Cellular-free MIMO low-complexity signal detection method based on deep Bayesian inference

Abstract

The invention discloses a cellular-free MIMO (Multiple Input Multiple Output) low-complexity signal detection method based on deep Bayesian inference, which adopts complex Gaussian distribution as approximation of prior probability, and takes Gaussian distribution parameters as learnable parameters to be incorporated into a network training process, thereby dynamically adapting to the requirements of different network levels, and improving the detection accuracy of the cellular-free MIMO low-complexity signal. The convergence problem caused by improper initial distribution selection in a traditional method is solved. Besides, the VID-Net determines and deploys the optimal weight to each AP through a pre-trained deep learning model, so that the operation from calculation to simple summation at a CPU end is simplified, the calculation complexity is remarkably reduced, and the adaptability to various channel conditions is enhanced. And the VID-Net is also integrated with a logarithmic function and a related function correction strategy. Under different modulation orders and system antenna configurations, the VID-Net shows excellent detection performance and robustness, and the performance of the VID-Net exceeds that of a suboptimal minimum mean square error detection algorithm.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless transmission, and in particular to a low-complexity collaborative signal detection method for a non-cellular large-scale MIMO system based on deep unfolded Bayesian inference. Background Art

[0002] In the context of the sixth generation of mobile communication (6G) technology, the non-cellular large-scale multiple-input multiple-output (MIMO) system has formed a new user-centric network architecture by integrating the advantages of distributed antenna system (DAS) and MIMO technology. A large number of distributed access points (AP) serve a small number of terminal users together, effectively avoiding the problem of interference between cells in traditional cellular networks and significantly improving the spectrum efficiency of the system. However, due to the distributed deployment of access points AP, compared with the centralized antenna system, the channel hardening characteristics of the system are weakened, so that the channel statistical characteristics are no longer stable, and the low-complexity linear detection scheme designed for the centralized antenna system cannot achieve satisfactory detection performance. In addition, large-scale MIMO makes the computational overhead of detection schemes with too high complexity unbearable. Therefore, the present invention aims to explore and design an efficient signal detection scheme suitable for non-cellular large-scale MIMO systems, designed from the perspective of distributed access point AP collaborative detection, in order to achieve a good balance between the complexity and performance of the detection algorithm.

[0003] Traditional signal detection algorithms, such as maximum likelihood detection (ML), linear detection, and approximate detection algorithms, are widely used in cell-free massive MIMO systems, but they also face challenges and limitations.

[0004] The computational complexity of the maximum likelihood detection algorithm grows exponentially with the increase in the number of antennas and the number of users. This makes the maximum likelihood detection algorithm have a huge computational overhead in practical applications, limiting its feasibility in large-scale MIMO systems. In order to balance complexity and performance, low-complexity suboptimal data detection methods such as linear minimum mean square error (MMSE) and zero forcing (ZF) can be used. However, they involve complex large-scale matrix inversion operations, which poses a great challenge to efficient implementation. In order to accelerate convergence and balance complexity and bit error rate, scholars have proposed Gauss-Seidel (GS) and Successive Over Relaxation (SOR) detectors. In addition, Lanczos (LA) method, Jacobi iteration, Coordinate Descent (CD), Steepest Descent, Conjugate Gradient (CG) and other algorithms have been used for low-complexity signal detection. However, although the above algorithms can approximate the detection performance of MMSE algorithm with second-order computational overhead, they are heavily dependent on channel hardening. Since the channel hardening characteristics are not significant in non-cellular massive MIMO systems, the performance of such algorithms is seriously degraded.

[0005] The Expectation Propagation (EP) algorithm and the Variational Bayesian Inference (VBI) algorithm based on Bayesian theory infer the estimated value of the transmitted symbol by approximating the posterior probability density of the transmitted symbol, achieving a near-optimal effect for Rayleigh fading channels. In addition, the Approximate Message Passing (AMP) algorithm based on the Approximate Message Passing theory has excellent performance and moderate complexity. The detector based on AMP uses the central limit theorem and Taylor expansion to approximate the posterior probability on the dense factor graph. When the elements of the channel matrix have an independent and identically distributed relationship, the Bayesian optimal performance can be achieved in a large-scale antenna system. However, for actual small-size MIMO systems, the performance of these iterative detectors still does not reach the Bayesian optimal solution, and the performance is seriously deteriorated when facing incomplete channel state information. However, this type of Bayesian detection algorithm is different from some linear algorithms for approximate inversion in that they rely on the statistical characteristics of the channel and the signal during signal transmission, and do not need to rely on the hardening characteristics of the channel. Therefore, the signal detection algorithm based on Bayesian theory has great potential in distributed antenna systems.

