Spectrum sensing method based on heterogeneous graph convolutional network

Abstract

The invention discloses a spectrum sensing method based on a heterogeneous graph convolutional network, and mainly solves the problems of poor detection effect, weak robustness and generalization and poor expandability in a non-ideal channel environment in the prior art. Comprising the following steps: 1) representing multi-modal information by using nodes in a multiplexing network, and modeling potential correlation of the multi-modal information by using an inner layer edge and an interlayer edge; 2) different modal information is mapped to a common dimension space by using a nonlinear mapping function, heterogeneous graph convolution used for multiplexing network node attribute aggregation is constructed, and a heterogeneous graph convolution network is obtained in combination with a global pooling layer and a classifier; 3) inputting the multiplexing network into the heterogeneous graph convolutional network, and training to converge; and 4) realizing spectrum sensing by using the trained heterogeneous graph convolutional network. According to the method, the multi-modal information and the heterogeneous graph convolutional network are combined, optimization is carried out in an end-to-end mode, performance fading caused by a non-ideal channel environment is effectively relieved, and meanwhile the robustness and generalization of a spectrum sensing algorithm are enhanced.

Description

Technical Field

[0001] The present 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 network, which can be used in the fields of spectrum detection, weak signal detection, etc. in the field of Internet of Things. Background Art

[0002] With the rapid development of IoT applications such as smart cities, smart homes, and smart agriculture, IoT devices have been widely studied and deployed. However, the spectrum resources designated for IoT services cannot meet the rapidly growing demand, resulting in competition with other RF solutions for shared spectrum resources. In addition, most of the remaining spectrum resources have been authorized to specific services, which has led to spectrum scarcity and spectrum congestion in IoT services. Cognitive radio technology is considered to be the most promising technology to alleviate spectrum scarcity. Spectrum sensing technology, which can detect available spectrum resources, is the most basic and important part of cognitive wireless technology.

[0003] In traditional spectrum sensing schemes, the test statistics are constructed based on theoretical statistical models, such as the received signal strength that obeys the chi-square distribution, the covariance matrix eigenvalue that obeys the Tracy-Widom distribution, the phase difference variance that obeys the uniform distribution, etc. However, traditional spectrum sensing schemes have certain drawbacks: 1) There is a deviation between the statistical model of the test statistics in the actual communication system and the theoretical statistical model; 2) The generalization and robustness of traditional algorithms are weak, and their detection performance is easily affected by non-ideal channel environments such as noise disturbance, channel fading and malicious interference.

[0004] Spectrum sensing algorithms based on neural networks can effectively overcome the impact of non-ideal channel environments on detection performance without the need to build complex statistical models. Although deep learning has developed rapidly in recent years, its development in the field of spectrum sensing is still in its initial stages. Spectrum sensing schemes based on deep learning have the following limitations: 1) Current deep learning-based methods use single-modal test statistics as input, such as original signals, covariance matrices, energy, and spatiotemporal correlations, and their detection performance is susceptible to non-ideal channel environments; 2) Current deep learning-based 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] In view of the shortcomings of the prior art, the present invention aims to propose a spectrum sensing method based on heterogeneous graph convolutional networks to solve the problems of poor detection effect, poor generalization and robustness, and the need to establish complex statistical models in existing detection schemes under non-ideal channel environments. The present invention uses a multiplexing network to model multimodal information while considering the potential correlation between multimodal information; the multiplexing network uses multimodal information as node attributes, and uses inter-layer edges 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 multiplexing networks and heterogeneous graph convolutional networks to optimize the spectrum sensing model in an end-to-end manner, thereby effectively improving detection accuracy, robustness, and generalization.

