Self-supervised learning algorithm based on OCT convolutional network

Abstract

The invention discloses a self-supervised learning algorithm based on an OCT (Optical Coherence Tomography) convolutional network, which is used for channel estimation assisted by an intelligent reflecting surface in a mine environment, and comprises the steps of constructing an IRS (Inter-Reference Signal) system model under a mine, constructing an OCT convolutional neural network model and performing self-supervised learning. And challenges brought by multipath attenuation and non-line-of-sight communication in a complex mine environment are solved. The performance of the SSL method is close to that of a traditional supervised learning method under different signal-to-noise ratio conditions, and efficient channel estimation can still be achieved under the condition that real label data is not used.

Description

Technical Field

[0001] The present invention relates to the technical field, and in particular to a self-supervised learning algorithm based on an OCT convolutional network. Background Art

[0002] Coal is still the mainstay of my country's energy consumption, accounting for two-thirds. This pattern is unlikely to change in the foreseeable future. At present, the informatization construction of my country's coal industry is still in its infancy, especially underground wireless communication technology is particularly backward compared to ground technology. my country's coal reserves are generally buried deep underground, and mining mainly relies on underground operations. The intricate tunnels and chambers in the mine make the propagation characteristics of electromagnetic waves in the mine very different from those in free space, and refraction, reflection, diffraction and other phenomena occur frequently. Coupled with the presence of miners and equipment, multipath attenuation and non-line-of-sight communication problems are particularly prominent. These factors together lead to large propagation losses and short communication distances of electromagnetic waves in mines, which have become the main obstacles to the effective implementation of underground wireless communications.

[0003] Channel estimation is the core link of communication technology, and its performance is directly related to the quality of the entire communication system. However, the unpredictable environment in the mine greatly increases the difficulty of channel estimation, and building a perfect wireless communication system that adapts to underground mines is facing huge challenges. Therefore, it is urgent to break the thinking constraints of traditional mine wireless communication technology research, use emerging wireless transmission theories and technical means, and innovatively solve the above problems to improve the efficiency and safety of underground communications.

[0004] In recent years, scientific research institutions and the communications industry around the world have actively engaged in the exploration and research of the sixth generation of mobile communication technology, which has become a research hotspot in this field. Against this background, a cutting-edge wireless communication technology called intelligent reflective surface (IRS) has emerged and quickly become the focus of this field.

[0005] IRS is composed of many passive reflective units. Through precise digital control, it can adjust the amplitude and / or phase of the incident signal, thereby actively shaping the wireless channel between the transmitter and the receiver. This breakthrough technology gives people the ability to change the wireless propagation environment, such as helping wireless signals bypass obstacles, which is not available in traditional wireless transmission technology.

[0006] IRS technology not only expands signal coverage, establishes virtual line-of-sight communication links, and overcomes propagation barriers, but also supports ubiquitous connectivity and enables ultra-high-speed data transmission. Compared with traditional active relay technology, IRS significantly reduces costs and energy consumption through full-duplex passive beamforming reflection, without the need for active RF links and self-interference elimination. Its practical advantages include easy deployment, low cost, sustainable operation, and high flexibility. It consists of metasurface passive scattering elements and is suitable for a variety of surfaces such as building exteriors, indoor walls, and ceilings. IRS achieves zero-power signal processing and promotes wireless communications to develop in a green and sustainable direction.

[0007] In an IRS-assisted communication system with sparse channels, cascade channel estimation can be converted into a sparse signal recovery problem, which can be effectively solved by compressed sensing methods. Due to the sparsity of the public IRS-BS base station (BS) channel, the cascade channel matrix of all users has a common row-column-block sparsity structure. IRS needs to obtain perfect channel state information (CSI) to fully exert its performance. Advanced methods such as machine learning can better optimize the transmission and reflection beamforming in IRS-assisted wireless systems. Unlike traditional wireless systems, IRS, as a passive device, cannot send, receive and process pilot signals, and the channel involving IRS is difficult to estimate. Therefore, directly estimating the cascaded user IRS-base station channel is a typical method.

[0008] In IRS systems, the dimension of the cascaded channel is often large, which leads to a large amount of pilot overhead. In order to solve these problems, many related studies have explored efficient channel estimation methods. For example, the estimation method based on the least squares (LS) method. The goal of the LS estimator is to obtain an unbiased cascaded channel in a single-user multiple-input single-output (SU-MISO) system.

