A method for constructing an intelligent signal detection model and application thereof

Abstract

The application discloses a kind of construction method and application of intelligent signal detection model, belong to mobile communication technical field;The application constructs a new type of graph neural network containing aggregation update module, the global feature of each node is aggregated with its neighbor node information, so as to extract the hidden feature state of data according to the feature information of neighbor node and itself, and then realize the effective capture of the feature of communication signal such non-euclidean space data;In addition, after each aggregation is completed, the intermediate feature obtained by aggregation is further extracted in the application Feature, to improve the expression ability of intermediate feature, so that the designed graph neural network has more efficient information interaction ability and more powerful relationship reasoning ability, greatly improve the precision of signal detection.The application effectively solves the problem that the performance and computational complexity of traditional signal detection method are difficult to balance, and has strong robustness to channel estimation error.

Description

Technical Field

[0001] This invention belongs to the field of mobile communication technology, and more specifically, relates to a method for constructing and applying an intelligent signal detection model. Background Technology

[0002] 6G mobile communication will possess significant characteristics such as ubiquity, socialization, and intelligence, with channel propagation ranges covering ultra-wide bandwidths, multiple frequency bands (high, medium, and low), and complex environments including air, land, sea, and airspace. Existing orthogonal frequency division multiplexing (OFDM) technology is insufficient to achieve high-performance transmission in the complex channel environments of 6G communication, necessitating the development of novel waveform modulation technologies.

[0003] However, existing signal detection methods face significant challenges in the face of novel waveform modulation techniques. Take Orthogonal Time-Frequency Space (OTFS) technology as an example. It is a novel waveform technique that utilizes the delayed Doppler domain to obtain multidimensional diversity, exhibiting a lower peak-to-average power ratio (PAPR), higher spectral efficiency, and stronger resistance to channel time-varying characteristics, making it one of the most promising modulation techniques for 6G communication. Due to the Doppler effect and the large dimension of the equivalent channel matrix, traditional signal detection methods are difficult to apply directly. Specifically, traditional linear detection methods mainly eliminate interference by constructing an equalization matrix, which has low complexity but suffers significant performance loss. Traditional nonlinear detection methods, such as message passing and expectation propagation, have detection performance closely related to computational complexity, making it difficult to achieve high-precision signal detection under low-complexity conditions.

[0004] On the other hand, with the successful application of deep learning in fields such as natural image processing and computer vision, intelligent signal detection algorithms based on this technology have received increasing attention. However, traditional deep learning techniques, such as deep neural networks (DNNs) and convolutional neural networks, are mainly designed for Euclidean space data and require translation invariance. But for communication signal detection problems, variable nodes and factor nodes naturally form a factor graph data model, which is a typical bipartite graph and does not satisfy translation invariance. This limits the performance of existing intelligent signal detection algorithms, resulting in low accuracy of detection results. Summary of the Invention

[0005] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a method for constructing an intelligent signal detection model and its application, so as to solve the technical problem of low detection accuracy of existing communication signal detection methods.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a method for constructing an intelligent signal detection model, comprising:

[0007] S1. Acquire multiple sets of communication system data, including transmitted signals, channel matrices, and received signals, and convert the channel matrices and received signals in each set of communication system data into corresponding graph structure data.

[0008] S2. Using graph structure data as input and the corresponding transmitted signal as output, train the graph neural network model to obtain the intelligent signal detection model;

[0009] The graph neural network model includes an aggregation update module, which is used to deeply aggregate the global features of each node with the information of its neighboring nodes through multiple iterations. In each iteration, the Laplacian operator is used to multiply the global features of each node in the previous iteration to obtain the intermediate features of each node, which are then input into the feature extraction unit for deep feature extraction, thereby obtaining the global features of each node in the current iteration.

[0010] The initial global features of each node are the result of convolution operations on the parameters of each node in the graph structure data; the node parameters are constructed based on the channel matrix, the received signal, and the system noise figure.