[0006] The specific steps of the present invention to achieve the above-mentioned purpose include the following:

[0007] (1) Building a multiplexing network:

[0008] (1.1) Using different modal mapping functions, the received signal is mapped into multiple modal information, and the modal information is used to construct the node attributes of the inner graph in the multiplexing network;

[0009] (1.2) Using radial basis functions and multiple modal information to construct the adjacency matrix of the intra-layer graph in the multiplexing network, and using coupling graphs and identity matrices to construct the adjacency matrix of the inter-layer graph in the multiplexing network;

[0010] (1.3) The reuse 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 nonlinear mapping functions to map different modal information into a common dimensional space and construct a NodeEmbedding module;

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

[0014] (2.3) Use global sum pooling to extract abstract features of the reused network and obtain a 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 into a decision statistic;

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

[0017] (3) Setting the initial hyperparameters, inputting the reused network into the heterogeneous graph convolutional network, iteratively training it until convergence, and obtaining the optimized heterogeneous graph convolutional network;

[0018] (4) The trained heterogeneous graph convolutional network is used to map the multiplexing network into a decision statistic. The Monte Carlo method is then used to construct a detection threshold with a fixed false alarm rate based on the decision statistic. The final spectrum occupancy state is then determined based on the binary hypothesis 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. The multiplexing network can not only express the multimodal information in the form of nodes, but also capture the hidden correlation in the multimodal information in the form of intra-layer adjacency matrix and inter-layer adjacency matrix.

[0021] Second, the present invention integrates multimodal information and heterogeneous graph convolutional networks into a unified framework. It can use the intra-layer graph convolution in the heterogeneous graph convolution to extract the high-order features embedded in the graph nodes in each modal layer from the spectral domain perspective, and can also use the inter-layer graph convolution in the heterogeneous graph convolution to adaptively aggregate the multimodal information in the multiplexing network from the spatial domain perspective, thereby effectively improving the detection accuracy, robustness and generalization performance of spectrum sensing. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is the overall implementation flow chart of the method of the present invention;

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

[0024] Figure 3 : It is a comparison diagram of the simulation results of the detection performance of the present invention and the existing method under the AWGN channel; wherein (a) is the ROC curve under the AWGN channel, and (b) is the PD-SNR curve under the AWGN channel;

[0025] Figure 4 It is a comparison diagram of the simulation results of the detection performance of the present invention and the existing method under the CGGN channel; wherein (a) is the ROC curve under the CGGN channel, and (b) is the PD-SNR curve under the CGGN channel;

[0026] Figure 5The simulation results of the detection performance of the present invention and the existing method under large-scale fading channels are compared; (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 results of the extended performance evaluation of the method of the present invention are shown in FIG. (a) is the ROC curve under the extended performance evaluation of the receiving sample length, and (b) is the PD-SNR curve under the extended performance evaluation of the number of receiving antennas.

[0028] Figure 7 Graph showing the detection performance evaluation results of the method of the present invention in a collaborative spectrum sensing scenario; (a) is a ROC curve of collaborative spectrum sensing based on an OFDM mechanism, and (b) is a PD-SNR curve of collaborative spectrum sensing based on an OFDM mechanism;

[0029] Figure 8 It is a diagram of the detection performance evaluation result of the method of the present invention under an imperfect reporting channel in a collaborative spectrum sensing scenario; wherein (a) is the ROC curve under an incomplete fading channel of collaborative spectrum sensing, and (b) is the PD-SNR curve under an incomplete fading channel of collaborative spectrum sensing. DETAILED DESCRIPTION

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

[0031] Example 1: Reference Figure 1 The present invention proposes a spectrum sensing method based on heterogeneous graph convolutional network, which integrates multimodal information, multiplexing network and heterogeneous graph convolutional network into a unified spectrum sensing framework to realize spectrum occupancy state detection in an end-to-end manner; the specific steps are as follows:

[0032] Step 1. Construct a reuse network: The reuse network consists of three parts: node attributes, intra-layer adjacency matrix, and inter-layer adjacency matrix, which are used to characterize multimodal information and the interaction between modal information. The construction steps are as follows:

[0033] (1.1) Using different modal mapping functions, the received signal is mapped into multiple modal information respectively, and the modal information is used to construct the node attributes of the inner graph in the multiplexed network; the multiple modal information includes at least any two or more of the four modal information of received signal strength (RSS), phase difference variance (PDV), phase difference kurtosis (PDK) and trajectory covariance matrix (TCM). This embodiment uses the mapping function FRSS (·),F PDV (·),F PDK (·) and F TCM (·) Map the received signal into received signal strength modal information Phase difference variance modal information Phase difference kurtosis modal information and trajectory covariance modal information These four modal information are used to construct the node attributes of the graphs in multiple modal layers in the reuse network, which are expressed as follows:

[0034] (1.1.1) Using F RSS (·) Map the received signal into RSS modal information Use all RSS modal information to build the in-layer graph G RSS Node properties

[0035]

[0036]

[0037] (1.1.2) Use F PDV (·) Map the received signal into PDV modal information Construct the intra-layer graph G using all PDV modality information PDV Node properties

[0038]

[0039] Among them, Re(y i (t)) and Im(y i (t)) represent the real part and imaginary part of the sampled value of the received signal at sampling time t respectively; is the instantaneous phase of the sampled value of the perceived signal at the tth moment on the i-th receiving antenna; the mod2π function normalized phase In the range [0,2π]; is the phase difference Θ i The mean of

[0040] (1.1.3) Use F PDK (·) Map the received signal into PDK modal information Construct the layer internal graph G using all PDK modal information PDK Node properties

[0041]

[0042]

[0043] (1.1.4) Using F TCM (·) Map the received signal into TCM modal information Use all TCM modality information to construct the intra-layer graph G TCM Node properties

[0044]

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

[0046] (1.2) Using radial basis functions and multiple modal information to construct the adjacency matrix of the intra-layer graph in the multiplexing network, and using coupling graphs and identity matrices to construct the adjacency matrix of the inter-layer graph in the multiplexing network. In this embodiment, radial basis functions are used to map multiple modal information into adjacency matrices of the intra-layer graphs of the corresponding modal layers, including the adjacency matrix A of the intra-layer graph of the RSS modal layer. RSS , the adjacency matrix A of the graph in the PDV modal layer PDV , the adjacency matrix A of the graph in the PDK modal layer PDK and the adjacency matrix A of the graph in the TCM modal layer TCM ; The coupling graph and the identity matrix are used to construct the inter-layer adjacency matrix in the multiplexing network It is expressed as follows:

[0047] (1.2.1) Calculate the modal information T using radial basis functions r i and T r j The real-valued distance between them is used to construct the intra-layer graph G r The adjacency matrix of

[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 reuse network consists of three parts: node attributes, intra-layer adjacency matrix and inter-layer adjacency matrix;

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

[0055] (2.1) The node attributes in the reuse network come from different modes. Nonlinear mapping functions are used to map different modal information to a common dimensional space to build a Node Embedding module. The implementation is as follows:

[0056] (2.1.1) Use the fully connected function to transform the node attributes T of the RSS, PDV, and PDK modal layers into RSS , T PDV , T PDK Mapping to node embeddings

[0057]

[0058] (2.1.2) Use 3D convolution with kernel size 2×τ×τ and stride 1 to transform the graph node attributes T in the TCM modal layer TCM Mapping to node embeddings

[0059]

[0060] Among them, W RSS ,W PDV , Respectively represent the weight mapping matrix of the fully connected layer in RSS, PDV and PDK modes; b RSS ,b PDK , is the bias parameter of the fully connected layer; Represents the convolution kernel of the 3D convolution 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. The intra-layer graph convolution uses graph spectral convolution to aggregate the same modality information on the receiving antenna array from the spectral domain perspective of node attributes; the inter-layer graph convolution uses the graph attention mechanism to assign corresponding weights to the information of different modalities on the same receiving antenna according to the inter-layer adjacency matrix and aggregate multi-modal information. The specific implementation steps are as follows:

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

[0063] Hyper-Laplacian Matrix of Decomposition Multiplexing Network is the intra-layer hyper-Laplacian matrix and the inter-layer hyper-Laplacian matrix The formula is as follows:

[0064]

[0065] in, is a diagonal excess matrix and The Laplacian matrix representing the graph within the r modal layer; is the inter-layer hyper-Laplacian matrix;

[0066] (2.2.2) Heterogeneous graph convolution construction:

[0067] The eigendecomposition of the hyper-Laplacian matrix constructs the eigenvalue matrix and the eigenvector matrix. The formula is as follows:

[0068]

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

[0070] The heterogeneous graph convolution is constructed using the shared feature basis U and feature spectrum Λ of the hyper-Laplacian matrix. The formula is as follows:

[0071]