[0006] Traditional Bayesian inference faces the challenges of complex probability models and high-dimensional parameter space when calculating the posterior distribution of unknown parameters, and it is often not feasible to accurately calculate the posterior distribution. To solve this problem, it introduces an approximate posterior distribution to approximate the true posterior distribution. The basic idea of ​​variational Bayesian inference VBI is to approximate the true posterior distribution by introducing a parameterized approximate distribution. The parameters of the approximate distribution are determined by minimizing the difference with the true posterior distribution. This process can be achieved by maximizing the posterior probability, so that the approximate distribution is as close to the true posterior distribution as possible.

[0007] The information disclosed in this background technology section is only intended to enhance the understanding of the overall background of the invention and should not be regarded as an acknowledgement or any form of suggestion that the information constitutes the prior art already known to ordinary technicians in this field. Summary of the invention

[0008] The technical problem to be solved by the present invention is to reduce the complexity of signal detection and improve the accuracy and convergence speed of signal detection in complex environments. The present invention proposes a low-complexity collaborative signal detection method for a non-cellular massive MIMO system based on deep unfolded Bayesian inference, comprising the following steps:

[0009] Step 1: The access point AP extracts signal data from the received signal to construct a signal data set;

[0010] Step 2: construct a variational reasoning and deep unfolding network VID-Net, and train the VID-Net to obtain a trained VID-Net;

[0011] Step 3: Input the signal data set into the trained variational inference and deep unfolding network VID-Net to obtain the estimated signal.

[0012] Further, in step 1, the signal data set includes a useful signal and a noise signal;

[0013] The expression for the signal data is defined as:

[0014] y l =h l x+n l

[0015] Among them, y l is the signal received by the lth access point AP, h l represents the channel between the lth access point AP and the user terminal UE, x is the signal sent by the user terminal UE, and nl is the additive white Gaussian noise;

[0016] Channel h l Defined as:

[0017]

[0018] in, is the complex small-scale fading coefficient, β l is the complex large-scale fading coefficient taking into account path loss and shadowing.

[0019] Furthermore, in step 2, the variational reasoning and deep unfolding network VID-Net is trained. Specifically, the variational backward inference method is used to train VID-Net. By minimizing the KL divergence between the posterior probability and the approximate posterior probability, the network parameters of VID-Net are gradually adjusted so that the posterior probability measured by the VID-Net network is close to the true posterior probability of the signal.

[0020] Furthermore, considering the impact of multi-layer network output on the results, the logarithmic function and correlation function are introduced to improve the network loss function, so that the front layer network in VID-Net can converge quickly. The loss function of VID-Net is:

[0021]

[0022] Where x is the signal sent by the UE, T iter represents the total number of layers of the VID-Net network, t represents the number of layers of the network, θ (t) is a trainable parameter, μ (t) The output value of the tth layer of the VID-Net network, r(x,μ (t) ) represents the true value x and the output value μ of the network layer t (t) The cosine similarity between ;

[0023]

[0024] Furthermore, step 2 also includes optimizing the VID-Net neural network, and the optimization objective function is expressed as:

[0025]

[0026] Where g(q,W) and W represent:

[0027]

[0028] W=[w 1 ,w 2 ,...,w L ]

[0029] x=[x 1 ,x 2 ,...,x i ,..x K ]

[0030] H=[h 1 ,h 2 ,...,h l ,..h :L ]

[0031] y=[y 1 ,y 2 ,...,y l ,..y :L ]

[0032] W is the introduced neural network parameter, x is the signal sent by UE, and x i represents the i-th signal sent by the UE, q(x) is the approximate posterior probability; the likelihood probability p(y|x) represents the probability of x given the signal y, represents the complex Gaussian probability distribution function with expectation μ and covariance Σ, σ 2 represents the variance of Gaussian noise; represents the distribution of p(x) that satisfies the normal distribution; δ i is a predetermined smaller value, M is the target number; f(w i ,x i ) represents the output of the constructed general neural network.

