Unsupervised MIMO channel estimation method and system fusing channel statistical characteristics

Abstract

The invention discloses an unsupervised MIMO channel estimation method and an unsupervised MIMO channel estimation system fusing channel statistical characteristics, and constructs an auto-encoder-like structure utilizing depth image prior network symmetric expansion in order to solve the problems that the traditional method is easily interfered and has high complexity and the channel statistical characteristics are difficult to utilize in the existing unsupervised learning. According to the structure, the encoder, the decoder and the jump connection are combined, so that effective extraction of channel characteristics and signal denoising are realized. The encoder is used for compressing and extracting signal features, and the utilization efficiency of low-dimensional code words is improved through a symmetric structure and jump connection. And the decoder adopts a DIP denoising strategy, and realizes signal denoising by iteratively fitting a signal matrix by using a low-dimensional code word. And carrying out LS estimation on the denoised signal matrix to finally obtain the channel CSI. According to the method, a staged training strategy is adopted: offline training improves the network feature extraction capability through a label-free data set; only decoder parameters are adjusted for denoising in online training, channel statistical characteristics can be fully utilized, and estimation precision and real-time performance are improved.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communications, and relates to channel estimation of a wireless communication receiver, and specifically to an unsupervised MIMO channel estimation method and system based on deep neural network fusion channel statistical characteristics. Background Art

[0002] In 5G and future 6G communications, Massive MIMO systems deploy a large number of antennas between base stations and user equipment to achieve large-scale spatial multiplexing, thereby improving spectrum efficiency and system capacity. With the increase in the number of antennas and the complexity of the communication environment, the challenges faced by real-time MIMO channel estimation are becoming increasingly prominent, and accurate channel estimation is crucial to optimizing system performance. Traditional channel estimation methods such as least squares (LS) and minimum mean square error (MMSE) methods: the LS method is not adaptable enough in dynamic environments and has low accuracy in low signal-to-noise ratio (SNR) environments; and although MMSE has ideal accuracy when processing high-dimensional signals, its computational complexity is too high to meet the requirements of real-time channel estimation; and both estimation methods perform poorly when dealing with channel noise and pilot contamination problems. In order to improve the performance of channel estimation, some studies have proposed compressed sensing (CS) methods. Based on the sparse characteristics of the channel matrix in certain transform domains, CS technology reconstructs high-dimensional channel information through a small amount of observation data to reduce sampling and computational complexity. However, the CS method assumes that the channel is sparse in a certain domain, which is difficult to accurately model in a complex and changing channel environment. In addition, the CS method is sensitive to noise and interference, and the channel reconstruction accuracy will drop significantly in a high-noise environment.

[0003] In recent years, with the development of deep learning technology, researchers have begun to try to apply it to channel estimation problems. By learning from a large amount of training data, deep learning models can automatically extract channel features, establish complex channel models, and achieve higher-precision estimation. However, these supervised learning-based methods rely on a large amount of labeled data, that is, they need to obtain the true value of the signal matrix in advance for training. In actual communication environments, the cost of obtaining these labeled data is very high, and due to the variability of the environment and equipment, the generalization ability of the model is limited. Therefore, although deep learning methods have the potential to improve channel estimation performance in theory, their costs and limitations in practical applications still need to be overcome. Summary of the invention

[0004] Purpose of the invention: Since the traditional channel estimation methods of existing large-scale MIMO systems have high computational complexity when processing high-dimensional signals and lack adaptability in dynamic environments, and the estimation methods based on deep learning are limited by the acquisition of labeled data sets, and the existing unsupervised learning methods are also difficult to fully utilize the statistical characteristics of the channel. The purpose of the present invention is to combine the autoencoder and the deep image prior technology (AE-DIP) to provide an unsupervised MIMO channel estimation method that integrates channel statistical characteristics. This method does not require a large number of labeled training data sets, has low complexity, is close to the LS estimator, and can achieve efficient and accurate channel estimation in complex dynamic environments, thereby improving the overall performance of large-scale MIMO systems.

