A super network assisted data model double-driven MIMO-OFDM signal detection method

Abstract

The application discloses a kind of super-network auxiliary data model double-drive MIMO-OFDM signal detection method, comprising: a signal detection network with a set of adjustable parameters, and a trainable super-network.First, using linear least mean square error algorithm obtains channel estimation, will estimated channel matrix H LMMSE As the input of super-network, adjustable parameter Ω hyper Provided to signal detection network;Then, channel matrix H LMMSE And received signal vector y are real value decomposition, together with the equivalent noise variance information derived and input to signal detection network, signal detection network is unfolded by the expectation propagation algorithm enhanced by graph neural network, in combination with real domain channel matrix real domain received signal vector environmental information and the parameter Ω hyper Provided by super-network, give the estimation to sending symbol vectorFinally, demapping is carried out to estimated symbol vector, and the estimation to original sending bit is obtainedThe application has high detection performance, consumes less time, and has strong generalization, and can adapt to scene change.

Description

TECHNICAL FIELD

[0001] The application relates to a super-network-aided data model double-driven MIMO-OFDM signal detection method and belongs to the technical field of wireless communication. BACKGROUND

[0002] Channel estimation and signal detection are important tasks of a receiving end of a MIMO-OFDM system. Among them, a channel estimation method based on linear minimum mean square error is widely applied, in addition, a signal detection algorithm based on expectation propagation achieves a good balance in terms of calculation complexity and performance, and is widely studied. However, the non-perfect channel state information obtained by the estimation algorithm will cause serious performance loss to the data inference method based on expectation propagation. In addition, in the scene where multi-user interference is strong, the expectation propagation detector still has a large performance gap compared with the optimal maximum likelihood detection baseline.

[0003] In recent years, deep learning technology has been widely applied in the fields of image recognition, speech recognition and natural language processing. The signal detection design assisted by artificial intelligence combines deep learning technology and expert knowledge, learns the data structural features embedded in multi-antenna detection by means of a large amount of labeled data, and is expected to make up for the shortcomings of the model-based scheme. In the related research direction, the graph neural network is integrated into the iterative detection process, which can effectively learn the multi-user interference information and compensate for the performance defects. However, the training of the detection network is often completed in a specific scene, and there is a problem of insufficient robustness. When the environment changes greatly, the network cannot adaptively adjust, and the performance will be seriously deteriorated. In addition, under the condition of non-perfect channel state information, the deviation of the iterative detection to the posterior probability distribution of the transmission symbol needs to be further solved, and the performance loss needs to be compensated. SUMMARY

[0004] The application provides a super-network-aided data model double-driven MIMO-OFDM signal detection method, which aims to use a data model double-driven neural network to effectively improve the detection performance, and introduces a super network to improve the adaptability of the signal detection network.

[0005] Technical scheme: The application adopts the following technical scheme:

[0006] The super-network-aided data model double-driven MIMO-OFDM signal detection method provided by the application comprises a signal detection network with a set of adjustable parameters and a trainable super network, and the method comprises the following steps:

[0007] (1) estimating channel state information by using a linear minimum mean square error algorithm, and inputting the channel state information into the super network as input, and outputting adjustable parameters Omega hyper to the signal detection network;

[0008] (2) Estimating the channel matrix H LMMSE Perform real-valued decomposition on the received signal vector y, and obtain the real-domain channel matrix. Real-domain received signal vector and the derived equivalent noise variance information The input is fed into a signal detection network, which consists of T interconnected sub-modules enhanced by graph neural networks, with adjustable parameters shared among different graph neural network modules; the signal detection is estimated based on the real-domain channel matrix. Real-domain received signal vector Environmental Information The parameters provided by the hypernetwork and the adjustable parameters of the submodules give an estimate of the transmitted symbol vector.

[0009] (3) For the estimated sign vector Perform demapping to obtain an estimate of the original transmitted bits.

