Deep learning-based breadth-first tree search signal detection method and system

Abstract

The invention provides a breadth-first tree search signal detection method and system based on deep learning, and the method comprises the steps: obtaining receiving data with noise through an antenna array, transmitting a modulation signal through a transmitting end, and obtaining a transmitting signal, a receiving signal and a channel matrix through a receiving end and a wireless channel; performing orthogonal triangular decomposition on the channel matrix, performing optimization deformation on a multiple-input-multiple-output detection model to obtain an equivalent channel matrix and a receiving signal, and splicing the channel matrix and the receiving signal to form a feature matrix; inputting the feature matrix into a pre-trained neural network, and outputting an initial detection result of a sending signal through the pre-trained converged neural network; and obtaining a search threshold value of each layer of a breadth-first tree search algorithm according to the initial detection result, and executing a tree search process based on the search threshold values to obtain a final detection result of the sent signal. According to the invention, the problems of high complexity and poor detection performance of existing signal detection are solved.

Description

Technical Field

[0001] The present invention relates to the field of signal detection technology, and in particular to a breadth-first tree search signal detection method and system based on deep learning. Background Art

[0002] Multiple-input multiple-output (MIMO) technology is the core technology for improving channel capacity and spectrum efficiency, and is also the key technology for existing wireless mobile communications. The transmitter and receiver are equipped with multiple antennas, which increases the difficulty of signal processing. Signal detection has also become an important problem to be solved in MIMO systems. The correctness of signal detection directly determines the quality of communication.

[0003] Traditional signal detection methods are mainly divided into two categories: linear detectors and nonlinear detectors. Traditional linear detectors mainly include zero forcing (ZF) detection and minimum mean square error (LMMSE) detection, etc., which have low detection complexity but poor detection performance; nonlinear detection includes maximum likelihood (ML) detection and other methods, which have high detection accuracy but high complexity. In particular, the ML detection algorithm has the best detection performance and is the performance lower bound of the detection algorithm, but it uses a brute force search method, and its complexity increases exponentially with the increase of the number of antennas, modulation methods, etc. Therefore, researchers are committed to seeking a detector that can provide similar detection performance to ML and can greatly reduce the detection complexity of ML. Since the sphere decoding method (SD) can meet the above two conditions, it has been widely studied by researchers.

[0004] The traditional SD method mainly consists of two search methods, namely the depth-first and breadth-first tree search methods. In the depth-first tree search process, there are two search directions, from top to bottom and from bottom to top. The method starts from the root node and searches only one node in each layer. After searching the node, the given initial radius and other information are used to calculate the searchable node range of the lower layer, and a child node of the node is searched within the range until the last layer is searched. Then, the searchable node of the upper layer is returned, and the search is continued in this way until the root node is returned. Since the search range of each layer is determined according to the search results of the previous layer, the search range is dynamically variable, so the complexity of the depth-first tree search method is fluctuating. Although the depth-first method can have a lower complexity under the condition of ensuring excellent detection performance in low-noise scenarios, in large-scale MIMO and high-order modulation scenarios, this method is often difficult to meet the requirements due to its high complexity. In the breadth-first tree search process, there is only one search direction, that is, searching from top to bottom. Compared with the depth-first strategy, the breadth-first strategy has similar performance, but because the breadth-first method can store information of multiple nodes at the same time and search multiple nodes in parallel, its search efficiency is higher.

[0005] In recent years, deep learning (DL) has been widely used in many fields such as computer vision, automatic speech recognition and wireless communication. The main advantages of DL are its excellent learning ability, excellent generalization characteristics and low computational complexity. Therefore, using DL to reduce the complexity of the breadth-first tree search algorithm while ensuring the detection accuracy of the SD algorithm is a very valuable direction. Summary of the invention

[0006] The present invention provides a breadth-first tree search signal detection method and system based on deep learning, which are used to solve the problems of high complexity and poor detection performance of existing signal detection.

[0007] The present invention provides a breadth-first tree search signal detection method based on deep learning, comprising:

[0008] The receiving data with noise is obtained through the antenna array, the modulated signal is sent through the transmitting end, and the transmitting signal, the receiving signal and the channel matrix are obtained through the receiving end and the wireless channel respectively;

[0009] Performing orthogonal triangular decomposition on the channel matrix, optimizing and deforming the multi-input multi-output detection model, obtaining equivalent channel matrices and received signals, and splicing them to form a feature matrix;

[0010] Inputting the feature matrix into a pre-trained neural network, and outputting an initial detection result of the transmitted signal through the pre-trained neural network after convergence;

[0011] A search threshold value of each layer of a breadth-first tree search algorithm is obtained according to the initial detection result, and a tree search process is performed based on the search threshold value to obtain a final detection result of the transmitted signal.

