Spectrum sensing method based on heterogeneous graph convolution network

Abstract

The application discloses a spectrum sensing method based on a heterogeneous graph convolution network, and mainly solves the problems of poor detection effect, weak robustness and generalization, and poor scalability of the prior art in a non-ideal channel environment. The method comprises the following steps: 1) using nodes in a multiplexing network to represent multi-modal information, and modeling the potential correlation of multi-modal information through intra-layer edges and inter-layer edges; 2) using a nonlinear mapping function to map different modal information to a common dimensional space, constructing a heterogeneous graph convolution used for attribute aggregation of multiplexing network nodes, and combining a global pooling layer and a classifier to obtain a heterogeneous graph convolution network; 3) inputting the multiplexing network into the heterogeneous graph convolution network and training until convergence; and 4) using the trained heterogeneous graph convolution network to realize spectrum sensing. The application combines multi-modal information and the heterogeneous graph convolution network, and optimizes in an end-to-end manner, effectively alleviates the performance decline caused by the non-ideal channel environment, and simultaneously enhances the robustness and generalization of the spectrum sensing algorithm.

Description

Technical Field

[0001] This invention belongs to the field of cognitive radio technology, and further relates to spectrum sensing technology, specifically a spectrum sensing method based on heterogeneous graph convolutional networks, which can be used in the fields of spectrum detection and weak signal detection in the Internet of Things. Background Technology

[0002] With the rapid development of IoT applications such as smart cities, smart homes, and intelligent agriculture, IoT devices have been extensively researched and deployed. However, the spectrum resources allocated for IoT services cannot meet the rapidly growing demand, leading to competition for shared spectrum resources with other RF solutions. Furthermore, most of the remaining spectrum resources have already been licensed to specific services, resulting in spectrum scarcity and congestion in IoT operations. Cognitive radio technology is considered the most promising technology for alleviating spectrum scarcity, and spectrum sensing technology, capable of detecting available spectrum resources, is the most fundamental and crucial part of cognitive radio technology.

[0003] Traditional spectrum sensing schemes rely on theoretical statistical models for their test statistics, such as received signal strength following a chi-square distribution, eigenvalues ​​of the covariance matrix following a Tracy-Widom distribution, and phase difference variance following a uniform distribution. However, traditional spectrum sensing schemes have certain drawbacks: 1) the statistical models of test statistics in actual communication systems deviate from the theoretical statistical models; 2) traditional algorithms have weak generalization and robustness, and their detection performance is easily affected by non-ideal channel environments such as noise disturbances, channel fading, and malicious interference.

[0004] Neural network-based spectrum sensing algorithms can effectively overcome the impact of non-ideal channel environments on detection performance without requiring the construction of complex statistical models. Although deep learning has developed rapidly in recent years, its application in spectrum sensing is still in its initial stages. Deep learning-based spectrum sensing schemes have the following limitations: 1) Current deep learning methods use single-modal test statistics as input, such as the original signal, covariance matrix, energy, and spatiotemporal correlation, making their detection performance susceptible to non-ideal channel environments; 2) Current deep learning methods have limited robustness and generalization performance. When performing spectrum sensing tasks in new test environments, fine-tuning or retraining is required to maintain good detection performance. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention aims to propose a spectrum sensing method based on heterogeneous graph convolutional networks, solving problems such as poor detection performance, poor generalization and robustness, and the need to establish complex statistical models in existing detection schemes under non-ideal channel environments. This invention models multimodal information using multiplexed networks while considering the potential correlations between multimodal information. The multiplexed network uses multimodal information as node attributes and uses inter-layer and intra-layer edges to describe the potential inter-layer and intra-layer correlations between multimodal information. Finally, a spectrum sensing framework is constructed using the multiplexed network and heterogeneous graph convolutional networks, optimizing the spectrum sensing model in an end-to-end manner, thereby effectively improving detection accuracy, robustness, and generalization.

[0006] The specific steps for achieving the above objectives according to the present invention include the following:

[0007] (1) Constructing a reused network:

[0008] (1.1) The received signal is mapped into multiple modal information by using different modal mapping functions, and the node attributes of the inner graph in the multiplexing network are constructed using the modal information;

[0009] (1.2) Construct the adjacency matrix of the inner graph of the multiplexing network using radial basis functions and multiple modal information, and construct the adjacency matrix of the inter-layer graph of the multiplexing network using coupling graph and identity matrix;

[0010] (1.3) The reused network consists of three parts: node attributes, intra-layer adjacency matrix, and inter-layer adjacency matrix;

[0011] (2) Constructing a heterogeneous graph convolutional network:

[0012] (2.1) Use a nonlinear mapping function to map different modal information to a common dimension space and construct the NodeEmbedding module;

[0013] (2.2) Heterogeneous graph convolution is composed of intra-layer graph convolution and inter-layer graph convolution. Intra-layer graph convolution uses graph convolution to aggregate the same mode information on the receiving antenna array from the perspective of the spectral domain of node attributes. Inter-layer graph convolution uses graph attention mechanism to assign corresponding weights to the information of different modes on the same receiving antenna and aggregate multi-mode information according to the inter-layer adjacency matrix.

[0014] (2.3) Use global summation pooling to extract the abstract features of the reused network and obtain the global representation of the reused network;

[0015] (2.4) A classifier is constructed by stacking multiple fully connected layers, and the output sequence of the global graph pooling layer is mapped to decision statistics;

[0016] (2.5) The Node Embedding module is cascaded with the heterogeneous graph convolution, global pooling layer and classifier in sequence to obtain the heterogeneous graph convolutional network;

[0017] (3) Set initial hyperparameters, input the reused network into the heterogeneous graph convolutional network, and iteratively train it until convergence to obtain the optimized heterogeneous graph convolutional network.