[0010] Furthermore, the supernetwork described in step (1) has trainable parameters Θ. hyper The neural network, the signal detection network has some adjustable parameters Ω hyper This is generated by the supernetwork; the channel state information input to the supernetwork is the channel matrix H estimated by the linear minimum mean square error algorithm. LMMSE The hypernetwork for the channel matrix H LMMSE Perform singular value decomposition, and then perform real-value decomposition on the obtained singular value vector s to obtain the singular value vector. The hypernetwork is specifically implemented using a deep neural network consisting of three fully connected layers, containing {256, 256, 4N} respectively. u} neurons, of which N u For the feature vector u (0) The length of the supernetwork; the parameters output by the supernetwork are a set of adjustable parameters Ω. hyper This corresponds to the input feature vector u of the first submodule t=1 of the signal detection network. (0) A set of adjustable parameters required during the calculation process.

[0011] Furthermore, step (2) specifically includes:

[0012] (2.1) For the channel matrix H LMMSE The channel matrix in the real domain is obtained by performing real-valued decomposition on the received signal vector y. and the received signal vector in the real number domain

[0013] (2.2) Correct the second-order statistical property R of the noise based on the estimated channel matrix error. z, and real-valued decomposition is performed to obtain a real-valued correlation matrix R e , and equivalent environmental information is extracted The expression is:

[0014]

[0015] where Re(·) represents taking the real part of a complex number, N r is the number of receive antennas, and tr(·) represents the trace of a matrix;

[0016] (2.3) The expectation propagation algorithm enhanced by the graph neural network is used as a signal detection network, which includes T serially connected sub-modules, and the adjustable parameters between different graph neural network modules are shared; the real-valued channel matrix The real-valued received signal vector and the equivalent environmental information are input into the signal detection network; the calculation process of the input feature vector u (0) of the first sub-module t = 1 of the signal detection network involves adjustable parameters Ω hyper provided by the super network; the input feature vector u (0) of the subsequent t = 2,...,T sub-modules is assigned a value according to the output feature vector u (L) of the graph neural network in the t-1 module, where L is the number of iterations of the feature vector in the graph neural network; the tth sub-module of the signal detection network completes the estimation according to the real-valued channel matrix The real-valued received signal vector The equivalent environmental information and the input feature vector u (0) The tth sub-module is divided into three parts, namely observation, graph neural network enhancement, and estimation; the observation part is based on the expectation propagation detection algorithm, which first constructs a Gaussian posterior approximation using an exponential cluster distribution, calculates the mean and the covariance of the exponential cluster distribution, and then calculates the likelihood function based on the Gaussian posterior function, the mean and the covariance of the likelihood function are used as prior information input into the graph neural network; the graph neural network enhancement part uses the characteristics of the likelihood distribution and the multi-user interference information to represent the latent function relationship between the variable nodes and the factor nodes, and strengthens the posterior probability inference, and the posterior probability is input into the estimation module, in addition, the feature vector is updated iteratively in the message passing process between the variable nodes and the factor nodes, and the output feature vector u (L) is obtained after L iterations; the estimation part performs soft decision on the transmission symbol according to the posterior probability, calculates the mean and the variance Combining prior information from the likelihood function with the parameter pairs (γ) required in the calculation of the exponential cluster distribution update process (t) ,Λ (t) ), and the output feature vector u (L) The inputs are fed into the (t+1)th submodule; the Tth submodule outputs the final estimated symbol vector.