[0009] Deep Learning (DL) provides a compromise between low complexity and high performance in channel estimation. Compared with traditional methods, DL can learn the mapping from input to output and find the local optimal solution. Its main advantages include: 1) only simple matrix multiplication is required during inference, without complex iterative optimization; 2) it has generalization ability, and the trained neural network can also perform well on unseen samples. This makes it possible to treat the channel estimation problem as a supervised learning problem. Some DL-based methods have shown excellent performance in the IRS channel estimation task. Synthetic Deep Neural Network (DNN) is used to estimate the cascade channel in real time to reduce training overhead. Multi-user cascade channel estimation is first described as a denoising problem and then solved under the deep residual learning framework based on Convolutional Neural Network (CNN) to refine the channel coefficients estimated from noise-based pilot observations. The wireless communication model is combined with deep learning to achieve channel estimation for IRS networks. The traditional multi-stage channel estimation method is used to construct a dataset to train the neural network.

[0010] Deep learning has obvious performance advantages in channel estimation, but it all relies on supervised learning to train neural networks. One problem is how to obtain the labels required for supervised learning in practice. If the labels are obtained through traditional channel estimation algorithms, the performance of supervised learning methods will be limited by traditional algorithms. End-to-end training of neural networks requires the use of real channel labels to calculate the loss function to update the neural network. However, in the real world, it is very difficult to obtain the real channel. Therefore, there is an urgent need for an effective and practical method that can utilize the excellent characteristics of deep learning (generalization ability, competitive online reasoning ability) without relying on real labels.

[0011] To solve this problem, the present application provides a self-supervised learning algorithm based on OCT convolutional network for channel estimation assisted by intelligent reflective surfaces in mine environments, so as to achieve efficient channel estimation in mine environments without the need for real labeled data. Summary of the invention

[0012] The purpose of the present invention is to provide a self-supervised learning algorithm based on OCT convolutional network to solve the problems of the prior art.

[0013] To achieve the above purpose, the present invention adopts the following technical solutions:

[0014] A self-supervised learning algorithm based on OCT convolutional network includes the following steps:

[0015] Step 1: Construct an IRS system model in the mine;

[0016] The Nakagami-g channel fading model is used to build a channel estimation simulation environment in a mine.

[0017] Nakagami distribution is a channel statistical model that describes the different characteristics of multipath signals by changing the value of g. The Rice fading channel model and the Rayleigh fading channel model are both obtained by changing the value of the Nakagami distribution parameter g. Its probability density function is:

[0018]

[0019] Among them, Γ(g) is the Gamma function, Ω represents the average power; a is the amplitude of the received multipath signal; g≥1 / 2, which represents the shape factor of the fading degree. The larger the value of g, the smaller the corresponding channel fading degree. By changing the value of g, different fading statistical channel models can be obtained. The Nakagami distribution can often match more experimental data in the actual wireless channel environment modeling, and thus can simulate the channel model under different communication environments. The channel model is used to simulate the communication system in the confined space of the mine and tunnel environment.

[0020] Install IRS on the tunnel wall of the mine, the signal can be propagated through the reflection of IRS, and can provide services to K single-antenna users at the same time when the user considers that BS and IRS use m antennas and n-element uniform planar arrays respectively;

[0021] Assume that the G matrix of size M×N represents the channel from IRS to BS, and the h matrix of size N×1 represents r,k represents the channel from the kth user to the IRS (k = 1, 2, ..., k), and the channel model G is:

[0022]

[0023] Among them, L G represents the number of paths between IRS and BS, and represents the path loss, the azimuth at the BS, and the azimuth at the IRS;

[0024] The channel between IRS and the user is expressed as:

[0025]

[0026] Among them, L r,k represents the number of paths between the kth user and the IRS, and represents the path gain and the azimuth of the IRS;

[0027] The M×N cascade channel of the kth user is:

[0028]

[0029] Assume that in T consecutive time slots, the transmitted signals of M terminals are represented as X∈C M×T , IRS has a total of N reflection units, the total number of available IRS phase configurations is P, forming the IRS phase configuration matrix Φ∈C P×N , then the received signal Y based on the pth (p=1,2,…,P)th IRS phase configuration p ∈￡ K×T Briefly expressed as:

[0030]