[0005] Technical solution: To achieve the above-mentioned invention object, the technical solution adopted by the present invention is:

[0006] An unsupervised MIMO channel estimation method integrating channel statistical characteristics is proposed. The method constructs a signal denoising network AE-DIP, uses a symmetrically extended autoencoder-like structure of a deep image prior network, combines an encoder, a decoder and a jump connection, and realizes effective extraction of channel features and signal denoising. The method includes the following steps:

[0007] (1) Convert the received signal matrix into a format suitable for neural network input;

[0008] (2) Offline training of the signal denoising network AE-DIP, inputting the dynamically received and transformed signal matrix into the AE-DIP network to train the network's overall signal statistical feature compression and signal reconstruction capabilities; AE-DIP includes an encoder and a decoder, the encoder is used to compress and extract features in the signal matrix, and the decoder is used to reconstruct the signal, using a structure that is completely symmetrical to the encoder and introducing jump connections;

[0009] (3) After offline training, update the parameters of the online AE-DIP overall network; during online training, fix the encoder and jump connection parameters, use the signal matrix statistics captured in the encoder hidden layer, and train the decoder with the DIP strategy for signal denoising;

[0010] (4) After restoring the denoised signal matrix to its original dimension, LS channel estimation is performed.

[0011] Furthermore, step (1) includes:

[0012] (1.1) Select the antenna dimension as the channel dimension of the input matrix and convert the received signal Convert the dimension to Where N f is the number of subcarriers, N p is the number of OFDM symbols, M is the number of antennas;

[0013] (1.2) The signal in the previous step The real part Y real = Re(Y) and the imaginary part Y imag =Im(Y) separated and superimposed to form a new matrix as input to the neural network.

[0014] Furthermore, the step of offline training the network in step (2) includes:

[0015] (2.1) Signal matrix Y T After the network encoder f encoder , is compressed into a low-dimensional feature tensor Z E , which contains the statistical information of the signal captured by the network; Z E Defined by the following formula:

[0016]

[0017] Where l represents the number of hidden layers, is the hidden layer structure of the i-th layer of the encoder, expressed as:

[0018]

[0019] In the formula is the input of the i-th hidden layer of the encoder, is the network parameter, symbol represents the convolution operation, Pooling, ReLU and BatchNorm represent the pooling operation, linear rectification activation function and batch normalization respectively;

[0020] (2.2) The low-dimensional feature tensor Z E The decoder f after the network decoder Restored to the original signal matrix, the channel reconstructed by the decoder network is expressed as:

[0021]

[0022] In the formula is the hidden layer structure of the i-th layer of the decoder, expressed as:

[0023]

[0024] In the formula represents the input of the i-th layer of the decoder, which contains the information of the previous layer and the jump connection from the corresponding encoder hidden layer, expressed as UpSample indicates upsampling operation;

[0025] (2.3) The mean square error between the reconstructed signal matrix and the original signal matrix is ​​used as the loss function for the overall network training, expressed as:

[0026]

[0027] Here, |||| represents the Euclidean norm.

[0028] Furthermore, the step of online training the network in step (3) includes:

[0029] (3.1) Fix the network parameters of the encoder, including the skip connection layers, to ensure that the training process starts with the same data baseline, i.e., the compressed feature vectors produced by the encoder;

[0030] (3.2) The parameters of the decoder network are trained according to the DIP strategy, and the loss function is:

[0031]

[0032] Where f decoder (Z E ) is the signal matrix output after the decoder reconstruction, the signal matrix Y T The feature vector Z compressed by the encoder E is input to the decoder and after being restored by the decoder, it is combined with Y T Compare them and optimize the parameters by calculating the mean square error between the two. Repeat the forward and backward propagation process until the signal matrix output by the decoder is close to the original signal matrix after denoising.

[0033] (3.3) Use the early stopping strategy to train the parameters of the decoder network.