[0018] (4) The trained heterogeneous graph convolutional network is used to map the reused network into decision statistics. Then, the Monte Carlo method is used to construct a detection threshold with a fixed false alarm rate based on the decision statistics. Finally, the final spectrum occupancy state is determined based on the binary assumptions to achieve spectrum sensing.

[0019] Compared with the prior art, the present invention has the following advantages:

[0020] First, the present invention uses a multiplexing network to model the multimodal information of the received signal. This multiplexing network can represent multimodal information in the form of nodes, and can also capture hidden correlations in multimodal information in the form of intra-layer adjacency matrix and inter-layer adjacency matrix.

[0021] Secondly, this invention integrates multimodal information and heterogeneous graph convolutional networks into a unified framework. It can extract high-order features embedded in the inner graph nodes of each modality from the spectral domain perspective using the inner graph convolution in the heterogeneous graph convolution, and can adaptively aggregate multimodal information in the reused network from the spatial domain perspective using the inter-layer graph convolution in the heterogeneous graph convolution, thereby effectively improving the detection accuracy, robustness and generalization performance of spectrum sensing. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the overall implementation of the method of the present invention;

[0023] Figure 2 This is a diagram of the heterogeneous graph convolutional network architecture constructed in this invention;

[0024] Figure 3 The figure shows a comparison of simulation results of the detection performance of the present invention and existing methods under AWGN channel; where (a) is the ROC curve under AWGN channel and (b) is the PD-SNR curve under AWGN channel.

[0025] Figure 4 The figure shows a comparison of simulation results of the detection performance of the present invention and existing methods under CGGN channel; where (a) is the ROC curve under CGGN channel and (b) is the PD-SNR curve under CGGN channel.

[0026] Figure 5The figure shows a comparison of simulation results of the detection performance of the present invention and existing methods under large-scale fading channels; where (a) is the ROC curve under large-scale fading channels and (b) is the PD-SNR curve under large-scale fading channels.

[0027] Figure 6 The figure shows the extended performance evaluation results of the method of the present invention; where (a) is the ROC curve under the extended performance evaluation with the received sample length, and (b) is the PD-SNR curve under the extended performance evaluation with the number of received antennas.

[0028] Figure 7 The figure shows the detection performance evaluation results of the method of the present invention in the cooperative spectrum sensing scenario; where (a) is the ROC curve of cooperative spectrum sensing based on OFDM mechanism, and (b) is the PD-SNR curve of cooperative spectrum sensing based on OFDM mechanism.

[0029] Figure 8 The figure shows the detection performance evaluation results of the method of the present invention in the cooperative spectrum sensing scenario under the imperfect reporting channel; where (a) is the ROC curve under the imperfect fading channel of cooperative spectrum sensing, and (b) is the PD-SNR curve under the imperfect fading channel of cooperative spectrum sensing. Detailed Implementation

[0030] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0031] Example 1: Refer to Figure 1 This invention proposes a spectrum sensing method based on heterogeneous graph convolutional networks. By integrating multimodal information, multiplexing networks, and heterogeneous graph convolutional networks into a unified spectrum sensing framework, it achieves spectrum occupancy state detection in an end-to-end manner. The specific steps are as follows:

[0032] Step 1. Constructing the Reuse Network: The reuse network consists of three parts: node attributes, intra-layer adjacency matrix, and inter-layer adjacency matrix, used to represent multimodal information and the interactions between modal information; the construction steps are as follows:

[0033] (1.1) Different modal mapping functions are used to map the received signal into multiple modal information, and this modal information is used to construct the node attributes of the inner graph in the multiplexing network. The multiple modal information includes at least two or more of the following four modal information: Received Signal Strength (RSS), Phase Difference Variance (PDV), Phase Difference Kurtosis (PDK), and Trajectory Covariance Matrix (TCM). This embodiment uses the mapping function F.RSS (·),F PDV (·),F PDK (·) and F TCM (·) Map the received signal into received signal strength mode information respectively. Phase difference variance modal information Phase difference kurtosis modal information Trajectory covariance modal information The node attributes of the multi-modal layer graph in the reused network are constructed using these four modal information, as shown below:

[0034] (1.1.1) Using F RSS (·) Map the received signal into RSS modal information Construct an in-layer graph G using all RSS modal information. RSS Node attributes

[0035]

[0036]

[0037] (1.1.2) Using F PDV (·) Map the received signal into PDV mode information Construct the in-layer graph G using all PDV modal information. PDV Node attributes

[0038]

[0039] Where Re(y) i (t)) and Im(y i (t) represents the real and imaginary parts of the sampled value of the received signal at sampling time t, respectively; Let be the instantaneous phase of the sampled value of the sensed signal on the i-th receiving antenna at time t; the phase is normalized by the modulo 2π function. Within the range [0, 2π]; Phase difference Θ i The mean;

[0040] (1.1.3) Using F PDK (·) Map the received signal into PDK mode information Construct the in-layer graph G using all PDK modal information. PDK Node attributes

[0041]

[0042]

[0043] (1.1.4) Using F TCM (·) Map the received signal into TCM mode information Construct the intralayer graph G using all TCM modal information. TCM Node attributes

[0044]

[0045] in, ξ is the sampling vector of a rectangular window function with length ξ and step size 1; τ represents the number of sampling vectors of the rectangular window function. This represents the trajectory matrix of the signal sensed on the i-th receiving antenna.