[0011] More preferably, the feature extraction unit includes: a cascaded DNN module and a GRU unit;

[0012] The DNN module is used to input the intermediate features of the nodes into K DNN networks respectively, and sum the outputs of the DNN networks and output them into the GRU unit; K≥2.

[0013] More preferably, the above-mentioned DNN network is a three-layer DNN network; each DNN network contains two hidden layers.

[0014] More preferably, the graph neural network model further includes: a convolutional module and an output module;

[0015] The convolution module is used to perform convolution operations on the parameters of each node in the graph structure data to obtain the initial global features of each node, and output them to the aggregation and update module.

[0016] The output module is used to perform feature mapping on the global features of each node output by the aggregation module to obtain the predicted value of the transmitted symbol corresponding to each node, and then obtain the detection result of the transmitted signal.

[0017] More preferably, the output module is a three-layer DNN network.

[0018] More preferably, the Laplace operator It is a diagonal matrix. Let e ​​be the edge feature matrix of the graph structure data.ij =h i T ×h j h i Let C be the i-th column of the channel matrix; C is... The number of columns.

[0019] More preferably, the method for constructing graph structure data in step S1 includes: obtaining the correspondence between transmitted symbols and received symbols based on the channel matrix of each set of communication system data, and constructing corresponding graph structure data by connecting all transmitted symbol nodes corresponding to the same received symbol in pairs, with the transmitted symbols as nodes.

[0020] More preferably, for the i-th node of the graph structure data, its node parameters are:

[0021] f i =[y T h i ,h i T h i ,σ 2 ]

[0022] Where y is the received signal; h i σ is the channel vector in the i-th column of the channel matrix; 2 The system noise figure is denoted as .

[0023] More preferably, step S1 includes: randomly generating multiple sets of transmitted signals x and channel matrix H according to the configuration information of the communication system, and obtaining the corresponding received signal y according to the communication system model y = Hx + n; where n is the system noise.

[0024] By simulating the communication system, massive training datasets can be generated directly and randomly without additional labels, enabling more efficient and cost-effective acquisition of sufficient training samples and ensuring the effectiveness of network training.

[0025] Secondly, the present invention provides an intelligent signal detection method, comprising: after acquiring the received signal and channel matrix at the receiving end of the communication system, converting them into corresponding graph structure data; and inputting the graph structure data into an intelligent signal detection model to obtain the detection result of the transmitted signal;

[0026] The aforementioned intelligent signal detection model is constructed using the intelligent signal detection model construction method provided in the first aspect of this invention.

[0027] Thirdly, the present invention provides an intelligent signal detection system, comprising: a memory and a processor, wherein the memory stores a computer program, and the processor executes the intelligent signal detection method provided in the second aspect of the present invention when executing the computer program.

[0028] Fourthly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed by a processor, it controls the device where the storage medium is located to execute the method for constructing the intelligent signal detection model provided in the first aspect of the present invention and / or the intelligent signal detection method provided in the second aspect of the present invention.

[0029] In summary, the above-described technical solutions conceived in this invention can achieve the following beneficial effects:

[0030] 1. This invention provides a method for constructing an intelligent signal detection model. It constructs a novel graph neural network containing an aggregation and update module, aggregating the global features of each node with the information of its neighboring nodes. This allows for the extraction of hidden feature states of the data based on the feature information of its neighboring nodes and its own features, thereby achieving effective capture of features of non-Euclidean space data such as communication signals. Furthermore, after each aggregation, this invention further extracts features from the intermediate features obtained from the aggregation to enhance the expressive power of the intermediate features. This results in a more efficient information interaction capability and a more powerful relational reasoning capability for the designed graph neural network, significantly improving the accuracy of signal detection.