[0017] Furthermore, the second-order statistical characteristic R of the noise z The correction is obtained by deriving the second moment of the channel matrix estimation error ΔH; using Indicates the n=1,...,N t The position of the pilot subcarrier occupied by the root transmitting antenna, using Indicates the position of the data subcarrier, where P is the pilot sequence length, D is the number of data subcarriers, and N is the number of data subcarriers. t The number of transmitting antennas is determined by the received pilot components. Known pilot sequence And the calculation of the frequency correlation characteristics of the channel for the m=1,...,N r LMMSE estimation of the channel coefficient between the nth receive antenna and the nth transmit antenna W LMMSE,n The interpolation matrix used in the linear minimum mean square error estimation calculation is represented by the true channel coefficient vector h. m,n The receive vector is located at the k-th data subcarrier. Based on the calculated estimated channel matrix And omitted Construct the following equivalent relationship:

[0018] y = H LMMSE x+z, z=ΔHx+w,

[0019] The variance of the white noise vector w is: Based on row index d k and column indexes From the interpolation matrix W LMMSE,n Extract column vector w LMMSE,n And calculate the matrix The second moment R of the equivalent noise vector z z =E{zz H} is represented as:

[0020] R Δh,n =R A +R D -R B +R C ,R A =R hh ,

[0021]

[0022] Among them, the average signal energy E s =1 / N t I is the identity matrix, w k′ and Corresponding to vectors w LMMSE,n and the elements in matrix M, Indicates that the main diagonal elements are diagonal matrix, Related information R hh , The definition is as follows:

[0023]

[0024] in, The above-mentioned relevant information was extracted using the MIMO-OFDM channel correlation matrix.

[0025] Furthermore, in the graph neural network enhancement part of the t-th submodule, a pairwise Markov random field model is used to represent the latent function relationship between variable nodes and factor nodes, and factor features are extracted. in and They refer to The column j, For the first submodule t=1, the variable eigenvector The calculation is based on the aforementioned latent function relationship and the adjustable parameter Ω provided by the hypernetwork. hyper Includes weight matrix W hyper and bias vector b hyper The eigenvectors of the variables are calculated as follows:

[0026]

[0027] Feature vector Length N u In the l=1,...,Lth iteration, the features With variable eigenvectors Input factor nodes are aggregated through a multilayer perceptron network 1, and output messages are generated. information Prior information related to the likelihood function After integration by the gated loop unit, the feature vector at the variable node is updated. Iterate L times to complete message passing between variable nodes and factor nodes, and finally output the feature vector. Computational graph neural network-enhanced posterior probability inference

[0028] Further, the complete neural network composed of the signal detection network and the super network is trained by using a small batch gradient descent algorithm for end-to-end supervised learning, and multiple rounds of training are performed; a small batch of the training set is represented as is a set composed of S randomly generated samples, wherein the transmitted real number field symbol vector is the real number field received symbol vector and the real number field accurate channel matrix are input features of the complete neural network; an adaptive momentum estimation optimizer is selected to optimize the trainable parameters Θ in the neural network, and the learning rate is set to 0.001; the training adopts a cross-entropy loss function, and the specific formula is:

[0029]

[0030] wherein, is a built-in function, which is 1 in the case of and 0 in other cases, is the posterior probability inference enhanced by the graph neural network output by the Tthsub-module of the signal detection network.

[0031] Beneficial effects: compared with the prior art, the present application has the following beneficial effects:

[0032] The present application constructs a signal detection neural network based on the expectation propagation algorithm, enhances the inference of the transmitted symbol posterior probability by means of the graph neural network, and avoids the deviation of the iterative detection algorithm caused by the imperfect channel state information through the correction of the environmental information. Compared with the existing scheme, the present application has achieved considerable performance gain. In addition, the present application introduces a super network to dynamically capture the changes of the channel characteristics, thereby adjusting a small number of key parameters in the signal detection network, and has the advantages of strong generalization and adaptability to scene changes. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 is a system block diagram of an embodiment of the present application;

[0034] Figure 2 is an architecture schematic diagram of the complete neural network used by the present application;

[0035] Figure 3 is an architecture schematic diagram of the signal detection network used by the present application. DETAILED DESCRIPTION

[0036] The present application will be specifically described below in combination with the embodiments of the MIMO system and the accompanying drawings.