[0046] (1.2) The adjacency matrix of the inner layer graph in the reused network is constructed using radial basis functions and multiple modal information, and the adjacency matrix of the inter-layer graph in the reused network is constructed using coupling graphs and identity matrices. In this embodiment, radial basis functions are used to map multiple modal information to the adjacency matrix of the corresponding modal inner layer graph, including the adjacency matrix A of the RSS modal inner layer graph. RSS The adjacency matrix A of the inner graph of the PDV modal layer PDV The adjacency matrix A of the PDK modal layer inner graph PDK The adjacency matrix A of the TCM modal layer inner graph TCM The method of constructing inter-layer adjacency matrices in a multiplexed network using coupling graphs and identity matrices is described. It is expressed as follows:

[0047] (1.2.1) Calculate modal information T using radial basis functions r i and T r j Constructing the intra-layer graph G using real-valued distances between layers r adjacency matrix

[0048]

[0049] r = {RSS, PDV, PDK, TCM}

[0050] i,j = 1,2,...,M;

[0051] (1.2.2) Constructing the adjacency matrix of the inter-layer graph using the coupling graph C and the identity matrix I

[0052]

[0053] (1.3) The reused network consists of three parts: node attributes, intra-layer adjacency matrix, and inter-layer adjacency matrix;

[0054] Step 2. Construct a heterogeneous graph convolutional network:

[0055] (2.1) Reusing the node attributes in the network from different modalities, a nonlinear mapping function is used to map the information of different modalities to a common dimension space, and a Node Embedding module is constructed; the implementation is as follows:

[0056] (2.1.1) Use fully connected functions to connect the node attributes T of the inner graphs of the RSS, PDV, and PDK modal layers respectively. RSS T PDV T PDK Mapping to node embedding

[0057]

[0058] (2.1.2) Use a 3D convolution with a kernel size of 2×τ×τ and a stride of 1 to extract the node attributes T of the TCM modality layer. TCM Mapping to node embedding

[0059]

[0060] Among them, W RSS W PDV , b represents the weight mapping matrix of the fully connected layer in RSS, PDV, and PDK modes, respectively; RSS ,b PDK , These are the bias parameters for the fully connected layer; Represents the convolution kernel of a 3D convolutional layer; is the output feature of the p-th output channel of the 3D convolution.

[0061] (2.2) Heterogeneous graph convolution is composed of intra-layer graph convolution and inter-layer graph convolution. Intra-layer graph convolution uses graph convolution to aggregate the same mode information on the receiving antenna array from the spectral domain perspective of node attributes. Inter-layer graph convolution uses graph attention mechanism to assign corresponding weights to information of different modes on the same receiving antenna and aggregate multi-mode information according to the inter-layer adjacency matrix. The specific implementation steps are as follows:

[0062] (2.2.1) Hyper-Laplace matrix decomposition:

[0063] The hyperLaplace matrix of decomposition and reuse networks For the intralayer hyperLaplace matrix and interlayer hyperLaplace matrix The formula is as follows:

[0064]

[0065] in, It is a diagonal hyperdegree matrix and Represents the Laplace matrix of the r-mode layer inner graph; For interlayer hyperLaplace matrix;

[0066] (2.2.2) Construction of heterogeneous graph convolution:

[0067] The eigenvalue matrix and eigenvector matrix are constructed from the eigenvalue decomposition of the hyperLaplace matrix, as shown in the following formula:

[0068]

[0069] Where U represents the eigenvector matrix in the spectral decomposition, and Λ is the diagonal eigenvalue matrix in the spectral decomposition;

[0070] Heterogeneous graph convolution is constructed using the shared feature basis U and feature spectrum Λ of the hyper-Laplacian matrix, as shown in the following formula:

[0071]

[0072] in These represent the spectral filter functions for intra-layer graph convolution and inter-layer graph convolution, respectively.

[0073] (2.2.3) Intra-layer graph convolution construction:

[0074] The in-layer graph convolutional spectrum filter function is approximated using a polynomial approximation technique. The formula is expressed as follows:

[0075]

[0076] Intralayer super Laplace matrix The graph convolution on each modal layer can be parameterized by taking the union of the Laplacian matrices of all the graphs within the modal layers, as shown in the following formula:

[0077]

[0078] r = {RSS, PDV, PDK, TCM}

[0079] In the formula, κ j =cos((j+0.5)π / (K+1)) represents the Chebyshev interpolation node; P k (·) represents the Chebyshev approximation polynomial, expressed as: P k (x)=2x·P k-1 (x)-P k-2 (x), P1 = x, P0 = 1; These represent the node embeddings of the in-layer graph convolution output on the r-mode in-layer graph, respectively. , respectively, represent the polynomial approximation coefficients of the convolution of the inner graph of the r-mode layer;

[0080] (2.2.4) Construction of inter-layer graph convolution:

[0081] Inter-layer graph convolution is constructed using a graph attention network, which performs a weighted summation of node embeddings within graphs of different modalities, as shown in the following formula:

[0082]

[0083] In the formula, ⊙ represents the dot product operation; Let represent the embedding weights of the i-th node in the graph of the r-th modality layer, respectively, and their calculation expressions are as follows:

[0084]

[0085] In the formula, This represents the node weight mapping matrix for the r-mode; Represents the attention vector of the r-th mode; This represents the embedding of the i-th node in the output of the inter-layer convolution;

[0086] (2.2.5) The output of a heterogeneous graph convolutional layer consists of intra-layer node embeddings and inter-layer node embeddings, represented as:

[0087]

[0088] (2.3) Global summation pooling is used to extract the abstract features of the reused network to obtain the global representation of the reused network; in this embodiment, the global representation is calculated according to the following formula:

[0089]

[0090] in, This represents the global node embedding of the graph pooling output; and represents the embedding of the i-th node in the RSS, PDV, PDK, and TCM modes of the heterogeneous graph convolution output, respectively.