[0012] According to a breadth-first tree search signal detection method based on deep learning provided by the present invention, the method obtains noisy received data through an antenna array, sends a modulated signal through a transmitting end, and obtains a transmitted signal, a received signal and a channel matrix through a receiving end and a wireless channel, respectively, specifically comprising:

[0013] The norm value of each column of the channel matrix obtained by the antenna array is sorted from small to large;

[0014] The signals sent by the transmitter are sorted according to the sorting index to determine the detection order of the signals.

[0015] According to a breadth-first tree search signal detection method based on deep learning provided by the present invention, the channel matrix is ​​orthogonally decomposed, the multi-input multi-output detection model is optimized and deformed, and an equivalent channel matrix and received signal are obtained, which are spliced ​​to form a feature matrix, specifically including:

[0016] Perform orthogonal triangular decomposition on the channel matrix, and optimize and transform the multi-input multi-output detection model based on the orthogonal triangular decomposition result;

[0017] The equivalent received signal and the upper triangular matrix after orthogonal triangular decomposition are combined into a characteristic matrix.

[0018] According to a breadth-first tree search signal detection method based on deep learning provided by the present invention, the neural network construction method is:

[0019] The neural network includes a plurality of deep neural network blocks, each transmitted signal is detected by using a deep neural network block, and the detection probability of all transmitted signals is obtained by superimposing the plurality of deep neural network blocks;

[0020] Each deep neural network block includes five neural network layers. The number of neurons in the input layer and hidden layer is 256. The activation function used is the Leaky ReLu function. The output layer has the same number of neurons as the modulation order, and the Softmax function is used as the activation function.

[0021] Adjacent deep neural network blocks are connected, and the output of each layer of the upper deep neural network block is used as the input of the next layer of the next deep neural network block, thereby constructing a neural network.

[0022] According to a breadth-first tree search signal detection method based on deep learning provided by the present invention, the feature matrix is ​​input into a pre-trained neural network, and the initial detection result of the transmitted signal is output through the neural network after the pre-training convergence, which specifically includes:

[0023] Inputting the feature matrix into a pre-trained neural network, and outputting a soft decision probability of each transmitted signal;

[0024] The detection probability of all transmitted signals is obtained according to the output of the deep neural network block corresponding to each transmitted signal, and the category with the highest probability is the initial detection result of the transmitted information.

[0025] According to a breadth-first tree search signal detection method based on deep learning provided by the present invention, a search threshold value of each layer of the breadth-first tree search algorithm is obtained according to the initial detection result, and a tree search process is performed based on the search threshold value to obtain a final detection result of the transmitted signal, which specifically includes:

[0026] Based on the obtained initial detection results, the accumulated distance of each layer is calculated to obtain the search threshold value of each layer of the tree search;

[0027] Nodes whose accumulated distance is greater than the threshold value are pruned, and nodes whose accumulated distance is less than or equal to the threshold value are retained to obtain candidate nodes;

[0028] In the search of the next layer, the child nodes of all candidate nodes are expanded, and the candidate node set of the current layer is obtained according to the accumulated distance calculated by the child nodes and the search threshold of the current layer;

[0029] When the search reaches the last layer, the node with the smallest cumulative distance is the optimal node, and all the nodes passed by the path from the root node to the optimal node are the final detection results.

[0030] The present invention also provides a breadth-first tree search signal detection system based on deep learning, the system comprising:

[0031] A signal receiving module is used to obtain noisy receiving data through an antenna array, send a modulated signal through a transmitting end, and obtain a transmitting signal, a receiving signal and a channel matrix through a receiving end and a wireless channel respectively;

[0032] A feature matrix generation module is used to perform orthogonal triangular decomposition on the channel matrix, optimize and deform the multi-input multi-output detection model, obtain equivalent channel matrices and received signals, and splice them to form a feature matrix;

[0033] An initial detection module, used to input the feature matrix into a pre-trained neural network, and output an initial detection result of the transmitted signal through the pre-trained neural network after convergence;

[0034] The final detection module is used to obtain the search threshold value of each layer of the breadth-first tree search algorithm according to the initial detection result, and perform a tree search process based on the search threshold value to obtain a final detection result of the transmitted signal.

