the superscript H indicates the matrix transpose;

[0023] 2) Keep the first N of the delay domain c rows, cut the matrix H to get the new N c Row, N t The complex matrix H of columns;

[0024] 3) Map the real and imaginary parts of the complex numbers at each position of the complex matrix H to the corresponding positions of two matrices of the same size as the complex matrix H, and obtain two real-valued matrices H i and H r ;

[0025] 4) The real-valued matrix H i and H r Expand the row into a one-dimensional vector, and concatenate the two vectors end to end to get a length of N c *N t *2 one-dimensional vector v m (m=1,2,…,M);

[0026] 5) The vector v m As the mth row of a new matrix, the matrix V of the data set after sample preprocessing is obtained.

[0027] The channel matrix encoder takes an input dimension of N c ×N t The channel matrix H × 2 is compressed into a matrix of length N c *N t The vector m of *2*Cr is processed as follows:

[0028] (I) Block Coding

[0029] The channel matrix H is divided into K blocks of equal size. Each block is flattened and mapped to a D-dimensional vector by the fully connected layer to obtain a K×D-dimensional matrix X, which is called the block coding of the channel matrix H.

[0030] (ii) Positional encoding

[0031] The block coding of the channel matrix H is added with position coding as follows:

[0032] Randomly initialize a learnable parameter matrix P with the same dimension as the channel matrix H, add the matrix P as the position code to the matrix X to obtain the matrix E, that is, P+X=E;

[0033] 3. Feature extraction of channel matrix

[0034] The TransBlock module is used to extract features from the channel matrix H. The TransBlock module is a core module containing d networks. A single TransBlock module of each network contains a multi-head attention mechanism module and a multi-layer perceptron module. The input K×D dimension matrix E is passed through one or more TransBlock modules to obtain an output K×D matrix (M T ) d , where d is the number of times TransBlock is repeated;

[0035] (IV) Compress codewords and output

[0036] The matrix (M T ) d After normalization through the reshape layer, we get a one-dimensional vector M with a length of K*D, which is passed through a fully connected layer whose output dimension is smaller than the input dimension to get a one-dimensional vector M with a length of N. c *N t A one-dimensional codeword vector m of *2*Cr, where Cr is the compression rate of the encoder.

[0037] The channel matrix decoder restores the codeword vector m to the channel matrix according to the following steps

[0038] (I) Decompressing codewords

[0039] Use the fully connected layer to decompress the one-dimensional codeword vector m to obtain a one-dimensional vector of length K*D The one-dimensional vector Transformed into a K×D dimensional matrix through the reshape layer

[0040] (II) Output of TransBlock module

[0041] Use a TransBlock module containing k networks, and the input of a single TransBlock module of each network is the matrix The output is a matrix of dimension K×D The single TransBlock module includes a multi-head attention mechanism module and a multi-layer perceptron module;

[0042] (III) Output module

[0043] The matrix Through layer standardization, nonlinear 1*1 convolution layer and Sigmoid function, the decompressed output matrix is ​​obtained, that is, the dimension is N c ×N t ×2 channel matrix

[0044] The training of the neural network uses a visualization tool to record the training data log, which is convenient for analyzing the training process after each training.

[0045] The hyperparameters and other parameters required for the initialization of the neural network model configuration are: compression rate Cr, number of module layers, number of model training cycles, number of batch samples, learning rate optimization method and number of random seeds. The compression rate Cr is set to 1 / 4, 1 / 8, 1 / 16, 1 / 32 or 1 / 64.

[0046] The TransBlock module specifically includes the following steps:

[0047] Step 1: Multi-head attention mechanism module

[0048] The multi-head attention mechanism module is used to add an attention mask to the matrix to be processed, marking the part containing the main information, and reshape the matrix output by the network to normalize it. The resulting K×D dimensional matrix M i = L(E) input multi-head attention layer, and finally get the output matrix M with attention mask o , the dimensions of the output and input matrices of the multi-head attention layer remain unchanged, the output matrix is ​​the input matrix after weight optimization, and the output is calculated as the weighted sum of the values, where the weight assigned to each value is calculated by the compatibility function of the query and the corresponding key. The implementation steps of the multi-head attention mechanism module are as follows:

[0049] Step 1.1: Input matrix M i By the linear mapping layer, it is mapped to Q = M i W Q , K=M i W K and V = M i W V Three different matrices, among which W Q , W K and W V is the linear mapping matrix;

[0050] Step 1.2: Calculate the output of the multi-head attention layer according to the following formula (b):

[0051]

[0052] Among them, M ois a K×D dimensional output matrix; d k is the block coding dimension in the block coding step; T is the matrix transpose; Q, K and V are query, key and value respectively; softmax is the softmax function of the following formula (c):

[0053]

[0054] Step 2: Multilayer Perceptron Module

[0055] The multi-layer perceptron module learns to record the characteristics of CSI in the training data, fits a nonlinear function that can complete the task of compressing the channel matrix, and obtains the K×D dimensional output matrix M of the multi-layer perceptron module. o_m ,

[0056] The processing sequence is: fully connected layer 1, GELU activation function, Dropout, fully connected layer 2 and Dropout;

[0057] Step 3: Output of TransBlock Module

[0058] The matrix M o_m After the dropout layer and M i ′ is added, and residual learning is performed using the residuals of the module input and output to obtain the output matrix of the entire TransBlock

[0059] Compared with the prior art, the present invention reduces the resource occupation and system overhead of the extraction interface. Without increasing the interface load, it can greatly improve the CSI extraction speed of the 5G positioning module, and effectively solves the problem that the positioning module extracts CSI from the base station too slowly in 5G positioning. It is suitable for 5G application scenarios that require fast CSI extraction, has good adaptability, and has a large application space and commercial value. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a CSI sample legend;

[0061] Figure 2 It is a structural diagram based on the transformer neural network model;

[0062] Figure 3 It is the structural block diagram of the TransBlock module;

[0063] Figure 4 It is the structural block diagram of the multi-layer perceptron module;

[0064] Figure 5 LOSS curve graph for model training;

[0065] Figure 6 The nmse variation curve for model testing;

[0066] Figure 7 A diagram of model deployment. DETAILED DESCRIPTION

[0067] The present invention is further described in detail below with reference to the accompanying drawings by taking an example of compression and decompression of CSI in a certain indoor 5.3 GHz frequency band scenario.

[0068] Example 1

[0069] Consider an indoor 5G positioning scenario of a single-cell FDD downlink. In order to achieve positioning, the 5G positioning module needs to obtain a very accurate signal TOA (arrival time) estimate, that is, a super-resolution TOA estimate. To obtain such a TOA estimate, the CSI information of the 5G physical layer measurement module in the base station is required. In this embodiment, a server is used instead of the positioning module for functional verification.

[0070] Step 1: Collect data in the scenario

[0071] The base station is set at the center of the square area, using a frequency of 5.3 GHz. The length and width of the square area are both 20 m. The base station antenna array is configured as a ULA (uniform linear array) with 32 antennas, and the number of OFDM (orthogonal frequency division multiplexing) subcarriers is 1024. Let the UE (user equipment) move randomly in the square area. During the UE movement, 150,000 downlink CSI (i.e., channel matrix) samples are continuously collected and saved at a certain frequency, and the space-frequency domain channel matrix in Table 1 is obtained. Example:

[0072] Table 1 Space-frequency domain channel matrix Example

[0073]

[0074]

[0075] Table 1 above shows the mth channel matrix sample: The number of rows indicates the number of subcarriers, and the number of columns indicates the number of antennas. is a complex value matrix with 1024 rows and 32 columns, where the value in the xth row and yth column represents the impulse response of the channel between the yth transmitting antenna and the receiving antenna on the xth subcarrier.

[0076] Step 2: Sample preprocessing and dataset division

[0077] right Multiply the left and right sides by a DFT matrix to transform from the space-frequency domain to the angle-delay domain, and obtain the matrix H. Since the matrix H is sparse, it can be trimmed to keep only the first 32 rows. These 32 rows retain most of the information in H, and a new 32-row 32-column complex value matrix is ​​obtained. Separate the real and imaginary parts of the complex value and map it into two real-valued matrices H i and H r ,

[0078] See also Figure 1 , shown as a matrix H r After conversion to grayscale values, we can see that it has a certain sparsity. i The characteristics are similar to those of H i and H r Considered as two channels of a picture, each channel matrix can be input into the neural network as a 2-channel 32-row 32-column grayscale image. The CSI samples preprocessed according to the above steps are divided into training set, test set, and validation set according to the number of 100,000, 20,000, and 30,000.