[0031] Among them, [Φ] p,: represents the p-th IRS phase configuration, corresponding to the p-th row of the IRS phase configuration matrix Φ, represents the channel matrix between IRS and base station, H s ∈￡ N×M represents the channel matrix between the terminal and the IRS, is additive complex Gaussian white noise;

[0032] Step 2: Build an OCT convolutional neural network model;

[0033] The OCT neural network consists of two sub-network modules and Each sub-network module is followed by a residual module and Each sub-network module contains N L convolutional layers, each using ReLu and a batch normalization;

[0034] The network consists of two sub-networks with skip connections, and the hidden layer receives the real and imaginary parts of the original wireless channel to enhance the information flow and alleviate the gradient vanishing problem;

[0035] The residual module uses the tanh activation function without BN, decomposing the complex matrix A into two channels, namely the real part and the imaginary part as input: X = [Re{A}, Im{A}]. In the training phase, the input is the received signal R; in the test phase, the input is the least squares channel estimation. In order to make full use of the global information of the original matrix A, the channel is linearly transformed, and the real and imaginary parts of X are exchanged as XU = [Im{A}, Re{A}], and averaged with the output of the first residual module, that is:

[0036]

[0037] Subtract the output of the second residual module from the original input to get the final output, which is The element-wise subtraction method is used for denoising. The scale of the above neural network is flexible by setting a small N L or large N L It is possible to obtain shallow or deep neural networks, which only use convolutional layers and no fully connected layers, with very few parameters and low storage overhead for the entire network.

[0038]

[0039] Step 3: Self-supervised learning;

[0040] The specific process of self-supervised learning is divided into two stages: training stage and testing stage. The specific steps of the training stage are:

[0041] Step 1: Input label definition. The training input is composed of the actual received signal and the added Gaussian white noise. The corresponding label is the original received signal without interference.

[0042] Step 2: Optimize the strategy by using the stochastic gradient descent method and combining iterative updates of network parameters with an appropriate learning rate;

[0043] Step 3: Iterative process: the entire training process is divided into multiple training cycles, and each cycle processes a certain amount of small batch data;

[0044] Step 4: Data perturbation: In each training iteration, new noise is injected into the input data to enhance the generalization and robustness of the model.

[0045] Step 5: Loss function. The mean square error is usually used as the loss function to evaluate the deviation between the model output and the label.

[0046] The task of the neural network in the training phase is to recover the clean signal from the noisy signal. Specifically, the optimization goal during the training process is to minimize the loss between the predicted signal and the true signal, expressed as:

[0047] arg w minE[L(F(X,w),H)];

[0048] Where L(·) is the loss function, and the network function g2 and h2 are two submodules in the network, corresponding to different parameters and

[0049] The precondition of the self-supervised loss function is denoising, i.e. cleaning the noisy version X of the received signal i +ε i , the expression is:

[0050]

[0051] The specific steps of the testing phase are:

[0052] Step 1: Model solidification: after training is completed, the model parameters are fixed and no longer updated;

[0053] Step 2: Input processing: the input data is the preliminary channel estimation result obtained by the least squares method;

[0054] Step 3: Iterative process, the output of the neural network is regarded as a further refinement and optimization of the LS estimation result, so as to obtain the final channel estimation result;

[0055] Using the trained network, the rough estimation result of the least squares estimator is used as input in the test phase, where the LS estimate is H LS =RP H (PP H ) -1 ;

[0056] The input of the training phase is the received signal with additional noise disturbance, and the output is the neural network's estimate of the original received signal. The input of the testing phase is the LS estimator's rough estimate of the channel, and the output is the neural network's optimized estimate of the IRS channel.

[0057] Compared with the prior art, the present invention has the following advantages: The present invention proposes a self-supervised learning algorithm based on an octave convolutional network for channel estimation assisted by intelligent reflective surfaces in a mine environment. The method solves the challenges of multipath fading and non-line-of-sight communication in a complex mine environment by constructing a channel estimation system that does not rely on real label data. The performance of the SSL method under different signal-to-noise ratio conditions is close to that of traditional supervised learning methods, and efficient channel estimation can still be achieved without using real label data. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a signal transmission model diagram of the IRS-assisted mine communication system of the present invention (a is an obstacle blocking signal transmission, and b is the IRS reflected signal for transmission);

[0059] Figure 2 This is a network architecture diagram of the present invention;