[0033] Furthermore, the objective function is solved by alternately optimizing q and W;

[0034] First, assuming that W is given, g(q,W) satisfies the following formula:

[0035]

[0036] Where μ and Σ are:

[0037]

[0038] Among them, Diag(x) represents a 1 ,x 2 ,...,x K The diagonal matrix composed of s = [s 1 ,..s m ,.,s M ] T ,s m Represents the set of possible signal symbols in a communication system;

[0039] You can get:

[0040]

[0041] Among them, V represents the unit ary matrix, including eigenvectors, and Λ represents the eigenvalue matrix;

[0042] Assuming that the parameter W is unknown, the objective is to minimize the function f to solve the parameter W. Since only the part of the objective function related to W needs to be considered, the objective function is simplified to the following form:

[0043]

[0044] The following formula is obtained:

[0045]

[0046] By Jensen's inequality, the solution to the above formula is:

[0047]

[0048] Therefore, by alternately optimizing q and W, a closed-form solution is obtained for each alternating minimization step.

[0049] Furthermore, a residual connection is introduced in VID-Net, which is expressed as:

[0050] μ (t+1) =t (t) μ (t) +(1-t (t) )μ (t-1)

[0051] t (t) is the residual network parameter, μ (t+1) Represents the output of each layer after adding the residual network in VID-Net;

[0052] Then perform nonlinear activation function mapping on the output values ​​of each layer of the network. The formula is as follows:

[0053] μ (t+1) =η(μ (t+1) )

[0054] Among them, η represents a nonlinear activation function.

[0055] Furthermore, the VID-Net neural network constructed in step 2 includes convolutional layers, pooling layers, and residual connections;

[0056] The convolutional layer is used to extract key features of the signal x received by the access point AP;

[0057] The pooling layer is used to reduce the dimension of the key features;

[0058] There is a residual connection between the convolution layer and the pooling layer for information skipping.

[0059] Beneficial effects:

[0060] 1. Improve signal detection accuracy: Through multi-level feature extraction, combination and optimization processing, VID-Net can significantly improve the accuracy of signal detection, especially in complex channel environments.

[0061] 2. Enhanced robustness: Modular design and residual connection enable VID-Net to maintain stable detection performance even under high noise and strong interference conditions, and has strong anti-interference ability.

[0062] 3. Adapt to complex channel environments: The VID-Net structure is applicable to a variety of complex channel environments, and can effectively handle scenarios such as dynamically changing channel conditions and multipath interference.

[0063] 4. Reduce computational complexity: The division of labor between module A and module B and the use of residual network optimize the computational process and reduce computational complexity, so that the network can run under limited computing resources while maintaining high performance.

[0064] 5. Reduce the risk of overfitting: The design of residual network and appropriate parameters reduces the risk of overfitting of the model, thereby improving the generalization ability of the network in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a flow chart of the present invention.

[0066] Figure 2 It is a structural diagram of VID-Net in the present invention.

[0067] Figure 3 It is the flowchart in module A. DETAILED DESCRIPTION

[0068] This specific implementation provides a low-complexity cooperative signal detection method for a non-cellular massive MIMO system based on deep unfolded Bayesian inference, such as Figure 1 As shown, including:

[0069] receiving a signal and obtaining a noise data set including an estimated signal;

[0070] The neural network is trained, including: based on the variational backward inference idea and the access point AP cooperation theory, the signal posterior probability is estimated by minimizing the KL divergence between the posterior probability and the approximate posterior probability, and each access point AP estimates the transmitted signal separately, and finally, it is handed over to the CPU for combined processing. Secondly, considering the impact of the multi-layer network output on the result, the logarithmic function and the correlation function are introduced to improve the network loss function, so that the front layer network can converge quickly. The signal data set is input into the trained variational inference and deep unfolding network VID-Net (Variational Inference and Deepunfolding Network), which is a deep neural network that captures the useful information in the signal data set and obtains the estimated signal.

[0071] Specifically, a low-complexity cooperative signal detection method for a non-cellular massive MIMO system based on deep unfolded Bayesian inference of the present invention comprises the following steps:

[0072] Step 1. System signal processing and signal data set generation

[0073] First, the access point AP extracts a signal data set from the received signal, where the signal data set contains a useful signal and a noise signal.

[0074] The process of extracting the signal data set by the access point AP involves data preprocessing after multiple receiving antennas receive the signal, including basic operations such as filtering and denoising, to ensure that the signal data set reflects the characteristics of the transmission channel and the actual impact of noise as realistically as possible.

