Deep learning-based CSI compression feedback method and system for multi-antenna MIMO system

Abstract

The present invention relates to the technical field of wireless communications. Disclosed are a deep learning-based CSI compression feedback method and system for a multi-antenna MIMO system. The method comprises: acquiring raw CSI data, and preparing a neural network model training set; on the basis of the acquired data, constructing a U-Net neural network model integrated with a Transformer; training the neural network model offline to obtain a trained neural network model; and performing metric evaluation on the trained neural network model. The present invention improves the CSI compression efficiency and reconstruction accuracy of the multi-antenna MIMO system, accelerates data processing and feedback speed, enhances model adaptability, and achieves accurate restoration of compressed data. In addition, the multi-scale convolution and skip connection design of the network preserves more details, thereby improving reconstruction quality. An optimized network structure not only reduces the amount of calculation, but also improves processing speed, so that the model can adapt to various channel environments and communication scenarios.

Description

A Deep Learning-Based CSI Compressed Feedback Method and System for Multi-Antenna MIMO Systems Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to a CSI compressed feedback method and system for multi-antenna MIMO systems based on deep learning. Background Technology

[0002] In fifth-generation (5G) and above wireless communication systems, MIMO antenna arrays and OFDM technology are widely considered two crucial technologies. In MIMO-OFDM systems, accurate downlink CSI feedback is essential to ensure the base station can achieve sufficient performance improvements for operations such as beamforming, subcarrier allocation, and power control. In frequency division duplex mode, downlink CSI estimation is typically performed by the user equipment side and fed back to the base station. In the aforementioned system scenarios, the increased number of antenna array elements and subcarriers significantly increases the cost of CSI feedback. This high-dimensional CSI data not only increases the overhead of feedback communication but may also lead to increased feedback latency and inaccurate channel state information. Furthermore, transmitting large amounts of CSI data consumes valuable radio resources and may cause system outages or instability. To address these issues, effective CSI compression techniques are needed to reduce the dimensionality and size of feedback data while retaining critical information. The goal of CSI compression is to maintain an accurate description of the channel state while minimizing the amount of data. Through CSI compression, the cost of feedback communication can be reduced, feedback latency can be decreased, and system efficiency and performance can be improved.

[0003] Traditional compression techniques have limitations when dealing with high-dimensional and complex CSI data. Deep learning models, however, can automatically learn feature representations suitable for channel states, eliminating the need for manually designed feature extractors. This approach can more accurately capture relevant information from the channel state and achieve more efficient compression while preserving essential data information. Due to their non-linear modeling capabilities and flexibility, deep learning models can adapt to different channel environments and communication scenarios. Furthermore, CSI data compressed using deep learning methods often achieves higher compression ratios and information fidelity, thereby reducing feedback communication costs, decreasing feedback latency, and improving system efficiency and performance.

[0004] Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the problem to be solved by this invention is: how to solve the limitations of traditional compression technology when dealing with high-dimensional and complex CSI data.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a CSI compression feedback method for multi-antenna MIMO systems based on deep learning, comprising: collecting raw CSI data and preparing a neural network model training set; constructing a U-Net neural network model with a joint Transformer based on the collected data; training the neural network model offline to obtain a trained neural network model; and evaluating the trained neural network model using metrics.

[0008] As a preferred embodiment of the deep learning-based CSI compression feedback method for multi-antenna MIMO systems described in this invention, the original GSI data includes historical communication data collected by the base station over a period of time, from which CSI information is extracted as the original CSI dataset for the neural network model.

