An Unsupervised MIMO Channel Estimation Method and System that Integrates Channel Statistical Characteristics

Abstract

This invention discloses an unsupervised MIMO channel estimation method and system that integrates channel statistical characteristics. Addressing the problems of traditional methods being susceptible to interference, high complexity, and the difficulty of utilizing channel statistical characteristics in existing unsupervised learning, this invention constructs a symmetric extension of a deep image prior network, resembling an autoencoder. This structure combines an encoder, a decoder, and skip connections to achieve efficient extraction of channel features and signal denoising. The encoder is used to compress and extract signal features, improving the efficiency of low-dimensional codeword utilization through its symmetric structure and skip connections. The decoder employs a DIP denoising strategy, using iterative fitting of the signal matrix with low-dimensional codewords to achieve signal denoising. The denoised signal matrix is ​​then subjected to LS estimation to finally obtain the channel CSI. This invention adopts a phased training strategy: offline training improves the network's feature extraction capability using unlabeled datasets; online training only adjusts the decoder parameters for denoising, fully utilizing channel statistical characteristics and improving estimation accuracy and real-time performance.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication and relates to channel estimation for wireless communication receivers. Specifically, it relates to an unsupervised MIMO channel estimation method and system based on deep neural networks that fuses channel statistical characteristics. Background Technology

[0002] In 5G and future 6G communications, Massive MIMO systems achieve large-scale spatial multiplexing by deploying a large number of antennas between base stations and user equipment, thereby improving spectral efficiency and system capacity. With the increasing number of antennas and the increasing complexity of the communication environment, the challenges of real-time MIMO channel estimation are becoming increasingly prominent, and accurate channel estimation is crucial for optimizing system performance. Traditional channel estimation methods, such as Least Squares (LS) and Minimum Mean Square Error (MMSE), present challenges: LS lacks adaptability in dynamic environments and has low accuracy in low signal-to-noise ratio (SNR) environments; while MMSE, although achieving relatively ideal accuracy when processing high-dimensional signals, suffers from excessive computational complexity, making it difficult to meet the requirements of real-time channel estimation; furthermore, both methods perform poorly in dealing with channel noise and pilot contamination. To improve channel estimation performance, compressed sensing (CS) methods have been proposed. CS technology is based on the sparsity of the channel matrix in certain transform domains, reconstructing high-dimensional channel information with a small amount of observation data, reducing 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 complex and variable channel environments. In addition, the CS method is sensitive to noise and interference, and the channel reconstruction accuracy will decrease significantly in high-noise environments.

[0003] In recent years, with the development of deep learning technology, researchers have begun to apply it to channel estimation problems. Deep learning models, through learning from large amounts of training data, can automatically extract channel features, build complex channel models, and achieve higher-precision estimations. However, these supervised learning-based methods rely on large amounts of labeled data, requiring the prior acquisition of the true values ​​of the signal matrix for training. In real-world communication environments, acquiring this labeled data is very costly, and the model's generalization ability is limited due to the variability of the environment and equipment. Therefore, while deep learning methods theoretically have the potential to improve channel estimation performance, their cost and limitations in practical applications still need to be overcome. Summary of the Invention

[0004] Purpose of the Invention: Existing traditional channel estimation methods for large-scale MIMO systems suffer from high computational complexity when processing high-dimensional signals and insufficient adaptability to dynamic environments. Deep learning-based estimation methods are limited by the availability of labeled datasets, while existing unsupervised learning methods struggle to fully utilize channel statistical characteristics. The purpose of this invention is to combine autoencoders and deep image prior techniques (AE-DIP) to provide an unsupervised MIMO channel estimation method that integrates channel statistical characteristics. This method requires no large labeled training datasets, has low complexity (approaching that of a simplex 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 objectives, the technical solution adopted by this invention is as follows:

[0006] An unsupervised MIMO channel estimation method incorporating channel statistical characteristics is proposed. This method constructs a signal denoising network AE-DIP, utilizes a symmetrically extended autoencoder-like structure of a deep image prior network, and combines an encoder, decoder, and skip connections to achieve 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 signal denoising network AE-DIP, input 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 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 skip connections.