[0031] 2. The intelligent signal detection model construction method provided by this invention first uses multiple DNN networks to simultaneously process the intermediate features after aggregating information from neighboring nodes. Then, the features obtained from each DNN network are fused to obtain more comprehensive and significant node features, thereby further improving the expressive power of the graph neural network. Finally, the fused features are input into a GRU unit for processing. Since there are some structural and contextual relationships hidden between nodes (detection symbols) in signal detection, by using the GRU recurrent neural network module, the contextual information contained between data and the hidden correlations between them can be further extracted, thereby improving the propagation ability of long-term information (referring to the memory unit storing historical information and long-term states) of the graph structure. Through the above methods, this invention greatly enhances the role of graph neural networks in signal detection, and can efficiently extract the hidden features of nodes at the cost of lower computational complexity and memory, further improving the accuracy of signal detection.

[0032] 3. The intelligent signal detection model construction method provided by the present invention represents the communication system data into a graph structure based on the correlation between the transmitted symbols and then inputs it into the graph neural network designed in the present invention for training. This method can better learn and utilize the relationship between the transmitted symbols for feature extraction, thereby achieving node feature updates with lower computational complexity and memory requirements, and ultimately obtaining detection results with better performance.

[0033] 4. This invention utilizes the powerful relational reasoning and information interaction capabilities of the constructed graph neural network to effectively solve the problem of balancing performance and computational complexity in traditional signal detection methods, and it has strong robustness to channel estimation errors. Furthermore, experiments show that the detection results obtained by the intelligent signal detection method based on graph neural networks provided by this invention are superior to those of traditional signal detection methods, and it also has strong robustness to imperfect channel state information. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the graph neural network model provided by the present invention;

[0035] Figure 2 A schematic diagram of the OTFS system provided in an embodiment of the present invention;

[0036] Figure 3 A flowchart illustrating the application of an intelligent signal detection method to an OTFS system, as provided in an embodiment of the present invention;

[0037] Figure 4 The following are simulation diagrams of the bit rate obtained by using different signal detection methods in embodiments of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0039] In this invention, the terms "first," "second," etc. (if present) in the invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0040] In a first aspect, the present invention provides a method for constructing an intelligent signal detection model, comprising:

[0041] S1. Acquire multiple sets of communication system data, including transmitted signals, channel matrices, and received signals, and convert the channel matrices and received signals in each set of communication system data into corresponding graph structure data.

[0042] It should be noted that when acquiring communication system data, data can be directly collected from actual communication systems in real-world scenarios, or massive amounts of communication system data can be randomly generated through simulated communication systems without additional labels. To obtain sufficient training samples more efficiently and at lower cost, and to ensure the effectiveness of network training, data collection is preferably achieved through simulation. Specifically, based on the configuration information of the communication system (including: the number of transmitting antennas N)... t Number of antennas at the receiving end N r Number of subcarriers M, number of time slots N, number of multipaths L, center frequency f c Bandwidth B, maximum delay τ max (Signal modulation method, channel signal-to-noise ratio, etc.) Multiple sets of transmitted signals x and channel matrices H are randomly generated, and the corresponding received signal y is obtained according to the communication system model y = Hx + n; where n is the system noise, determined according to the signal noise figure σ. 2 Calculated.

[0043] Furthermore, in constructing the graph structure data, this invention obtains the correspondence between transmitted and received symbols based on the channel matrix of each set of communication system data. Using transmitted symbols as nodes, the corresponding graph structure data is constructed by connecting all transmitted symbol nodes corresponding to the same received symbol pairwise. The edge features in the graph structure data are obtained based on the channel vectors corresponding to the two connected nodes. Specifically, the edge feature e between the nodes corresponding to the i-th and j-th transmitted symbols is... ij =h i T ×h j h i Let f be the i-th column of the channel matrix. The node parameters in the graph-structured data are constructed based on the channel matrix, the received signal, and the system noise figure. Alternatively, traditional signal detection results can be used as initial features of the nodes, followed by optimization using the graph network. For example, the detection results of the near-zero algorithm or the minimum mean square error algorithm can be used as initial features of the nodes. For instance, when using the near-zero algorithm, the node parameters of the i-th node in the graph-structured data are: f i =[H -1 y i When using the least mean square error algorithm, the node parameter of the i-th node in the graph structure data is: f i =H H (H H +σ 2 I) -1 y i ; where y i σ is the i-th received symbol in the received signal; 2 The system noise figure is denoted as .