[0079] In order to save the data set, these samples need to be saved in three data set matrices. Each sample is processed as follows: First, two 32-row and 32-column matrices H i and H r Both are expanded row by row into a one-dimensional vector of length 1024, and then the two vectors are concatenated end to end to obtain a new one-dimensional vector v of length 2048 m . In this way, each sample will become a one-dimensional vector v m . Then the one-dimensional vector v m As the mth row of a new matrix, the preprocessed data set matrix is ​​obtained. The saved test set is shown in the following table 2:

[0080] Table 2 Test set storage format examples

[0081]

[0082]

[0083] Table 2 above shows a saved test set, which is a matrix of 20,000 rows and 2,048 columns, where each row represents a sample. The training set and validation set have the same format, except for the number of rows.

[0084] Step 3: Train the model

[0085] See also Figure 2 , use the data set obtained in the previous step to train the transformer-based neural network model.

[0086] See also Figure 3-4 , configure the network model initialization hyperparameters (such as the number of TransBlocks), configure the number of training cycles, the number of batch samples, the learning rate optimization method, the random seed and other parameters. The hyperparameters of this embodiment are: 3 TransBlocks for the encoder and 1 for the decoder. The number of training cycles is set to 1000 epochs, the number of batch samples is 200, and the ADAM optimizer with preheated cosine annealing learning rate optimization is used for training optimization. Then, the data set obtained in step 1 is loaded for the model, and the training program is run. During the training process, the LOSS function value after each epoch is recorded. Every time the training set is used for 10 epochs of training, the current model is tested once using the test set, and the NMSE value obtained by the test is recorded. NMSE is the normalized mean square error, which can be used to characterize the error degree between the matrix of the input model and the output matrix. The smaller the value, the smaller the error, and the higher the decompression accuracy of the model. After each training, the hyperparameters of the model are continuously adjusted according to the changes in the CSI recovery accuracy obtained, until a test NMSE value is obtained to achieve a hyperparameter configuration that reaches the training target.

[0087] See also Figure 5-6 The figure shows the test nmse curve and training LOSS curve obtained after a model training with a compression rate of 1 / 8. The horizontal axis is the number of training cycles. Its NMSE reaches -19.25, which meets the training goal, that is, it meets the requirements of the 5G positioning system for compression and decompression accuracy. Then, the hyperparameter configuration of this training is fixed and the trained model file is saved, so that the CSI compression and decompression model for system deployment is obtained.

[0088] Step 4: Deployment of the model

[0089] Deploy the encoder part of the CSI compression and decompression model in the base station, configure the pre-trained parameter file of the model, and connect the CSI output interface measured by the 5G physical layer in the base station to the input interface of the model's encoder, and connect the output of the encoder to the base station-server communication interface. Then, in a similar manner, deploy the decoder part of the model in the server that needs to extract the CSI in the base station, connect the CSI input interface of the server to the input interface of the model decoder in the server, and connect the output of the decoder to the part used for super-resolution TOA\TDOA estimation in the positioning server. Finally, configure the model hyperparameters obtained in step 3 that meet the system requirements, and load the model file in the program, thus completing the deployment of the model.

[0090] See also Figure 7It can be seen that the 5G physical layer measurement module in the 5G base station inputs the CSI into the encoder for compression, and then the compressed codeword is decompressed by the decoder deployed in the 5G positioning module. In this way, when the data interface transmits CSI, it only needs to spend the resources required to transmit a shorter codeword, which greatly reduces the load on the interface.

[0091] Step 5: Application of 5G positioning

[0092] The UE is placed at a certain position in the scene and its coordinates are recorded. The 5G positioning server sends a request to the base station through the communication interface. After receiving the request, the base station starts to read the CSI of the downlink. The CSI data is first input into the model encoder and compressed by the encoder to form a section that occupies a number of bytes far less than the number of bytes of the original CSI information. In this example, the channel matrix size is 32*32*2, the number of bytes is 2048, and the compression rate is set to 1 / 8. The compressed codeword length is 256 bytes. This codeword is transmitted to the model decoder in the server through the base station-server communication interface, and then decompressed by the decoder to restore the CSI information with high precision for the server to estimate the high-precision TOA and further complete the positioning of the UE. After testing and verification, the recovery accuracy of the CSI extracted and recovered using this method, NMSE reached -19.25 on the logarithmic scale, and the cosine similarity of the channel matrix before and after compression and decompression was 0.9914, which can meet the needs of 5G indoor positioning scenarios.