[0037] I. System model adopted by the present embodiment

[0038] Consider a MIMO-OFDM system with spatial multiplexing, the transmitter has N t receive antennas, the receiver has N r receive antennas, and the number of subcarriers is K. The training pilot length of each transmit antenna is P, and the remaining D = K - PN t subcarriers are used to transmit data. Use to indicate the pilot subcarrier position of the nth transmit antenna, and use to indicate the data subcarrier position. In this embodiment, N t = N r = 8, K = 512, P = 16, and D = 384. In this embodiment, the modulation method uses QPSK constellation modulation. In one symbol time, the data bits of the nth transmit antenna have 384 x 2 = 768 bits, and after QPSK modulation, the transmitted data symbol vector and the transmitted pilot symbol vector form an OFDM symbol, where A is the QPSK constellation symbol set. At the mth receive antenna, use to indicate the received data component, and use to indicate the received pilot component corresponding to the nth transmit antenna. The signal model of the MIMO-OFDM system can be represented as:

[0039]

[0040] where is a diagonal matrix with the main diagonal elements being , and similarly and correspond to the frequency domain channels through which the pilot and data are transmitted, and are additive white Gaussian noise vectors, and the noise variance is Using the received pilot component , the transmitted pilot sequence , and the frequency correlation characteristics of the channel, the linear minimum mean square error estimate of the channel coefficient between the mth receive antenna and the nth transmit antenna can be calculated as:

[0041]

[0042] where W LMMSE,n is the LMMSE interpolation matrix. In addition, at the d k data subcarrier position, there is a MIMO subsystem, and the signal model can be represented as:

[0043]

[0044] wherein, omit Simplify the vector representation in MIMO subsystem.

[0045] In this embodiment, the MIMO-OFDM channel is generated by using the tap delay line model, and the spatial correlation characteristics are simulated by using the matrix R, and the specific formula is:

[0046]

[0047] wherein, N r ×N r dimensional matrix R r and N t ×N t dimensional matrix R t are the correlation matrices of the receiving end and the sending end respectively, generated by using the exponential correlation model, and the (i, j) element r ij of the matrix satisfies:

[0048]

[0049] wherein, ρ represents the correlation coefficient of the channel matrix.

[0050] Second, the specific steps of this embodiment

[0051] As Figure 1 shown, the embodiment of the present application provides a system block diagram of a hypernetwork-aided data model double-driven MIMO-OFDM signal detection method, which includes real value decomposition and equivalent environment information correction, a signal detection neural network with a set of adjustable parameters, a hypernetwork for generating part of the adjustable parameters in the signal detection network, and a demapping module. In this embodiment, the receiving end obtains the imperfect channel state information by using the linear minimum mean square error algorithm; the complete neural network composed of the signal detection network and the hypernetwork is as shown in Figure 2 , and specifically, the signal detection network is composed of T graph neural network enhanced submodules connected in series, and the adjustable parameters are shared between different graph neural network modules; the hypernetwork is a deep neural network composed of three fully connected layers, has trainable parameters Θ hyper , and generates a set of adjustable parameters Ω hyper according to the channel state information, which corresponds to a set of adjustable parameters required in the process of calculating the input features of the first submodule of the signal detection network.

[0052] The complete neural network composed of the signal detection network and the hypernetwork is trained by end-to-end supervised learning in offline, and the trainable parameters Θ hyperThe adjustable parameters in the graph neural network module are estimated using the trained network during online deployment. A single forward estimation consists of four steps: generating adjustable parameters via the supernetwork, real-valued decomposition and equivalent environment information correction, forward signal detection, and symbolic demapping.

[0053] (1) Hypernetwork generates adjustable parameters

[0054] For each channel realization, the hypernetwork generates partially adjustable parameters for the signal detection network based on the channel state information. The channel state information input to the hypernetwork is the channel matrix estimated by the linear least mean square error algorithm. The hypernetwork first processes the channel matrix H LMMSE Singular value decomposition yields singular value vectors. Then perform real-value decomposition to obtain singular value vectors. like Figure 2 As shown, the hypernet will Input a deep neural network, output a set of adjustable parameters Ω hyper Includes weight matrix W hyper and bias vector b hyper The specific calculation formula is as follows:

[0055]

[0056] g(·) contains three fully connected layers, with neurons set to {256, 256, 4N} respectively. u}, N u For the feature vector u (0) Length; Output Ω hyper The first submodule of the signal detection network, t=1, is used to realize the input feature u from external information. (0) The mapping.