[0044] Furthermore, considering that traditional linear signal detection algorithms obtain the signal detection value x by constructing various equalization matrices and multiplying them with the received signal y, and y... T h, h T h and the noise coefficient are key parameters for constructing these equalization matrices. Therefore, to highlight these key parameters and obtain better detection results, in another alternative approach, the node parameters for the i-th node of the graph-structured data are:

[0045] f i =[y T h i ,h i T h i ,σ 2 ]

[0046] Where y is the received signal; h i σ is the channel vector in the i-th column of the channel matrix; 2 The system noise figure is denoted as .

[0047] It should be noted that, given a communication system (the model can be expressed as y = Hx + n), the transmitted symbol x can be obtained based on the channel matrix H. i With the received symbol y j The factor graph (correspondence between transmitting and receiving symbols) allows for the deduction of relationships between transmitting symbols by using the receiving symbols as intermediate nodes. This invention fully utilizes the inherent structural relationships of the communication system, transforming the factor graph containing both transmitting and receiving symbols into a graph structure containing only transmitting symbol nodes. This facilitates the analysis of the relationships between receiving symbols by the designed graph neural network. By representing the communication system data as a graph structure based on the relationships between transmitting symbols and then inputting it into the designed graph neural network for training, this invention can better learn and utilize the relationships between transmitting symbols for feature extraction. This achieves node feature updates with lower computational complexity and memory requirements, ultimately resulting in superior detection performance.

[0048] S2. Using graph structure data as input and the corresponding transmitted signal as output, train the graph neural network model to obtain the intelligent signal detection model;

[0049] Among them, such as Figure 1 As shown, the graph neural network model includes cascaded convolutional modules, aggregation update modules, and output modules;

[0050] 1) Convolutional module:

[0051] The convolution module performs convolution operations on the parameters of each node in the graph structure data to obtain the initial global features of each node, which are then output to the aggregation and update module. In one optional implementation, each node parameter in the graph structure data is encoded through a single fully connected layer to obtain the initial global features of each node. Taking the second node parameter construction method mentioned above as an example, the initial global features of the i-th node are: Where W1 is the dimension R n The learnable parameter matrix of a ×3 single-layer network, where b1 is its learnable bias vector.

[0052] 2) Aggregated Update Module:

[0053] The aggregation update module is used to deeply aggregate the global features of each node with the information of its neighboring nodes through multiple iterations. It includes an aggregation unit and a feature extraction unit. In each iteration, the aggregation unit uses the Laplacian operator to multiply the global features of each node in the previous iteration to obtain the intermediate features of each node. These intermediate features are then input into the feature extraction unit for deep feature extraction, thereby obtaining the global features of each node in the current iteration.

[0054] Specifically, the intermediate features of the i-th node in the t-th iteration. Among them, u i (t-1) The global feature of the i-th node in the (t-1)-th iteration; Laplace operator It is a diagonal matrix. Let e ​​be the edge feature matrix of the graph structure data. ij =h i T ×h j h i Let C be the i-th column of the channel matrix; C is... The number of columns.

[0055] In one alternative implementation, the total number of iterations in the aggregation update module is 10.

[0056] To further enhance the role of neural networks in signal detection algorithms, a "multi-head" concept is introduced in one optional implementation. First, multiple DNN networks are used simultaneously to process the intermediate features obtained after aggregating information from neighboring nodes. Then, feature fusion is used to obtain more significant node features. Finally, a GRU unit is used to further enhance the expressive power of the graph neural network. Specifically, the feature extraction unit includes a cascaded DNN module and a GRU unit. The DNN module is used to input the intermediate node features into K DNN networks respectively, sum the outputs of the DNN networks, and output the sum to the GRU unit; K ≥ 2. In a preferred embodiment, the DNN network is a three-layer DNN network; each DNN network contains two hidden layers, containing N1 and N2 neurons respectively, and the number of neurons in the output layer is R. n Furthermore, both the hidden and output layers use ReLU activation functions.