[0091] (2.4) A classifier is constructed by stacking multiple fully connected layers, and the output sequence of the global graph pooling layer is mapped to decision statistics; the processing formula of the classifier is expressed as follows:

[0092]

[0093] in, These represent the weights of the two fully connected layers, respectively. and The bias vectors are respectively; σ(·) represents the ReLU activation function.

[0094] (2.5) The Node Embedding module is cascaded with the heterogeneous graph convolution, global pooling layer and classifier in sequence to obtain the heterogeneous graph convolutional network;

[0095] Step 3. Set initial hyperparameters, input the reused network into the heterogeneous graph convolutional network, and iteratively train it until convergence to obtain the optimized heterogeneous graph convolutional network. In this embodiment, the cross-entropy function is used to train the heterogeneous graph convolutional network, and the Adam optimizer is used to iteratively update the network parameters until the convolutional network converges.

[0096] Step 4. Use the trained heterogeneous graph convolutional network to map the reused network into decision statistics. Then, use the Monte Carlo method to construct a detection threshold with a fixed false alarm rate based on the decision statistics. Finally, determine the final spectrum occupancy state based on the binary assumptions to achieve spectrum sensing.

[0097] In this embodiment, the decision statistic T is specifically obtained by processing the output of the heterogeneous graph convolutional network as follows:

[0098]

[0099] in, G represents the class probabilities under hypotheses H1 and H0 in the decision vector, respectively; θ (·) represents the expression for a heterogeneous graph neural network.

[0100] Example 2: The overall implementation steps of the method proposed in this example are the same as in Example 1, and will now be referred to... Figure 2 The following is a more detailed description of the implementation process of using multiplexed networks to represent multimodal information and model the potential correlations between multimodal information in this invention:

[0101] Step A. Building node attributes:

[0102] a1) Using the modality mapping function F RSS (·) Transfer the sensing signal y on the i-th receiving antenna i Mapped to RSS modal information Construct an inner graph G of the RSS mode layer using all RSS mode information on the receiving antenna array. RSS Node attributes

[0103]

[0104]

[0105] a2) Using the modality mapping function F PDV (·) Transfer the sensing signal y on the i-th receiving antenna i Mapped to PDV modal information Construct the PDV mode layer inner diagram G using all PDV mode information from the receiving antenna array. PDV Node attributes

[0106]

[0107] Where Re(y) i (t)) and Im(y i (t) represents the real and imaginary parts of the sampled value of the sensing signal at sampling time t, respectively; Let be the instantaneous phase of the sampled value of the sensed signal at time t on the i-th receiving antenna; the phase is normalized by the modulo 2π function. Within the range [0, 2π]; Phase difference Θ i The mean.

[0108] a3) Use the modality mapping function F PDK (·) Transfer the sensing signal y on the i-th receiving antenna i Mapped to PDK modal information Construct the PDK mode layer inner graph G using all PDK mode information on the receiving antenna array. PDK Node attributes

[0109]

[0110]

[0111] a4) Use the modality mapping function F TCM (·) Transfer the sensing signal y on the i-th receiving antenna i Mapped to TCM modal information Construct the TCM mode layer inner diagram G using all TCM mode information from the receiving antenna array. TCM Node attributes

[0112]

[0113] in, ξ represents the sampling vector using a rectangular window function with length ξ and step size 1; τ represents the number of sampling vectors using the rectangular window function. This represents the trajectory matrix of the signal sensed on the i-th receiving antenna; Let M be the trajectory covariance matrix of the signal sensed on the i-th receiving antenna, which also represents the TCM mode information of the signal sensed on the i-th receiving antenna. The specific parameter settings for constructing the node attributes are as follows: M = 8; τ = 8; ξ = N - τ.

[0114] Step B. Adjacency Matrix Construction: The real-valued distances between modal information are calculated using radial basis functions to construct the adjacency matrix of the intra-layer graph. The specific process of constructing the adjacency matrix of each modality's intra-layer graph can be represented as follows:

[0115] b1) Calculate RSS modal information using radial basis functions The real-valued distance between the nodes is used to construct the intra-modal graph G of the RSS. RSS adjacency matrix

[0116]

[0117] b2) Calculate PDV modal information using radial basis functions Constructing the intra-modal graph G of PDV using real-valued distances between them PDV adjacency matrix

[0118]

[0119] b3) Calculate PDK modal information using radial basis functions Constructing the PDK modal layer inner graph G using real-valued distances between them PDK adjacency matrix

[0120]

[0121] b4) Calculate TCM modal information using radial basis functions The real-valued distance between the points is used to construct the intra-modal graph G of the TCM. TCM adjacency matrix

[0122]

[0123] b5) Inter-layer adjacency matrix It is a set of binary relations representing the inter-layer connectivity between nodes in the graphs of different modal layers, consisting of a coupling graph C and an identity matrix I. Since multimodal information is constructed from different mapping transformations of the same received signal, the coupling graph C can be predefined as:

[0124]

[0125] Step C. Node Embedding Module Construction: The Node Embedding module uses a non-linear mapping function to map node features to a common dimensional space to facilitate the effective aggregation of multimodal information, as follows:

[0126] c1) Use a fully connected function to plot the RSS mode layer in-plane graph G on the receiving antenna array. RSS Node attribute TRSS Mapping to node embedding