[0057] (2) Real value decomposition and equivalent environmental information correction

[0058] To facilitate algebraic operations and the use of deep learning methods, we first utilize the estimated channel matrix H LMMSE The system represented by equation (3) is equivalently constructed as follows:

[0059] y = H LMMSE x+z,z=ΔHx+w (4)

[0060] Wherein, the channel matrix estimation error ΔH = HH LMMSE Based on row index d k and column indexes From the interpolation matrix W LMMSE,n Extract column vector w LMMSE,n And calculate the matrix The second moment R of the equivalent noise vector z z =E{zzH} is expressed as follows:

[0061] R Δh,n = R A + R D - R B + R C , R A = R hh ,

[0062]

[0063] where the signal average energy E s = 1 / N t , I is a unit matrix, w k′ and correspond to the elements in the vector w LMMSE,n and the matrix M, respectively, tr(·) represents the trace of a matrix, and the correlation information R hh , is defined as follows:

[0064]

[0065] where, The above correlation information is extracted using the MIMO-OFDM channel correlation matrix.

[0066] Then, the equation (4) is real-valued decomposed as follows:

[0067]

[0068] Re(·) and Im(·) represent taking the real part and the imaginary part of a complex number, respectively, and (·) T denotes the transposition operation. The equivalent real form of the MIMO subsystem obtained after real-valued decomposition is as follows:

[0069]

[0070] where the variance of the equivalent noise is The expression is:

[0071]

[0072] (3) Forward signal detection

[0073] For signal detection, the graph neural network enhanced expectation propagation algorithm is used as the signal detection network. This deep neural network contains T serially connected sub-modules, and the adjustable parameters between different graph neural network modules are shared, as shown in Figure 3 The signal detection network is based on the real-valued channel matrix Real-domain received signal vector and equivalent environmental information Solving the system represented by equation (4) completes the process of transmitting the symbol vector. The estimate.

[0074] like Figure 3 As shown, the t-th submodule of the signal detection network significantly improves detection performance through posterior probability inference enhanced by a graph neural network. The input feature vector u of the first submodule of the signal detection network at t=1 is... (0) The calculation process involves the adjustable parameter Ω hyper Provided by the hypernetwork; the t=1,...,T submodule is based on the real-domain channel matrix. Real-domain received signal vector Equivalent environmental information and the input feature vector u (0) The estimation for this layer is completed. Each submodule has the same structure and shares adjustable parameters, and can be divided into three parts: observation, graph neural network enhancement, and estimation. The following uses the t-th submodule as an example to illustrate the workflow of these three parts:

[0075] (3.1) Observation section: The observation section includes a minimum mean square error calculation unit and an extrinsic information calculation unit; firstly, the mean of the Gaussian posterior distribution is constructed using the exponential cluster distribution. With covariance In essence, it is the minimum mean square error estimation:

[0076]

[0077] Among them, (γ) (t) ,Λ (t) ) Parameters according to Perform initialization. Then, the mean of the likelihood function is calculated based on the Gaussian posterior function. With covariance diagonal matrix The first of the main diagonals element and mean The element The specific formula used for the calculation is as follows:

[0078]

[0079] in Corresponding Gaussian posterior distribution covariance matrix The first of the main diagonals One element, Corresponding mean the first element of the external mean and covariance as prior information into the graph neural network part.