[0034] Furthermore, step (4) includes:

[0035] (4.1) The dimension of the signal matrix after network denoising is expressed as Recombines its real and imaginary parts and transforms them back to the original matrix dimensions.

[0036] (4.2) According to the channel setting, the first several OFDM symbols of the received signal are selected as pilot signals, and the pilot matrix X is extracted. The known pilot matrix X is used as the transmission signal for subsequent channel estimation. The estimated value of the channel matrix is

[0037] Based on the same inventive concept, the present invention also provides an unsupervised MIMO channel estimation system based on deep learning that integrates channel statistical characteristics, comprising the following modules:

[0038] A preprocessing module, used to convert the received signal matrix into a format suitable for neural network input;

[0039] AE-DIP offline training module is used for offline training of the signal denoising network AE-DIP. The dynamically received converted signal matrix is ​​input into the AE-DIP network to train the overall signal statistical feature compression and signal reconstruction capabilities of the network. AE-DIP includes an encoder and a decoder. The encoder is used to compress and extract features from the signal matrix, and the decoder is used for signal reconstruction. It adopts a structure that is completely symmetrical with the encoder and introduces jump connections.

[0040] The network update module is used to update the parameters of the online AE-DIP overall network after offline training;

[0041] AE-DIP online training module, which is used to fix the encoder and jump connection parameters during online training, and uses the signal matrix statistics captured in the encoder hidden layer to train the decoder with the DIP strategy for signal denoising;

[0042] The channel estimation module is used to restore the denoised signal matrix to its original dimension and then perform LS channel estimation.

[0043] The present invention also provides a computer system, comprising a memory, a processor, and a computer program / instruction stored in the memory and executable on the processor, wherein when the computer program / instruction is executed by the processor, the steps of the unsupervised MIMO channel estimation method integrating channel statistical characteristics are implemented.

[0044] The present invention also provides a computer program product, comprising a computer program / instruction, wherein when the computer program / instruction is executed by a processor, the steps of the unsupervised MIMO channel estimation method integrating channel statistical characteristics are implemented.

[0045] Beneficial effects: The present invention proposes a novel unsupervised MIMO channel estimation method that integrates channel statistical characteristics, combined with a quasi-autoencoder structure and deep image prior (DIP) technology, which can effectively improve the accuracy and real-time performance of channel estimation without relying on a large amount of labeled data. The method makes full use of the intrinsic statistical characteristics of the signal in the time domain, spatial domain and frequency domain to achieve efficient denoising and estimation of the channel. Specifically, the present invention compresses the signal matrix into a low-dimensional feature representation by introducing an autoencoder structure, extracts the main features of the channel, and retains the information transfer between different levels through jump connections to ensure that the signal maintains key spatiotemporal correlation during the denoising process. In addition, the present invention innovatively introduces DIP technology into the field of signal processing. DIP technology was originally used for image denoising, and its core idea is to use the intrinsic prior characteristics of the signal through an unsupervised learning model to achieve effective denoising processing. The present invention uses DIP technology for signal denoising tasks in channel estimation, and combines it with an autoencoder structure to perform fast and efficient denoising without relying on labeled data, thereby improving the real-time performance and accuracy of channel estimation and meeting the real-time requirements in large-scale MIMO systems.

[0046] Compared with the prior art, the significant effect of the present invention is that it overcomes the defect that supervised learning estimation requires a large-scale labeled data set and the defect that unsupervised learning estimation does not reasonably utilize channel statistical information, improves the performance of the LS estimator, and achieves the performance of MMSE estimation while maintaining low complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flow chart of a channel estimation method according to an embodiment of the present invention;

[0048] Figure 2 is an overall network structure diagram of an embodiment of the present invention;

[0049] Figure 3 A schematic diagram of the process and connection of offline training and online training in an embodiment of the present invention;

[0050] Figure 4 This is a comparison chart of channel estimation performance between the method mentioned in the embodiment of the present invention and LS estimation and MMSE estimation in a single-cell system;

[0051] Figure 5 This is a comparison chart of channel estimation performance between the method mentioned in the embodiment of the present invention and LS estimation and MMSE estimation in a multi-cell system;