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

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

[0011] Furthermore, step (1) includes:

[0012] (1.1) Select the antenna dimension as the channel dimension of the input matrix, and input the received signal. Transformation Dimensions Where N f N is the number of subcarriers. p M represents the number of OFDM symbols and M represents the number of antennas.

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

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

[0015] (2.1) Signal matrix Y T Encoder f via network encoder It is compressed into a low-dimensional feature tensor Z. E It contains statistical information about signals captured by the network; Z E Defined by the following formula:

[0016]

[0017] In the formula, l represents the number of hidden layers. The hidden layer structure of the encoder's i-th layer is represented as follows:

[0018]

[0019] In the formula It is the input of the i-th hidden layer of the encoder. These are network parameters, symbols. The symbol represents the convolution operation, while Pooling, ReLU, and BatchNorm represent pooling, linear rectified activation function, and batch normalization, respectively.

[0020] (2.2) Convert the low-dimensional feature tensor Z E The decoder f after passing through the network decoder The channel representation after reconstructing the original signal matrix and using the decoder network is as follows:

[0021]

[0022] In the formula Let the hidden layer structure of the i-th layer of the decoder be represented as:

[0023]

[0024] In the formula This represents the input to the i-th layer of the decoder, which contains information from the previous layer and skip connections from the corresponding encoder hidden layer, and is represented in the form of: UpSample indicates an 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 training of the network, expressed as:

[0026]

[0027] In the formula, |||| represents the Euclidean norm.

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

[0029] (3.1) Fix the network parameters of the encoder, including the skip connection layers therein, to ensure that the training process starts with the same data benchmark, i.e. the compressed feature vector generated 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] In the formula f decoder (Z E Y is the signal matrix output after reconstruction by the decoder. T The feature vector Z compressed by the encoder E It is input into the decoder, and after being restored by the decoder, it is compared with Y. T The parameters are optimized by comparing the mean square error of the two signals and iterating this forward and backward propagation process repeatedly until the signal matrix output by the decoder is close to the original signal matrix after denoising.

[0033] (3.3) Train the parameters of the decoder network using the early stopping strategy.

[0034] Furthermore, step (4) includes:

[0035] (4.1) The dimension of the signal matrix after network denoising is expressed as follows: Recombine the real and imaginary parts and rearrange their order to obtain the original matrix dimensions.

[0036] (4.2) Select the first few OFDM symbols of the received signal as pilot signals according to the channel settings, extract the pilot matrix X, and use the known pilot matrix X as the transmission signal for subsequent channel estimation. The estimated value of the channel matrix is...

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

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

[0039] The AE-DIP offline training module is used to train the AE-DIP signal denoising network offline. It inputs 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 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 skip connections.

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

[0041] The AE-DIP online training module is used to train the decoder for signal denoising by fixing the encoder and skip connection parameters during online training and utilizing the statistical information of the signal matrix captured in the encoder's hidden layer and the DIP strategy.

[0042] The channel estimation module is used to perform LS channel estimation after restoring the denoised signal matrix to its original dimensions.

[0043] The present invention also provides a computer system, including a memory, a processor, and a computer program / instructions stored in the memory and executable on the processor, wherein the computer program / instructions, when executed by the processor, implement the steps of the unsupervised MIMO channel estimation method for fusing channel statistical characteristics.

[0044] The present invention also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the unsupervised MIMO channel estimation method that integrates channel statistical characteristics.