[0035] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the deep learning-based breadth-first tree search signal detection method as described in any one of the above-mentioned methods is implemented.

[0036] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the breadth-first tree search signal detection methods based on deep learning as described above.

[0037] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements any of the above-mentioned breadth-first tree search signal detection methods based on deep learning.

[0038] The present invention provides a breadth-first tree search signal detection method and system based on deep learning. Through the breadth-first tree search signal detection technology based on deep learning, the detection system model is optimized by using a sorting strategy, and the received signal and the upper triangular matrix are used as the input of the network, and the true value of the transmitted signal is used as the label of the network training. Through training, the network parameters converge and the loss function value is minimized. The trained network assists the breadth-first tree search process, which can achieve detection accuracy close to that of the traditional sphere decoding algorithm and reduce its detection complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0040] Figure 1 It is one of the flow charts of a breadth-first tree search signal detection method based on deep learning provided by the present invention;

[0041] Figure 2 This is the second flow chart of a breadth-first tree search signal detection method based on deep learning provided by the present invention;

[0042] Figure 3 This is a flowchart of a breadth-first tree search signal detection method based on deep learning provided by the present invention;

[0043] Figure 4 This is a fourth flow chart of a breadth-first tree search signal detection method based on deep learning provided by the present invention;

[0044] Figure 5 This is a flowchart of a breadth-first tree search signal detection method based on deep learning provided by the present invention;

[0045] Figure 6 It is a module connection diagram of a breadth-first tree search signal detection system based on deep learning provided by the present invention;

[0046] Figure 7 It is a schematic diagram of the overall architecture of the neural network provided by the present invention;

[0047] Figure 8 This is the processing process of module A provided by the present invention;

[0048] Fig. 9 It is a schematic diagram showing the performance of the tree search algorithm provided by the present invention and the traditional method as the signal-to-noise ratio changes;

[0049] Fig.10 It is a schematic diagram showing the variation of the number of search nodes of the tree search algorithm provided by the present invention and the traditional method with the signal-to-noise ratio;

[0050] Fig.11 It is a structural schematic diagram of the electronic device provided by the present invention.

[0051] Reference numerals:

[0052] 110: signal receiving module; 120: feature matrix generating module; 130: initial detection module; 140: final detection module;

[0053] 1110: processor; 1120: communication interface; 1130: memory; 1140: communication bus. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] Combine the following Figure 1-Figure 5 The present invention describes a breadth-first tree search signal detection method based on deep learning, comprising:

[0056] S100, obtaining noisy received data through an antenna array, sending a modulated signal through a transmitting end, and obtaining a transmitted signal, a received signal, and a channel matrix through a receiving end and a wireless channel respectively;

[0057] S200, performing orthogonal triangular decomposition on the channel matrix, optimizing and deforming the multi-input multi-output detection model, obtaining an equivalent channel matrix and a received signal, and concatenating them to form a feature matrix;

[0058] S300, inputting the feature matrix into a pre-trained neural network, and outputting an initial detection result of the transmitted signal through the neural network after the pre-training convergence;

[0059] S400, obtaining a search threshold value of each layer of a breadth-first tree search algorithm according to the initial detection result, executing a tree search process based on the search threshold value, and obtaining a final detection result of the transmitted signal.

[0060] The breadth-first tree search algorithm based on deep learning designed in the present invention has detection accuracy comparable to that of the traditional tree search algorithm and has lower computational complexity than the traditional algorithm. It is suitable for signal detection algorithms in fields such as wireless communications under various signal-to-noise ratio conditions.

[0061] The receiving data with noise is obtained through the antenna array, the modulated signal is sent through the transmitter, and the transmission signal, the receiving signal and the channel matrix are obtained through the receiver and the wireless channel respectively, which specifically includes:

[0062] S101, sorting the norm value of each column of the channel matrix obtained by the antenna array from small to large;

[0063] S102: Sort the signals sent by the transmitter according to the sorting index to determine the detection order of the signals.

[0064] In a specific example, in a large-scale transmitting and receiving antenna array scenario, the transmitter sends a modulated signal, the antenna array receiver and the wireless channel respectively obtain the transmitted signal s, the received signal x and the channel matrix H, x = Hs + v; the detection order is rearranged, that is, the channel matrix is ​​arranged from small to large according to the norm value of each column of the channel matrix to obtain the sorted channel matrix.