[0127] X RSS =σ(T) RSS W RSS +b RSS )

[0128] c2) Use a fully connected function to plot the PDV mode layer in-plane G on the receiving antenna array. PDV Node attribute T PDV Mapping to node embedding

[0129] X PDV =σ(T) PDV W PDV +b PDV )

[0130] c3) Use a fully connected function to connect the PDK mode layer in-plane graph G on the receiving antenna array. PDK Node attribute T PDK Mapping to node embedding

[0131] X PDK =σ(T) PDK W PDK +b PDK )

[0132] c4) Use convolution and 3D convolution of size 2×τ×τ with stride 1 to plot the TCM mode layer in-matrix G on the receiving antenna array. TCM Node attribute T TCM Mapping to node embedding

[0133]

[0134] in, These represent the weight mapping matrices of the fully connected layer in RSS, PDV, and PDK modes, respectively. These represent the bias parameters of the fully connected layer in RSS, PDV, and PDK modes, respectively. This represents the convolution kernel of a 3D convolutional layer, with a size of 2×τ×τ and a stride of 1. Let p be the output feature of the p-th output channel of the 3D convolution. Embed the node sensing the signal on the i-th receiving antenna; These represent the embedding matrices of node attributes in the inner graphs of the RSS, PDV, PDK, and TCM modal layers, respectively.

[0135] Example 3: The overall implementation steps of the spectrum sensing method proposed in this example are the same as in Example 1, and will now be referred to... Figure 2 The specific construction steps of the spectrum sensing framework based on heterogeneous graph convolutional networks are further described in detail below:

[0136] Step 1. Construct heterogeneous graph convolution: Decompose the hyperLaplacian matrix into intra-layer hyperLaplacian matrix and inter-layer hyperLaplacian matrix; use intra-layer graph convolution on the intra-layer hyperLaplacian matrix to aggregate the same modal information in the intra-layer graphs of each modality; use inter-layer graph convolution on the inter-layer hyperLaplacian matrix to aggregate node embeddings in different intra-layer graphs.

[0137] (1.1) Hyper-Laplace matrix spectral decomposition: decomposing the hyperadjacency matrix of a reused network Intra-layer hyperadjacency matrix and inter-layer hyperadjacency matrix

[0138]

[0139] The hyperLaplace matrix of decomposition and reuse networks For the intralayer hyperLaplace matrix and interlayer hyperLaplace matrix

[0140]

[0141] The super Laplacian matrix is ​​defined as follows: It is a diagonal hyperdegree matrix and This represents the adjacency matrix of the graph within the r-th modal layer. Let be the adjacency matrix of the inter-layer graphs of the r-th and k-th modes; Represents the Laplace matrix of the inner graph of the r-th mode layer; Let be the Laplace matrix of the inter-layer diagrams for modes r and k; the inter-layer hyperLaplace matrix is ​​defined as follows: Denotes the Laplace matrix of the coupling graph C. It is an identity matrix;

[0142] The eigenvalue matrix and eigenvector matrix are obtained by performing eigendecomposition on the hyperLaplace matrix, as shown in the following formula:

[0143]

[0144] Here, U represents the eigenvector matrix in the spectral decomposition, and Λ is the diagonal eigenvalue matrix in the spectral decomposition. Therefore, heterogeneous graph convolutions are constructed using the shared eigenbase U of the hyperLaplacian matrix and the eigenspectrum Λ.

[0145]

[0146] in These represent the spectral filter functions for intra-layer graph convolution and inter-layer graph convolution, respectively.

[0147] (1.2) Intra-layer graph convolution: Intra-layer graph convolution uses a polynomial approximation technique to approximate the spectral filter function to reduce computational complexity. The expression for the polynomial approximation is:

[0148]

[0149] In the formula, θ k These are the approximation coefficients in the polynomial. The formula for calculating the in-layer graph convolution based on polynomial approximation is:

[0150]

[0151] Intralayer super Laplace matrix The graph convolution on each modal layer inner graph can be parameterized by taking the union of the Laplacian matrices of all the inner graphs of the modal layers. Therefore, the graph convolution on each inner graph of the modal layer is expressed as:

[0152]

[0153] In the formula, κ j =cos((j+0.5)π / (K+1)) represents the Chebyshev interpolation node; P k (·) represents the Chebyshev approximation polynomial, and its expression is: P k (x)=2x·P k-1 (x)-P k-2 (x), P1 = x, P0 = 1; and These represent the output node embeddings of in-layer graph convolutions on the in-layer graphs of the RSS, PDV, PDK, and TCM modalities, respectively. represents the polynomial approximation coefficients of the in-layer graph convolution on the in-layer graphs of the RSS, PDV, PDK, and TCM modes, respectively.

[0154] (1.3) Inter-layer graph convolution: Inter-layer graph convolution is constructed using a graph attention network, representing a weighted sum of node embeddings of graphs within different modalities, as shown in the following formula:

[0155]

[0156] In the formula, ⊙ represents the dot product operation; These represent the embedding weights of the i-th node in the inner graphs of the RSS, PDV, PDK, and TCM modal layers, respectively, and their calculation expressions are as follows:

[0157]

[0158] In the formula, These represent the node weight mapping matrices of the graph within each modal layer; and These represent the attention vectors in the graph of each modal layer; This indicates the embedding of the output node of the interlayer convolution on the RSS modality interlayer graph.

