A low-complexity deep learning MIMO detection method based on partial MAP

Abstract

The application discloses a low-complexity deep learning MIMO detection method based on partial MAP, comprising the following steps: obtaining multiple groups of training and test samples, and performing hierarchical and data preprocessing on the training and test data samples through QR decomposition to obtain K training sublayers and calculate the input and output labels of each training sublayer; training a DNN for generating local LLR in each detection stage of the K training sublayers based on a partial MAP method; performing layer-by-layer inspection from the Kth layer to the first layer, for each detection stage, calculating the local LLR by using the trained DNN, updating the LLR result of the current layer in combination with the prior information transmitted by the previous layer, and transmitting the LLR result to the next layer as prior information until the final LLR result is obtained; inputting the final LLR result into an FEC decoder to perform error correction and obtain a final signal estimation result; and the partial MAP directly outputs log-likelihood ratio soft information, and the combination of the partial MAP and a forward error correction code can obtain better detection performance.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of signal detection, and particularly relates to a low-complexity deep learning MIMO detection method based on partial MAP. BACKGROUND

[0002] Multiple-input multiple-output (MIMO) can realize diversity multiplexing by using signal space dimension, and can multiply improve signal transmission reliability and transmission rate without increasing power and spectrum resources. In a MIMO system, a signal detection technology recovers the original transmitted signal from the received signal transmitted through multiple paths, thereby ensuring communication performance, so the multi-antenna signal detection is of great significance. The classical maximum likelihood (ML) detection method and maximum a posteriori (MAP) detection method can achieve the best detection performance, but the calculation complexity increases exponentially with the number of transmitting antennas, which seriously affects the practical application, so it is urgent to study a low-complexity MIMO detection method. Existing linear detectors, such as minimum mean square error (MMSE) detector and zero forcing (ZF) detector, can significantly reduce the calculation complexity through weighted matrix multiplication, but the detection performance is poor. In order to better balance the calculation complexity and the detection performance, suboptimal methods such as approximate message passing (AMP) and semidefinite relaxation (SDR) are proposed. In addition, detection methods based on successive interference cancellation (SIC) are also widely studied. The performance of SIC is limited by error propagation, but the influence of error propagation can be effectively reduced by a soft estimation method, in which the idea of partial MAP can reduce the MIMO detection to several sub-detection problems, realizing the trade-off between detection performance and calculation complexity.

[0003] At present, deep learning has been widely applied in MIMO signal detection. The existing deep learning signal detection can be divided into data-driven method and model-driven method. The data-driven detector regards the signal detection process as a black box, and can realize the end-to-end mapping from the received signal to the transmitted signal in a complex environment through the automatic feature extraction capability of deep neural networks (DNNs), but a large amount of training data and training time are required. In comparison, the model-driven detection method utilizes domain knowledge and physical models to construct neural networks, and usually increases trainable parameters in traditional iterative detection algorithms, thereby obtaining faster convergence speed. Recently, a hybrid strategy combining model-based systems with data-driven methods is proposed, which utilizes domain knowledge to design network structure, so that learning can be performed on a relatively small data set to achieve better performance. However, the existing deep learning assisted signal detection methods are often based on expansion or improvement of traditional iterative detection algorithms, such as the classic data-driven MIMO detection network DetNet, the model-driven MIMO detection network OAMP-Net and the like, the complexity of which is related to the number of iterations, although the size of the training data and the learning parameters is controlled, but still faces the challenge of large computational complexity. SUMMARY

[0004] Therefore, the present application aims to provide a low-complexity deep learning MIMO detection method based on partial MAP, which can solve the above technical problems.