[0065] The channel matrix is ​​orthogonally decomposed, and the multi-input multi-output detection model is optimized and deformed to obtain an equivalent channel matrix and a received signal, which are then spliced ​​to form a feature matrix, specifically including:

[0066] S201, performing orthogonal triangular decomposition on the channel matrix, and optimizing and deforming the multi-input multi-output detection model based on the orthogonal triangular decomposition result;

[0067] S202, forming a characteristic matrix by combining the equivalent received signal and the upper triangular matrix after orthogonal triangular decomposition.

[0068] In the present invention, QR decomposition is an orthogonal triangular decomposition, and the sorted channel matrix is ​​decomposed by QR to obtain H = QR, and the system model is converted into y = Rs + n, where R is an upper triangular matrix, y = Q T x,n=Q T v.

[0069] The feature matrix is ​​input into a pre-trained neural network, and the initial detection result of the transmitted signal is output by the neural network after the pre-training convergence, specifically including:

[0070] S301, inputting the feature matrix into a pre-trained neural network, and outputting a soft decision probability of each transmitted signal;

[0071] S302. Obtain the detection probability of all transmitted signals according to the output of the deep neural network block corresponding to each transmitted signal, and the category with the highest probability is the initial detection result of the transmitted information.

[0072] In the present invention, the feature matrix composed of y and R is used as the input of the neural network, and the soft decision probability of each transmitted signal is output.

[0073] Among them, the method of establishing the neural network is:

[0074] The neural network includes multiple deep neural network blocks, each transmitted signal is detected by using a deep neural network block, and the detection probability of all transmitted signals is obtained by superimposing multiple deep neural network blocks;

[0075] Each deep neural network block includes five neural network layers. The number of neurons in the input layer and hidden layer is 256. The activation function used is the Leaky ReLu function. The output layer has the same number of neurons as the modulation order, and the Softmax function is used as the activation function.

[0076] Adjacent deep neural network blocks are connected, and the output of each layer of the upper deep neural network block is used as the input of the next layer of the next deep neural network block, thereby constructing a neural network.

[0077] In the present invention, the received signal of the receiving antenna array and the channel state information of the transmitting and receiving antenna array are obtained, and then the training data set and the test data set composed of the received signal and the like are obtained.

[0078] In the training phase, the classification neural network uses the cross entropy loss function. The classification neural network uses an adaptive gradient descent algorithm to update the parameters of the network. After the training is completed, the parameters of the network remain unchanged. The initial value of the detection is obtained using the input of the test set. The classification network performance evaluation is achieved by calculating the error between the initial detection value and the label value. The output of the neural network is used to perform a tree search process, and the symbol error rate between the output of the tree search and the true value of the transmitted signal is calculated to achieve detection performance evaluation.

[0079] Each training session inputs 128 sets of data, and the training is performed until the neural network converges, and the network parameters and model are saved. After the training, the test set is used as the input of the neural network, and the model is directly loaded into the tree search process. The output of the model is used to assist the tree search process and obtain the signal detection results. The training data set is trained under a single signal-to-noise ratio, and the obtained neural network model can be tested under different signal-to-noise ratios.

[0080] Obtaining a search threshold value of each layer of a breadth-first tree search algorithm according to the initial detection result, and performing a tree search process based on the search threshold value to obtain a final detection result of the transmitted signal, specifically including:

[0081] S401, based on the obtained initial detection results, calculate the accumulated distance of each layer, and obtain the search threshold value of each layer of the tree search;

[0082] S402, nodes whose accumulated distance is greater than a threshold value are pruned, and nodes whose accumulated distance is less than or equal to the threshold value are retained, and candidate nodes are obtained;

[0083] S403, in the search of the next layer, expand the child nodes of all candidate nodes, and obtain the candidate node set of the current layer according to the accumulated distance calculated by the child nodes and the search threshold of the current layer;

[0084] S404, when the search reaches the last layer, the node with the smallest cumulative distance is the optimal node, and all the nodes passed by the path from the root node to the optimal node are the final detection results.

[0085] In the present invention, the initial detection results output by the neural network are used to calculate the cumulative distance of each layer in the tree search, that is,

[0086] In the tree search process, nodes within the cumulative distance are retained, and nodes greater than the cumulative distance retain num nodes with the smallest cumulative distance. These nodes serve as expansion nodes of subsequent nodes, that is, candidate nodes; the search process adopts the breadth-first method, and according to the above candidate nodes, the child nodes of all candidate nodes are expanded, and the cumulative distances of all child nodes are calculated; all expanded child nodes are selected, and the candidate nodes of the current layer are obtained according to the cumulative distance of each layer and num output by the neural network; when expanding to the last layer, the node with the smallest cumulative distance is retained, and all nodes passed by the path from the root node to the node are read in turn, which is the final detection result.