[0009] As a preferred embodiment of the deep learning-based CSI compressed feedback method for multi-antenna MIMO systems described in this invention, the preparation of the network model training set includes the sparse properties of the CSI matrix H in the angular delay domain. Using discrete Fourier transform, the H matrix in the obtained original GSI dataset is transformed into the angular and delay domains, resulting in a matrix of size N. c ×N t matrix Where, N c and N t These represent the number of user subcarriers and the number of base station antennas, respectively, and are used to obtain the matrix. The first N a Rows as CSI compression matrix H a The obtained data is preprocessed and divided into training and testing sets; the CSI matrix H has sparse properties in the angular delay domain, and the angular domain matrix F is obtained using the DFT transformation matrix. c and delay field matrix F t And calculate the channel matrix in the angular delay domain. Represented as:

[0010] Among them, F c It is a angular field matrix with size N. c ×N c F t Let N be the delay field matrix. t ×N t For the angular delay domain channel matrix H%, take the first N elements of matrix H%. a Rows as CSI compression matrix H a N a <N cz-score data standardization is performed on all CSI compression matrices Ha, which involves subtracting the mean of the entire dataset from each data point and then dividing by the standard deviation of the dataset. This accelerates model convergence during deep learning model training. Next, min-max scaling is applied to constrain the dataset's value range to between -1 and 1, allowing the network model's output layer to use a suitable activation function. The min-max scaling is expressed as follows:

[0011] Where, x scaled The data is after minimum-maximum scaling, where x is the input data. min For the minimum data, x max To maximize the data size, the preprocessed dataset is divided into training and test sets in a 4:1 ratio.

[0012] As a preferred embodiment of the deep learning-based CSI compression feedback method for multi-antenna MIMO systems described in this invention, the neural network model comprises a user-side CSI encoding part and a base station-side CSI decoding part. The user-side encoding part includes a U-Net encoding path concatenated with a Transformer module, and the base station-side decoding part includes a Transformer module concatenated with a U-Net decoding path. The user-side encoding part compresses the CSI compression matrix H... a Feature encoding yields a one-dimensional vector of length v. The base station decoding section then decomposes the compressed encoded vector to obtain H. a The compression ratio at this point is η, expressed as:

[0013] η=v / (N c ×N t ×2)

[0014] Where η is the compression ratio and v is the length of the one-dimensional vector obtained by feature encoding.

[0015] As a preferred embodiment of the deep learning-based CSI compression feedback method for multi-antenna MIMO systems described in this invention, the U-Net encoding path has 5 downsampling convolutional blocks for compressing and extracting feature information; the U-Net decoding path contains 5 upsampling convolutional blocks for restoring the compressed information; unlike the ordinary U-Net model, the U-Net encoding and decoding paths do not use skip connections between them. The first two downsampling convolutional blocks in the encoding path employ multi-scale convolutions of 3×3 and 5×5 sizes, with feature maps extracted from convolutions of different kernel sizes and then superimposed. The latter three layers use conventional 3×3 downsampling convolutions. In the decoding path, the first three upsampling convolutional blocks use conventional 3×3 upsampling convolutions due to the small feature map size, while the latter two upsampling convolutional blocks consist of a 3×3 and a 5×5 multi-scale convolution.

[0016] As a preferred embodiment of the deep learning-based CSI compressed feedback method for multi-antenna MIMO systems described in this invention, the offline training neural network model includes merging and connecting two parts of the joint network to form an end-to-end model, inputting training data into the entire network, minimizing the overall L1 loss function through the backpropagation algorithm and optimizer, and updating the parameters of the entire network.

[0017] As a preferred embodiment of the deep learning-based CSI compressed feedback method for multi-antenna MIMO systems described in this invention, the performance evaluation includes online testing of the neural network model. The trained network model is evaluated using the normalized mean square error (NMSE), expressed as follows:

[0018] Among them, H a For the true value, These are the predicted values ​​from the neural network model, where NMSE is the normalized mean square error. The expected value operator is ∥·∥2, which represents the L2 norm. A normalized mean squared error (NMSE) threshold is set. If the NMSE is greater than the threshold, the model is considered substandard. In-depth performance analysis is conducted, including data quality and model diagnostics. Model parameters are adjusted and optimized, including hyperparameter tuning, adding regularization, and adjusting the model structure. If the NMSE is less than the threshold, the model performs well. Finally, the optimal model selected after evaluation is deployed.

