A MIMO signal detection method based on neural network message passing

Abstract

The present invention discloses a MIMO signal detection method based on neural network message passing, which relates to the field of communication technology and includes the following steps: S1: a base station receives a signal sent by a user terminal; S2: a MIMO system is modeled using a neural network, a neural network architecture is designed based on the signal flow of the MIMO system, and a deep neural network model is designed to simulate hardware defects and multi-user interference; S3: an efficient Bayesian detector MP‑NN based on message passing is developed; S4: a turbo receiver is implemented. The present invention adopts the above method to realize Bayesian signal detection for a communication system with a complex input-output relationship; a MIMO system is modeled using a neural network, a neural network architecture is designed based on the signal flow of the MIMO system, the number of neural network layers and parameters is minimized, a factor graph is used to represent the trained neural network, and an efficient Bayesian signal detector based on message passing is designed using the unitary approximate message passing UAMP algorithm.

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to a MIMO signal detection method based on neural network message transmission. Background Art

[0002] In the field of wireless communications, MIMO (Multiple-Input Multiple-Output) refers to the technology of using multiple antennas to send and receive signals. It is mainly used in the field of mobile communications and can effectively improve system capacity, coverage and signal-to-noise ratio. However, with the increase in the number of users and base station (BS) antennas and the increase in the complexity of received signals, the imbalance of in-phase and quadrature (I / Q) and the nonlinearity of power amplifiers (PA), it has become increasingly difficult to effectively detect signals at the BS. Therefore, how the receiver at the BS can detect the transmitted data signals of all users has become a focus of attention.

[0003] There are two detection methods in total. The first is direct detection, that is, the signal detector is trained directly using pilot signals, where the input is the received signal and the output is the predicted signal. Since the system transfer function S() is complex, direct detection seems to be sensible. In order to deal with nonlinearity, techniques based on polynomials and deep neural networks (DNNs) are used. However, low-order polynomials have limited ability to combat nonlinearity. Although high-order polynomials have better performance, it is difficult to determine the polynomial coefficients due to numerical instability. DNN technology handles nonlinearity more effectively, but it is difficult to train DNNs with a limited number of pilot signals; the second is indirect detection, which uses pilot signals to identify the system function S(), and then develops a signal detector based on the system function. This strategy allows the design of a powerful Bayesian detector, but the implementation of indirect detection is challenging. First, to identify S() with pilot signals, we need to simultaneously estimate the I / Q imbalance and PA nonlinear models as well as the parameters of the MIMO channel, which is a difficult task due to nonlinearity. Second, even if we assume that S() is known, it is still difficult to develop a detector, especially a Bayesian detector, due to the nonlinearity of S().

[0004] Therefore, it is necessary to provide a MIMO signal detection method based on neural network message passing to solve the above problems. Summary of the invention

[0005] The purpose of the present invention is to provide a MIMO signal detection method based on neural network message passing, so as to realize Bayesian signal detection for a communication system with a complex input-output relationship; the MIMO system is modeled by using a neural network, a neural network architecture is designed based on the signal flow of the MIMO system, the number of neural network layers and parameters is minimized, the trained neural network is represented by a factor graph, and an efficient Bayesian signal detector based on message passing is designed by using the unitary approximate message passing UAMP algorithm.

[0006] To achieve the above object, the present invention provides a MIMO signal detection method based on neural network message passing, comprising the following steps:

[0007] S1: The base station receives the signal sent by the user terminal;

[0008] S2: Model the MIMO system using neural networks, design a neural network architecture based on the signal flow of the MIMO system, and design a deep neural network model to simulate hardware defects and multi-user interference;

[0009] S3: Develop an efficient Bayesian detector MP-NN based on message passing;

[0010] S4: Implementation of turbo receiver.