[0060] Figure 3 Schematic diagram of the OCT convolution module of the present invention;

[0061] Figure 4 Schematic diagram of the OCT convolution module of the present invention;

[0062] Figure 5 The loss comparison diagram of the two algorithms of the present invention at different signal-to-noise ratios (a is the loss of SL and SSL when SNR=-5dB, and b is the loss of SL and SSL when SNR=0dB);

[0063] Figure 6 It is a comparison diagram of NMSE of two algorithms with different signal-to-noise ratios of the present invention (a is the NMSE of SL and SSL when SNR=0dB, and b is the NMSE of SL and SSL when SNR=0dB);

[0064] Figure 7 This is a NMSE comparison chart of different methods of the present invention. DETAILED DESCRIPTION

[0065] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0066] IRS system model in mines;

[0067] Aiming at the complex environment of mines, IRS-assisted channel estimation in mines is proposed. IRS technology assists channel estimation and can improve the accuracy of channel estimation. The Nakagami-g channel fading model is used to build a simulation environment for channel estimation in mines.

[0068] Nakagami distribution is a relatively flexible channel statistical model, which can describe the different characteristics of multipath signals by changing the value of g. Both the Rice fading channel model and the Rayleigh fading channel model can be obtained by changing the value of the Nakagami distribution parameter g. Its probability density function is:

[0069]

[0070] Among them, Γ(g) is the Gamma function, Ω represents the average power; a is the amplitude of the received multipath signal; g≥1 / 2, which represents the shape factor of the fading degree. The larger the value of g, the smaller the corresponding channel fading degree. Because different fading statistical channel models can be obtained by simply changing the value of g, the Nakagami distribution can often match more experimental data in the modeling of actual wireless channel environments, and thus can better simulate the channel models under different communication environments. The channel model was used to simulate the communication system in confined spaces such as mines and tunnel environments.

[0071] Due to the complex mine environment, electromagnetic waves are not only affected by their own fading during propagation, but also by reflection, refraction and scattering from the tunnel walls, which can easily lead to multipath fading, signal distortion, inter-code interference and other problems during electromagnetic wave propagation. Figure 1As shown in the figure, there is an obstacle between the user and the base station LOS (Line-Of-Sight) link, so the signal cannot propagate. Install IRS on the tunnel wall of the mine, and the signal can propagate through the reflection of IRS. Consider that the BS and IRS use m antennas and n-element uniform planar arrays respectively to provide services to K single-antenna users at the same time.

[0072] Assume that the G matrix of size M×N represents the channel from IRS to BS, and the h matrix of size N×1 represents r,k represents the channel from the kth user to the IRS (k = 1, 2, ..., k), and the channel model G is:

[0073]

[0074] Among them, L G represents the number of paths between IRS and BS, and represents the path loss, the azimuth (elevation) at the BS, and the azimuth (elevation) at the IRS.

[0075] The channel between IRS and the user is expressed as:

[0076]

[0077] Among them, L r,k represents the number of paths between the kth user and the IRS, and Indicates the path gain and the azimuth (elevation) angle of the IRS.

[0078] Furthermore, the M×N cascade channel representing the kth user is:

[0079]

[0080] Assume that in T consecutive time slots, the transmitted signals of M terminals are represented as X∈C M×T , IRS has a total of N reflection units, the total number of available IRS phase configurations is P, forming the IRS phase configuration matrix Φ∈C P×N , then the received signal Y based on the pth (p=1,2,…,P)th IRS phase configuration p ∈￡ K×T It can be briefly expressed as:

[0081]

[0082] Among them, [Φ] p,: represents the p-th IRS phase configuration, corresponding to the p-th row of the IRS phase configuration matrix Φ, H r ∈￡ K×Nrepresents the channel matrix between IRS and base station, H s ∈￡ N×M represents the channel matrix between the terminal and IRS, W p ∈￡ K×T is additive complex white Gaussian noise.

[0083] OCT convolutional neural network model;

[0084] The OCT neural network consists of two sub-network modules and Each sub-network module is followed by a residual module and Each sub-network module contains N L Convolutional layers, each using ReLu and a batch normalization (BN).

[0085] The network consists of two sub-networks with skip connections, and the hidden layer receives the real and imaginary parts of the original wireless channel to enhance the information flow and alleviate the gradient vanishing problem. Figure 2 shown.