[0057] The inputs to the GRU unit are the node features d after the current fusion processing (i.e., the sum of the outputs of K DNN networks). i (t) And the node features d retained from the previous iteration after fusion processing. i (t-1) The output is a dimension of size R. n ×1 new feature.

[0058] Furthermore, the aforementioned feature extraction unit also includes a scale transformation unit, used to transform the feature size output by the GRU unit so that the size of the global features obtained in each iteration remains consistent.

[0059] 3) Output module:

[0060] The output module performs feature mapping on the global features of each node output by the aggregation module to obtain the predicted value of the transmitted symbol corresponding to each node, thereby obtaining the detection result of the transmitted signal. In a preferred embodiment, the output module is a three-layer DNN network, which contains two hidden layers and uses the ReLU function as its corresponding activation function; the number of neurons in the output layer is determined by the modulation method of the communication system, and the activation function is the softmax function.

[0061] This invention utilizes a small DNN network as the output module, which can effectively classify nodes to achieve signal detection at the cost of relatively low computational complexity. Test results show that this output module can achieve good detection accuracy.

[0062] In summary, this invention constructs a novel graph neural network with efficient information interaction and powerful relational reasoning capabilities as an intelligent signal detector. By deeply aggregating the global features of each node with the information of its neighboring nodes, it can extract the hidden feature states of the data based on the feature information of its neighboring nodes and itself. This enables the effective capture of features of non-Euclidean space data such as communication signals, achieving high signal detection accuracy while balancing signal detection performance and computational complexity. Furthermore, thanks to the efficient nonlinear representation capabilities of deep learning, it exhibits strong robustness against channel estimation errors.

[0063] Secondly, the present invention provides an intelligent signal detection method, comprising:

[0064] After the receiving end of the communication system obtains the received signal and channel matrix, it converts them into corresponding graph structure data; and then inputs the graph structure data into the intelligent signal detection model to obtain the detection result of the transmitted signal.

[0065] The aforementioned intelligent signal detection model is constructed using the intelligent signal detection model construction method provided in the first aspect of this invention.

[0066] It should be noted that the method for converting the received signal and channel matrix into corresponding graph structure data during intelligent signal detection is the same as the conversion method used in the construction method of the intelligent signal detection model provided in the first aspect of this invention. Related technical solutions are the same as those in the construction method of the intelligent signal detection model provided in the first aspect of this invention, and will not be elaborated upon here.

[0067] Thirdly, the present invention provides an intelligent signal detection system, comprising: a memory and a processor, wherein the memory stores a computer program, and the processor executes the intelligent signal detection method provided in the second aspect of the present invention when executing the computer program.

[0068] The related technical solutions are the same as the intelligent signal detection method provided in the second aspect of this invention, and will not be described in detail here.

[0069] Fourthly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed by a processor, it controls the device where the storage medium is located to execute the method for constructing the intelligent signal detection model provided in the first aspect of the present invention and / or the intelligent signal detection method provided in the second aspect of the present invention.

[0070] The relevant technical solutions are the same as the construction method of the intelligent signal detection model provided in the first aspect and the intelligent signal detection method provided in the second aspect, and will not be repeated here.

[0071] To further illustrate the technical solution provided by the present invention, detailed descriptions are provided below with reference to specific embodiments:

[0072] Existing intelligent detection methods mainly rely on traditional deep learning techniques and are primarily designed for Euclidean space data. However, in communication signal detection problems, variable nodes and factor nodes form a factor graph relationship, which does not satisfy translation invariance, leading to performance degradation of existing intelligent detection methods. Therefore, this embodiment provides an intelligent signal detection method based on graph neural networks for OTFS systems, which can effectively balance signal detection performance and computational complexity, thereby achieving efficient detection of OTFS signals.

[0073] Consider a case such as Figure 2 The OTFS system model is shown. Information bits are first modulated by amplitude and phase to generate M×N constellation symbols, where M is the number of subcarriers and N is the time slot size. Then, they are converted from the time-delay-Doppler domain to time-frequency domain symbols using a two-dimensional inverse symptotic Fourier transform. The transmitted signal is then obtained by processing the time-frequency domain signal using the Heisenberg transform and window function.