[0080] (3.2) Graph neural network enhancement part: based on the pair-wise Markov random field model to represent the latent function relationship between variable nodes and factor nodes in the graph neural network, and extract factor features wherein and respectively refer to the first element of the j-th column, for the first sub-module t = 1, the variable feature vector is calculated according to the above latent function relationship and the adjustable parameter Ω provided by the super network hyper ={W hyper ,b hyper} is calculated as follows:

[0081]

[0082] the feature vector has a length of N u ; the message passing process between the variable nodes and the factor nodes in the l = 1,...,L iterations and the iterative update calculation of the feature vector are as follows:

[0083] a. The feature and the vector are input into the factor node, aggregated by the multi-layer perceptron network 1, and output the message

[0084] wherein the multi-layer perceptron f MLP1 (·) is realized by a three-layer deep neural network;

[0085] b. The message and the likelihood function prior information are integrated by the gated recurrent unit to update the feature vector at the variable node

[0086]

[0087] wherein and are the previous and current hidden states of the gated recurrent unit f GRU (·), and f D ense1(·) is a single-layer deep neural network.

[0088] After L iterations, the output feature vector Computing posterior probability inference

[0089] Multilayer perceptron f MLP2 (·) is realized by a three-layer deep neural network, Ω is a modulation symbol set consisting of real parts of three-layer deep neural network; in addition, according to the output feature vector The variable feature vector of the t+1th sub-module is assigned a value.

[0090] (3.3) Estimation part: the estimation part contains soft decision calculation and parameter pair update; first, according to the posterior probability Soft decision is made on the transmitted symbol, in this embodiment, Each component of is from a modulation symbol set consisting of real parts of QPSK constellation points Therefore, the mean value The first element of the diagonal matrix The first element of the main diagonal The formula used for the calculation of the first element of the main diagonal is as follows:

[0091]

[0092] The above soft estimation result is used together with the prior information of the likelihood function to update the parameter pair (γ (t) ,Λ (t) ); the T sub-modules all perform operations according to the above steps (3.1)-(3.3), and finally output the estimated symbol vector

[0093] The complete neural network composed of the above signal detection network and super network is trained end-to-end using the mini-batch gradient descent algorithm, in this embodiment, a total of 2000 rounds of training are performed, and each round contains 64 mini-batches; a mini-batch of the training set is denoted as is a set consisting of S randomly generated samples, wherein the transmitted real symbol vector is used as the label, the real received symbol vector and the real accurate channel matrix are used as the input features of the complete neural network, and i is the sample serial number, in this embodiment, S is taken as 500, and the correlation coefficient ρ of the channel matrix of the training set is randomly selected within [0.3, 0.6]. The adaptive momentum estimation optimizer is selected to optimize the trainable parameters Θ in the neural network, and the learning rate is set to 0.001; the training adopts the cross-entropy loss function, and the specific formula is as follows:

[0094] ​

[0095] wherein, is an inbuilt function, and 1 in case, and 0 otherwise, is the graph neural network enhanced posterior probability inference of the Tthsub-module output of the signal detection network; in training, 500 samples are fed into the network for forward propagation each time, and after calculating the loss function, back propagation is performed for optimizing the parameters Θ.

[0096] After the above training is completed online, the network is deployed online, and in this embodiment, the correlation coefficient p of the spatial correlation channel used to evaluate the performance of the network is taken as 0.8. Since the super network learns the rule of feature initialization mapping of the signal detection network in different scenes in training, it can still generate appropriate parameters for the signal detection network in testing, and the signal detection network can maintain considerable detection performance based on these parameters. Thus, the adaptability of the detection method is greatly enhanced.

[0097] (4) Symbol demapping

[0098] After signal detection, the estimated symbol vector is restored to the complex domain, and then demapping is performed, and in this embodiment, hard decision is performed to obtain the estimation of the original transmitted bits

[0099] The above only illustrates one preferred embodiment of the present application in combination with the drawings, and cannot be used to limit the scope of rights contained in the present application, and it should be understood that any equivalent changes made without departing from the purpose of the present application are within the scope of protection of the claims of the present application.