[0019] Another objective of this invention is to provide a system for a deep learning-based CSI compressed feedback method for multi-antenna MIMO systems, which solves the CSI compressed feedback problem of deep learning-based multi-antenna MIMO systems by constructing a deep learning-based CSI compressed feedback system.

[0020] To address the aforementioned technical problems, this invention provides the following technical solution: a deep learning-based multi-antenna MIMO system CSI compression feedback system, comprising a data acquisition module, a model building module, a model training module, and an evaluation module; the data acquisition module is used to acquire raw CSI data and prepare a neural network model training set; the model building module constructs a U-Net neural network model with a joint Transformer based on the acquired data; the model training module is used to train the neural network model offline to obtain a trained neural network model; and the evaluation module is used to evaluate the trained neural network model using metrics.

[0021] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the deep learning-based CSI compressed feedback method for multi-antenna MIMO systems as described above.

[0022] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the deep learning-based CSI compressed feedback method for multi-antenna MIMO systems as described above.

[0023] The beneficial effects of this invention are as follows: The deep learning-based CSI compression feedback method for multi-antenna MIMO systems provided by this invention combines U-Net and Transformer modules, achieving efficient CSI compression while maintaining information accuracy. The design of the U-Net encoding and decoding paths allows the network to fully utilize spatial and frequency domain features, effectively extracting and reconstructing CSI information. Furthermore, the Transformer module can capture long-range dependencies, further improving compression efficiency. More accurate reconstruction: The U-Net model, combined with Transformer, can more accurately recover CSI information. The information flow between the U-Net encoding and decoding paths is effectively interacted and integrated through the Transformer module, helping to reduce information loss and distortion. Simultaneously, through the design of multi-scale convolution and skip connections, the network can better preserve important details and spatial information, improving reconstruction accuracy. Faster feedback speed: The optimized network structure and feature extraction method accelerate the processing and feedback of CSI data. The parallel computation and parameter sharing mechanism of the U-Net and Transformer modules reduce computational load and model complexity, improving feedback speed. In addition, the lightweight design of the network also helps reduce computational and communication latency, further enhancing feedback speed. Better adaptability: The U-Net model, combined with Transformer, exhibits strong adaptability and versatility. The flexibility and adjustability of its network structure enable it to adapt to different channel environments and communication scenarios, thereby improving the system's stability and performance under complex conditions. Furthermore, the end-to-end training method allows the model to automatically learn to adapt to different data distributions and feature variations, further enhancing its adaptability. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 is a flowchart of the CSI compressed feedback method for a multi-antenna MIMO system based on deep learning provided in the first embodiment of the present invention.

[0026] Figure 2 is a block diagram of the CSI compressed feedback system of the deep learning-based multi-antenna MIMO system CSI compressed feedback method provided in the first embodiment of the present invention.

[0027] Figure 3 is a block diagram of the joint model of the user side and the base station side of the CSI compressed feedback method for a multi-antenna MIMO system based on deep learning provided in the first embodiment of the present invention.

[0028] Figure 4 is a network model structure diagram of the user-side coding part of the CSI compression feedback method for a multi-antenna MIMO system based on deep learning provided in the first embodiment of the present invention.

[0029] Figure 5 is a network model structure diagram of the base station-side decoding part of the CSI compression feedback method for a multi-antenna MIMO system based on deep learning provided in the first embodiment of the present invention.

[0030] Figure 6 is a structural diagram of the CSI compressed feedback system of a multi-antenna MIMO system based on deep learning provided in the second embodiment of the present invention.

[0031] Figure 7 is a performance comparison chart of the CSI compressed feedback method for a multi-antenna MIMO system based on deep learning provided in the third embodiment of the present invention.