[0045] Beneficial Effects: This invention proposes a novel unsupervised MIMO channel estimation method that integrates channel statistical characteristics. By combining an autoencoder-like structure and Deep Image Prior (DIP) technology, it effectively improves the accuracy and real-time performance of channel estimation without relying on large amounts of labeled data. This method fully utilizes the inherent statistical characteristics of signals in the time, spatial, and frequency domains to achieve efficient channel denoising and estimation. Specifically, this invention introduces an autoencoder structure to compress the signal matrix into a low-dimensional feature representation, extracting the main features of the channel. By using skip connections to preserve information transfer between different levels, it ensures that the signal maintains crucial spatiotemporal correlation during denoising. Furthermore, this invention innovatively introduces DIP technology into the field of signal processing. DIP technology was initially used for image denoising; its core idea is to utilize the inherent prior characteristics of signals through an unsupervised learning model to achieve effective denoising. This invention applies DIP technology to the signal denoising task in channel estimation, combining 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 of large-scale MIMO systems.

[0046] Compared with the prior art, the significant advantage of the present invention is that it overcomes the shortcomings of supervised learning estimation, which requires a large-scale labeled dataset, and unsupervised learning estimation, which does not make reasonable use of channel statistics. It improves the performance of the LS estimator and achieves the performance of MMSE estimation while maintaining low complexity. Attached Figure Description

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

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

[0049] Figure 3 This is a schematic diagram illustrating the process and connection between offline training and online training in an embodiment of the present invention;

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

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

[0052] Figure 6 This is a comparison chart of the denoising timeliness in a single-cell environment based on whether or not an autoencoder network structure is used to update network parameters in this embodiment of the invention.

[0053] Figure 7This is a comparison chart showing the denoising timeliness of updating network parameters in a multi-cell environment with or without using an autoencoder network structure in this embodiment of the invention. Detailed Implementation

[0054] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] This invention discloses an unsupervised MIMO channel estimation method that integrates channel statistical characteristics. Combining autoencoders and deep image prior techniques, it extracts and compresses the signal matrix through an autoencoder network structure, and performs denoising starting from 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 specific embodiments, format conversion is performed according to the specific MIMO communication model. This embodiment takes the LTE-EPA model, which considers spatial, frequency, and time domain information, as an example. The signal received by the base station can be represented as follows: Where N f N represents the number of subcarriers. p Where M 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 deep neural networks, the channel dimension should be set to the 0th dimension, and the spatial dimension, i.e., the antenna dimension, should be selected as the channel dimension of the input matrix. Specifically, the received signal is processed by... Transformation Dimensions

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

[0060]

[0061] Step (2): Train the signal denoising network AE-DIP offline. Input the dynamically received and converted signal matrix into the AE-DIP network to train the network's overall signal statistical feature compression and signal reconstruction capabilities.

[0062] In this step, the signal matrix Y after transformation in step (1) is... T As input to the overall network, the autoencoder network is trained for data compression and recovery. To improve the network's feature extraction and compression capabilities, this embodiment also introduces skip connections based on the autoencoder network architecture. The overall network structure diagram is as follows: Figure 2 The channel reconstructed by the overall autoencoder network can be represented 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] In the formula, l represents the number of hidden layers. The hidden layer structure of the encoder's i-th layer can be represented as:

[0066]

[0067] In the formula It is the input of the i-th hidden layer of the encoder. These are network parameters, symbols. Representing the convolution operation, in signal processing it actually represents cross-correlation. It uses a 1×1 convolution as a cross-correlator, multiplying each element in the time-frequency grid with the same shared parameter matrix to obtain a new spatial vector for the next hidden layer. Pooling, ReLU, and BatchNorm represent pooling, linear rectified activation function, and batch normalization, respectively.

[0068] Step (2.2): Convert the low-dimensional feature tensor Z E The decoder part of the network recovers the original signal matrix, training the encoder network's channel recovery capability. The channel reconstructed by the decoder network can be represented as:

[0069]

[0070] In the formula The hidden layer structure of the decoder can be represented as follows:

[0071]

[0072] In the formula This represents the input to the i-th layer of the decoder, which contains information from the previous layer and skip connections from the corresponding encoder hidden layer, and is represented in the form of: 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 training of the network, expressed as:

[0074]

[0075] In the formula, |||| represents the Euclidean norm. This method is used to train the autoencoder network to improve the feature extraction and compression capabilities of the encoder part, as well as the channel recovery capabilities of the decoder part.