[0011] Preferably, in step S1, characterization modeling is performed around I / Q imbalance and nonlinear distortion of the power amplifier PA during transmission, the number of users is K, the number of antennas at the base station is N, and the signal of user k is expressed as:

[0012]

[0013] Among them, ξ k and k is the two-way unbalance coefficient of I and Q, which is specifically expressed as:

[0014] ξ k =cos(θ k / 2)+jλ k sin(θ k / 2)

[0015] ζ k =λ k cos(θ k / 2)+j sin(θ k / 2)

[0016] λ k and θ k It is the imbalance parameter of amplitude and phase. After the signal is subjected to I / Q imbalance, it enters the PA and is characterized by amplitude distortion and phase distortion:

[0017]

[0018]

[0019] where α a , α φ , β φ 、x sat , σ a , q1 and q2 are model parameters, and the distorted signal is expressed as:

[0020]

[0021] The received signal at time m is expressed as:

[0022] y(m)=Hs(m)+ω(m)

[0023] in, is the MIMO channel matrix, y(m)=[y1(m),y2(m),...,y N (m)] T ,s(m)=f(x a (m)), ω(m) represents the white Gaussian noise vector, and the received signal is rewritten into real form:

[0024]

[0025] Due to the influence of I / Q imbalance and PA nonlinearity, the input-output relationship of the MIMO system is expressed as:

[0026]

[0027] Where S() is the system transfer function.

[0028] Preferably, step S2 specifically includes the following sub-steps:

[0029] 2a: First, a neural network model is established for the transmission signal after I / Q orthogonal imbalance and PA nonlinear distortion. The neural network consists of an input layer, two non-fully connected hidden layers and a fully connected output layer. All user signals are input into the NN. The output of the NN is the predicted received signal. The real and imaginary parts of the signal are separated, making the NN a real-valued NN.

[0030] The real and imaginary parts of the signal are shared by two sub-NNs, which are called sub-NN pairs. 2K sub-NNs are input, and there are K sub-NN pairs, which are indexed by (l, k), where l = 1, 2, k = 1, ..., K. The output of one sub-NN is approximately Another approximation

[0031] The input of the k-th sub-NN pair is expressed as:

[0032]

[0033] The output of the (l, k)th sub-NN is:

[0034]

[0035] in and are the corresponding weight matrices and bias vectors of the sub-NNs. Each sub-NN has an output node. The output of the (l, k)th sub-NN is expressed as:

[0036]

[0037] in, is the output weight of the sub-NN;

[0038] 2b: The second hidden layer and output layer in the neural network are used to model multi-user interference; the real channel matrix H' is modeled using a fully connected network and a linear activation function, and the weights of the fully connected layer are expressed as:

[0039]

[0040] Where W 31 and W 32 is a sub-weight matrix with dimension N×K. The output of the neural network is expressed as:

[0041]

[0042] Where s′(m) is the output vector of 2K sub-NNs, is an output vector of length 2N, where The predicted signal of the nth receiving antenna is expressed as:

[0043]

[0044] 2c: Train the constructed neural network model. The length of the training pilot signal is M0. There are M0 training samples {(p(m), t(m)), m=1,...,M0}, where p(m) and t(m) represent the pilot signal and the received signal respectively. Construct the loss function:

[0045]

[0046] Determine the weight function through back propagation of the trained neural network and the bias vector The following model is obtained:

[0047]

[0048] Where NN() represents the trained neural network and ω'(m) represents the noise vector.

[0049] Preferably, step S3 includes the following sub-steps:

[0050] 3a: In the data transmission stage, Bayesian reasoning is performed on the transmitted signal based on the trained NN. The Bayesian method converts the deep neural network graph structure trained by S2 into a probabilistic graph model.

[0051] y′=W 3 s′+ω′

[0052] Performing unitary transformation on the above formula yields:

[0053]

[0054] Where Φ=U H W 3 =ΛV, noise has the same distribution as ω′, U is a unitary matrix, the accuracy of the noise is represented by ∈, and the auxiliary vector is defined as:

[0055] z=Φs′

[0056] The joint distribution of the send and receive models is expressed as:

[0057] p(x,z,s′,∈∣r)∝p(∈)p(r∣z,∈)p(z∣s′)p(s′∣x)p(x)

[0058] Among them, ∈ represents the accuracy of additive Gaussian white noise, the local function in the formula is called a factor, according to the factor decomposition of the global function, the factor node is represented by a square, the variable node is represented by a circle, and the graph connecting the variable node and the factor node is called a factor graph;

[0059] 3b: Develop a message passing algorithm UAMP based on the transformed factor graph. The iterative process includes several rounds of forward and backward recursion. UAMP is used to process the densely connected parts of the graph; belief propagation BP and variational message passing VMP are used to process various factor nodes. The message transmitted from node A to node B is expressed by m A→B (c) indicates that multiple iterations are performed until the algorithm converges.