[0032] Figure 8 is a complexity comparison chart of the CSI compressed feedback method for a multi-antenna MIMO system based on deep learning provided in the third embodiment of the present invention. Detailed Implementation

[0033] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0034] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0035] Example 1

[0036] Referring to Figures 1 to 5, the first embodiment of the present invention provides a CSI compression feedback method for a multi-antenna MIMO system based on deep learning, including: collecting raw CSI data and preparing a neural network model training set; constructing a U-Net neural network model with a joint Transformer based on the collected data; training the neural network model offline to obtain a trained neural network model; and evaluating the trained neural network model based on metrics.

[0037] As shown in Figure 2, the block diagram of the CSI compressed feedback system mainly includes a transmitter and a receiver. At the transmitter, the channel state is first estimated using a channel estimator, and the estimated CSI information is then compressed. The compression process typically includes selective extraction of key information, quantization, and encoding steps to reduce the number of bits required for feedback. The compressed CSI information is sent to the receiver through the feedback channel. Upon receiving the compressed CSI information, the receiver decompresses it to restore the original CSI information. Then, this restored CSI information is used to optimize modulation, coding, and other signal processing strategies to adapt to the current channel environment, thereby improving system performance. The entire process achieves effective utilization and feedback control of channel information to optimize the performance and resource utilization of the communication system.

[0038] S1. Collect raw CSI data and prepare a training set for the neural network model.

[0039] The raw CSI data is collected by the base station, which gathers historical communication data over a period of time and extracts the CSI information to serve as the raw CSI dataset for the neural network model. Furthermore, this embodiment uses a 5.3GHz indoor scenario as an example. The dataset is generated according to the default settings in COST2100. The base station is set to N... t The array is a uniform linear array of 32, with a single receiving antenna at the user end. For a frequency division duplex system, the frequency domain is N. c Given 1024, take N in the corner domain. a The 321,500 unique CSI feedback data were divided into training and testing datasets, containing 120,000 and 30,000 channel matrices, respectively. The batch size in the experiment was set to 32, the learning rate was 0.001, and the optimizer used was Adam.

[0040] The preparation of the network model training set involves using the Discrete Fourier Transform (DFT) to transform the H matrix from the original dataset into both the angular and delay domains, due to the sparsity of the CSI matrix H in the angular delay domain. This transformation yields a training set of size N. c ×N t matrix Where N c and N t These represent the number of user subcarriers and the number of base station antennas, respectively. Take the matrix. The first N a row (N) a <N c H is the CSI compression matrix. a The obtained data is preprocessed and then divided into training and testing sets.

[0041] The original CSI data was processed to obtain the network model dataset, which was then divided into a training set (80%) and a test set (20%). The specific steps for processing the original CSI dataset are as follows:

[0042] S1.1 The CSI matrix H has sparse properties in the angular delay domain. Using the DFT transformation matrix, the angular domain matrix F is obtained. c and delay field matrix F t And calculate the channel matrix in the angular delay domain. Represented as:

[0043] Among them, F c and F t The matrix sizes are N c ×N c and N t ×N t ;

[0044] S1.2, For the angular delay domain channel matrix Only the first Na rows contain large values, while the remaining rows consist of elements close to zero. These elements can be ignored, and there is not much information loss in the compression of the overall CSI. Therefore, the matrix is ​​used. The first N a row (N) a <N c H is the CSI compression matrix. a .

[0045] S1.3 First, compress all CSI matrices H a z-score standardization involves subtracting the mean of the entire dataset from each data point and then dividing by the standard deviation of the dataset. This accelerates model convergence during deep learning model training.

[0046] S1.4 Next, perform min-max scaling on the data from the above steps to constrain the dataset's value range to between -1 and 1, so that the output layer of the network model can use a suitable activation function. The formula for min-max scaling is expressed as:

[0047] Where, x scaled The data is after minimum-maximum scaling, where x is the input data. min For the minimum data, x max This represents the maximum data.