[0075] Channel modeling is a key link in the entire signal detection process. In the present invention, L single-antenna access points AP are used, which are deployed in an area of ​​1*1km to serve K randomly distributed single-antenna UEs. The calculation formula of the signal data is defined as:

[0076] y l =h l x+n l

[0077] Among them, y l The signal received by the lth access point AP, h l represents the channel between the lth access point AP (Access Point) and the user terminal UE, x is the signal sent by the UE, n l is additive white Gaussian noise.

[0078] Here, channel h l Specifically, it includes two main parts: small-scale fading coefficient and large-scale fading coefficient, which are used to describe the microscopic and macroscopic environmental fading conditions encountered by the signal during transmission.

[0079] The non-cellular massive MIMO simulation setting used in this specific implementation is: In the 3GPP model, the large-scale fading coefficient is expressed by the following formula:

[0080]

[0081] Among them, d l,k is the distance between the lth access point AP and the kth UE, F l,k As the shadows fade,

[0082] The channel impulse response of the kth user, that is, channel h l Defined as:

[0083]

[0084] in, is the complex small-scale fading coefficient, β l is the complex large-scale fading coefficient taking into account path loss and shadowing.

[0085] In order to ensure the accuracy of channel modeling, the present invention further considers the channel consistency assumption within the time-frequency block, that is, within a short time and frequency range, the channel h l can be regarded as a constant. This assumption greatly simplifies the difficulty of channel modeling while still effectively describing the characteristics of the signal in the actual transmission environment.

[0086] Step 2: Training and optimization of the VID-Net (Variational Inference and Deepunfolding Network) neural network

[0087] The training of the VID-Net neural network adopts the variational backward inference method. Specifically, during the training process of the network, based on the access point AP cooperation theory, by minimizing the KL divergence between the posterior probability and the approximate posterior probability, the VID-Net neural network parameters are gradually adjusted so that the posterior probability measured by the network can approach the true posterior probability of the signal, and the signal posterior probability is estimated. Each access point AP estimates the transmitted signal separately, and finally, it is handed over to the CPU for combined processing.

[0088] During the training and optimization process of the VID-Net neural network, the access point AP (Access Point) nodes will use the VID-Net neural network to independently estimate the received signal, and the estimation results of each access point AP node will be transmitted to the central processing unit (CPU) in the form of probability distribution. The CPU combines the probability distributions of the estimation results of all access point AP nodes to generate the final signal estimation result; the combination processing, for example, can be weighted and averaged to merge all independent estimation results into a comprehensive probability distribution. Through this multi-level inference and combination, the error of a single node can be effectively reduced, while the overall estimation capability of the signal can be enhanced.

[0089] (1) In order to further improve the performance of the network, the loss function of the present invention also introduces a logarithmic function and a correlation function to improve the loss function of the network. Specifically, by logarithmic transformation of the output of each layer of the network, the nonlinear characteristics in the data can be better captured during the optimization process. The introduction of the correlation function can effectively enhance the correlation of the output results between the layers of the network, so that the network can converge faster during the training process and avoid falling into the local optimal solution. The correlation function refers to the mean square error. In addition, in order to further improve the generalization ability of the network, L1 and L2 regularization terms are added to the loss function to prevent the network from overfitting.

[0090] In order to improve the accuracy of signal detection, the loss function not only considers the traditional mean square error, but also introduces cosine similarity as a regularization term.

[0091] Specifically, the loss function is calculated as:

[0092]

[0093] Where x is the signal sent by the UE, T iter represents the total number of layers of the VID-Net network, t represents the number of layers of the VID-Net network, θ (t) is a trainable parameter used to control the cosine similarity r(x,μ (t) ) in the loss function, generally 0<θ (t) <1, initialized to 0.5 in the network, μ (t) The output value of the tth layer of the network; in the calculation formula of the loss function, ||x-μ (t) The || part represents the traditional mean square error, r(x,μ (t) ) part represents the regularization term.

[0094] Based on the original mean square error function, a weight coefficient lg(t+1) is given to each layer of the network. The weight coefficient lg(t+1) not only takes into account the impact of the output value of each layer of the signal detection network on the final output, ensuring the convergence of the network, but also makes the subsequent layer network closer to the actual output value. Further considering the estimated result μ of the AP node (t) The correlation between the true value x and the cosine similarity is introduced as a regularization term with reference to L1 regularization and L2 regularization. The formula is as follows:

[0095]

[0096] Among them, r(x,μ (t) ) represents the true value x and the output value μ of the tth layer of the VID-Net network (t) The cosine similarity between them focuses more on the matching in the vector direction.