[0093] The present invention is based on the 5G large-scale antenna array channel state information (CSI) compression and decompression method of transformer deep neural network, which solves the problem of slow speed of the positioning module extracting CSI from the base station in 5G positioning research, and reduces the resource occupation and system overhead of the extraction interface. It has more satisfactory performance for CSI compression and recovery tasks, especially the deep neural network can automatically learn the statistical characteristics of the channel during the training process, which has obvious advantages over traditional methods.

[0094] The above is only a preferred implementation of the present invention and does not limit the present invention in any form. Therefore, any simple modification, equivalent change and modification made to the above implementation method based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall be included in the scope of the claims of this patent as long as it is an equivalent implementation of the present invention.

Claims

1. A channel state information compression and decompression method for 5G positioning, characterized in that The dataset created by CSI data is trained using a transformer-based deep neural network to obtain a CSI compression and decompression model, which is deployed in the base station and the positioning module where CSI needs to be extracted to achieve compression and decompression of channel state information. The specific steps include the following: Step 1: Sample collection The user equipment to be located is placed in the vicinity of the base station and moves randomly. No less than 100,000 CSI samples are losslessly collected using uncompressed and decompressed methods as a data set for training the neural network. Step 2: Sample pretreatment The collected CSI samples are preprocessed by performing discrete Fourier transform, cropping and saving as sample matrices, so that the collected CSI samples conform to the input format of neural network model training; Step 3: Dataset Division The preprocessed samples are randomly divided into training set, validation set and test set according to the proportion. The number of validation set and test set is the same, and the number of samples in the two sets does not exceed 1 / 2 of the number of samples in the training set. Step 4: Network model training Configure the hyperparameters and other parameters required for initialization of the neural network model, load the data set, start training the network model, and adjust the hyperparameters of the model according to the CSI recovery accuracy after each training until a hyperparameter configuration with a recovery accuracy that meets the system requirements is obtained, and then fix it and save the model file to obtain a CSI compression and decompression model for system deployment, wherein the CSI compression and decompression model is an automatic encoder and decoder structure including a channel matrix encoder and a channel matrix decoder; Step 5: Deployment of CSI compression and decompression models Deploy the channel matrix encoder in the base station, configure the pre-trained parameter file of the network model, connect the CSI output interface in the base station with the input interface of the network model, and connect the output of the network model with the communication interface between the base station and the positioning module; The channel matrix decoder is deployed in the CSI server in the base station, and its CSI output is connected to the input interface of the decoder to complete the deployment of the CSI compression and decompression model; Step 6: Application of 5G positioning When the server needs to read the CSI data of the base station, the communication interface sends a request to the base station. After receiving the request, the base station inputs the read CSI information into the encoder through the output module of the base station. The compressed codeword is input into the decoder through the communication interface between the base station and the server for decompression and restored to the CSI information for the server to call; The sample preprocessing specifically includes the following steps: 1) Assuming that the number of collected samples is M, for each collected CSI sample, that is, each channel matrix Using two-dimensional discrete Fourier transform, the channel matrix The following formula (a) is used to transform from the space-frequency domain to the angle-delay domain to obtain a sparse matrix H: Where: For one Row, N t A complex-valued matrix of columns; F a 、F d is the discrete Fourier transform matrix; the superscript H indicates the matrix transpose; 2) Keep the first N of the delay domain c rows, cut the matrix H to get the new N c Row, N t The complex matrix H of columns; 3) Map the real and imaginary parts of the complex numbers at each position of the complex matrix H to the corresponding positions of two matrices of the same size as the complex matrix H, and obtain two real-valued matrices H i and H r ; 4) The real-valued matrix H i and H r Expand the row into a one-dimensional vector, and concatenate the two vectors end to end to get a length of N c *N t *2 one-dimensional vector v m (m=1,2,…,M); 5) The vector v m As the mth row of a new matrix, the matrix V of the data set after sample preprocessing is obtained; The channel matrix