[0048] S1.5 Finally, the preprocessed dataset is divided into training and test sets in a 4:1 ratio.

[0049] S2. Construct a U-Net neural network model with a joint Transformer based on the collected data.

[0050] A U-Net neural network combining Transformer and CSI is constructed. This network model consists of a user-side CSI encoding part and a base station-side CSI decoding part, as shown in Figure 3. The user-side encoding part includes a U-Net encoding path concatenated with a Transformer module, while the base station-side decoding part includes a Transformer module concatenated with a U-Net decoding path. The user-side encoding part compresses the CSI matrix H... a Feature encoding yields a one-dimensional vector of length v. The base station decoding section then decomposes the compressed encoded vector to obtain H. a The compression ratio at this point is η, expressed as:

[0051] η=v / (N c ×N t ×2)

[0052] Where η is the compression ratio and v is the length of the one-dimensional vector obtained by feature encoding.

[0053] Furthermore, the U-Net neural network model with a combined Transformer described in S2 includes an encoder part on the user side and a decoder part on the base station side. As shown in Figure 4, the encoder part on the user side consists of a U-Net encoding path connected in series with a Transformer module. The U-Net encoding path has 5 downsampling convolutional blocks used for compressing and extracting feature information. As shown in Figure 5, the decoder part on the base station side is constructed by connecting a Transformer module in series with a U-Net decoding path. The U-Net decoding path contains 5 upsampling convolutional blocks to restore the compressed information.

[0054] Unlike the standard U-Net model, the U-Net encoding and decoding paths in S2 do not use skip connections. The encoding path consists of five downsampling convolutional blocks. The first two use multi-scale convolutions of 3×3 and 5×5 kernel sizes, extracting feature maps from convolutions of different kernel sizes and then stacking the outputs. The last three layers use standard 3×3 downsampling convolutions. Similarly, the decoding path consists of five upsampling convolutional blocks. The first three layers use standard 3×3 upsampling convolutions due to the smaller feature map size, while the last two upsampling convolutional blocks consist of a 3×3 and a 5×5 multi-scale convolution.

[0055] S3. Train the neural network model offline to obtain a trained neural network model.

[0056] Offline training of neural network models is performed using the PyTorch framework. The L1 loss function is used during the training phase, and the training settings include hyperparameters such as input sample batch size, network iterative optimizer, learning rate, and number of iterations.

[0057] In step S3, the two parts of the joint network are merged and connected to form an end-to-end model. Training data is input into the entire network, and the overall L1 loss function is minimized through backpropagation and an optimizer, thereby updating the parameters of the entire network.

[0058] S4. Evaluate the trained neural network model using metrics.

[0059] Online testing of neural network models involves evaluating the trained network model using the normalized mean squared error (NMSE) metric, expressed as:

[0060] Among them, H a For the true value, These are the predicted values ​​from the neural network model, where NMSE is the normalized mean square error. The expected value operator is ∥·∥2, which represents the L2 norm. A normalized mean squared error (NMSE) threshold is set. If the NMSE is greater than the threshold, the model is considered substandard. In-depth performance analysis is conducted, including data quality and model diagnostics. Model parameters are adjusted and optimized, including hyperparameter tuning, adding regularization, and adjusting the model structure. If the NMSE is less than the threshold, the model performs well. Finally, the optimal model selected after evaluation is deployed.

[0061] Example 2

[0062] Referring to Figure 6, which is the second embodiment of the present invention, it differs from the previous embodiment in that it provides a deep learning-based multi-antenna MIMO system CSI compression feedback system, including: a data acquisition module, a model building module, a model training module, and an evaluation module.

[0063] The data acquisition module is used to collect raw CSI data and prepare a training set for the neural network model.

[0064] The model building module constructs a U-Net neural network model with a joint Transformer based on the collected data.

[0065] The model training module is used to train neural network models offline, resulting in well-trained neural network models.