[0060] Preferably, in step S4, the turbo receiver is composed of a Bayesian detector and a SISO decoder, and the turbo receiver exchanges a posteriori log-likelihood ratios (LLRs) of coded bits.

[0061] Preferably, step S4 includes the following sub-steps:

[0062] 4a: The Bayesian detector uses the extrinsic LLRs from the SISO decoder as prior information to compute the posterior LLRs for each coded bit;

[0063] 4b: The SISO decoder uses code constraints to refine the LLRs.

[0064] Preferably, in step 4a, the information bits are first encoded and interleaved before mapping, and each symbol Mapped from a subsequence of the coded bit sequence, the subsequence being It means that each λ a Corresponding to the length log|a| binary sequence, given by express.

[0065] Therefore, the present invention adopts the above-mentioned MIMO signal detection method based on neural network message passing, which has the following beneficial effects:

[0066] (1) The present invention uses a neural network to characterize and model the MIMO system receiving system, thereby solving the problems of I / Q imbalance, PA nonlinearity, and multi-user interference at the receiving end of the MIMO system.

[0067] (2) The signal flow-based neural network architecture designed by the present invention reduces the number of layers and parameters of the neural network, thereby achieving efficient training.

[0068] (3) The present invention represents the neural network as a factor graph and develops a UAMP-based message passing inference algorithm (MP-NN), realizing Bayesian signal detection based on neural network.

[0069] (4) The MP-NN Bayesian detector proposed in the present invention can work with a soft-input soft-output (SISO) decoder to produce a more powerful turbo receiver.

[0070] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a flowchart of an implementation of a MIMO signal detection method based on neural network message transmission of the present invention;

[0072] Figure 2 It is a deep neural network model of a MIMO signal detection method based on neural network message passing of the present invention;

[0073] Figure 3 It is a factor graph of a neural network modeling system of a MIMO signal detection method based on neural network message passing of the present invention;

[0074] Figure 4 It is a block diagram of a turbo receiver of a MIMO signal detection method based on neural network message passing of the present invention;

[0075] Figure 5 It is a performance comparison diagram of different training lengths of a MIMO signal detection method based on neural network message passing of the present invention;

[0076] Figure 6 It is a comparison of modeling performance of a MIMO signal detection method based on neural network message passing of the present invention;

[0077] Figure 7 It is a comparison diagram of the bit error rate performance of the receiver in the coding system of the MIMO signal detection method based on neural network message passing of the present invention;

[0078] Figure 8 It is a message passing specific iterative algorithm of a MIMO signal detection method based on neural network message passing of the present invention; DETAILED DESCRIPTION

[0079] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0080] Unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0081] Example

[0082] like Figure 1-Figure 4 As shown, the present invention provides a MIMO signal detection method based on neural network message passing, comprising the following steps:

[0083] S1: The base station receives the signal sent by the user end. In step S1, the I / Q imbalance and nonlinear distortion of the power amplifier PA in the transmission process are characterized and modeled. The number of users is K, the number of antennas at the base station is N, and the signal of user k is expressed as:

[0084]

[0085] Among them, ξ k and k is the two-way unbalance coefficient of I and Q, which is specifically expressed as:

[0086] ξ k =cos(θ k / 2)+jλ k sin(θ k / 2)

[0087] ζ k =λ k cos(θ k / 2)+j sin(θ k / 2)

[0088] λ k and θ k It is the imbalance parameter of amplitude and phase. After the signal is subjected to I / Q imbalance, it enters the PA and is characterized by amplitude distortion and phase distortion:

[0089]

[0090]

[0091] where α a , α φ , β φ 、x sat , σ a , q1 and q2 are model parameters, and the distorted signal is expressed as:

[0092]

[0093] The received signal at time m is expressed as:

[0094] y(m)=Hs(m)+ω(m)

[0095] in, is the MIMO channel matrix, y(m)=[y1(m),y2(m),...,y N (m)] T ,s(m)=f(x a (m)), ω(m) represents the white Gaussian noise vector, rewriting the received signal into real form:

[0096]

[0097] Due to the influence of I / Q imbalance and PA nonlinearity, the input-output relationship of the MIMO system is expressed as:

[0098]

[0099] Where S() is the system transfer function.

[0100] S2: Modeling the MIMO system using a neural network, designing a neural network architecture based on the signal flow of the MIMO system, and designing a deep neural network model to simulate hardware defects and multi-user interference; step S2 specifically includes the following sub-steps;

[0101] 2a: First, a neural network model is established for the transmission signal after I / Q orthogonal imbalance and PA nonlinear distortion. The neural network consists of an input layer, two non-fully connected hidden layers and an output layer. All user signals are input into the NN. The output of the NN is the predicted received signal. The real and imaginary parts of the signal are separated, making the NN a real-valued NN.

[0102] The real and imaginary parts of the signal are shared by two sub-NNs, which are called sub-NN pairs. 2K sub-NNs are input, and there are K sub-NN pairs, which are indexed by (l, k), where l = 1, 2, k = 1, ..., K. The output of one sub-NN is approximately Another approximation

[0103] The input of the k-th sub-NN pair is expressed as:

[0104]

[0105] The output of the (l, k)th sub-NN is:

[0106]

[0107] in and are the corresponding weight matrices and bias vectors of the sub-NNs. Each sub-NN has an output node. The output of the (l, k)th sub-NN is expressed as:

[0108]

[0109] in, is the output weight of the sub-NN.

[0110] 2b: The second hidden layer and output layer in the neural network are used to model multi-user interference; the real channel matrix H' is modeled using a fully connected network and a linear activation function, and the weights of the fully connected layer are expressed as:

[0111]

[0112] Where W 31and W 32 is a sub-weight matrix with dimension N×K. The output of the neural network is expressed as:

[0113]

[0114] Where s′(m) is the output vector of 2K sub-NNs, is an output vector of length 2N, where The predicted signal of the nth receiving antenna is expressed as:

[0115]

[0116] 2c: Train the constructed neural network model. The length of the training pilot signal is M0. There are M0 training samples {(p(m), t(m)), m=1,...,M0}, where p(m) and t(m) represent the pilot signal and the received signal respectively. Construct the loss function:

[0117]

[0118] Determine the weight function through back propagation of the trained neural network and the bias vector The following model is obtained:

[0119]

[0120] Where NN() represents the trained neural network and ω'(m) represents the noise vector.

[0121] S3: Develop an efficient Bayesian detector MP-NN based on message passing; Step S3 includes the following sub-steps,

[0122] 3a: In the data transmission stage, Bayesian reasoning is performed on the transmitted signal based on the trained NN. The Bayesian method converts the deep neural network graph structure trained by S2 into a probabilistic graph model.

[0123] y′=W 3 s′+ω′

[0124] Performing unitary transformation on the above formula yields:

[0125]

[0126] Where Φ=U H W 3 =ΛV, noise has the same distribution as ω′, U is a unitary matrix, the accuracy of the noise is represented by ∈, and the auxiliary vector is defined as:

[0127] z=Φs′

[0128] The joint distribution of the send and receive models is expressed as:

[0129] p(x,z,s′,∈∣r)∝p(∈)p(r∣z,v)p(z∣s′)p(s′∣x)p(x)

[0130] Among them, ∈ represents the additive Gaussian white noise accuracy, the local function in the formula is called a factor, according to the factor decomposition of the global function, the factor nodes are represented by squares, the variable nodes are represented by circles, and the graph connecting the variable nodes and the factor nodes is called a factor graph; the symbols are introduced in Table 1 to show the correspondence between the factor labels and the corresponding distributions they represent.