[0097] (2) Objective function construction and optimization problem solving

[0098] The solution to the VID-Net neural network optimization problem in the present invention is based on the basic principle of Bayesian inference. The objective function for optimizing the VID-Net neural network can be expressed as:

[0099]

[0100] Where g(q,W) and W represent:

[0101]

[0102] W=[w 1 ,w 2 ,...,w L ]

[0103] x=[x 1 ,x 2 ,...,x i ,..x K ]

[0104] H=[h 1 ,h 2 ,...,h l ,..h :L ]

[0105] y=[y 1 ,y 2 ,...,y l ,..y :L ]

[0106] W is the introduced neural network parameter, x is the signal sent by UE, and x iIndicates that UE sends the i-th signal.

[0107] Among them, q(x) is the approximate posterior probability, and the likelihood probability p(y|x) represents the probability of x given the signal y. represents the complex Gaussian probability distribution function with expectation μ and covariance Σ, σ 2 represents the variance of Gaussian noise, f(w i ,x i ) represents the output of the constructed neural network, and y represents the entire signal received by the access point AP.

[0108]

[0109] in, Indicates that the mean is s j , with covariance δ i The complex Gaussian probability distribution function, represents a distribution that approximates p(x) and satisfies the normal distribution, δ i is a predetermined small value, and M is the target number.

[0110] Solve the problem,

[0111] First, the system model can tell us the likelihood probability

[0112] So, first assume that W is given, g(q,W) satisfies the following formula:

[0113]

[0114] Where μ and Σ are:

[0115]

[0116] Among them, Diag(x) represents a 1 ,x 2 ,...,x K The diagonal matrix composed of s = [s 1 ,..s m ,.,s M ] T ,s m Represents the set of possible signal symbols in a communication system.

[0117] You can get:

[0118]

[0119] Among them, V represents the unit ary matrix, including eigenvectors, and Λ represents the eigenvalue matrix.

[0120] Assuming that the parameter W is unknown, the objective is to minimize the function f to solve the parameter W. Since only the part of the objective function related to W needs to be considered, the objective function can be simplified to the following form:

[0121]

[0122] Then we can get the following formula:

[0123]

[0124] By Jensen's inequality, the solution to the above formula is:

[0125]

[0126] In summary, the problem of optimizing q and W can be solved alternately, so that a closed-form solution can be obtained for each alternating minimization step.

[0127] (3) Optimization of residual networks and deep networks

[0128] In order to overcome the gradient vanishing and gradient exploding problems that are prone to occur during the training of deep neural networks, the present invention introduces a residual network. The residual network uses skip connections to allow information to be directly transferred from the input layer to the output layer, thus avoiding the layer-by-layer attenuation of information in the deep network.

[0129] The calculation formula for finding the signal with the maximum estimated function is:

[0130]

[0131] The basic processing formula of the residual network is:

[0132]

[0133] in, represents a parameter value in the t+1th iteration, It represents an estimate of a certain observed value y of the signal, α l Represents the weight coefficient.

[0134] In addition, in order to prevent the problem of deep network degradation caused by increasing the number of network layers, this paper introduces the parameter t based on the residual network idea. (t) The core idea of ​​the residual network is to allow information to be directly transferred from the input layer to the output layer through skip connections, thereby solving the gradient vanishing and gradient exploding problems in deep network training. As mentioned above, the residual network only adds a small number of parameters compared to the traditional deep network, which greatly reduces the risk of model overfitting. The output of VID-Net is processed as follows:

[0135] μ (t+1) =t (t) μ (t) +(1-t (t) )μ (t-1)

[0136] μ (t+1) Represents the output of each layer after adding the residual network in VID-Net;

[0137] Finally, the output values ​​of each layer of the network are mapped with a nonlinear activation function. The formula is as follows:

[0138] μ (t+1) =η(μ (t+1) )

[0139] Among them, η represents a nonlinear activation function.