[0066] The evaluation module is used to evaluate the trained neural network model using metrics.

[0067] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0068] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0069] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0070] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0071] Example 3

[0072] Referring to Figures 7 and 8, the third embodiment of the present invention differs from the previous two embodiments in that it is used to verify and explain the technical effects adopted in the present invention, so as to verify the real effect of the method.

[0073] Figure 7 shows a comparison of the NMSE performance of CSI information recovery when the length of the compressed vector v is 64, 128, 256, and 512. The figure demonstrates that the method described in this invention, by incorporating the Transformer module for constructing long-term dependencies and extracting global feature information, exhibits good NMSE performance at different compression lengths. Compared to traditional benchmark methods, it still achieves higher NMSE even with shorter compression lengths, and at compression lengths of 64, 128, and 256, the method of this invention shows a significant advantage of approximately 10 dB.

[0074] Figure 8 shows the number of network model parameters and computational complexity (FLOPs) for different methods. It can be seen from the figure that the method involved in this invention only increases the number of model parameters by a small amount, and the FLOPs do not increase significantly compared to other methods. Overall, the method of this invention only introduces a small increase in the number of parameters and FLOPs, while significantly improving the NMSE performance of CSI compression feedback. It overcomes the problem of the decline in NMSE performance of traditional methods with increasing compression ratio during CSI compression.

[0075] In summary, the deep learning-based CSI compression feedback method for multi-antenna MIMO systems proposed in this invention can accurately and efficiently compress and feedback CSI information in frequency division duplex mode, while ensuring no degradation of system communication quality. This significantly reduces the amount of information required for CSI feedback, thereby reducing hardware overhead and complexity, and further improving system performance and spectral efficiency. This invention employs a U-Net neural network structure with a combined Transformer architecture to achieve effective CSI compression and reconstruction. This network consists of a user-side encoder and a base station-side decoder. The user-side uses the encoder to compress the CSI data, while the base station uses the decoder to decode and reconstruct the CSI. This invention achieves high-precision CSI compression and reconstruction, providing an efficient and stable CSI compression feedback method for multi-antenna MIMO systems.

[0076] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A CSI compressed feedback method for multi-antenna MIMO systems based on deep learning, characterized in that: include, Collect raw CSI data and prepare a training set for the neural network model; A U-Net neural network model with a joint Transformer is constructed based on the collected data; Train the neural network model offline to obtain a trained neural network model; The trained neural network model is evaluated using metrics.

2. The CSI compressed feedback method for a multi-antenna MIMO system based on deep learning as described in claim 1, characterized in that: The raw GSI data includes historical communication data collected by the base station over a period of time, from which CSI information is extracted as the raw CSI dataset for the neural network model.

3. The CSI compressed feedback method for a deep learning-based multi-antenna MIMO system as described in claim 2, characterized in that: The network model training set includes the sparse properties of the CSI matrix H in the angular delay domain. Using Discrete Fourier Transform, the H matrix in the original GSI dataset is transformed into the angular and delay domains, resulting in a set of size N. c ×N t matrix Where, N c and N t These represent the number of user subcarriers and the number of base station antennas, respectively, and are used to obtain the matrix. The first N a Rows as CSI compression matrix H a The obtained data was preprocessed and then divided into training and testing sets. The CSI matrix H exhibits sparsity in the angular delay domain. Using the DFT transformation matrix, the angular domain matrix F is obtained. c and delay domain matrix F t And calculate the channel matrix in the angular delay domain. Represented as, Among them, F c It is a angular field matrix with size N. c ×N c F t Let N be the delay field matrix. t ×N t ; For the channel matrix in the angular delay domain Take matrix The first N a Rows as CSI compression matrix H a N a <N c ; z-score data standardization is performed on all CSI compression matrices Ha, which involves subtracting the mean of the entire dataset from each data point and then dividing by the standard deviation of the dataset. This accelerates model convergence during deep learning model training. Next, the data is min-max scaled to constrain the dataset's value range to between -1 and 1, allowing the network model's output layer to use a suitable activation function. The min-max scaling is represented as follows: Where, x scaled The data is after minimum-maximum scaling, where x is the input data. min For the minimum data, x max For the largest data; The preprocessed dataset was divided into training and test sets in a 4:1 ratio.