[0131] Table 1

[0132]

[0133] 3b: Develop a message passing algorithm UAMP based on the transformed factor graph. The iterative process includes several rounds of forward and backward recursion. UAMP is used to process the densely connected parts of the graph; belief propagation BP and variational message passing VMP are used to process various factor nodes. The message transmitted from node A to node B is expressed by m A→B (c) indicates that it is a function of c. The specific steps are as follows Figure 8 As shown, multiple iterations are performed until the algorithm converges.

[0134] S4: Implementation of turbo receiver.

[0135] In step S4, the turbo receiver is composed of a Bayesian detector and a SISO decoder, and the turbo receiver exchanges the posterior log-likelihood ratios (LLRs) of the coded bits.

[0136] Step S4 includes the following sub-steps:

[0137] 4a: The Bayesian detector uses the extrinsic LLRs from the SISO decoder as prior information to compute the posterior LLRs for each coded bit;

[0138] 4b: The SISO decoder uses code constraints to refine the LLRs.

[0139] In step 4a, the information bits are first encoded and then interleaved before mapping. Mapped from a subsequence of the coded bit sequence, the subsequence being It means that each λ a Corresponding to the length log|a| binary sequence, given by After obtaining the trained deep neural network, Bayesian reasoning is performed based on the neural network to correct the multi-source nonlinear distortion and achieve stable and efficient communication at the MIMO receiving end.

[0140] Embodiment 1

[0141] The BS is equipped with a uniform linear antenna array, N = 10 and K = 5. The modulation scheme used is 16-QAM, where the channel vector h between the kth user and the N receiving antennas is k It is expressed as:

[0142]

[0143] where θ kq is the incident angle of the qth path, θ kq The value range is (-0.5π, 0.5π], and the direction vector is N in length

[0144]

[0145] The antenna spacing is d, λ is the wavelength of the carrier, d = λ / 2, Q k is the number of paths for user k, Q = 3, β kq is the complex gain of the qth path, following a Gaussian distribution with zero mean and unit variance, and the parameter for the PA nonlinearity is α a =4.65,α φ =2560,β φ =0.114,σ a =0.81,x sat =0.58, q1=2.4, q2=2.3; the parameters for I / Q imbalance are: θ k =4°, λ k =0.05; the signal-to-noise ratio is the ratio of the signal transmission power to the noise power at the receiver, denoted as

[0146] Three simulation experiments were carried out respectively. Experiment 1: The proposed MP-NN detector was compared with the direct detector based on DNN and the direct detector based on RMP under different training length conditions, and the performance of different types of detectors was tested to verify the high performance of the proposed MP-NN detector, which can effectively realize signal detection; Experiment 2: The neural network modeling method and the traditional polynomial modeling method were compared under different signal-to-noise ratios to verify the advantages of the proposed neural network in modeling; Experiment 3: The performance of the system with a turbo receiver and that without a turbo receiver was compared. The experiment proved that the system with a turbo receiver enhanced the performance of the detector.

[0147] Experiment 1: The proposed MP-NN detector is compared with a direct detector based on a deep neural network (DNN) and a direct detector based on a real-valued memory polynomial (RMP) under different training length conditions, where the parameters of the sub-NN pairs are bound. The simulation results are shown in Figure 5 As shown in Figure 3, the proposed MP-NN detector always performs better than other detectors. When the training length is reduced from 3000 to 500, the performance of MP-NN changes only slightly, but for the direct detector D-RMP based on real-valued memory polynomial and the direct detector D-DNN based on deep neural network, the impact of training length on performance is significant, and their performance decreases rapidly with the reduction of training length. It can be seen that the proposed detector can be trained more efficiently and the Bayesian detector is more powerful.

[0148] Experiment 2: The modeling performance of the neural network modeling method and the traditional polynomial modeling method under different signal-to-noise ratios is evaluated using the normalized mean square error (NMSE). The traditional polynomial-based technology is modeled using 5th and 12th order polynomials. The simulation results are shown in Figure 2. Figure 6 shown.