[0087] In a specific embodiment, assuming that the number of array antennas at the receiving end is N and the number of array antennas at the transmitting end is K, the expression of the signal received by the array at the receiving end is:

[0088] x=Hs+v,

[0089] Where x is the received signal, H is the channel matrix between the transmit and receive antenna arrays, and v is the noise vector.

[0090] In order to improve the detection accuracy of the signal, the detection order of the transmitted signal is sorted. Therefore, according to the column norm of the channel matrix, it is arranged in order from small to large, and the sorted channel matrix is ​​obtained as follows:

[0091] H o =[ho,1 ,h o,2 ,…,h o,K-1 ,h o,K ],

[0092] Among them, ||h o,1 || 2 ≤||h o,2 || 2 ≤…||h o,K-1 || 2 ≤||h o,K || 2 .

[0093] Sort the symbols at the transmitter according to the channel sorting index, and the system model can be expressed as:

[0094] x=H o s o +v,

[0095] Among them, s o Indicates the sent signal after sorting. For the convenience of subsequent representation, the subscript of this model is omitted.

[0096] Perform QR decomposition on the channel matrix, that is, H = QR, then the MIMO system model can be transformed into:

[0097] y=Rs+n,

[0098] Where y = Q T x,n=Q T v.

[0099] Figure 7 This is the neural network structure diagram used in the present invention, Figure 8 is the neural network architecture of module A. Each signal is detected by a deep neural network DNN. Each DNN block consists of 5 neural network layers. By inputting the feature matrix, each DNN block outputs the classification probability of the symbol. When detecting symbol i, the output of the tth layer of the i-th DNN block can be expressed as:

[0100]

[0101] Where t = 0, 1, ..., 4, represents the input of this layer, and Represent the weight and bias of this layer respectively, Represents the activation function of this layer.

[0102] Define the output of the i-th DNN block as When t=1, that is, the first layer of the DNN block, the input of the network is:

[0103]

[0104] where y i is the i-th element of y, R i, : represents the i-th row element of matrix R.

[0105] represents the classification probability output by the i+1th DNN block. In other words, when t=1, we can get

[0106] When t={1,2,3,4}, the input of the tth layer of the i-th DNN block is:

[0107]

[0108] Among them, con[] represents the concatenation of two matrices, represents the input of the t-1th layer of the i+1th DNN block, represents the input of the t-1th layer of the i-th DNN block.

[0109] When t={0,1,2,3}, Leaky ReLu is used as the activation function, and when t=4, softmax is used as the activation function.

[0110] According to the output of the neural network, the symbol probabilities of all symbols can be obtained as:

[0111]

[0112] For symbol i, its initial detection probability can be obtained as:

[0113]

[0114] Where L represents the modulation order. Therefore, the initial detection results of all symbols are:

[0115]

[0116] After obtaining the initial detection results of each symbol, the cumulative distance can be calculated as:

[0117]

[0118] At the kth level, the recursive form of the accumulated distance is:

[0119]

[0120] in, Expressed as A vector consisting of the kth to Kth elements of .

[0121] Assume that the set of nodes retained in the k+1th layer is:

[0122]

[0123] The cumulative distance of the corresponding nodes is:

[0124]

[0125] Among them, a represents the number of nodes retained in the k+1th layer, and there is

[0126]

[0127] At the kth level, traverse The L child nodes of all nodes in the layer, then the set of candidate nodes retained by this layer can be expressed as:

[0128]

[0129] Where c = La. The accumulated distance of each node can be obtained:

[0130]

[0131] Sort the accumulated distance from small to large, and sort the node set according to the sort index to obtain:

[0132]

[0133]

[0134] Among them, I(i) represents the sorting index.

[0135] according to The size of , find its index value d such that At the same time, define the num parameter and set the pruning threshold of the kth layer to:

[0136]

[0137] in, Indicates the cumulative distance value of the d+numth node.

[0138] According to the pruning threshold, all candidate nodes in the kth layer are screened, and the retained nodes and their corresponding cumulative distances are obtained as follows:

[0139]

[0140]

[0141] When k=1, the final detection result is:

[0142]

[0143] in, Representing a collection The first node of the tree, the path from this node to the root node is the optimal path.