4. The CSI compressed feedback method for a deep learning-based multi-antenna MIMO system as described in claim 3, characterized in that: The neural network model consists of a CSI encoding part on the user side and a CSI decoding part on the base station side. The encoding part on the user side includes a U-Net encoding path connected in series with a Transformer module, and the decoding part on the base station side includes a Transformer module connected in series with a U-Net decoding path. The user-side encoding part will compress the CSI matrix H a Feature encoding yields a one-dimensional vector of length v. The base station decoding section then decomposes the compressed encoded vector to obtain H. a The compression ratio at this point is η, expressed as: η=v / (N c ×N t ×2) Where η is the compression ratio and v is the length of the one-dimensional vector obtained by feature encoding.

5. The CSI compressed feedback method for a multi-antenna MIMO system based on deep learning as described in claim 4, characterized in that: The U-Net encoding path has 5 downsampling convolutional blocks for compressing and extracting feature information; The U-Net decoding path contains 5 upsampled convolutional blocks to restore the compressed information. Unlike the ordinary U-Net model, the encoding and decoding paths of U-Net do not use skip connections. The first two downsampling convolutional blocks in the encoding path use multi-scale convolutions of 3×3 and 5×5 sizes, with convolutions of different kernel sizes extracting feature maps and then stacking them. The last three layers use conventional 3×3 downsampling convolutions. The first three upsampling convolutional blocks in the decoding path use conventional 3×3 upsampling convolutions because the feature map size is small. The last two upsampling convolutional blocks consist of a multi-scale convolution of 3×3 and 5×5 sizes.

6. The CSI compressed feedback method for a multi-antenna MIMO system based on deep learning as described in claim 5, characterized in that: The offline training neural network model involves merging and connecting two parts of a joint network to form an end-to-end model, inputting training data into the entire network, minimizing the overall L1 loss function through a backpropagation algorithm and an optimizer, and updating the parameters of the entire network.

7. The CSI compressed feedback method for a multi-antenna MIMO system based on deep learning as described in claim 6, characterized in that: The evaluation criteria include online testing of the neural network model. The trained network model is evaluated using the Normalized Mean Squared Error (NMSE), denoted as [example needed]. Among them, H a is the true value, These are the predicted values ​​from the neural network model, where NMSE is the normalized mean square error. For expectation value operators, ||·||2 is the L2 norm; If a normalized mean squared error (NMSE) threshold is set, and the NMSE is greater than the threshold, the model settings are considered substandard. A thorough analysis of the performance reasons is then conducted, including data quality and model diagnostics. Model parameters are adjusted and optimized, including hyperparameter tuning, adding regularization, and adjusting the model structure. If the NMSE is less than the threshold, the model performs well. Finally, the optimal model selected after evaluation is deployed.

8. A system employing the deep learning-based CSI compressed feedback method for multi-antenna MIMO systems as described in any one of claims 1 to 7, characterized in that: It includes a data acquisition module, a model building module, a model training module, and an evaluation module; The data acquisition module is used to collect raw CSI data and prepare a training set for the neural network model. The model building module constructs a U-Net neural network model that combines Transformer based on the collected data; The model training module is used to train the neural network model offline to obtain a trained neural network model; The evaluation module is used to evaluate the trained neural network model using metrics.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the deep learning-based CSI compressed feedback method for multi-antenna MIMO systems as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the CSI compressed feedback method for a deep learning-based multi-antenna MIMO system as described in any one of claims 1 to 7.

Images