[0005] To achieve the above purpose, the present application provides the following technical solutions:

[0006] A low-complexity deep learning MIMO detection method based on partial MAP comprises:

[0007] A plurality of training and test samples are obtained, and the training and test data samples are layered and preprocessed through QR decomposition, to obtain K training sub-layers, and the input and output labels of each training sub-layer are calculated;

[0008] Based on the partial MAP method, a DNN for generating local LLR in each detection stage of the K training sub-layers is trained;

[0009] From the Kth layer to the 1st layer, layer-by-layer detection is performed, for each detection stage, the trained DNN is used to calculate the local LLR, the LLR result of the current layer is updated in combination with the prior information transmitted from the previous layer, and is transmitted to the next layer as prior information, until the final LLR result is obtained;

[0010] The final LLR result is input to an FEC decoder for error correction, to obtain the final signal estimation result.

[0011] Further, a plurality of sets of training and test samples are obtained, and the training and test data samples are stratified and preprocessed by QR decomposition to obtain K training sub-layers, including:

[0012] It is assumed that the current detection communication system is equipped with K transmit antennas at the transmitting end and L receive antennas at the receiving end.

[0013] According to the communication system, a plurality of sets of training and test samples are transmitted and received.

[0014] The data received on the lth receive antenna is represented as:

[0015]

[0016] wherein n l ~ CN(0, N0) is the Gaussian white noise on the lth receive antenna, h l，k is the channel coefficient between the kth transmit antenna and the lth receive antenna, s k is the data transmitted by the kth transmit antenna, s k ∈ S, S is a constellation point set, the size of the constellation point set is M, i.e., |S| = M, l ∈ [1,..., L], k ∈ [1,..., K];

[0017] The transmission process of the data received on the receive antenna y is represented in matrix form as:

[0018] y = Hs + n

[0019] wherein H is the channel matrix, s is the data transmitted by the transmit antenna, and n is the Gaussian white noise on the receive antenna;

[0020] According to QR decomposition, H is decomposed as:

[0021] H = QR

[0022] Q is obtained is a sub-matrix composed of the first K columns of Q, and R is obtained is a sub-matrix composed of the first K rows of R, to obtain the matrix representation of each training sub-layer:

[0023]

[0024] According to the equation y l and the equation z, the entire signal detection process is converted into K training sub-layers.

[0025] wherein the equation corresponding to the kth training sub-layer is:

[0026]

[0027] wherein r k，i is the matrix the element in the kth column and ith row of matrix S, s i the data transmitted by the ith transmit antenna, n k the equivalent received AWGN of the kth sub-layer.

[0028] Further, the input and output labels of each training sub-layer are calculated, including:

[0029] Let denote the LLR soft information of signal s i in the kth layer detection, where i∈{k,..., K}, b∈{1,..., B}, and B=log2 M is the number of information bits carried by a transmission symbol;

[0030] Let denote the local LLR of signal s i in the kth layer detection;

[0031] The input and output labels of each training sub-layer are calculated according to the following formula:

[0032]

[0033] where s -i ={s1,..., s i-1 , s i+1 ,..., s K}, v + =[r k，k , r k，k+1 ,..., r k，K ] T , s k =[s k ,..., s K ];

[0034] An extended input set I k，ex containing 2((K-k+1)M+1) real-valued elements is adopted to train and test each layer of DNN:

[0035] I k，ex ={z k , r k，k ⊙S,..., r k，K ⊙S}

[0036] where,

[0037] The input and output labels of each training sub-layer are calculated according to and I k，ex

[0038] ​Further, based on the partial MAP method, the DNN for generating local LLR in each layer of the K training sub-layers is trained, including:

[0039] When K=2, the LLR soft information of the relevant bits in s2 is obtained, and Q(s2) is calculated and used for transmission to the first layer as prior information for its detection; the calculation formula is:

[0040]

[0041] wherein c b，k represents the bth bit of s k , represents the prior probability of c b，k obtained from the previous layer;

[0042] In the detection of the first layer, z1 is processed and calculated:

[0043] z1=r 1，1 s1+r 1，2 s2+n1

[0044] After z1 and the prior probability of c b，k are given, the LLR soft information of c b，k is calculated; the calculation formula is:

[0045]

[0046] wherein Θ b，k is the prior information obtained from the previous layer;

[0047] The local LLR is obtained, which is represented as:

[0048]

[0049] wherein s -k ={s1,..., s k-1 , s k+1 ,..., s K};

[0050] According to Λ b，k and ∑ b，k , the LLR update process of the first layer detection is determined as:

[0051] Λ b，1 =Σ b，1

[0052] Λ b，k =Σ b，2 +Θ b，2 , b=0,..., B-1

[0053] Since

[0054] The local LLR in the 1st layer detection is represented as a function of the following input:

[0055]

[0056] wherein represents taking the real part of a complex number, S N represents the N-th Cartesian product of S;

[0057] The output of the function is represented as:

[0058] O = {∑ 0，1 ,..., ∑ B-1，1 , ∑ 0，2 ,..., ∑ B-1，2}

[0059] According to the input function and the output function, the sub-detection problem of the current layer is regarded as a regression problem with input I and output O, and the DNN is used to solve the regression problem, thereby generating the DNN of the local LLR of the detection stage of the current layer.

[0060] Further, the layer-by-layer detection is performed from the Kth layer to the 1st layer, for each detection stage, the trained DNN is used to calculate the local LLR, the LLR result of the current layer is updated in combination with the prior information transmitted from the previous layer, and is transmitted to the next layer as the prior information, until the final LLR result is obtained, including:

[0061] The updating process of the kth layer detection is:

[0062]

[0063] wherein, is the prior information from the detection result of the previous layer:

[0064]

[0065] After the iteration and updating of the K layers, the final decoding LLR result is obtained

[0066] According to the LLR result, the hard decoding result is directly obtained:

[0067]

[0068] wherein, sign(·) represents a sign function, sign(x) = 1 when x ≥ 0, otherwise sign(x) = -1.

[0069] The present application has the following advantages:

[0070] The application applies a DNN-assisted data-model hybrid driven deep learning method to a partial MAP algorithm, realizes low complexity MIMO detection by means of a hierarchical detection idea of the partial MAP, and proposes a low complexity deep learning signal detection network PMAP-Net based on the partial MAP; wherein, since the partial MAP algorithm can be reduced to multiple sub-detection stages, each layer of detection can be regarded as a regression problem, because a learning method assisted by a DNN can be used to replace the exponential calculation complexity brought by traversal; in addition, unlike most existing MIMO detection methods learning hard bits, the partial MAP adopted by the application directly outputs log-likelihood ratio soft information, and can obtain better detection performance in combination with a forward error correction code, so that low complexity and high detection performance are taken into account.

[0071] Additional advantages, objects, and features of the application will be set forth in part by the description that follows, and in part will become apparent to those skilled in the art upon examination of the following or can be learned from practice of the application. The objects and other advantages of the application can be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings.

[0072] The technical solutions of the application will be further described in detail below with the aid of the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0073] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and serve to explain the application, and do not limit the application. In the drawings:

[0074] Figure 1 A method flowchart of a low complexity deep learning MIMO detection method based on a partial MAP in an embodiment of the application;

[0075] Figure 2 A deep learning signal detection structure diagram based on a partial MAP in a low complexity deep learning MIMO detection method based on a partial MAP in an embodiment of the application;

[0076] Figure 3 A BER performance diagram of a PMAP-Net under a fixed channel in a low complexity deep learning MIMO detection method based on a partial MAP in an embodiment of the application;

[0077] Figure 4 A BER performance diagram of a PMAP-Net under a changing channel in a low complexity deep learning MIMO detection method based on a partial MAP in an embodiment of the application. DETAILED DESCRIPTION

[0078] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0079] like Figure 1 , Figure 2 , Figure 3 and Figure 4 As shown, the present invention proposes a low-complexity deep learning MIMO detection method based on partial MAP, comprising:

[0080] S101, obtain multiple sets of training and test samples, and stratify and preprocess the training and test data samples through QR decomposition to obtain K training sub-layers, and calculate the input and output labels of each training sub-layer;

[0081] S102, based on the partial MAP method, training the DNN of each layer in the K training sub-layers to generate local LLRs during the detection phase;

[0082] S103, perform layer-by-layer inspection from the Kth layer to the 1st layer. For each layer inspection stage, use the trained DNN to calculate the local LLR, combine the prior information passed by the previous layer to update the LLR result of the current layer, and pass it to the next layer as prior information until the final LLR result is obtained;

[0083] S104, obtaining the final LLR result and inputting it into the FEC decoder for error correction to obtain the final signal estimation result;

[0084] Among them, multiple sets of training and test samples are obtained, and the training and test data samples are stratified and preprocessed by QR decomposition to obtain K training sub-layers, including:

[0085] Assume that the communication system used for current detection is equipped with K transmitting antennas at the transmitting end and L receiving antennas at the receiving end;

[0086] transmitting and receiving multiple sets of training and testing samples according to the communication system;

[0087] The data received on the lth receiving antenna is expressed as:

[0088]

[0089] Among them, n l ~CN(0, N0) is the Gaussian white noise on the lth receiving antenna, h l，k is the channel coefficient between the kth transmitting antenna and the lth receiving antenna, s k is the data transmitted by the kth transmitting antenna, s k∈ S, S is a set of constellation points, the size of the set of constellation points is M, that is, |S| = M, l ∈ [1, …, L], k ∈ [1, …, K];

[0090] The transmission process of the received data y on the receiving antenna is expressed in matrix form as:

[0091] y = Hs + n

[0092] Where H is a channel matrix, s is the data transmitted by the transmitting antenna, and n is the Gaussian white noise on the receiving antenna;

[0093] According to QR decomposition, H is decomposed as:

[0094] H = QR

[0095] Obtain Qk is a submatrix composed of the first K columns of Q, and obtain Rk is a submatrix composed of the first K rows of R, and the matrix representation of each training sublayer is obtained:

[0096]

[0097] According to the equation y l and the equation z, the entire signal detection process is converted into K training sublayers;

[0098] Wherein, the equation corresponding to the kth training sublayer is:

[0099]

[0100] Wherein, r k，i is the element in the kth column and the lth row of the matrix s i is the data transmitted by the ith transmitting antenna, n k is the equivalent received Gaussian white noise of the kth sublayer;

[0101] The input and output labels of each training sublayer are calculated, including:

[0102] Let denote the LLR soft information of the signal s l in the kth layer detection, wherein l ∈ {k, …, K}, b ∈ {1, …, B}, and B = log2 M is the number of information bits carried by a transmission symbol;

[0103] Let denote the local LLR of the signal s l in the kth layer detection;

[0104] The following formula is used to calculate:

[0105]

[0106] where s -i = {s1,..,s i-1i ,..,s i+1} and v K = [r k ,..,r k，k ] k，k+1 k，K T , s k = [s k ,..,s K ].

[0107] An extended input set I k，ex containing 2((K-k+1)M+1) real-valued elements is adopted

[0108] I k，ex = {z k , r k，k ⊙S,..,r k，K ⊙S}

[0109] where

[0110] According to and I k，ex , the input and output labels of each training sub-layer are calculated

[0111] Based on the partial MAP method, the DNN for generating local LLRs in the detection stage of each of the K training sub-layers is trained, including:

[0112] When K=2, the LLR soft information of the relevant bits in s2 is obtained, and Q(s2) is calculated and used for transmission to the first layer as the prior information for its detection; the calculation formula is:

[0113]

[0114] where c b，k represents the bth bit of s k , and represents the prior probability of c b，k obtained from the previous layer;

[0115] In the detection of the first layer, z1 is processed and calculated:

[0116] z1 = r 1，1 s1 + r 1，2 s2 + n1

[0117] After z1 and the prior probability of c b，k are given, c​​b，k LLR soft information; the calculation formula is:

[0118]

[0119] wherein Θ b，k is the prior information obtained from the previous layer;

[0120] The local LLR is obtained, denoted as:

[0121]

[0122] wherein s -k = { {s1,..., s k-1 , s k+1 ,..., s K}} ;

[0123] According to Λ b，k and ∑ b，k , the LLR updating process of the 1st layer detection is determined as:

[0124] Λ b，1 = ∑ b，1

[0125] Λ b，k = ∑ b，2 + Θ b，2 , b = 0,..., B-1

[0126] Since

[0127] The local LLR in the 1st layer detection is denoted as a function of the following input:

[0128]

[0129] wherein represents taking the real part of a complex number, and S N represents the Nth Cartesian product of S.

[0130] The output of the function is denoted as:

[0131] O = {∑ 0，1 ,..., ∑ B-1，1 , ∑ 0，2 ,..., ∑ B-1，2}

[0132] According to the input function and the output function, the sub-detection problem of the current layer is regarded as a regression problem with input I and output O, and the DNN is generated based on the DNN to solve the regression problem, thereby generating the DNN of the local LLR in the detection stage of the current layer;

[0133] The layer-by-layer detection is performed from the Kth layer to the 1st layer, for each layer detection stage, the local LLR is calculated by using the trained DNN, the LLR result of the current layer is updated in combination with the prior information transmitted from the previous layer, and is transmitted to the next layer as the prior information, until the final LLR result is obtained, including:

[0134] The updating process of the kth layer detection is as follows:

[0135]

[0136] wherein, is the prior information, which is from the detection result of the previous layer:

[0137]

[0138] After K layers of iteration and updating, the final decoding LLR result is obtained

[0139] The hard decoding result is directly obtained according to the LLR result:

[0140]

[0141] wherein, sign(·) represents a sign function, sign(x) = 1 when x is greater than or equal to 0, otherwise sign(x) = -1;

[0142] The working principle of the above technical solution is as follows: step S1, data preprocessing is performed through QR decomposition, including: preparing N T group training data samples wherein s (j) , H (j) , y (j) are respectively data transmitted on the jth sample transmitting antenna, a channel matrix and data received on the receiving antenna; QR decomposition is performed on the channel matrix of all training data, and the data is converted into K training sub-layers; the input and output labels of each training sub-layer are calculated according to formula (13) and formula (15)

[0143] Step S2, training the DNN network for generating the local LLR in each layer detection, including: each training sub-layer can be independently trained; the DNN of the kth layer detection is a full connection structure with 2((K-k+1)M+1) neurons as input and (K-k+1)B neurons as output, wherein M is the constellation point set size, and B=log2M is the number of bits carried in a symbol. Each DNN network has n hidden layers and adopts a ReLU activation function; a batch processing scheme is adopted, the minimum mean square error (MSE) of the training data is used as the loss function for each batch, and the ADAM optimizer is used to train and optimize the parameters through gradient descent;

[0144] Step S3, after the network is trained, a layer-by-layer detection updating process is performed from the Kth layer to the first layer, including: for each layer detection, calculating the local LLR using the trained deep learning network, and updating the LLR result of the layer using the local LLR and the prior information passed from the previous layer;

[0145] Step S4, inputting the final LLR soft information into the FEC decoder for error correction to obtain the final signal detection result;

[0146] Specifically, considering that the transmitting end of the communication system is equipped with K transmitting antennas, and the receiving end is equipped with L receiving antennas; let s k ∈S, where S is a set of constellation points; let h l，k be the channel coefficient between the kth transmitting antenna and the lth receiving antenna, so the data received on the lth receiving antenna can be represented as formula (1):

[0147]

[0148] where n l ~CN(0, N0) is the Gaussian white noise on the lth antenna; this transmission process can be represented in matrix form as formula (2):