[0074] After the above processing, the signal s(t) will be wirelessly transmitted in a time-varying channel, and the receiving end will receive the signal r(t) = ∫∫h(τ,v)s(t-τ)e j2πv(t-τ) dτdv. Where h(τ, v) represents the impulse response of the channel, and τ and v represent the time delay and Doppler frequency offset, respectively. Then, a Wigner transform is performed on r(t) to obtain the time-frequency domain signal. in, This is the matching window function at the receiving end. Finally, the time-frequency domain signal Y is transformed using the symmetric Fourier transform. mn Transformed into the time-delay-Doppler domain, Therefore, the matrix form corresponding to the receiver section of the system can be expressed as: in, F N and These represent the Fast Fourier Transform and Inverse Fourier Transform at N points, respectively. It is the Kronecker product, H tv It is the matrix of wireless channels, G rx and G tx These are diagonal matrices constructed from the window functions at the receiving and transmitting ends, respectively. The equivalent channel H represented by OTFS eff , Equivalent noise

[0075] The detection of OTFS signals can be equivalent to solving the following problem:

[0076]

[0077] In the formula, p(y|x)=p y|z (y|H eff ·x), and z=H eff ·x.

[0078] This invention applies an intelligent signal detection method based on graph neural networks to the OTFS system, and its overall process is as follows: Figure 3 As shown, the specific steps include:

[0079] Step 1: Offline Training Phase

[0080] S1. Based on the OTFS system configuration information, generate the received signal matrix and the equivalent channel matrix as inputs to the training dataset, and generate the transmitted signal matrix as the label of the training dataset.

[0081] Specifically, firstly, based on the configuration information of the OTFS system, including: number of subcarriers M, number of time slots N, number of paths L, and carrier frequency f. c Maximum Delay Index Doppler index range The relative velocity v between the transmitting and receiving devices c The transmitted signal is randomly generated based on the signal modulation method and the signal-to-noise ratio of the communication system channel. Then, according to the formula... Calculate the OTFS equivalent channel matrix. Then, based on the established system model... The received signal y is calculated from the input signal, the equivalent channel matrix, and the system noise, where This corresponds to the equivalent noise. Finally, the received signal y and the equivalent channel matrix H are compared. eff The signal x from the transmitting end is used as the input to the training dataset, while the signal x from the transmitting end is used as the label.

[0082] S2. Construct a graph neural network model, which mainly includes three functional modules: aggregation, update, and output. The users of the OTFS system are used as nodes in the graph neural network. The initial feature information of the nodes will be processed through the above three modules in sequence to obtain the signal detection results.

[0083] Specifically, the graph neural network constructed in this invention combines the advantages of deep learning and Markov random fields, capturing the structural information of the data as a feature vector, and then updating it through message passing between nodes. A set of random variables x = {x0, ..., x...} N-1 The structure signal can be represented by an undirected graph. We will use this model, where V and E represent the sets of nodes and edges, respectively. Each node corresponds one-to-one with a random variable, and satisfies p(x) = ... i |x\x i )=p(x i|ne(i)), where \ refers to the exclusion operation, and ne(i) is the set of neighboring nodes of the i-th node. In a pairwise Markov random field, the self-potential is... Assign it to the i-th node, and adjust the potential accordingly. Assign the edge connecting the i-th node and the j-th node. The corresponding posterior probability can be expressed as: Where C is a normalized constant, and Where, σ 2 h is the noise figure of the OTFS system. i It is the equivalent channel matrix H eff The i-th column, p i (x i ) represents the prior probability of a variable. Graph neural networks primarily infer the posterior probability p through nodes and edges. GNN (x|y), and use the self potential and the relative potential as the feature vectors of the nodes and edges.