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

[0077] In practice, network parameters can be periodically updated to determine if the offline training has reached a preset time. If it has, the overall parameters of the online denoising network are updated; otherwise, the overall network parameters are not updated. Alternatively, if the channel state changes significantly, the network can be retrained using the signal matrix dynamically captured during online training, updating the network parameters. This includes calculating the newly obtained channel matrix H′ and the previously estimated channel matrix. The Frobenius norm of the difference between the features is calculated, and a threshold is set; when the difference exceeds the threshold, the channel state is considered to have changed significantly. Then, the encoder part of the network is fixed, meaning that the encoder parameters are not updated with gradients during denoising. Adopting the idea of ​​deep image prior, only the decoder is used to compress the features z in low dimension. E Channel denoising is performed starting from the current point. The specific steps are as follows:

[0078] Step (3.1): Fix the network parameters of the encoder, including the skip connection layers, to ensure that the training process starts from the same data baseline, namely 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 preserve the feature extraction capability of the AE-DIP model trained in the previous stage and avoid destroying the channel statistics information captured by the encoder during the optimization process.

[0079] Step (3.2): Since the DIP method utilizes the idea that the network structure contains natural image priors—that is, because the network itself has low passivity for natural image information but high impedance to random information such as noise—it generates a clean, noise-free sample during the process of fitting noisy samples. Simultaneously, the signal matrix also possesses natural information similar to that of an image, making signal denoising similar to image denoising. Therefore, the parameters of the decoder network will be trained according to the DIP strategy, selecting the loss function...

[0080]

[0081] In the formula f decoder (Z E Y is the signal matrix output after reconstruction by the decoder. T The feature vector Z compressed by the encoder E It is input to the decoder, and after being restored to the original channel dimension by the decoder, it is combined with Y. T The parameters are optimized by comparing the two and calculating their mean squared errors. This forward and backward propagation process is iterated repeatedly until the decoder's output signal matrix is ​​close to the denoised original signal matrix. Figure 3 The offline training phase first trains the network's overall signal statistical feature compression and signal reconstruction capabilities, using the original signal matrix Y. T The feature vector Z is obtained after compression by the encoder. E The feature vector is then input into the decoder network for reconstruction. During the online training phase, after fixing the encoder and skip connection parameters, the decoder parameters are continuously optimized through forward and backward propagation processes, making the reconstructed signal matrix approach the clean signal matrix, thereby achieving signal denoising. (See figure) This is the loss function mentioned above. D According to the gradient of the loss function Update the parameters for each layer.

[0082] Step (3.3): Train the parameters of the decoder network using an early stopping strategy. Leveraging the characteristics of deep image prior neural networks, the network prioritizes simulating natural information over noise information during iteration; that is, the network exhibits low passivity 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 dimensions, perform LS channel estimation. The specific steps are as follows:

[0084] Step (4.1): The dimensions of the signal matrix after network denoising are represented as follows: Recombine the real and imaginary parts and rearrange their order to obtain the original matrix dimensions.

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

[0086] After processing the current signal matrix, the process moves on to the next received signal matrix and repeats the above process for channel estimation and denoising to ensure the accuracy and stability of channel estimation.

[0087] In summary, the unsupervised MIMO channel estimation method integrating channel statistical characteristics provided by this 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 used 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 using the mean squared error (MSE) loss function, ensuring that the encoder can compress the channel information into a low-dimensional feature vector containing sufficient statistical information, thereby improving the decoder's channel recovery capability. This stage uses an unsupervised learning method, directly using the channel matrix received during communication for training, reducing dependence on large datasets and enhancing the model's flexibility and practicality. In the channel denoising training stage, the training focus shifts to enhancing the decoder's denoising capability. The encoder parameters are fixed to maintain consistency in feature extraction, allowing the decoder to perform denoising training from the same starting point of compressed feature vectors, further optimizing channel recovery. Through repeated forward and backward propagation, the decoder network gradually improves its ability to recover low-dimensional feature vectors, ultimately achieving efficient denoising and reconstruction of the original channel.