[0159] (1.4) The output node embedding of a heterogeneous graph convolutional layer consists of intra-layer node embedding and inter-layer node embedding, as shown in the following formula:

[0160]

[0161] Step 2. Constructing Global Graph Pooling: Global summation graph pooling is used to extract reusable network structure information and generate a fixed-size global representation. The processing flow is as follows:

[0162]

[0163] in, This represents the global node embedding of the graph pooling output; represents the embedding of the i-th node in the RSS, PDV, PDK, and TCM modes of the heterogeneous graph convolution output, respectively.

[0164] Step 3. Construct the classifier: The classifier maps the global representation of the reusable network to decision statistics. The classifier consists of two stacked fully connected layers, and its processing formula can be expressed as:

[0165] X F =softmax(σ(X) S W F (1) +b1)W F (2) +b2)

[0166] in, These represent the weight matrices of the two fully connected layers, respectively. and The bias vectors are respectively; σ(·) represents the ReLU activation function.

[0167] Step 4. Constructing Decision Statistics: Map the received signal on the receiving antenna array to the multiplexing network using mode mapping functions and radial basis functions; input the multiplexing network into a heterogeneous graph convolutional network to construct decision vectors.

[0168]

[0169] in, G represents the class probabilities under hypotheses H1 and H0 in the decision vector, respectively; θ(·) denotes the expression of the proposed model; according to the Neiman-Pearson criterion, the optimal decision statistic is set as follows:

[0170]

[0171] Step 5. Construct the detection threshold: The spectrum occupancy state can be determined by comparing the index value corresponding to the maximum probability value in the decision vector. The calculation process is as follows:

[0172]

[0173] The IEEE 802.22 protocol specifies that the detection threshold needs to be constrained by a fixed false alarm rate. The steps for constructing a detection threshold with a fixed false alarm rate based on the Monte Carlo method are as follows:

[0174] (4.1) Collect sensing signals under hypothesis H0 to construct a multiplexing network To construct a new dataset;

[0175]

[0176] (4.2) Calculate the decision statistic T for the new dataset using the trained heterogeneous graph convolutional network. H0 ;

[0177] (4.3) Sort the decision statistics in descending order and construct the decision statistics dataset Ω;

[0178] (4.4) Use Obtain the constant false alarm rate (CFAR) detection threshold, where, Indicates the false alarm rate;

[0179] (4.5) The spectral occupancy status of the target frequency band can be expressed as:

[0180] The effects of the present invention will be further explained below with reference to experiments.

[0181] 1. Experimental conditions:

[0182] The experiments of this invention used a high-performance server as the hardware platform. The server was equipped with an Intel(R) Xeon(R) Gold 5218 processor and an NVIDIA GeForce RTX 2080Ti graphics card. The operating system was Windows Server 2019, the deep learning environment was PyCharm 2021, the PyTorch version was 1.10.2, and the CUDA version was 10.2.

[0183] 2. Experiment Content:

[0184] The performance of the spectrum sensing algorithm of this invention was tested: test data was input into the optimized heterogeneous graph convolutional network, and PD-SNR curves and ROC curves at signal-to-noise ratios of -30dB to -12dB and a false alarm rate of 0.1 were plotted based on the detection results. The detection probability and the area under the ROC curve (AUC) were used as evaluation metrics to measure the accuracy, robustness, and generalization performance of the detection algorithm under various non-ideal channel environments.

[0185] This experiment evaluates the performance of the proposed method under various scenarios, including AWGN, CGGN, FWGN, large-scale fading, robustness and scalability, OFDM-based cooperative spectrum sensing, and imperfect reporting channels. In each channel environment, the detection performance of the proposed method is compared with mainstream methods from the perspectives of ROC curves and PD-SNR curves. The mainstream methods include ME-GM (paper: Performance of statistical tests for single-source detection using random matrix theory), PDVD (paper: Phase difference variance-based low complexity spectrum sensing scheme), CM-CNN (paper: Deep cm-cnn for spectrum sensing in cognitive radio), DS2MA (paper: DS2MA: A Deep Learning-Based Spectrum Sensing Scheme for a Multi-Antenna Receiver), GCN (paper: A Graph Convolution Network Based Adaptive Cooperative Spectrum Sensing in Cognitive Radio Network), and SM-GIN (paper: Graph neural network based cooperative spectrum sensing for cognitive radio).

[0186] 3. Simulation Results and Analysis:

[0187] Figure 3 and Figure 4 This is a comparison chart of the detection results of the method of this invention and the benchmark method in a noisy channel environment, wherein... Figure 3 (a) and Figure 4 (a) ROC curves for AWGN and CGGN channels, respectively. Figure 3 (b) and Figure 4 (b) PD-SNR curves for AWGN and CGGN channels, respectively. By comparing existing spectrum sensing methods and the method of this invention on two evaluation curves, it can be seen that the method of this invention performs the best and is significantly superior to the mainstream benchmark methods in the prior art. Figure 4 In the figure, "HGCN-SS-AWGN" represents the ROC curve and PD-SNR curve of the detection performance under AWGN channel pre-training and under CGGN channel environment. The spectrum sensing algorithm proposed in this invention does not require retraining or fine-tuning and can directly predict the spectrum occupancy status under CGGN environment, demonstrating robustness to CGGN channel fading environment.

[0188] Figure 5 The figures show the detection performance of the method of this invention and mainstream spectrum sensing methods in large-scale fading environments, where (a) represents the PD-SNR curve and (b) represents the ROC curve. By comparing the mainstream methods with the method of this invention, the method of this invention effectively adapts to large-scale fading scenarios and mitigates the impact of path loss and shadow fading on detection performance. Figure 5 In the figure, "HGCN-SS-AWGN" represents the ROC curve and PD-SNR curve of the pre-trained model under AWGN channel conditions in a large-scale channel environment. The spectrum sensing algorithm proposed in this invention does not require retraining or fine-tuning and shows detection performance comparable to the model trained under large-scale fading scenarios, indicating that the method of this invention has good generalization ability.