[0084] In this embodiment, Let's define the initial eigenvector of the i-th node, which is composed of the received signal y and the i-th column channel matrix h. i and system noise figure σ 2 This is determined jointly. The initial features of a node are obtained by encoding the node parameters through a single-layer fully connected layer. And serve as the initial global feature of the node; where W1 is the dimension of this single-layer network of dimension R. n The learnable parameter matrix is ​​3×3, and b1 is the learnable parameter vector. Furthermore, the idiomatic representation of the edge between the i-th node and the j-th node is...

[0085] Subsequently, the node features and edge features are processed sequentially by the graph network's aggregation and update module and output module to obtain the final detection results.

[0086] Specifically, the aggregation process first calculates the corresponding Laplacian operator. in It is a diagonal matrix, and Then, based on the calculated Laplacian operator, graph convolution operations are used to aggregate the information between nodes. in the formula It is the node feature matrix (intermediate features) after aggregating neighbor node information in the t-th iteration; in the first iteration, Initial global features In subsequent iterations, This is the output of the feature extraction unit in the (t-1)th iteration.

[0087] Specifically, the feature extraction unit consists of K three-layer DNN networks and GRU units (gated recursive units). Each DNN network contains two hidden layers with N1 and N2 neurons respectively, and the number of neurons in the output layer is R. n Furthermore, the activation functions for both the hidden and output layers are ReLU. To further enhance the role of neural networks in signal detection algorithms, this invention introduces the concept of multi-head processing from graph attention networks. This involves using K DNN networks to simultaneously process the aggregated node features, then fusing them to obtain more comprehensive new features. GRU units are then used to improve the expressive power of the graph neural network. The input to this module consists of the node features processed by the K DNN networks. and the features of the nodes in the previous iteration The output is a dimension of size R. n ×1 new feature.

[0088] Specifically, the output module consists of a three-layer DNN module. It contains two hidden layers and uses the ReLU function as its corresponding activation function. The number of neurons in the output layer is determined by the modulation scheme of the OTFS system, and the activation function is the softmax function.

[0089] S3. Use randomly generated training set data to train the constructed network offline and adjust the network's weight parameters.

[0090] Specifically, the offline training process includes core steps such as generating the training dataset, configuring network hyperparameters, and setting the loss function. The training dataset is randomly generated (x, H...). eff , σ 2 The signal is composed of the received signal y and the effective channel matrix H. eff and system noise figure σ 2 The input to the training dataset is denoted by x, and the label is denoted by x. Each training loop contains 30,000 training data sets, with 24,000 sets used as the training set and the remaining 6,000 sets used as the validation set. The batch size is set to 100. Furthermore, the hyperparameters of the network are configured as follows: the number of loops in the graph network is set to 10, and the number of neurons in the network, R... n =8, N1=32, N2=16, number of heads K=3, learning rate 1×10 -4 The total number of training iterations is 3000. Finally, this invention uses cross-entropy as the corresponding loss function:

[0091]

[0092] in, It is the final output of the graph neural network, p s (x i) is the true probability distribution of x.

[0093] Each training iteration calculates the corresponding gradient using the obtained loss function value, then updates the weights and bias parameters of the graph neural network according to the selected gradient descent rule, and saves the optimal weight values ​​and bias parameters.

[0094] This embodiment records the weight and bias parameters of the network during each training iteration and saves the parameters corresponding to the best training result. By setting predefined network hyperparameters, such as batch size, learning rate, number of training iterations, and loss function, efficient network training is achieved, ensuring the network's convergence and effectiveness.

[0095] Step Two: Online Detection Phase

[0096] S4. Use the trained network to perform online detection on the information received by the OTFS system receiver to obtain the actual detection results of the network.

[0097] The following comparison uses existing minimum mean square error message passing signal detection methods as a basis to further verify the beneficial effects of the graph neural network-based signal detection method provided in this embodiment of the invention compared to existing technologies. Figure 4 In the graph, the horizontal axis represents the signal-to-noise ratio (SNR) in dB, and the vertical axis represents the bit error rate (BER). The smaller the BER value, the better the performance of the corresponding algorithm. Figure 3 The relationship between bit error rate and signal-to-noise ratio obtained by three signal detection methods is provided: minimum mean square error signal detection, message passing signal detection, and the signal detection method based on graph neural network provided in this invention.