[0052] Figure 6 A comparison diagram of denoising timeliness in a single cell environment when an autoencoder network structure is used to update network parameters in an embodiment of the present invention;

[0053] Figure 7This is a comparison diagram of the denoising timeliness in a multi-cell environment when the autoencoder network structure is used to update the network parameters in an embodiment of the present invention. DETAILED DESCRIPTION

[0054] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0055] The embodiment of the present invention discloses an unsupervised MIMO channel estimation method integrating channel statistical characteristics, combining autoencoders and deep image prior technology, extracting and compressing signal matrices through the autoencoder network structure, and performing denoising based on low-dimensional feature tensors, thereby improving the timeliness and accuracy of channel estimation. Figure 1 As shown, the method mainly includes the following steps:

[0056] Step (1): Convert the received signal matrix into a format suitable for neural network input.

[0057] In the specific implementation, the format conversion is performed according to the specific MIMO communication model. In this embodiment, the LTE-EPA model considering the spatial domain, frequency domain and time domain information is taken as an example. The signal received by the base station can be expressed as Where N f is the number of subcarriers, N p is the number of OFDM symbols, and M is the number of antennas. The specific steps are as follows:

[0058] Step (1.1): Considering the input characteristics of the deep neural network, the channel dimension should be set to the 0th dimension, and the spatial dimension, that is, the antenna dimension, should be selected as the channel dimension of the input matrix. Specifically, the received signal is Convert the dimension to

[0059] Step (1.2): Since the received signal is a complex matrix, the deep neural network cannot directly process complex signals, so it is necessary to separate the real and imaginary parts of the signal and superimpose them to form a new matrix. Specifically, the signal in the previous step The real part Y real = Re(Y) and the imaginary part Y imag =Im(Y) separated and superimposed to form a new matrix Y T as input to the neural network.

[0060]

[0061] Step (2): Offline training of the signal denoising network AE-DIP, inputting the dynamically received transformed signal matrix into the AE-DIP network to train the overall signal statistical feature compression and signal reconstruction capabilities of the network.

[0062] In this step, the signal matrix Y converted in step (1) is T As the input of the overall network, the autoencoder network is trained to perform data compression and recovery. In order to improve the feature extraction and compression capabilities of the network, this embodiment also introduces jump connections based on the network architecture of the autoencoder. Figure 2 The channel after reconstruction by the overall autoencoder network can be expressed as The specific steps are as follows:

[0063] Step (2.1): Signal matrix Y T After passing through the encoder part of the network, it is compressed into a low-dimensional feature tensor Z E . Z E Defined by the following formula:

[0064]

[0065] Where l represents the number of hidden layers, is the hidden layer structure of the i-th layer of the encoder, which can be expressed as:

[0066]

[0067] In the formula is the input of the i-th hidden layer of the encoder, is the network parameter, symbol Represents the convolution operation, which actually represents cross-correlation in signal processing, that is, using 1×1 convolution as a cross-correlator to multiply each element in the time-frequency grid by the same shared parameter matrix to obtain a new spatial vector for the next hidden layer. Pooling, ReLU, and BatchNorm represent pooling operations, linear rectification activation functions, and batch normalization, respectively.

[0068] Step (2.2): Convert the low-dimensional feature tensor Z E The decoder part of the network is restored to the original signal matrix to train the channel recovery ability of the encoder network. The channel reconstructed by the decoder network can be expressed as:

[0069]

[0070] In the formula That is the hidden layer structure of the decoder, which can be expressed as:

[0071]

[0072] In the formula represents the input of the i-th layer of the decoder, which contains the information of the previous layer and the jump connection from the corresponding encoder hidden layer, expressed as UpSample indicates an upsampling operation.

[0073] Step (2.3): Use the mean square error between the reconstructed channel and the original channel as the loss function for the overall network training, expressed as:

[0074]

[0075] In the formula, |||| represents the Euclidean norm. In this way, the feature extraction and compression capabilities of the encoder part of the autoencoder network and the channel recovery capabilities of the decoder part are trained.