[0189] Figure 6 The ROC and PD-SNR curves of the pre-trained model are shown under three different numbers of modulation symbols and three different numbers of receiving antennas. The algorithm proposed in this invention does not require retraining or fine-tuning and directly completes the detection of the spectrum state. The detection performance increases with the increase of the number of modulation symbols and the number of receiving antennas.

[0190] Figure 7 The results show the detection performance of the proposed method compared to mainstream methods in OFDM-based cooperative spectrum sensing. By comparing mainstream methods with the proposed method, the proposed method demonstrates a significant performance improvement over other benchmark methods. The proposed method leverages the multipath effect of the transmission channel in cooperative spectrum sensing and incorporates multiple sub-users into the decision-making process, thereby enhancing detection performance.

[0191] Figure 8 This chart compares the detection results of the proposed method and mainstream methods in an imperfect channel environment within a cooperative spectrum sensing scenario. By comparing existing cooperative spectrum sensing methods and the proposed method on two evaluation curves, it can be seen that the proposed method has the highest area under the ROC curve and the highest detection gain.

[0192] The parts of this invention not described in detail are common knowledge to those skilled in the art.

[0193] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Obviously, those skilled in the art, after understanding the content and principle of the present invention, may make various modifications and changes in form and detail without departing from the principle and structure of the present invention. However, these modifications and changes based on the concept of the present invention are still within the scope of protection of the claims of the present invention.

Claims

1. A spectrum sensing method based on heterogeneous graph convolutional networks, characterized in that... This approach integrates multimodal information, multiplexed networks, and heterogeneous graph convolutional networks into a unified spectrum sensing framework to achieve spectrum occupancy state detection in an end-to-end manner. The specific steps are as follows: (1) Constructing a reused network: (1.1) The received signal is mapped into multiple modal information by using different modal mapping functions, and the node attributes of the inner graph in the multiplexing network are constructed using the modal information; (1.2) Construct the adjacency matrix of the inner graph of the multiplexing network using radial basis functions and multiple modal information, and construct the adjacency matrix of the inter-layer graph of the multiplexing network using coupling graph and identity matrix; (1.3) The reused network consists of three parts: node attributes, intra-layer adjacency matrix, and inter-layer adjacency matrix; (2) Constructing a heterogeneous graph convolutional network: (2.1) Use a nonlinear mapping function to map different modal information to a common dimension space and construct the NodeEmbedding module; (2.2) Heterogeneous graph convolution is composed of intra-layer graph convolution and inter-layer graph convolution. Intra-layer graph convolution uses graph convolution to aggregate the same mode information on the receiving antenna array from the perspective of the spectral domain of node attributes. Inter-layer graph convolution uses graph attention mechanism to assign corresponding weights to the information of different modes on the same receiving antenna and aggregate multi-mode information according to the inter-layer adjacency matrix. (2.3) Use global summation pooling to extract the abstract features of the reused network and obtain the global representation of the reused network; (2.4) A classifier is constructed by stacking multiple fully connected layers, and the output sequence of the global graph pooling layer is mapped to decision statistics; (2.5) The Node Embedding module is cascaded with the heterogeneous graph convolution, global pooling layer and classifier in sequence to obtain the heterogeneous graph convolutional network; (3) Set initial hyperparameters, input the reused network into the heterogeneous graph convolutional network, and iteratively train it until convergence to obtain the optimized heterogeneous graph convolutional network. (4) The trained heterogeneous graph convolutional network is used to map the reused network into decision statistics. Then, the Monte Carlo method is used to construct a detection threshold with a fixed false alarm rate based on the decision statistics. Finally, the final spectrum occupancy state is determined based on the binary assumptions to achieve spectrum sensing.

2. The method according to claim 1, characterized in that: The multiple modal information mentioned in step (1.1) includes any two or more of the following four modal information: received signal strength RSS, phase difference variance PDV, phase difference kurtosis PDK, and trajectory covariance matrix TCM.

3. The method according to claim 2, characterized in that: In step (1.1), different mode mapping functions are used to map the received signal into multiple mode information, including using the mapping function F. RSS (·),F PDV (·),F PDK (·) and F TCM (·) Map the received signal into received signal strength mode information respectively. Phase difference variance modal information Phase difference kurtosis modal information Trajectory covariance modal information The node attributes of the multi-modal layer graph in the reused network are constructed using these four modal information, as shown below: (1.1.1) Using F RSS (·) Map the received signal into RSS modal information Construct an in-layer graph G using all RSS modal information. RSS Node attributes (1.1.2) Using F PDV (·) Map the received signal into PDV mode information Construct the in-layer graph G using all PDV modal information. PDV Node attributes Where Re(y i (t)) and Im(y i (t) represents the real and imaginary parts of the sampled value of the received signal at sampling time t, respectively; Let be the instantaneous phase of the sampled value of the sensed signal on the i-th receiving antenna at time t; the phase is normalized by the modulo 2π function. Within the range [0, 2π]; Phase difference Θ i The mean; (1.1.3) Using F PDK (·) Map the received signal into PDK mode information Construct the in-layer graph G using all PDK modal information. PDK Node attributes (1.1.4) Using F TCM (·) Map the received signal into TCM mode information Construct the intralayer graph G using all TCM modal information. TCM Node attributes in, ξ is the sampling vector of a rectangular window function with length ξ and step size 1; τ represents the number of sampling vectors of the rectangular window function. This represents the trajectory matrix of the signal sensed on the i-th receiving antenna.