[0140] The present invention estimates the signal posterior probability by minimizing the KL divergence between the posterior probability and the approximate posterior probability. Each access point AP estimates the transmitted signal respectively, and finally, the CPU performs combined processing. When used, it not only has extremely low computational complexity, but also converges in only 2 to 3 layers. In terms of detection performance, it is superior to the suboptimal MMSE detection algorithm under different modulation orders and different numbers of system antennas, and exhibits excellent detection performance and robustness.

[0141] (4).VID-Net network architecture and parameter initialization

[0142] The VID-Net network architecture is an important part of the present invention. In order to achieve efficient signal detection, VID-Net adopts a modular design, which divides the signal detection process into multiple independent functional modules. Each module contains several layers of neural networks, each layer is unfolded layer by layer, the output of the previous layer is used as the input of the next layer, and deep features are gradually extracted to jointly optimize the signal detection effect and achieve deep processing of the input signal.

[0143] The initial parameters of the VID-Net network are set to a full 0 column vector to ensure that the network does not produce any offset to the input signal in the initial state. During the training process, the network gradually adjusts these parameters, including the eigenvalue δ (0) =20, auxiliary parameters Variation parameter θ (0) =0.5, for the lth access point AP receiving signal x, the initial value of the network in the VID-Net network is μ (0) , is a full 0 column vector, the parameter to be trained is t (0) =0.5,δ (0) =20, θ (0)= 0.5. The input of each layer is the expected value of the noise data set The output is the estimated expectation

[0144] The VID-Net neural network includes module A and module B, such as Figure 2-3 As shown in the figure, in the VID-Net neural network, the structure of module A includes convolution layer, pooling layer and residual connection: 1. Convolution layer: used to extract key features from the signal received by the access point (AP). The convolution operation can extract local features, such as changes in frequency and amplitude, which are very important for signal detection and estimation. 2. Pooling layer: usually used for dimensionality reduction, reducing the size of the feature map, thereby reducing the computational complexity, while retaining the main information. This layer can effectively reduce the impact of noise and improve the robustness of the features. 3. Residual connection: used to transfer information jumps between different layers of module A, prevent gradient disappearance, and maintain the integrity and coherence of signal features. Module A extracts the main features of the transmitted signal through these structures, and provides high-quality feature representation for the signal combination and optimization of subsequent module B. Module B is set at the CPU, and combined with the access point AP cooperation theory, the output signal of module A is optimized and combined, and finally an accurate signal detection result is obtained. The optimization and combination refers to weighted averaging of the output signal of module A; the connection between module A and module B is realized through the residual network to ensure efficient transmission and processing of the signal in the entire network.

[0145] Through the collaborative work of the above multiple modules, VID-Net can not only effectively capture useful information in the signal, but also significantly improve the accuracy and robustness of signal detection through multi-level optimization and combined processing. In practical applications, this network architecture can not only adapt to various complex channel environments, but also maintain good detection performance under conditions of high noise and strong interference.

[0146] In summary, the present invention provides a low-complexity cooperative signal detection method for a non-cellular large-scale MIMO system based on deep unfolding Bayesian inference. For the cooperative signal detection mechanism between access points (APs), the third-level cooperative level based on the AP four-level cooperative theoretical framework realizes the preliminary signal estimation performed independently by each AP and the subsequent aggregation and integration at the central processing unit (CPU) level. The core innovation of VID-Net lies in its selection of prior probability distribution in variational Bayesian inference. The present invention adopts complex Gaussian distribution as an approximation of prior probability, and incorporates Gaussian distribution parameters as learnable parameters into the network training process to dynamically adapt to the needs of different network levels. This innovation overcomes the convergence problem caused by improper initial distribution selection in traditional methods. In addition, VID-Net determines and deploys the optimal weights to each AP through a pre-trained deep learning model, simplifies the calculation on the CPU side to a simple summation operation, significantly reduces the computational complexity, and enhances the adaptability to diverse channel conditions. In order to further optimize performance, VID-Net incorporates logarithmic function and correlation function correction strategies to tap the potential of deep network output and optimize the loss function. The simulation results show that the VID-Net detection algorithm has an excellent convergence speed, and only 2 to 3 layers of networks are needed to reach the convergence state. Under different modulation orders and system antenna configurations, VID-Net shows excellent detection performance and robustness, and its performance exceeds the suboptimal minimum mean square error (MMSE) detection algorithm.