[0076] Step (3): After offline training, update the parameters of the online AE-DIP overall network; during online training, fix the encoder and jump connection parameters, use the signal matrix statistics captured in the encoder hidden layer (such as spatial and frequency domain correlations), and train the decoder with the DIP strategy for signal denoising.

[0077] In specific implementation, the network parameters can be updated periodically to determine whether the offline training has reached the preset time. If it has, the overall parameters of the online denoising network are updated; if it has not yet arrived, the parameters of the overall network are not updated. Or if the channel state changes significantly, the signal matrix dynamically captured during the online network training process is used to retrain the entire network and update the network parameters. For example, the newly obtained channel matrix H′ and the channel matrix obtained after the original estimation are calculated. The Frobenius norm of the difference between the two is calculated, and a threshold is set. When the difference exceeds the threshold, it is considered that the channel state has changed significantly. Then the encoder part of the network is fixed, that is, during the denoising process, the encoder parameters are not updated with gradients. The idea of ​​deep image prior is adopted, and only the decoder is used to compress the feature z in a low-dimensional manner. E The specific operation is as follows:

[0078] Step (3.1): Fix the network parameters of the encoder, including the skip connection layer, to ensure that the training process starts with the same data benchmark, that is, the compressed feature vector generated by the encoder. Fixing the parameters of the encoder network means that the denoising of the decoder will start from the same starting point, that is, from the compressed feature vector of the original channel. This training strategy can retain the feature extraction ability of the AE-DIP model trained in the previous stage and avoid destroying the channel statistics captured by the encoder during the optimization process.

[0079] Step (3.2): Since the DIP method uses the idea that the network structure contains natural image priors, that is, because the network itself has low permeability to natural image information and high impedance to random information such as noise, a clean sample without noise will be generated during the process of fitting the noisy sample. At the same time, the signal matrix also has natural information similar to that of the image, and signal denoising is similar to the image denoising task. Therefore, the parameters of the decoder network will be trained according to the DIP strategy, and the selected loss function is

[0080]

[0081] Where f decoder (Z E ) is the signal matrix output after the decoder reconstruction, the signal matrix Y T The feature vector Z compressed by the encoder E is input to the decoder and restored to its original channel dimension after being decoded and Y T Compare them and optimize the parameters by calculating the mean square error between the two. Repeat this forward and backward propagation process until the signal matrix output by the decoder is close to the original signal matrix after denoising. Figure 3 In the offline training phase, the network is first trained to compress the overall signal statistics and reconstruct the signal. The original signal matrix Y T After the encoder is compressed, the feature vector Z is obtained E , the feature vector is input into the decoder network for reconstruction. In the online training phase, after fixing the encoder and jump connection parameters, the decoder parameters are continuously optimized through the forward and back propagation processes, so that the reconstructed signal matrix is ​​close to the clean signal matrix, thereby achieving the effect of signal denoising. That is the above loss function Loss D , according to the gradient of the loss function Update the parameters of each layer.

[0082] Step (3.3): Use the early stopping strategy to train the parameters of the decoder network. Using the characteristics of the deep image prior neural network, the network will prioritize simulating natural information rather than noise information during the iteration process, that is, the network has low permeability to natural information and high impedance to noise information. The specific number of early stopping rounds depends on the specific MIMO system.

[0083] Step (4): After restoring the denoised signal matrix to its original dimension, perform LS channel estimation. The specific steps are as follows:

[0084] Step (4.1): The dimension of the signal matrix after network denoising is expressed as Recombines its real and imaginary parts and transforms them back to the original matrix dimensions.

[0085] Step (4.2): Select the first several OFDM symbols of the received signal as the pilot signal according to the channel setting, and extract the pilot matrix X. The signal received at the receiving end can be expressed as Y = HX + N, where H is the channel matrix and N is Gaussian white noise. Use the known frequency matrix X as the transmission signal for subsequent channel estimation. The estimated value of the channel matrix is

[0086] After completing the processing of the current signal matrix, it transfers to the next received signal matrix and repeats the above process to perform channel estimation and denoising to ensure the accuracy and stability of channel estimation.