4. The method according to claim 2, characterized in that: In step (1.2), radial basis functions are used to map multiple modal information into the adjacency matrix of the corresponding modal layer inner graph, including the adjacency matrix A of the RSS modal layer inner graph. RSS The adjacency matrix A of the inner graph of the PDV modal layer PDV The adjacency matrix A of the PDK modal layer inner graph PDK The adjacency matrix A of the TCM modal layer inner graph TCM The method of constructing inter-layer adjacency matrices in a multiplexed network using coupling graphs and identity matrices is described. It is expressed as follows: (1.2.1) Using radial basis functions to calculate modal information and Constructing the intra-layer graph G using real-valued distances between layers r adjacency matrix r = {RSS, PDV, PDK, TCM} i,j = 1,2,...,M; (1.2.2) Constructing the adjacency matrix of the inter-layer graph using the coupling graph C and the identity matrix I 5. The method according to claim 2, characterized in that: In step (2.1), the Node Embedding module uses a non-linear mapping function to map node attributes to a common dimension space, as follows: (2.1.1) Use fully connected functions to connect the node attributes T of the inner graphs of the RSS, PDV, and PDK modal layers respectively. RSS T PDV T PDK Mapping to node embedding X RSS =σ(T RSS W RSS +b RSS ) X PDV =σ(T PDV W PDV +b PDV ) X PDK =σ(T PDK W PDK +b PDK ); (2.1.2) Use a 3D convolution with a kernel size of 2×τ×τ and a stride of 1 to extract the node attributes T of the inner graph of the TCM modality layer. TCM Mapping to node embedding in, These represent the weight mapping matrices of the fully connected layer in RSS, PDV, and PDK modes, respectively. These are the bias parameters for the fully connected layer; Represents the convolution kernel of a 3D convolutional layer; is the output feature of the p-th output channel of the 3D convolution. f 6. The method according to claim 1, characterized in that: The heterogeneous graph convolution composed of intra-layer graph convolution and inter-layer graph convolution described in step (2.2) is implemented as follows: (2.2.1) Hyper-Laplace matrix decomposition: The hyperLaplace matrix of decomposition and reuse networks For the intralayer hyperLaplace matrix and interlayer hyperLaplace matrix The formula is as follows: in, It is a diagonal hyperdegree matrix and Represents the Laplace matrix of the r-mode layer inner graph; For interlayer hyperLaplace matrix; (2.2.2) Construction of heterogeneous graph convolution: The eigenvalue matrix and eigenvector matrix are constructed from the eigenvalue decomposition of the hyperLaplace matrix, as shown in the following formula: Where U represents the eigenvector matrix in the spectral decomposition, and Λ is the diagonal eigenvalue matrix in the spectral decomposition; Heterogeneous graph convolution is constructed using the shared feature basis U and feature spectrum Λ of the hyper-Laplacian matrix, as shown in the following formula: in and These represent the spectral filter functions for intra-layer graph convolution and inter-layer graph convolution, respectively. (2.2.3) Intra-layer graph convolution construction: The in-layer graph convolutional spectrum filter function is approximated using a polynomial approximation technique. The formula is expressed as follows: Intralayer super Laplace matrix The graph convolution on each modal layer can be parameterized by taking the union of the Laplacian matrices of all the graphs within the modal layers, as shown in the following formula: In the formula, κ j =cos((j+0.5)π / (K+1)) represents the Chebyshev interpolation node; P k (·) represents the Chebyshev approximation polynomial, expressed as: P k (x)=2x·P k-1 (x)-P k-2 (x), P1 = x, P0 = 1; These represent the node embeddings of the in-layer graph convolution output on the r-mode in-layer graph, respectively. , respectively, represent the polynomial approximation coefficients of the convolution of the inner graph of the r-mode layer; (2.2.4) Construction of inter-layer graph convolution: Inter-layer graph convolution is constructed using a graph attention network, which performs a weighted summation of node embeddings within graphs of different modalities, as shown in the following formula: In the formula, ⊙ represents the dot product operation; Let represent the embedding weights of the i-th node in the graph of the r-th modality layer, respectively. Their calculation expressions are as follows: In the formula, This represents the node weight mapping matrix for the r-mode; Represents the attention vector of the r-th mode; This represents the embedding of the i-th node in the output of the inter-layer convolution; (2.2.5) The output of a heterogeneous graph convolutional layer consists of intra-layer node embeddings and inter-layer node embeddings, represented as:

7. The method according to claim 2, characterized in that: The pooling representation described in step (2.3) is calculated according to the following formula: in, This represents the global node embedding of the graph pooling output; and represents the embedding of the i-th node in the RSS, PDV, PDK, and TCM modes of the heterogeneous graph convolution output, respectively.

8. The method according to claim 1, characterized in that: The classifier described in step (2.4) has the following processing formula: X F =softmax(σ(X S W F (1) +b1)W F (2) +b2), in, and These represent the weights of the two fully connected layers, respectively. and The bias vectors are respectively; σ(·) represents the ReLU activation function.

9. The method according to claim 1, characterized in that: The iterative training described in step (3) specifically involves training the heterogeneous graph convolutional network using the cross-entropy function and iteratively updating the network parameters using the Adam optimizer until the convolutional network converges.

10. The method according to claim 1, characterized in that: The decision statistic T mentioned in step (4) is obtained by processing the output of the heterogeneous graph convolutional network as follows: in, and G represents the class probabilities under hypotheses H1 and H0 in the decision vector, respectively; θ (·) represents the expression for a heterogeneous graph neural network; T is the decision statistic.

Images