Claims

1. A method for detecting dual-drive MIMO-OFDM signals using a supernetwork-assisted data model, characterized in that, A MIMO-OFDM system includes a signal detection network with a set of adjustable parameters and a trainable supernetwork. The method includes the following steps: (1) Channel state information is estimated using the linear minimum mean square error algorithm. The hypernetwork takes the channel state information as input and outputs an adjustable parameter Ω. hyper Provided to the signal detection network; (2) Estimating the channel matrix H LMMSE Perform real-valued decomposition on the received signal vector y, and obtain the real-domain channel matrix. Real-domain received signal vector and the derived equivalent noise variance information The input is fed into a signal detection network, which consists of T interconnected sub-modules enhanced by graph neural networks, with adjustable parameters shared among different graph neural network modules; the signal detection is estimated based on the real-domain channel matrix. Real-domain received signal vector Equivalent noise variance information The parameters provided by the hypernetwork and the adjustable parameters of the submodules give an estimate of the transmitted symbol vector. (3) For the estimated sign vector Perform demapping to obtain an estimate of the original transmitted bits. The supernetwork in step (1) has trainable parameters Θ hyper The neural network, the signal detection network has some adjustable parameters Ω hyper This is generated by the supernetwork; the channel state information input to the supernetwork is the channel matrix H estimated by the linear minimum mean square error algorithm. LMMSE The hypernetwork for the channel matrix H LMMSE Perform singular value decomposition, and then perform real-value decomposition on the obtained singular value vector s to obtain the singular value vector. The hypernetwork is specifically implemented using a deep neural network consisting of three fully connected layers, containing {256, 256, 4N} respectively. u } neurons, of which N u For the feature vector u (0) The length of the supernetwork; the parameters output by the supernetwork are a set of adjustable parameters Ω. hyper This corresponds to the input feature vector u of the first submodule t=1 of the signal detection network. (0) A set of adjustable parameters required during the calculation process; Step (2) specifically includes: (2.1) For the channel matrix H LMMSE The channel matrix in the real domain is obtained by performing real-valued decomposition on the received signal vector y. and the received signal vector in the real number domain (2.2) Correct the second-order statistical property R of the noise based on the estimated channel matrix error. z The correlation matrix R in the real field is obtained through real-valued decomposition. e And extract equivalent noise variance information. The expression is: Where Re(·) denotes taking the real part of the complex number, and N r The number of receiving antennas is given, and tr(·) represents the trace of the matrix; (2.3) The expectation propagation algorithm enhanced by graph neural networks is used as the signal detection network. This deep neural network contains T cascaded sub-modules, and adjustable parameters are shared among different graph neural network modules; the real-domain channel matrix is ​​used. Real-domain received signal vector and equivalent noise variance information As the input to the signal detection network; the input feature vector u of the first submodule of the signal detection network at t=1. (0) The calculation process involves the adjustable parameter Ω hyper The input feature vector u of the subsequent t=2,...,T sub-modules is provided by the hypernetwork. (0) Then, based on the output feature vector u of the graph neural network in the (t-1)th module... (L) Assign values, where L is the number of iterations of the feature vector in the graph neural network; the t-th submodule of the signal detection network is based on the real-domain channel matrix. Real-domain received signal vector Equivalent noise variance information and the input feature vector u (0) To complete the estimation for this layer, the t-th submodule is divided into three parts: observation, graph neural network augmentation, and estimation. The observation part is based on the expectation propagation detection algorithm. First, it uses the exponential cluster distribution to construct a Gaussian posterior approximation and calculates its mean. With covariance Then, the likelihood function is calculated based on the Gaussian posterior function, and the mean of the likelihood function is... With covariance The prior information is input into the graph neural network; the enhancement part of the graph neural network utilizes the likelihood distribution characteristics and multi-user interference information to represent the latent function relationship between variable nodes and factor nodes, thereby strengthening the posterior probability inference. The input estimation module, in addition, continuously iterates and updates the feature vector during message passing between variable nodes and factor nodes, obtaining the output feature vector u after L iterations. (L) The estimation part performs soft decision-making on the transmitted symbols based on the posterior probability and calculates the mean. and variance Combining prior information from the likelihood function with the parameter pairs (γ) required in the calculation of the exponential cluster distribution update process (t) ,Λ (t) ), and the output feature vector u (L) The inputs are fed into the (t+1)th submodule; the Tth submodule outputs the final estimated symbol vector.