Claims

1. A massive MIMO system signal detection method, characterized in that: The steps include: Step 1: The access point AP extracts signal data from the received signal to construct a signal data set; Step 2: construct a variational reasoning and deep unfolding network VID-Net, and train the VID-Net to obtain a trained VID-Net; Step 3: Input the signal data set into the trained variational inference and deep unfolding network VID-Net to obtain the estimated signal.

2. The method for detecting signals in a massive MIMO system according to claim 1, characterized in that: In step 1, the signal data set includes a useful signal and a noise signal; The expression for the signal data is defined as: y l =h l x+n l Among them, y l is the signal received by the lth access point AP, h l represents the channel between the lth access point AP and the user terminal UE, x is the signal sent by the user terminal UE, and nl is the additive white Gaussian noise; Channel h l Defined as: in, is the complex small-scale fading coefficient, β l is the complex large-scale fading coefficient taking into account path loss and shadowing.

3. The method for detecting signals in a massive MIMO system according to claim 1, wherein: In step 2, the variational reasoning and deep unfolding network VID-Net is trained. Specifically, the variational backward inference method is used to train VID-Net. By minimizing the KL divergence between the posterior probability and the approximate posterior probability, the network parameters of VID-Net are gradually adjusted so that the posterior probability measured by the VID-Net network is close to the true posterior probability of the signal.

4. The method for detecting signals in a massive MIMO system according to claim 1, wherein: Considering the impact of multi-layer network output on the results, the logarithmic function and correlation function are introduced to improve the network loss function, so that the front layer network in VID-Net converges quickly; the loss function of VID-Net is: Where x is the signal sent by the UE, T iter represents the total number of layers of the VID-Net network, t represents the number of layers of the network, θ (t) is a trainable parameter, μ (t) The output value of the tth layer of the VID-Net network, r(x,μ (t) ) represents the true value x and the output value μ of the network layer t (t) The cosine similarity between ; 5. The method for detecting signals in a massive MIMO system according to claim 1, wherein: Step 2 also includes optimizing the VID-Net neural network, and the optimization objective function is expressed as: Where g(q,W) and W represent: In=[in1,in2,...,in L ] x=[x1,x2,…,x i ,..x K ] H=[h1,h2,...,h l ,..h :L ] y=[y1,y2,...,y l ,..y :L ] W is the introduced neural network parameter, x is the signal sent by UE, and x i represents the i-th signal sent by the UE, q(x) is the approximate posterior probability; the likelihood probability p(y|x) represents the probability of x given the signal y, represents the complex Gaussian probability distribution function with expectation μ and covariance Σ, σ 2 represents the variance of Gaussian noise; represents the distribution of p(x) that satisfies the normal distribution; δ i is a predetermined smaller value, M is the target number; f(w i ,x i ) represents the output of the constructed general neural network.

6. The method for detecting signals in a massive MIMO system according to claim 5, characterized in that: By alternately optimizing q and W, the objective function is solved; First, assuming that W is given, g(q,W) satisfies the following formula: Where μ and Σ are: Among them, Diag(x) represents a system consisting of x1,x2,...,x K The diagonal matrix composed of s = [s1, ..s m ,.,s M ] T ,s m Represents the set of possible signal symbols in a communication system; get: Among them, V represents the unit ary matrix, including eigenvectors, and Λ represents the eigenvalue matrix; Assuming that the parameter W is unknown, the objective is to minimize the function f to solve the parameter W. Since only the part of the objective function related to W needs to be considered, the objective function is simplified to the following form: The following formula is obtained: By Jensen's inequality, the solution to the above formula is: Therefore, by alternately optimizing q and W, a closed-form solution is obtained for each alternating minimization step.

7. The method for detecting signals in a massive MIMO system according to claim 4, characterized in that: The residual connection is introduced in VID-Net, which is expressed as: m (t+1) =t (t) m (t) +(1-t (t) )m (t-1) t (t) is the residual network parameter, μ (t+1) Represents the output of each layer after adding the residual network in VID-Net; Then perform nonlinear activation function mapping on the output values ​​of each layer of the network. The formula is as follows: m (t+1) =η(μ (t+1) ) Among them, η represents a nonlinear activation function.

8. The method for detecting signals in a massive MIMO system according to claim 7, characterized in that: The VID-Net neural network constructed in step 2 includes convolutional layers, pooling layers, and residual connections; The convolutional layer is used to extract key features of the signal x received by the access point AP; The pooling layer is used to reduce the dimension of the key features; There is a residual connection between the convolution layer and the pooling layer for information skipping.

Images