[0088] The AE-DIP method of this invention demonstrates excellent channel estimation performance and convergence speed in both single-cell and multi-cell scenarios. For example... Figure 4 As shown, 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. In particular, it achieves performance comparable to MMSE estimation under high signal-to-noise ratio conditions. Figure 5 The performance comparison in a multi-cell interference environment is shown. The AE-DIP method still maintains strong robustness. Even under severe interference, its estimation accuracy is better than the traditional LS method and comparable to the MMSE method.

[0089] Regarding convergence speed, Figure 6 and Figure 7 This shows a comparison between AE-DIP and the traditional DIP method. For example... Figure 6In single-cell scenarios, the AE-DIP method exhibits faster convergence speed, significantly reducing mean square error (MSE) in the initial iteration phase and substantially improving channel recovery efficiency. For example... Figure 7 Even in multi-cell interference environments, 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 system's real-time response capability.

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

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

[0092] The AE-DIP offline training module is used to train the AE-DIP signal denoising network offline. The transformed signal matrix is ​​input 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 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 skip connections.

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

[0094] The AE-DIP online training module is used to train the decoder for signal denoising by fixing the encoder and skip connection parameters during online training and utilizing the statistical information of the signal matrix captured in the encoder's hidden layer and the DIP strategy.

[0095] The channel estimation module is used to perform LS channel estimation after restoring the denoised signal matrix to its original dimensions.

[0096] This invention also discloses a computer system, including a memory, a processor, and a computer program / instructions stored in the memory and executable on the processor. When the computer program / instructions are executed by the processor, they implement the steps of the unsupervised MIMO channel estimation method that integrates channel statistical characteristics.

[0097] This 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 that integrates channel statistical characteristics.

Claims

1. An unsupervised MIMO channel estimation method incorporating channel statistical characteristics, characterized in that, This method constructs a signal denoising network AE-DIP, which utilizes a symmetrically extended autoencoder-like structure of a depth image prior network, combining an encoder, decoder, and skip connections to achieve effective extraction of channel features and signal denoising. The method includes 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: The dynamically received, transformed signal matrix is ​​input 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 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 skip connections. The steps of offline training of the network include: (2.1) Signal matrix Y T Encoder f via network encoder It is compressed into a low-dimensional feature tensor Z. E It contains statistical information about signals captured by the network; Z E Defined by the following formula: In the formula, l represents the number of hidden layers. The hidden layer structure of the encoder's i-th layer is represented as follows: In the formula It is the input of the i-th hidden layer of the encoder. These are network parameters, symbols. The symbol represents the convolution operation, while Pooling, ReLU, and BatchNorm represent pooling, linear rectified activation function, and batch normalization, respectively. (2.2) Convert the low-dimensional feature tensor Z E After passing through the network decoder f decoder The channel representation after reconstructing the original signal matrix and using the decoder network is as follows: In the formula Let the hidden layer structure of the i-th layer of the decoder be represented as: In the formula This represents the input to the i-th layer of the decoder, which contains information from the previous layer and skip connections from the corresponding encoder hidden layer, and is represented in the form of: UpSample indicates an 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 training of the network; (3) After offline training, update the parameters of the overall online AE-DIP network; during online training, fix the encoder and skip connection parameters, and use the statistical information of the signal matrix captured in the encoder's hidden layer to train the decoder for signal denoising using the DIP strategy; the steps of online network training include: (3.1) Fix the network parameters of the encoder, including the skip connection layers therein, to ensure that the training process starts with the same data benchmark, i.e. the compressed feature vector generated by the encoder; (3.2) The parameters of the decoder network are trained according to the DIP strategy, and the loss function is: In the formula f decoder (Z E () represents the signal matrix output after reconstruction by the decoder. Signal matrix Y T The feature vector Z compressed by the encoder E It is input into the decoder, and after being restored by the decoder, it is compared with Y. T The parameters are optimized by comparing the mean square error of the two signals and iterating this forward and backward propagation process repeatedly until the signal matrix output by the decoder is close to the original signal matrix after denoising. (3.3) Train the parameters of the decoder network using the early stopping strategy; (4) After restoring the denoised signal matrix to its original dimensions, perform LS channel estimation.