2. The method for detecting dual-drive MIMO-OFDM signals using a supernetwork-assisted data model according to claim 1, characterized in that: Second-order statistical properties of noise R z The correction is obtained by deriving the second moment of the channel matrix estimation error ΔH; using Indicates the n=1, ..., Nth t The position of the pilot subcarrier occupied by the root transmitting antenna, using Indicates the position of the data subcarrier, where P is the pilot sequence length, D is the number of data subcarriers, and N is the number of data subcarriers. t The number of transmitting antennas is determined by the received pilot components. Known pilot sequence And the frequency correlation characteristics of the channel are calculated for the m=1,…,N r Linear minimum mean square error estimation of the channel coefficients between the nth receiving antenna and the nth transmitting antenna W LMMSE,n The interpolation matrix used in the linear minimum mean square error estimation calculation is represented by the true channel coefficient vector h. m,n The receive vector is located at the k-th data subcarrier. Based on the calculated estimated channel matrix And omitted Construct the following equivalent relationship: y=H LMMSE x+z,z=ΔHx+w, The variance of the white noise vector w is: Based on row index d k and column indexes From the interpolation matrix W LMMSE,n Extract column vector w LMMSE,n And calculate the matrix The second moment R of the equivalent noise vector z z =E{zz H } is represented as: Among them, the average signal energy E s =1 / N t I is the identity matrix, w k′ and Corresponding to vectors w LMMSE,n and the elements in matrix M, Indicates that the main diagonal elements are diagonal matrix, Related information R hh , The definition is as follows: in, The above-mentioned relevant information was extracted using the MIMO-OFDM channel correlation matrix.

3. The method for detecting dual-drive MIMO-OFDM signals using a supernetwork-assisted data model according to claim 1, characterized in that: In the graph neural network enhancement part of the t-th submodule, a pairwise Markov random field model is used to represent the latent function relationship between variable nodes and factor nodes, and factor features are extracted. in and They refer to The column j, For the first submodule t=1, the variable eigenvector The calculation is based on the latent function relationship and the adjustable parameter Ω provided by the hypernetwork. hyper Includes weight matrix W hyper and bias vector b hyper The eigenvectors of the variables are calculated as follows: Feature vector Length N u In the l=1,...,Lth iteration, the features With variable eigenvectors Input factor nodes are aggregated through a multilayer perceptron network 1, and output messages are generated. information Prior information related to the likelihood function After integration by the gated loop unit, the feature vector at the variable node is updated. Iterate L times to complete message passing between variable nodes and factor nodes, and finally output the feature vector. Computational graph neural network-enhanced posterior probability inference 4. The method for detecting dual-drive MIMO-OFDM signals using a supernetwork-assisted data model according to claim 1, characterized in that: The complete neural network consisting of the signal detection network and the supernetwork is trained end-to-end using the mini-batch gradient descent algorithm, and trained for multiple epochs. A mini-batch of the training set is represented as... It is a set consisting of S randomly generated samples, where the transmitted real-field symbol vector As a label, the real number field receives the symbol vector. and the real-domain exact channel matrix The input features of the complete neural network are used; an adaptive momentum estimation optimizer is selected to optimize the trainable parameter Θ in the neural network, with a learning rate set to 0.001; training uses the cross-entropy loss function, the specific formula of which is: in, As a built-in function, in It takes the value 1 in the first case and 0 in the others. It is the posterior probability inference output by the Tth submodule of the signal detection network, enhanced by a graph neural network.

Images