[0087] In summary, an unsupervised MIMO channel estimation method integrating channel statistical characteristics provided by an embodiment of the present invention mainly includes a feature extraction training stage and a channel denoising training stage. In the feature extraction training stage, a global training strategy is adopted to update the parameters of the entire network. The goal is to enable the network to learn how to map the input channel to the output through the mean square error (MSE) loss function, so as to ensure that the encoder can compress the channel information into a low-dimensional feature vector containing sufficient statistical information, thereby improving the channel recovery ability of the decoder. In this stage, the channel matrix received during the communication process is directly used for training through an unsupervised learning method, which reduces the dependence on large data sets and enhances the flexibility and practicality of the model. In the channel denoising training stage, the training focus is turned to enhancing the denoising ability of the decoder. The parameters of the encoder are fixed to maintain the consistency of feature extraction, so that the decoder performs denoising training on the same compressed feature vector starting point, and further optimizes channel recovery. Through repeated forward and back propagation, the decoder network gradually improves the ability to recover low-dimensional feature vectors, and finally achieves efficient denoising reconstruction of the original channel.

[0088] The AE-DIP method of the present invention demonstrates excellent channel estimation performance and convergence speed in both single-cell and multi-cell scenarios. Figure 4 As shown in the figure, in a single-cell environment, the AE-DIP method has better estimation accuracy than the traditional LS method, and exhibits better stability and anti-interference ability, especially under high signal-to-noise ratio conditions, and achieves performance comparable to that of the MMSE estimation. Figure 5 The performance comparison in a multi-cell interference environment is demonstrated. The AE-DIP method still maintains strong robustness. Even in the case of severe interference, its estimation accuracy is better than the traditional LS method and is comparable to the performance of the MMSE method.

[0089] In terms of convergence speed, Figure 6 and Figure 7 The comparison between AE-DIP and traditional DIP method is shown. Figure 6In a single-cell scenario, the AE-DIP method shows a faster convergence speed, significantly reducing the mean square error (MSE) in the initial iteration stage, and significantly improving the efficiency of channel recovery. Figure 7 ,In a multi-cell interference environment, AE-DIP still maintains its advantage of fast convergence, indicating that it can ,quickly adapt to channel changes in complex interference scenarios and ,improve the real-time response capability of the system.

[0090] Based on the same inventive concept, an embodiment of the present invention also discloses an unsupervised MIMO channel estimation system based on deep learning that integrates channel statistical characteristics, including the following modules:

[0091] A preprocessing module, used to convert the received signal matrix into a format suitable for neural network input;

[0092] AE-DIP offline training module, used for offline training of the signal denoising network AE-DIP, inputting the converted signal matrix into the AE-DIP network to train the overall signal statistical feature compression and signal reconstruction capabilities of the network; AE-DIP includes an encoder and a decoder, the encoder is used to compress and extract features in the signal matrix, and the decoder is used for signal reconstruction, adopting a structure that is completely symmetrical with the encoder and introducing jump connections;

[0093] The network update module is used to update the parameters of the online AE-DIP overall network after offline training;

[0094] AE-DIP online training module, which is used to fix the encoder and jump connection parameters during online training, and uses the signal matrix statistics captured in the encoder hidden layer to train the decoder with the DIP strategy for signal denoising;

[0095] The channel estimation module is used to restore the denoised signal matrix to its original dimension and then perform LS channel estimation.

[0096] An embodiment of the present invention also discloses a computer system, including a memory, a processor, and a computer program / instruction stored in the memory and executable on the processor. When the computer program / instruction is executed by the processor, the steps of the unsupervised MIMO channel estimation method integrating channel statistical characteristics are implemented.

[0097] The embodiment of the present invention also discloses a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the unsupervised MIMO channel estimation method integrating channel statistical characteristics.