[0098] exist Figure 4 In this study, the OTFS system operates in the 4GHz band with 8 subcarriers (M=8), 8 time slots (N=8), 4 paths (L=4), a relative speed of 150 km / h between the transmitter and receiver, and QPSK modulation. The signal detection method based on graph neural networks presented in this application is significantly superior to the other two traditional methods, and its advantages gradually increase with the signal-to-noise ratio.

[0099] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing an intelligent signal detection model, characterized in that, include: S1. Acquire multiple sets of communication system data, including transmitted signals, channel matrices, and received signals, and convert the channel matrices and received signals in each set of communication system data into corresponding graph structure data; the communication system is an OTFS system, and the transmitted signals are OTFS signals; S2. Using graph structure data as input and the corresponding transmitted signal as output, train the graph neural network model to obtain the intelligent signal detection model. The graph neural network model includes an aggregation update module, which is used to deeply aggregate the global features of each node with the information of its neighboring nodes through multiple iterations. In each iteration, the Laplacian operator is used to multiply the global features of each node in the previous iteration to obtain the intermediate features of each node, which are then input into the feature extraction unit for deep feature extraction, thereby obtaining the global features of each node in the current iteration. The initial global features of each node are the results of convolution operations performed on the parameters of each node in the graph structure data; the node parameters are constructed based on the channel matrix, the received signal, and the system noise figure. The feature extraction unit includes: a cascaded DNN module and a GRU unit; The DNN module is used to input the intermediate features of the nodes into... K In each DNN network, the results output by the DNN network are summed and then output to the GRU unit; .

2. The method for constructing the intelligent signal detection model according to claim 1, characterized in that, The graph neural network model further includes: a convolution module and an output module; The convolution module is used to perform convolution operations on the parameters of each node in the graph structure data to obtain the initial global features of each node, and output them to the aggregation and update module. The output module is used to perform feature mapping on the global features of each node output by the aggregation update module to obtain the predicted value of the transmitted symbol corresponding to each node, and then obtain the detection result of the transmitted signal.

3. The method for constructing the intelligent signal detection model according to claim 1, characterized in that, The Laplace operator ; It is a diagonal matrix. ; The edge feature matrix of graph structure data, its elements ; The channel matrix is ​​the first i Column; C is The number of columns.

4. The method for constructing the intelligent signal detection model according to any one of claims 1-3, characterized in that, The method for constructing graph structure data in step S1 includes: obtaining the correspondence between transmitted symbols and received symbols based on the channel matrix of each set of communication system data, and constructing the corresponding graph structure data by connecting all transmitted symbol nodes corresponding to the same received symbol in pairs, with the transmitted symbols as nodes.

5. The method for constructing the intelligent signal detection model according to any one of claims 1-3, characterized in that, For graph structure data, the first i There are 1 node, and its node parameters are: in, To receive signals; The first in the channel matrix i Column channel vector; The system noise figure is denoted as .

6. The method for constructing the intelligent signal detection model according to any one of claims 1-3, characterized in that, Step S1 includes: randomly generating multiple sets of transmission signals based on the configuration information of the communication system. and channel matrix And according to the communication system model Obtain the corresponding received signal ;in, This represents system noise.

7. A method for intelligent signal detection, characterized in that, include: After the receiving end of the communication system obtains the received signal and channel matrix, it is converted into corresponding graph structure data. The graph structure data is then input into the intelligent signal detection model to obtain the detection results of the transmitted signal. The intelligent signal detection model is constructed using the construction method of the intelligent signal detection model according to any one of claims 1-6.

8. An intelligent signal detection system, characterized in that, include: A memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the intelligent signal detection method of claim 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed by a processor, it controls the device where the storage medium is located to perform the method for constructing the intelligent signal detection model according to any one of claims 1-6 and / or the intelligent signal detection method according to claim 7.

Images