[0144] The loss function used in neural network training is the cross entropy loss function:

[0145]

[0146] Among them, p(s i ) means s i The neural network is trained at 30dB, and the trained model can be tested and used under other SNR conditions. At the same time, the batch size used is 128, and the network is trained until convergence, and the number of iterations is set to 40000. The number of neurons used in the hidden layer is 256, and the number of neurons in the output layer is the modulation order. The ADAM optimization is used to optimize the network. After calculating the loss function value, the network parameters are updated by the back propagation method. When the neural network training is completed, the trained model is loaded into the tree search algorithm to complete the search process and output the final detection result.

[0147] Fig. 9 This is a schematic diagram of the performance of the tree search algorithm provided by the present invention and the traditional method as the signal-to-noise ratio changes, where the simulated channel scenario is a Rayleigh channel, the signal-to-noise ratio range is [0,15]dB, and a simulation is performed every five signal-to-noise ratios. The proposed algorithm sets num=10. As can be seen from the figure, when the transceiver antenna array is (16,16) and QPSK modulation is used, the performance of the tree search algorithm based on the neural network is comparable to that of the traditional M-best algorithm.

[0148] Fig.10 This is a schematic diagram of the number of search nodes of the tree search algorithm provided by the present invention and the traditional method as the signal-to-noise ratio changes, where the simulated channel scenario is a Rayleigh channel, the signal-to-noise ratio range is [0,15]dB, and a simulation is performed every five signal-to-noise ratios. The proposed algorithm sets num=10. As can be seen from the figure, when the transceiver antenna array is (16,16) and QPSK modulation is used, the complexity of the tree search algorithm based on the neural network is smaller than that of the traditional M-best algorithm.

[0149] The invention provides a breadth-first tree search signal detection method based on deep learning. The invention uses a breadth-first tree search signal detection technology based on deep learning and a sorting strategy to optimize the detection system model. The received signal and the upper triangular matrix are used as the input of the network. The true value of the transmitted signal is used as the label of the network training. The network parameters converge and the loss function value is minimized through training. The trained network-assisted breadth-first tree search process can achieve a detection accuracy close to that of the traditional spherical decoding algorithm and reduce its detection complexity.

[0150] refer to Figure 6 The present invention also discloses a breadth-first tree search signal detection system based on deep learning, the system comprising:

[0151] The signal receiving module 110 is used to obtain noisy received data through the antenna array, send a modulated signal through the transmitting end, and obtain a transmitted signal, a received signal and a channel matrix through the receiving end and the wireless channel respectively;

[0152] The feature matrix generation module 120 is used to perform orthogonal triangular decomposition on the channel matrix, optimize and transform the multi-input multi-output detection model, obtain equivalent channel matrices and received signals, and splice them to form a feature matrix;

[0153] An initial detection module 130, used to input the feature matrix into a pre-trained neural network, and output an initial detection result of the transmitted signal through the pre-trained neural network after convergence;

[0154] The final detection module 140 is used to obtain the search threshold value of each layer of the breadth-first tree search algorithm according to the initial detection result, and perform a tree search process based on the search threshold value to obtain a final detection result of the transmitted signal.

[0155] The signal receiving module sorts the norm value of each column of the channel matrix obtained by the antenna array from small to large;

[0156] The signals sent by the transmitter are sorted according to the sorting index to determine the detection order of the signals.

[0157] The feature matrix generation module performs orthogonal triangular decomposition on the channel matrix and optimizes and transforms the multi-input multi-output detection model based on the orthogonal triangular decomposition results;

[0158] The equivalent received signal and the upper triangular matrix after orthogonal triangular decomposition are combined into a characteristic matrix.

[0159] An initial detection module inputs the feature matrix into a pre-trained neural network and outputs a soft decision probability of each transmitted signal;

[0160] The detection probability of all transmitted signals is obtained according to the output of the deep neural network block corresponding to each transmitted signal, and the category with the highest probability is the initial detection result of the transmitted information.

[0161] The final detection module calculates the accumulated distance of each layer based on the initial detection results obtained, and obtains the search threshold value of each layer of the tree search;

[0162] Nodes whose accumulated distance is greater than the threshold value are pruned, and nodes whose accumulated distance is less than or equal to the threshold value are retained to obtain candidate nodes;

[0163] In the search of the next layer, the child nodes of all candidate nodes are expanded, and the candidate node set of the layer is obtained according to the accumulated distance calculated by the child nodes and the search threshold of the layer;

[0164] When the search reaches the last layer, the node with the smallest cumulative distance is the optimal node, and all the nodes passed by the path from the root node to the optimal node are the final detection results.