2. The unsupervised MIMO channel estimation method based on fused 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 input the received signal. Transformation Dimensions Where N f N represents the number of subcarriers. p M represents the number of OFDM symbols and M represents the number of antennas. (1.2) The signal from the previous step The real part Y real =Re(Y) and the imaginary part Y imag =Im(Y) are separated and superimposed to form a new matrix. As input to a neural network.

3. The unsupervised MIMO channel estimation method based on fused 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 follows: Recombine the real and imaginary parts and rearrange their order to obtain the original matrix dimensions. (4.2) Select the first few OFDM symbols of the received signal as pilot signals according to the channel settings, extract the pilot matrix X, and use the known pilot matrix X as the transmission signal for subsequent channel estimation. The estimated value of the channel matrix is...

4. A deep learning-based channel estimation system for unsupervised MIMO that integrates channel statistical characteristics, characterized in that, This system constructs a signal denoising network AE-DIP, which utilizes a symmetrically extended autoencoder-like structure of a depth image prior network, combined with an encoder, decoder, and skip connections, to achieve effective extraction of channel features and signal denoising. It includes the following modules: The preprocessing module is used to convert the received signal matrix into a format suitable for neural network input; The AE-DIP offline training module is used to train the AE-DIP signal denoising network offline. It inputs 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 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 skip connections. The steps involved in offline network training include: Matrix Y T Encoder f via network encoder It is compressed into a low-dimensional feature tensor Z. E It contains statistical information about signals captured by the network; Z E Defined by the following formula: In the formula, l represents the number of hidden layers. The hidden layer structure of the encoder's i-th layer is represented as follows: In the formula It is the input of the i-th hidden layer of the encoder. These are network parameters, symbols. The symbol represents the convolution operation, while Pooling, ReLU, and BatchNorm represent pooling, linear rectified activation function, and batch normalization, respectively. The low-dimensional feature tensor Z E After passing through the network decoder f decoder The channel representation after reconstructing the original signal matrix and using the decoder network is as follows: In the formula Let the hidden layer structure of the i-th layer of the decoder be represented as: In the formula This represents the input to the i-th layer of the decoder, which contains information from the previous layer and skip connections from the corresponding encoder hidden layer, and is represented in the form of: UpSample indicates an 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 training of the network. The network update module is used to update the parameters of the online AE-DIP network after offline training. The AE-DIP online training module is used to train the decoder for signal denoising by fixing the encoder and skip connection parameters during online training and utilizing the statistical information of the signal matrix captured in the encoder's hidden layer, using the DIP strategy. The steps of online network training include: The network parameters of the encoder are fixed, including the skip connection layers, to ensure that the training process begins with the same data benchmark, namely the compressed feature vector generated by the encoder. The parameters of the decoder network are trained according to the DIP strategy, and the loss function is: In the formula f decoder (Z E () represents the signal matrix output after reconstruction by the decoder. Signal matrix Y T The feature vector Z compressed by the encoder E It is input into the decoder, and after being restored by the decoder, it is compared with Y. T The parameters are optimized by comparing the mean square error of the two signals and iterating this forward and backward propagation process repeatedly 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; The channel estimation module is used to perform LS channel estimation after restoring the denoised signal matrix to its original dimensions.

5. A computer system, comprising a memory, a processor, and computer programs / instructions stored in the memory and executable on the processor, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the unsupervised MIMO channel estimation method based on the fusion of channel statistical characteristics as described in any one of claims 1-3.

6. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the unsupervised MIMO channel estimation method based on the fusion of channel statistical characteristics as described in any one of claims 1-3.

Images