Claims

1. An unsupervised MIMO channel estimation method integrating channel statistical characteristics, characterized in that: This method constructs a signal denoising network AE-DIP, uses the autoencoder-like structure symmetrically extended by the deep image prior network, combines the encoder, decoder and jump connection, and realizes the effective extraction of channel features and signal denoising, including the following steps: (1) Convert the received signal matrix into a format suitable for neural network input; (2) Offline training of the signal denoising network AE-DIP, inputting the dynamically received and transformed signal matrix into the AE-DIP network to train the network's overall signal statistical feature compression and signal reconstruction capabilities; AE-DIP includes an encoder and a decoder, the encoder is used to compress and extract features in the signal matrix, and the decoder is used to reconstruct the signal, using a structure that is completely symmetrical to the encoder and introducing jump connections; (3) After offline training, update the parameters of the online AE-DIP overall network; during online training, fix the encoder and jump connection parameters, use the signal matrix statistics captured in the encoder hidden layer, and train the decoder with the DIP strategy for signal denoising; (4) After restoring the denoised signal matrix to its original dimension, LS channel estimation is performed.

2. The unsupervised MIMO channel estimation method integrating channel statistical characteristics according to claim 1, characterized in that: Step (1) includes: (1.1) Select the antenna dimension as the channel dimension of the input matrix and convert the received signal Convert the dimension to Where N f is the number of subcarriers, N p is the number of OFDM symbols, M is the number of antennas; (1.2) The signal in the previous step The real part Y real = Re(Y) and the imaginary part Y imag =Im(Y) separated and superimposed to form a new matrix as input to the neural network.

3. The unsupervised MIMO channel estimation method integrating channel statistical characteristics according to claim 1, characterized in that: The steps of offline training the network in step (2) include: (2.1) Signal matrix Y T After the network encoder f encoder , is compressed into a low-dimensional feature tensor Z E , which contains the statistical information of the signal captured by the network; Z E Defined by the following formula: Where l represents the number of hidden layers, is the hidden layer structure of the i-th layer of the encoder, expressed as: In the formula is the input of the i-th hidden layer of the encoder, is the network parameter, symbol represents the convolution operation, Pooling, ReLU and BatchNorm represent the pooling operation, linear rectification activation function and batch normalization respectively; (2.2) The low-dimensional feature tensor Z E The decoder f after the network decoder Restored to the original signal matrix, the channel reconstructed by the decoder network is expressed as: In the formula is the hidden layer structure of the i-th layer of the decoder, expressed as: In the formula represents the input of the i-th layer of the decoder, which contains the information of the previous layer and the jump connection from the corresponding encoder hidden layer, expressed as UpSample indicates upsampling operation; (2.3) The mean square error between the reconstructed signal matrix and the original signal matrix is ​​used as the loss function for the overall network training, expressed as: In the formula, ‖‖ represents the Euclidean norm.

4. The unsupervised MIMO channel estimation method integrating channel statistical characteristics according to claim 1, characterized in that: The steps of online training the network in step (3) include: (3.1) Fix the network parameters of the encoder, including the skip connection layers, to ensure that the training process starts with the same data baseline, i.e., the compressed feature vectors produced by the encoder; (3.2) The parameters of the decoder network are trained according to the DIP strategy, and the loss function is: Where f decoder (Z E ) is the signal matrix output after the decoder reconstruction, the signal matrix Y T The feature vector Z compressed by the encoder E is input to the decoder and after being restored by the decoder, it is combined with Y T Compare them and optimize the parameters by calculating the mean square error between the two. Repeat the forward and backward propagation process until the signal matrix output by the decoder is close to the original signal matrix after denoising. (3.3) Use the early stopping strategy to train the parameters of the decoder network.