[0149] y = Hs + n

[0150] In detection, first perform QR decomposition on H, represented as formula (3):

[0151] H = QR

[0152] Let be the submatrix composed of the first K columns of Q, be the submatrix composed of the first K rows of R, then formula (4) is obtained:

[0153]

[0154] Therefore, the detection process can be converted into K equations, where the kth equation is represented as formula (5):

[0155]

[0156] Taking K = 2 as an example, in SIC detection, first detect s2 from z2 using the maximum likelihood strategy, and then eliminate s2 as interference before detecting s1 in the next layer; without considering the complexity of QR decomposition, the complexity of the SIC detection algorithm increases linearly with the number of transmitting antennas K; however, the SIC algorithm is affected by error propagation; to solve this problem, consider using a partial MAP method in detection, represented as formula (6):

[0157]

[0158] Where Q(s2) = cf(z2|s2), c>0 is the normalization coefficient; consider Q(s2) as the prior information of s2, then the above formula can be regarded as a partial MAP problem; different from the standard MAP method, the partial MAP method only requires the prior information of part of the signal, rather than the prior information of the entire signal; in the SIC detection process of the entire MIMO system, the z k The relevant detection is called the k-th layer detection of the partial MAP; therefore, the entire detection process is carried out in descending order, that is, from the K-th layer detection to the l-th layer detection; the result of each layer detection is passed to the next layer as part of the prior information of the current detection;

[0159] This partial MAP-based SIC detection process can be divided into multiple sub-stages, thereby reducing computational complexity. Compared with traditional SIC methods, interference cancellation is performed using a soft message passing method. Each detection layer can update the prior information of some signals, thus alleviating the problem of error propagation. In addition, because the detection process at each layer can be regarded as a regression problem, it can be solved using low-complexity deep learning methods.

[0160] Taking K = 2 as an example, we first perform the second layer test to obtain the soft information of the log-likelihood ratios (LLRs) of the relevant bits in s2; based on this layer of test, we can calculate Q(s2) and pass it to the first layer as its prior information for testing; using c b，k Indicates s k Assuming that they are independent of each other, we have formula (7):

[0161]

[0162] in Represents c obtained from the previous layer b，k Prior probability; in the test of the first level, z1 is processed and calculated to obtain formula (8):

[0163] z1=r 1，1 s1+r 1，2 s2+n1

[0164] So, given z1 and c b，k After the prior probability of c b，k The LLR soft information is obtained from formula (9):

[0165]

[0166] where Θb，k is the a priori information obtained from the previous layer, while the first term is the local LLR, which can be expressed as equation (10):

[0167]

[0168] where s -k = {s1,..., s k-11 , s k+1 ,..., s K}; thus the LLR updating process of the 1st layer detection can be summarized as equation (11):

[0169] b，1 =∑ b，1

[0170] b，k =∑ b，2 +Θ b，2 , b = 0,..., B - 1

[0171] Since the local LLR in the 1st layer detection can be expressed as a function of the input (12) as follows:

[0172]

[0173] where denotes taking the real part of a complex number, S N denotes the N-th Cartesian product of S.

[0174] The output of the function is O = {∑ 0，1 ,..., ∑ B-1，1 , ∑ 0，2 ,..., ∑ B-1，2}; thus the sub-detection problem of this layer can be regarded as a regression problem with the input I and the output O, which can be solved by means of DNNs;

[0175] When K > 2, we can generalize the detection process. In the k-th layer detection, we can obtain the LLR soft information of s k ,..., s k from z K ; let denote the LLR soft information of signal s i in the k-th layer detection, where i ∈ {k,..., K}, b ∈ {1,..., B}, and B = log2 M is the number of information bits carried by a transmission symbol; let denote the local LLR of signal s i in the k-th layer detection, which can be calculated by equation (13) as follows:

[0176] ​​

[0177] where v k = [r k，k , r k，k+1 ,..., r k，K ] T , s k = [s k ,..., s K ].