[0165] The invention provides a breadth-first tree search signal detection system based on deep learning. Through the breadth-first tree search signal detection technology based on deep learning, the detection system model is optimized by using a sorting strategy, and the received signal and the upper triangular matrix are used as the input of the network, and the true value of the transmitted signal is used as the label of the network training. Through training, the network parameters converge and the loss function value is minimized. The trained network-assisted breadth-first tree search process can achieve a detection accuracy close to that of the traditional spherical decoding algorithm and reduce its detection complexity.

[0166] Fig.11 An example of a physical structure diagram of an electronic device is shown in FIG. Fig.11 As shown, the electronic device may include: a processor 1110, a communication interface 1120, a memory 1130 and a communication bus 1140, wherein the processor 1110, the communication interface 1120 and the memory 1130 communicate with each other through the communication bus 1140. The processor 1110 may call the logic instructions in the memory 1130 to execute a breadth-first tree search signal detection method based on deep learning, the method comprising: acquiring noisy received data through an antenna array, sending a modulated signal through a transmitting end, and acquiring a transmitted signal, a received signal and a channel matrix through a receiving end and a wireless channel respectively;

[0167] Performing orthogonal triangular decomposition on the channel matrix, optimizing and deforming the multi-input multi-output detection model, obtaining equivalent channel matrices and received signals, and splicing them to form a feature matrix;

[0168] Inputting the feature matrix into a pre-trained neural network, and outputting an initial detection result of the transmitted signal through the pre-trained neural network after convergence;

[0169] A search threshold value of each layer of a breadth-first tree search algorithm is obtained according to the initial detection result, and a tree search process is performed based on the search threshold value to obtain a final detection result of the transmitted signal.

[0170] In addition, the logic instructions in the above-mentioned memory 1130 can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0171] On the other hand, the present invention further provides a computer program product, the computer program product includes a computer program, the computer program can be stored on a non-transitory computer-readable storage medium, when the computer program is executed by a processor, the computer can execute a breadth-first tree search signal detection method based on deep learning provided by the above methods, the method comprising: acquiring noisy received data through an antenna array, sending a modulated signal through a transmitting end, and acquiring a transmitted signal, a received signal and a channel matrix through a receiving end and a wireless channel respectively;

[0172] Performing orthogonal triangular decomposition on the channel matrix, optimizing and deforming the multi-input multi-output detection model, obtaining equivalent channel matrices and received signals, and splicing them to form a feature matrix;

[0173] Inputting the feature matrix into a pre-trained neural network, and outputting an initial detection result of the transmitted signal through the pre-trained neural network after convergence;

[0174] A search threshold value of each layer of a breadth-first tree search algorithm is obtained according to the initial detection result, and a tree search process is performed based on the search threshold value to obtain a final detection result of the transmitted signal.

[0175] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented when the computer program is executed by a processor to perform a breadth-first tree search signal detection method based on deep learning provided by the above methods, the method comprising: acquiring noisy received data through an antenna array, sending a modulated signal through a transmitting end, and acquiring a transmitted signal, a received signal and a channel matrix through a receiving end and a wireless channel respectively;

[0176] Performing orthogonal triangular decomposition on the channel matrix, optimizing and deforming the multi-input multi-output detection model, obtaining equivalent channel matrices and received signals, and splicing them to form a feature matrix;

[0177] Inputting the feature matrix into a pre-trained neural network, and outputting an initial detection result of the transmitted signal through the pre-trained neural network after convergence;

[0178] A search threshold value of each layer of a breadth-first tree search algorithm is obtained according to the initial detection result, and a tree search process is performed based on the search threshold value to obtain a final detection result of the transmitted signal.

[0179] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0180] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0181] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A breadth-first tree search signal detection method based on deep learning, characterized in that: include: The receiving data with noise is obtained through the antenna array, the modulated signal is sent through the transmitting end, and the transmitting signal, the receiving signal and the channel matrix are obtained through the receiving end and the wireless channel respectively; Performing orthogonal triangular decomposition on the channel matrix, optimizing and deforming the multi-input multi-output detection model, obtaining equivalent channel matrices and received signals, and splicing them to form a feature matrix; Inputting the feature matrix into a pre-trained neural network, and outputting an initial detection result of the transmitted signal through the pre-trained neural network after convergence; A search threshold value of each layer of a breadth-first tree search algorithm is obtained according to the initial detection result, and a tree search process is performed based on the search threshold value to obtain a final detection result of the transmitted signal.