5. The unsupervised MIMO channel estimation method integrating channel statistical characteristics according to claim 2, characterized in that: Step (4) includes: (4.1) The dimension of the signal matrix after network denoising is expressed as Recombines its real and imaginary parts and transforms them back to the original matrix dimensions. (4.2) According to the channel setting, the first several OFDM symbols of the received signal are selected as pilot signals, and the pilot matrix X is extracted. The known pilot matrix X is used as the transmission signal for subsequent channel estimation. The estimated value of the channel matrix is 6. An unsupervised MIMO channel estimation system based on deep learning that integrates channel statistics, characterized in that: The system constructs a signal denoising network AE-DIP, which uses the autoencoder-like structure symmetrically extended by the deep image prior network, combines the encoder, decoder and jump connection to achieve effective extraction of channel features and signal denoising, including the following modules: A preprocessing module, used to convert the received signal matrix into a format suitable for neural network input; AE-DIP offline training module is used for offline training of the signal denoising network AE-DIP. The dynamically received converted signal matrix is ​​input into the AE-DIP network to train the overall signal statistical feature compression and signal reconstruction capabilities of the network. AE-DIP includes an encoder and a decoder. The encoder is used to compress and extract features from the signal matrix, and the decoder is used for signal reconstruction. It adopts a structure that is completely symmetrical with the encoder and introduces jump connections. The network update module is used to update the parameters of the online AE-DIP overall network after offline training; AE-DIP online training module, which is used to fix the encoder and jump connection parameters during online training, and uses the signal matrix statistics captured in the encoder hidden layer to train the decoder with the DIP strategy for signal denoising; The channel estimation module is used to restore the denoised signal matrix to its original dimension and then perform LS channel estimation.

7. The unsupervised MIMO channel estimation system integrating channel statistics according to claim 6, characterized in that: In the AE-DIP offline training module, the steps of offline training the network include: Signal matrix Y T After the network encoder f encoder , is compressed into a low-dimensional feature tensor Z E , which contains the statistical information of the signal captured by the network; Z E Defined by the following formula: Where l represents the number of hidden layers, is the hidden layer structure of the i-th layer of the encoder, expressed as: In the formula is the input of the i-th hidden layer of the encoder, is the network parameter, symbol represents the convolution operation, Pooling, ReLU and BatchNorm represent the pooling operation, linear rectification activation function and batch normalization respectively; The low-dimensional feature tensor Z E The decoder f after the network decoder Restored to the original signal matrix, the channel reconstructed by the decoder network is expressed as: In the formula is the hidden layer structure of the i-th layer of the decoder, expressed as: In the formula represents the input of the i-th layer of the decoder, which contains the information of the previous layer and the jump connection from the corresponding encoder hidden layer, expressed as UpSample indicates upsampling operation; The mean square error between the reconstructed signal matrix and the original signal matrix is ​​used as the loss function for the overall network training, expressed as: In the formula, ‖‖ represents the Euclidean norm.

8. The unsupervised MIMO channel estimation system integrating channel statistics according to claim 6, characterized in that: In the AE-DIP online training module, the steps of online training the network include: The network parameters of the encoder, including the skip connection layers, are fixed to ensure that the training process starts with the same data baseline, i.e., the compressed feature vectors produced by the encoder. The parameters of the decoder network are trained according to the DIP strategy, and the loss function is: Where f decoder (Z E ) is the signal matrix output after the decoder reconstruction, the signal matrix Y T The feature vector Z compressed by the encoder E is input to the decoder and after being restored by the decoder, it is combined with Y T Compare them and optimize the parameters by calculating the mean square error between the two. Repeat the forward and backward propagation process until the signal matrix output by the decoder is close to the original signal matrix after denoising. The parameters of the decoder network are trained using an early stopping strategy.

9. A computer system comprising a memory, a processor, and a computer program / instruction stored in the memory and executable on the processor, characterized in that: When the computer program / instructions are executed by a processor, the steps of an unsupervised MIMO channel estimation method integrating channel statistical characteristics according to any one of claims 1 to 5 are implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of an unsupervised MIMO channel estimation method integrating channel statistical characteristics according to any one of claims 1 to 5 are implemented.

Images