[0178] This calculation process needs to traverse all M K constellation points, which will produce an unbearable amount of calculation as the number of antennas increases; this method regards the calculation process of local LLR as a regression problem, and realizes low-complexity detection through training of a DNN network; similarly, the general input set I k of the kth layer detection consists of 2M K-k+1 real-valued variables, and is expressed as formula (14):

[0179]

[0180] In order to further reduce the number of input variables of DNN training, and thus reduce the training complexity, we use an extended input set I k，ex consisting of 2((K-k+1)M+1) real-valued elements to train and test each layer of DNN, which is expressed as formula (15):

[0181] I k，ex = {z k , r k，k ⊙S,..., r k，K ⊙S}

[0182] where, At this time, the size of the DNN input set grows linearly with K and M, greatly reducing the training cost;

[0183] In the detection of the kth layer, (K-k+1)B local LLRs, i.e. are obtained by a low-complexity DNN deep neural network;

[0184] After the local LLR of the kth layer detection is calculated, the LLR of each signal of the kth layer can be updated in combination with the prior information obtained from the k+1th layer; the updating process of the kth layer detection can be expressed as formula (16):

[0185]

[0186] where the prior information comes from the detection result of the previous layer, which is expressed as (17):

[0187]

[0188] After the iteration and update of K layers, the final decoding LLR result is obtained The hard decoding result (18) can be obtained directly according to the LLR:

[0189]

[0190] Wherein, sign(·) represents a sign function, sign(x) = 1 when x >= 0, otherwise sign(x) = -1;

[0191] And with the help of forward error correction code (FEC), the hard bit information shown in (18) or the LLR soft decoding information The final decision result after error correction can be obtained by inputting the FEC decoder; most existing deep learning-based MIMO detectors are involved in learning hard bit decisions, so they can only directly support hard FEC decoders; and the partial MAP-based deep learning signal detection network PMAP-Net provided by the present application decodes through the update of LLR soft information, so it can be directly combined with soft FEC decoders without special processing; since soft decoding can provide a 1.5-2dB error correction performance gain compared to hard decoding, PMAP-Net can obtain significant performance gain under FEC assistance;

[0192] The beneficial effects of the above technical solutions are: the low-complexity MIMO detector based on partial MAP proposed by the present application adopts a data-model hybrid driven deep learning detection architecture assisted by DNN inference, which decouples the large-scale MIMO detection process into multiple low-dimensional sub-detection problems, wherein each layer of the sub-detection process can be regarded as a regression problem, thereby using DNN learning instead of traversal calculation exponentially growing with the number of antennas; by using an extended input set with a dimension linearly growing with the number of antennas, the training cost and computational complexity of the DNN are greatly reduced; in addition, since the partial MAP algorithm relied on by the present application transmits LLR soft information between layers and constantly updates the prior information, it can alleviate the error propagation problem of the SIC decoding method; at the same time, the DNN network directly learns the LLR soft information rather than the hard bit, so that the detection accuracy can be greatly improved through FEC soft decoding assistance;

[0193] In the simulation, we tested the performance of the PMAP-Net proposed by the present application in a fixed channel and a changing channel with an input antenna number of 4, as shown in Figure 3 and Figure 4It can be seen that the performance of PMAP-Net not only exceeds that of the traditional MMSE and SIC detectors, but also has obvious performance gain compared with DetNet and OMAP-Net which are also based on deep learning; under the variable channel, although the performance advantage of the uncoded PMAP-Net is no longer as obvious as that under the fixed channel at high SNR, the algorithm proposed in the application still performs well at low SNR; the uncoded PMAP-Net can quickly reduce the BER level to 10-4~10-5, which enables the soft FEC decoder to fully play its role, so that the encoding-aided PMAP-Net exhibits good performance advantage.

[0194] Finally, it should be pointed out that the above preferred embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present application.