encoder takes an input dimension of N c ×N t The channel matrix H × 2 is compressed into a matrix of length N c *N t The vector m of *2*Cr is processed as follows: (I) Block Coding The channel matrix H is divided into K blocks of equal size. Each block is flattened and mapped into a D-dimensional vector by the fully connected layer to obtain a K×D-dimensional matrix X, which is called the block coding of the channel matrix H. (ii) Positional encoding The block coding of the channel matrix H is added with position coding as follows: Randomly initialize a learnable parameter matrix P with the same dimension as the channel matrix H, add the matrix P as the position code to the matrix X to obtain the matrix E, that is, P+X=E; 3. Feature extraction of channel matrix The TransBlock module is used to extract features from the channel matrix H. The TransBlock module is a core module containing d networks. A single TransBlock module of each network contains a multi-head attention mechanism module and a multi-layer perceptron module. The input K×D dimension matrix E is passed through one or more TransBlock modules to obtain an output K×D matrix (M T ) d , where d is the number of times TransBlock is repeated; (IV) Compress codewords and output The matrix (M T ) d After normalization through the reshape layer, we get a one-dimensional vector M with a length of K*D, which is passed through a fully connected layer whose output dimension is smaller than the input dimension to get a one-dimensional vector M with a length of N. c *N t *2*Cr one-dimensional codeword vector m, where Cr is the compression rate of the encoder; The channel matrix decoder restores the codeword vector m to the channel matrix according to the following steps (I) Decompressing codewords Use the fully connected layer to decompress the one-dimensional codeword vector m to obtain a one-dimensional vector of length K*D The one-dimensional vector Transformed into a K×D dimensional matrix through the reshape layer (II) Output of TransBlock module Use a TransBlock module containing k networks, and the input of a single TransBlock module of each network is the matrix The output is a matrix of dimension K×D The single TransBlock module includes a multi-head attention mechanism module and a multi-layer perceptron module; (III) Output module The matrix Through layer standardization, nonlinear 1*1 convolution layer and Sigmoid function, the decompressed output matrix is ​​obtained, that is, the dimension is N c ×N t ×2 channel matrix The TransBlock module specifically includes the following steps: Step 1: Multi-head attention mechanism module The multi-head attention mechanism module is used to add an attention mask to the matrix to be processed, marking the part containing the main information, and reshape the matrix output by the network to normalize it. The resulting K×D dimensional matrix M i = L(E) input multi-head attention layer, and finally get the output matrix M with attention mask o , the dimensions of the output and input matrices of the multi-head attention layer remain unchanged, the output matrix is ​​the input matrix after weight optimization, and the output is calculated as the weighted sum of the values, where the weight assigned to each value is calculated by the compatibility function of the query and the corresponding key. The implementation steps of the multi-head attention mechanism module are as follows: Step 1.1: Input matrix M i By the linear mapping layer, it is mapped to Q = M i W Q , K=M i W K and V = M i W V Three different matrices, among which W Q , W K and W V is the linear mapping matrix; Step 1.2: Calculate the output of the multi-head attention layer according to the following formula (b): Among them, M o is a K×D dimensional output matrix; d k is the block coding dimension in the block coding step; T is the matrix transpose; Q, K and V are query, key and value respectively; softmax is the softmax function of the following formula (c): Step 2: Multilayer Perceptron Module The multi-layer perceptron module learns to record the characteristics of CSI in the training data, fits a nonlinear function that can complete the task of compressing the channel matrix, and obtains the K×D dimensional output matrix M of the multi-layer perceptron module. o_m , The processing sequence is: fully connected layer 1, GELU activation function, Dropout, fully connected layer 2 and Dropout; Step 3: Output of TransBlock Module The matrix M o_m After the dropout layer and M i ′ is added, and residual learning is performed using the residuals of the module input and output to obtain the output matrix of the entire TransBlock 2. According to claim 1, the channel state information compression and decompression method for 5G positioning is characterized in that The training of the neural network uses a visualization tool to record the training data log, which is convenient for analyzing the training process after each training.

3. According to claim 1, the channel state information compression and decompression method for 5G positioning is characterized in that The hyperparameters and other parameters required for the initialization of the neural network model configuration are: compression rate Cr, number of module layers, number of model training cycles, number of batch samples, learning rate optimization method and number of random seeds. The compression rate Cr is set to 1 / 4, 1 / 8, 1 / 16, 1 / 32 or 1 / 64.

Images