[0086] The residual module uses the tanh activation function and does not contain BN. The complex matrix A is decomposed into two channels, namely the real part and the imaginary part as input: X = [Re{A}, Im{A}]. In the training phase, the input is the received signal R; in the test phase, the input is the least squares channel estimate. In order to make full use of the global information of the original matrix A, the channel is linearly transformed, the real part and the imaginary part of X are exchanged as XU = [Im{A}, Re{A}], and it is averaged with the output of the first residual module. That is:

[0087]

[0088] Subtract the output of the second residual module from the original input to get the final output, which is The denoising is done by element-wise subtraction. The scale of the above neural network is flexible by setting a small N L or large N L You can get shallow or deep neural networks. The neural network only uses convolutional layers, no fully connected layers, and very few parameters, so the storage overhead of the entire network is also very small.

[0089]

[0090]

[0091] OCT convolution is a convolution operation that directly acts on the feature expression. Its core idea is to directly process high-frequency components and low-frequency components separately. Figure 3 As shown, it includes the update of each frequency's own state (dotted arrow) and the information interaction between frequencies (straight arrow).

[0092] OCT convolution can capture both the rough features and subtle differences of the channel, thus providing more comprehensive channel information and helping to estimate the channel state information more accurately. The methods are orthogonal and complementary. This approach can reduce the computational burden of the network while improving the performance of the network. Through the effective combination of high-frequency and low-frequency components, OCT convolution helps improve the accuracy of the estimation. The low-frequency components can capture the main trends and changes of the channel, while the high-frequency components can depict details such as edges and mutations, which may be lost in traditional convolution.

[0093] Self-supervised learning;

[0094] The specific process of self-supervised learning is divided into two stages: training stage and testing stage. The specific steps of the training stage are:

[0095] Step 1: Input label definition. The training input is composed of the actual received signal and the added Gaussian white noise. The corresponding label is the original received signal without interference.

[0096] Step 2: Optimize the strategy by using the stochastic gradient descent method and combining iterative updates of network parameters with an appropriate learning rate;

[0097] Step 3: Iterative process: the entire training process is divided into multiple training cycles, and each cycle processes a certain amount of small batch data;

[0098] Step 4: Data perturbation: In each training iteration, new noise is injected into the input data to enhance the generalization and robustness of the model.

[0099] Step 5: Loss function. The mean square error is usually used as the loss function to evaluate the deviation between the model output and the label.

[0100] The task of the neural network in the training phase is to recover the clean signal from the noisy signal. Specifically, the optimization goal during the training process is to minimize the loss between the predicted signal and the true signal, expressed as:

[0101] arg w minE[L(F(X,w),H)];

[0102] Where L(·) is the loss function, and the network function Here g2 and h2 are two submodules in the network, corresponding to different parameters and

[0103] The precondition of the self-supervised loss function is denoising, i.e. cleaning the noisy version X of the received signal i +ε i , the expression is:

[0104]

[0105] The specific steps of the testing phase are:

[0106] Step 1: Model solidification: after training is completed, the model parameters are fixed and no longer updated;

[0107] Step 2: Input processing: the input data is the preliminary channel estimation result obtained by the least squares method;

[0108] Step 3: Iterative process, the output of the neural network is regarded as a further refinement and optimization of the LS estimation result, so as to obtain the final channel estimation result;

[0109] Using the trained network, the rough estimation result of the least squares estimator is used as input in the test phase, where the LS estimate is H LS =RP H (PP H ) -1 ;

[0110] The input of the training phase is the received signal with additional noise disturbance, and the output is the neural network's estimate of the original received signal. The input of the testing phase is the LS estimator's rough estimate of the channel, and the output is the neural network's optimized estimate of the IRS channel.

[0111] Using the trained network, the rough estimation result of the least squares estimator is used as input in the test phase, where the LS estimate is H LS =RP H (PP H ) -1 .

[0112] Although the training and testing phases are similar in approach, the data types and objectives are fundamentally different. The training phase aims to recover the original signal from the noise-affected data, while the testing phase uses the trained model to optimize the channel parameters from preliminary estimates, demonstrating the potential of deep learning in an unsupervised setting.

[0113] The input of the training phase is the received signal with additional noise perturbation, such as Figure 4 As shown, the output is the neural network's estimate of the original received signal (indicated by the blue line). The input of the test phase is the LS estimator's rough estimate of the channel, and the output is the neural network's optimized estimate of the IRS channel (indicated by the red line).