Claims

1. A low complexity deep learning MIMO detection method based on partial MAP, characterized in that, Comprising: A plurality of groups of training and test samples are acquired, and the training and test data samples are stratified and data preprocessed by QR decomposition to obtain training sub-layers, and the input and output labels of each training sub-layer are calculated. Based on the partial MAP method, training a DNN that generates local LLRs at each layer of the training sub-layers; From The LLR of each layer is calculated by using the trained DNN, and the LLR result of the current layer is updated in combination with the prior information passed by the previous layer, and then passed to the next layer as the prior information until the final LLR result is obtained. Obtaining the final LLR result input to the FEC decoder for error correction to obtain the final signal estimation result; Wherein, a plurality of groups of training and test samples are obtained, and the training and test data samples are stratified and data preprocessed through QR decomposition to obtain a training sublayer, comprising: The communication system transmitting end used for current detection is provided with a root transmitting antenna, and the receiving end is provided with a root receiving antenna; According to the communication system, transmitting and receiving multiple groups of training and test samples; No. The data received on the root receiving antenna is expressed as: in, It is Gaussian white noise on the receiving antenna, For the The first transmitting antenna to the The channel coefficients between the root receiving antennas, For the The data transmitted by the transmitting antenna, , is the constellation point set, and the size of the constellation point set is , , , ; The received data on the receiving antenna The transmission process is represented in matrix form as: wherein H is a channel matrix, s is the data transmitted by the transmit antennas, n is the Gaussian white noise on the receive antennas; Decompose the matrix A according to QR decomposition Decomposition: acquire for the preceding columns of the sub-matrix, and acquire the preceding rows of the sub-matrix, to obtain the matrix representation of each training sub-layer: According to equation and equation the entire signal detection process is converted to training sub-layers; wherein the first training sub-layer corresponds to the equation: wherein the first training sub-layer corresponds to the equation: wherein is a matrix of the column of the element of the is the data transmitted by the is the equivalent received white Gaussian noise of the Calculate the input and output labels of each training sub-layer, including: Let denote the LLR soft information of the signal in the detection of the layer, where is the number of information bits carried by a transmission symbol. Let denote the local LLR of the signal in the layer detection ; According to the following formula: wherein , , ; An extended input set comprising real-valued elements Training and testing of each DNN is performed: wherein , denotes the complex field; According to and The input and output labels of each training sub-layer are calculated ; Based on the partial MAP method, training a DNN for generating local LLRs in each layer of the training sub-layers, comprising: When the LLR soft information of the relevant bits is obtained , and the following formula is used to calculate the LLR soft information of the relevant bits: for delivery to the first layer as prior information for its verification; the calculation formula is: wherein, represents the first bit of the prior probability of from the previous layer; and represents the second bit of the In the Layer 1 test, the following are processed and calculated: performing processing and calculations: After given the prior probability of and the LLR soft information of is calculated; the calculation formula is: wherein is the prior information obtained from the previous layer; Obtain the local LLR, denoted as: wherein ; According to and , the LLR updating procedure for layer 1 detection is determined as: Due to ; The local LLR in the first layer detection is represented as a function of the following input: wherein denotes taking the real part of a complex number, denotes of second Cartesian product; The output of the function is represented as: According to the input function and the output function, the sub-detection problem of the current layer is regarded as a regression problem with input as and output as , and the DNN is used to solve the regression problem and generate the DNN of the local LLR of the detection stage of the current layer. From The LLR of each layer is calculated by using the trained DNN. The LLR result of the current layer is updated by combining the prior information passed by the previous layer and passed to the next layer as prior information until the final LLR result is obtained, including: No. The update process of layer detection is: wherein, is the prior information, from the detection result of the previous layer: After iteration and update of layers, resulting in final decoded LLR results ; According to the LLR result, directly obtain the hard decoding result: wherein denotes the sign function, when , otherwise .​

Images