2. The breadth-first tree search signal detection method based on deep learning according to claim 1, characterized in that: The method of acquiring noisy received data through an antenna array, sending a modulated signal through a transmitting end, and acquiring a transmitted signal, a received signal, and a channel matrix through a receiving end and a wireless channel respectively includes: The norm value of each column of the channel matrix obtained by the antenna array is sorted from small to large; The signals sent by the transmitter are sorted according to the sorting index to determine the detection order of the signals.

3. The breadth-first tree search signal detection method based on deep learning according to claim 1, characterized in that: The orthogonal triangular decomposition of the channel matrix is ​​performed, and the multi-input multi-output detection model is optimized and deformed to obtain an equivalent channel matrix and a received signal, and the equivalent channel matrix and received signal are spliced ​​to form a feature matrix, which specifically includes: Perform orthogonal triangular decomposition on the channel matrix, and optimize and transform the multi-input multi-output detection model based on the orthogonal triangular decomposition result; The equivalent received signal and the upper triangular matrix after orthogonal triangular decomposition are combined into a characteristic matrix.

4. The breadth-first tree search signal detection method based on deep learning according to claim 1, characterized in that: The neural network construction method is: The neural network includes a plurality of deep neural network blocks, each transmitted signal is detected by using a deep neural network block, and the detection probability of all transmitted signals is obtained by superimposing the plurality of deep neural network blocks; Each deep neural network block includes five neural network layers. The number of neurons in the input layer and hidden layer is 256. The activation function used is the Leaky ReLu function. The output layer has the same number of neurons as the modulation order, and the Softmax function is used as the activation function. Adjacent deep neural network blocks are connected, and the output of each layer of the upper deep neural network block is used as the input of the next layer of the next deep neural network block, thereby constructing a neural network.

5. The breadth-first tree search signal detection method based on deep learning according to claim 1, characterized in that: The feature matrix is ​​input into a pre-trained neural network, and the initial detection result of the transmitted signal is output by the neural network after the pre-training convergence, specifically including: Inputting the feature matrix into a pre-trained neural network, and outputting a soft decision probability of each transmitted signal; The detection probability of all transmitted signals is obtained according to the output of the deep neural network block corresponding to each transmitted signal, and the category with the highest probability is the initial detection result of the transmitted information.

6. The breadth-first tree search signal detection method based on deep learning according to claim 1, characterized in that: Obtaining a search threshold value of each layer of a breadth-first tree search algorithm according to the initial detection result, and performing a tree search process based on the search threshold value to obtain a final detection result of the transmitted signal, specifically including: Based on the obtained initial detection results, the accumulated distance of each layer is calculated to obtain the search threshold value of each layer of the tree search; Nodes whose accumulated distance is greater than the threshold value are pruned, and nodes whose accumulated distance is less than or equal to the threshold value are retained to obtain candidate nodes; In the search of the next layer, the child nodes of all candidate nodes are expanded, and the candidate node set of the current layer is obtained according to the accumulated distance calculated by the child nodes and the search threshold of the current layer; When the search reaches the last layer, the node with the smallest cumulative distance is the optimal node, and all the nodes passed by the path from the root node to the optimal node are the final detection results.

7. A breadth-first tree search signal detection system based on deep learning, characterized in that: The system comprises: A signal receiving module is used to obtain noisy receiving data through an antenna array, send a modulated signal through a transmitting end, and obtain a transmitting signal, a receiving signal and a channel matrix through a receiving end and a wireless channel respectively; A feature matrix generation module is used to perform orthogonal triangular decomposition on the channel matrix, optimize and deform the multi-input multi-output detection model, obtain equivalent channel matrices and received signals, and splice them to form a feature matrix; An initial detection module, used to input the feature matrix into a pre-trained neural network, and output an initial detection result of the transmitted signal through the pre-trained neural network after convergence; The final detection module is used to obtain the search threshold value of each layer of the breadth-first tree search algorithm according to the initial detection result, and perform a tree search process based on the search threshold value to obtain a final detection result of the transmitted signal.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the breadth-first tree search signal detection method based on deep learning is implemented as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the breadth-first tree search signal detection method based on deep learning is implemented as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the breadth-first tree search signal detection method based on deep learning is implemented as described in any one of claims 1 to 6.

Images