[0114] Experimental simulation;

[0115] The experiment was simulated using MATLAB and Python. An IRS-assisted mine channel estimation system model was built in MATLAB 2022a, and training and test data sets were generated. The experimental settings are as follows: the number of IRS reflection units is 64, the number of base station antennas is 16, the number of users is 6, and the channel model is Nakagami-m channel.

[0116] First, a data set was generated using an IRS system in a mine under the conditions of -5dB and 0dB signal-to-noise ratio. Then the neural network was trained for 100 cycles. Figure 5 As shown in Figure 3, whether it is SL or SSL, as the training cycle increases, the test loss gradually decreases and tends to stabilize after reaching a certain range.

[0117] The purpose of the experiment is to perform channel estimation without relying on real labeled data. To this end, we compared the performance of SL and SSL in terms of Normalized Mean Square Error (NMSE). Under the conditions of SNR of 0dB and 5dB, we used the same neural network structure, learning rate scheduler and optimizer for channel estimation, and set the experimental cycle to 100 times.

[0118] Specifically, the input of the SL method in the training phase is the received signal, the output is the channel estimated by the neural network, and the loss is calculated using the real channel as the label; in the test phase, the input is the received signal. The SSL method is trained with noisy received signals and does not rely on real label data.

[0119] like Figure 6 As shown in the figure, under the conditions of SNR of 0dB and 5dB, as the experimental period increases, the NMSE performance of SL and SSL methods is very close, and SSL is slightly smaller than SL. This shows that without using the true label, the SSL method can achieve channel estimation performance comparable to the SL method.

[0120] To demonstrate the advantages of SSL, we also compare it with other methods.

[0121] like Figure 7 As shown in Figure 1, the comparison of NMSE performance among LS, LMMSE, SL and SSL is shown. The figure shows that SL and SSL methods perform well under various SNR conditions, especially under high SNR conditions, where their NMSE is significantly lower than that of traditional LS and LMMSE methods.

[0122] SSL can still achieve performance close to SL without relying on real labels. SL and SSL perform well in channel estimation, especially under high SNR conditions. This shows that the SSL method has important practical value in practical applications, especially in environments where it is difficult to obtain real labels.

[0123] Aiming at the multipath attenuation and non-line-of-sight communication problems in mines, a self-supervised learning (SSL) algorithm based on Octave Convolution (OCT) convolutional network was proposed. The algorithm trains the neural network by adding noise without the need of real labeled data.

[0124] The NMSE performance of the SSL method under different signal-to-noise ratio conditions is close to that of the SL method, and the test loss gradually decreases and tends to be stable with the increase of training cycles.

[0125] The above is a preferred embodiment of the present invention. For ordinary technicians in this field, according to the teachings of the present invention, without departing from the principles and spirit of the present invention, changes, modifications, substitutions and variations made to the implementation methods are still within the scope of protection of the present invention.