[0149] Although the use of higher-order polynomials can improve the modeling capabilities of polynomial techniques, it leads to difficulties in determining the polynomial parameters due to numerical instability. Figure 6 It can be seen that the proposed neural network significantly outperforms the fifth-order polynomial, indicating that the proposed network has better modeling ability. When the polynomial order is increased to 12, the performance of the polynomial method becomes very poor due to numerical instability and overfitting. The results demonstrate the advantages of the proposed neural network in modeling.

[0150] Experiment 3: The performance of the system with and without turbo receiver is compared to evaluate the performance of the detector in the coding system. A 2 / 3 rate convolutional code with a generator [23, 35], a random interleaver and 16-QAM modulation are used, where Gray mapping is used in symbol mapping and the BCJR algorithm is used to implement the SISO decoder. Since it is not possible to directly implement turbo receivers based on direct detectors D-RMP and D-DNN, non-iterative receivers are implemented for them, where the output of the detector is fed to the Viterbi decoder after the hard decision. The bit error rate (BER) is used to evaluate the performance of different receivers. The bit error rate performance results of different receivers are shown in Figure 3. Figure 7 shown.

[0151] exist Figure 7It can be clearly seen that the performance of the D-RMP receiver is slightly better than that of the D-DNN receiver, but the performance of the proposed MP-NN detector is significantly better than the other two receivers. The results show that the proposed MP-NN Bayesian detector working together with the soft-input soft-output (SISO) decoder can produce a more powerful turbo receiver.

[0152] In summary, the simulation experiments have verified the correctness, effectiveness and reliability of the present invention.

[0153] Therefore, the present invention adopts the above-mentioned MIMO signal detection method based on neural network message passing to realize Bayesian signal detection for communication systems with complex input-output relationships; uses neural networks to model the MIMO system, designs a neural network architecture based on the signal flow of the MIMO system, minimizes the number of neural network layers and parameters, uses factor graphs to represent the trained neural network, and uses the unitary approximate message passing (UAMP) algorithm to design an efficient Bayesian signal detector based on message passing.