Claims

1. A self-supervised learning algorithm based on OCT convolutional network, characterized in that: The steps include: Step 1: Construct an IRS system model in the mine; The Nakagami-g channel fading model is used to build a channel estimation simulation environment in a mine. Nakagami distribution is a channel statistical model that describes the different characteristics of multipath signals by changing the value of g. The Rice fading channel model and the Rayleigh fading channel model are both obtained by changing the value of the Nakagami distribution parameter g. Its probability density function is: Among them, Γ(g) is the Gamma function, Ω represents the average power; a is the amplitude of the received multipath signal; g≥1 / 2, which represents the shape factor of the fading degree. The larger the value of g, the smaller the corresponding channel fading degree. By changing the value of g, different fading statistical channel models can be obtained. The Nakagami distribution can often match more experimental data in the actual wireless channel environment modeling, and thus can simulate the channel model under different communication environments. The channel model is used to simulate the communication system in the confined space of the mine and tunnel environment. Install IRS on the tunnel wall of the mine, the signal can be propagated through the reflection of IRS, and can provide services to K single-antenna users at the same time when the user considers that BS and IRS use m antennas and n-element uniform planar arrays respectively; Assume that the G matrix of size M×N represents the channel from IRS to BS, and the h matrix of size N×1 represents r,k represents the channel from the kth user to the IRS (k = 1, 2, ..., k), and the channel model G is: Among them, L G represents the number of paths between IRS and BS, and represents the path loss, the azimuth at the BS, and the azimuth at the IRS; The channel between IRS and the user is expressed as: Among them, L r,k represents the number of paths between the kth user and the IRS, and represents the path gain and the azimuth of the IRS; The M×N cascade channel of the kth user is: Assume that in T consecutive time slots, the transmitted signals of M terminals are represented as X∈C M×T , IRS has a total of N reflection units, the total number of available IRS phase configurations is P, forming the IRS phase configuration matrix Φ∈C P×N , then the received signal Y based on the pth (p=1,2,…,P)th IRS phase configuration p ∈￡ K×T Briefly expressed as: Among them, [Φ] p,: represents the p-th IRS phase configuration, corresponding to the p-th row of the IRS phase configuration matrix Φ, H r ∈￡ K×N represents the channel matrix between IRS and base station, H s ∈￡ N×M represents the channel matrix between the terminal and IRS, W p ∈￡ K×T is additive complex Gaussian white noise; Step 2: Build an OCT convolutional neural network model; The OCT neural network consists of two sub-network modules and Each sub-network module is followed by a residual module and Each sub-network module contains N L convolutional layers, each using ReLu and a batch normalization; the network consists of skip connections between two sub-networks, and the hidden layer receives the real and imaginary parts of the original wireless channel to enhance the information flow and alleviate the gradient vanishing problem; The residual module uses the tanh activation function without BN, decomposing the complex matrix A into two channels, namely the real part and the imaginary part as input: X = [Re{A}, Im{A}]. In the training phase, the input is the received signal R; in the test phase, the input is the least squares channel estimation. In order to make full use of the global information of the original matrix A, the channel is linearly transformed, and the real and imaginary parts of X are exchanged as XU = [Im{A}, Re{A}], and averaged with the output of the first residual module, that is: Subtract the output of the second residual module from the original input to get the final output, which is The element-wise subtraction method is used for denoising. The scale of the above neural network is flexible by setting a small N L or large N L It is possible to obtain shallow or deep neural networks, which only use convolutional layers and no fully connected layers, with very few parameters and low storage overhead for the entire network. Step 3: Self-supervised learning; The specific process of self-supervised learning is divided into two stages: training stage and testing stage.

2. A self-supervised learning algorithm based on OCT convolutional network as claimed in claim 1, characterized in that: The specific steps of the training phase are: Step 1: Input label definition. The training input is composed of the actual received signal and the added Gaussian white noise. The corresponding label is the original received signal without interference. Step 2: Optimize the strategy by using the stochastic gradient descent method and combining iterative updates of network parameters with an appropriate learning rate; Step 3: Iterative process: the entire training process is divided into multiple training cycles, and each cycle processes a certain amount of small batch data; Step 4: Data perturbation: In each training iteration, new noise is injected into the input data to enhance the generalization and robustness of the model. Step 5: Loss function. The mean square error is usually used as the loss function to evaluate the deviation between the model output and the label. The task of the neural network in the training phase is to recover the clean signal from the noisy signal. Specifically, the optimization goal during the training process is to minimize the loss between the predicted signal and the true signal, expressed as: arg w minE[L(F(X,w),H)]; Where L(·) is the loss function, and the network function g2 and h2 are two submodules in the network, corresponding to different parameters and The precondition of the self-supervised loss function is denoising, i.e. cleaning the noisy version X of the received signal i +ε i , the expression is:

3. A self-supervised learning algorithm based on OCT convolutional network as claimed in claim 1, characterized in that: The specific steps of the testing phase are: Step 1: Model solidification: after training is completed, the model parameters are fixed and no longer updated; Step 2: Input processing: the input data is the preliminary channel estimation result obtained by the least squares method; Step 3: Iterative process, the output of the neural network is regarded as a further refinement and optimization of the LS estimation result, so as to obtain the final channel estimation result; Using the trained network, the rough estimation result of the least squares estimator is used as input in the test phase, where the LS estimate is H LS =RP H (PP H ) -1 ; The input of the training phase is the received signal with additional noise disturbance, and the output is the neural network's estimate of the original received signal. The input of the testing phase is the LS estimator's rough estimate of the channel, and the output is the neural network's optimized estimate of the IRS channel.

Images