[0154] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A MIMO signal detection method based on deep neural network, characterized in that: Contains the following steps: S1: The base station receives the signal sent by the user terminal; S2: Use neural networks to model MIMO systems, design neural network architecture based on the signal flow of MIMO systems, and design a deep neural network model to simulate hardware defects and multi-user interference; S3: We develop an efficient message-passing-based Bayesian detector MP-NN by representing the trained neural network as a factor graph and leveraging approximate message passing. S4: Implementation of turbo receiver; Step S2 specifically includes the following sub-steps: 2a: First, a neural network model is established after the transmission signal is distorted due to I / Q orthogonal imbalance and PA nonlinearity. The neural network consists of an input layer, two non-fully connected hidden layers and an output layer. All user signals are input into the neural network NN. The output of NN is the predicted received signal, in which the real and imaginary parts of the signal are separated, making NN a real-valued NN. The real and imaginary parts of the signal are shared by two sub-NNs, which are called sub-NN pairs. The input is 2K sub-NNs, and there are K sub-NN pairs in total, which are indexed by (l, k), where l = 1, 2, k = 1, … ,K, the output of one of the sub-NNs is approximately Another approximation s(m) is the distorted signal, Represents the operation of the reality department, represents the operation of finding the imaginary part; The input of the k-th sub-NN pair is expressed as: The output of the (l, k)th sub-NN is: in and is the corresponding weight matrix and bias vector of the sub-NN, g1( ) represents the first hidden layer activation function of the neural network NN, and each sub-NN has an output node. The output of the (l, k)th sub-NN is expressed as: in, is the output weight of the sub-NN; 2b: The second hidden layer and output layer in the neural network are used to model multi-user interference. Since the real MIMO channel model is a linear operation, it is modeled as a real channel matrix H′ using a fully connected network and a linear activation function. The weight of the fully connected layer is expressed as: Where W 31 and W 32 is a sub-weight matrix with dimension N×K. At this time, the output of the neural network is expressed as: Where s′(m) is the output vector of 2K sub-NNs, is an output vector of length 2N, where Any element of this vector is represented by v l，n , where l∈{1, 2}, n∈(1, 2, 3, ..., N}. Therefore, the predicted signal of the nth receiving antenna is expressed as: 2c: Train the constructed deep neural network model. The length of the training pilot signal is M0, that is, it consists of M0 training samples {(p(m), t(m)), m=1,...,M0}, where p(m) and t(m) represent the pilot signal and the corresponding received signal respectively; construct the loss function: Determine the weight function through back propagation of the trained neural network and the bias vector The following model is obtained: Where NN(·) represents the trained neural network, ω′(m) represents the noise vector, taking into account the training and modeling errors; Step S3 specifically includes the following sub-steps: 3a: During the data transmission phase, Bayesian reasoning is performed on the transmitted signal based on the trained NN; the Bayesian method requires converting the deep neural network graph structure trained by S2 into a probabilistic graph model; y′=W 3 s′+ω′ For simplicity, the time index m is discarded and the above formula is transformed into: Where Φ=U H W 3 =ΛV, noise has the same distribution as ω′, U is a unitary matrix, and the auxiliary vector is defined as: z=Φs′ The joint distribution of the send and receive models is expressed as: p(x,z,s',∈|r)∝p(∈)p(r|z,∈)p(z|s′)p(s'|x)p(x) Where ∈ represents the accuracy of additive Gaussian white noise, and the local function in the formula is also called a factor. According to the factor decomposition of the global function, the factor nodes are represented by squares, the variable nodes are represented by circles, and the graph connecting the related variable nodes and factor nodes is called a factor graph; 3b: Develop a message passing algorithm UAMP based on the factor graph converted from the trained neural network. The iterative process includes several rounds of forward and backward recursion, and use UAMP to process the densely connected parts of the graph; Belief propagation BP and variational message passing VMP are used to process various factor nodes. The message transmitted from node A to node B is expressed by m A→B (c) indicates that it is a function of c and is iterated multiple times until the algorithm converges; The specific implementation process of step S4 is as follows: The turbo receiver consists of a UAMP-based Bayesian detector and a SISO decoder. The Bayesian detector uses the external log-likelihood ratio LLR from the SISO decoder as prior information to calculate the external LLRs for each coded bit. The SISO decoder uses the code constraints to refine the LLR. In the turbo system, the information bits are first encoded and then interleaved before mapping. Each symbol x k ∈A=[λ1,...,λ |A| ] is mapped from a subsequence of the coded bit sequence, which is composed of It means that each λ a Corresponding to the length log|a| binary sequence, given by express.

2. The MIMO signal detection method based on deep neural network according to claim 1, characterized in that: In step S1, characterization modeling is performed around I / Q imbalance and nonlinear distortion of the power amplifier PA during transmission, assuming that the number of user terminals is K and the number of base station antennas is N; The signal of user k is expressed as: Among them, ξ k and k is the two-way unbalance coefficient of I and Q, which is specifically expressed as: x k =cos(θ k / 2)+jλ k sin(θ k / 2) g k =λ k cos(θ k / 2)+j sin(θ k / 2) λ k and θ k is the imbalance parameter of amplitude and phase. After the signal is subjected to I / Q imbalance, it enters PA. j represents the imaginary unit of the complex number. and phase distortion Characterization: where α φ , β φ 、x sat , σ a , q1 and q2 are model parameters, and the distorted signal is expressed as: The received signal at time m is expressed as: y(m)=Hs(m)+ω(m) in, is the MIMO channel matrix, y(m)=[y1(m),y2(m),...,y N (m)] T is the received signal vector of length N, s(m)=f(x a (m)), ω(m) represents the white Gaussian noise vector, and the received signal is rewritten into real form: in, Represents the operation of the reality department, Represents the imaginary part operation. Due to the combined effects of I / Q imbalance and PA nonlinearity, the input-output relationship of the MIMO system is expressed as: Where S(.) is the system transfer function.

Images