[0072] in Represent the spectral filter functions of intra-layer graph convolution and inter-layer graph convolution respectively;

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

[0074] Approximating the in-layer graph convolutional spectral filter function using polynomial approximation techniques The formula is as follows:

[0075]

[0076] In-layer hyper-Laplacian matrix It is composed of the union of the Laplacian matrices of the graphs in all modal layers. The graph convolution on the graphs in each modal layer can be parameterized separately, as follows:

[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 (·) is the Chebyshev approximation polynomial, and its expression is: P k (x) = 2x·P k-1 (x)-P k-2 (x), P 1 =x,P 0 =1; They represent the node embeddings of the intra-layer graph convolution output on the intra-layer graph of the r-modal layer respectively; They represent the polynomial approximation coefficients of the convolution of the inner graph of the r-modal layer respectively;

[0080] (2.2.4) Inter-layer graph convolution construction:

[0081] The inter-layer graph convolution is constructed using a graph attention network, which is a weighted summation of the node embeddings of the graphs in different modal layers. The formula is as follows:

[0082]

[0083] In the formula, ⊙ represents the dot product operation; They represent the embedding weight of the i-th node in the graph of the r-th modal layer, and their calculation expression is:

[0084]

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

[0086] (2.2.5) The output of the heterogeneous graph convolution layer consists of intra-layer node embedding and inter-layer node embedding, expressed as:

[0087]

[0088] (2.3) Use global sum pooling to extract abstract features of the multiplexing network and obtain a global representation of the multiplexing network. In this embodiment, the global representation is calculated according to the following formula:

[0089]

[0090] in, Represents the global node embedding of the graph pooling output; and They represent the i-th node embedding on the RSS, PDV, PDK and TCM modalities of the heterogeneous graph convolution output respectively.

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

[0092]

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

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

[0095] Step 3. Set the initial hyperparameters, input the reused network into the heterogeneous graph convolutional network, iteratively train it until convergence, and 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 multiplexing network into a decision statistic, and then use the Monte Carlo method to construct a detection threshold with a fixed false alarm rate based on the decision statistic. Then, determine the final spectrum occupancy state based on the binary hypothesis to achieve spectrum sensing.

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

[0098]

[0099] in, Respectively represent the hypothesis H in the decision vector 1 and H 0 The class probability under G θ (·) represents the expression of heterogeneous graph neural network.

[0100] Embodiment 2: The overall implementation steps of the method proposed in this embodiment are the same as those in Embodiment 1. Figure 2 , the implementation process of using a multiplexed network to characterize multimodal information and model the potential correlation between multimodal information in the present invention is further described in detail:

[0101] Step A. Node attribute construction:

[0102] a1) Use the modal mapping function F RSS (·) The perceived signal y on the i-th receiving antenna i Mapped to RSS modal information Use all RSS modal information on the receiving antenna array to construct the RSS modal layer internal graph G RSS Node properties

[0103]

[0104]

[0105] a2) Use the modal mapping function F PDV (·) The perceived signal y on the i-th receiving antenna i Mapped into PDV modal information Use all PDV modal information on the receiving antenna array to construct the PDV modal layer internal graph G PDV Node properties

[0106]

[0107] Among them, Re(y i (t)) and Im(y i (t)) represent the real and imaginary parts of the sampled value of the perception signal at sampling time t, respectively; is the instantaneous phase of the sampled value of the perceived signal at time t on the i-th receiving antenna; the mod2π function normalized phase In the range [0,2π]; is the phase difference Θ i The mean of .

[0108] a3) Use the modal mapping function F PDK (·) The perceived signal y on the i-th receiving antenna i Mapped into PDK modal information Use all PDK modal information on the receiving antenna array to build the PDK modal layer internal graph G PDK Node properties

[0109]

[0110]

[0111] a4) Use the modal mapping function F TCM (·) The perceived signal y on the i-th receiving antenna i Mapped into TCM modal information Use all TCM modal information on the receiving antenna array to construct the TCM modal layer internal graph G TCM Node properties

[0112]

[0113] in, is a vector sampled using a rectangular window function with a length of ξ and a step size of 1; τ represents the number of rectangular window function sampling vectors; represents the trajectory matrix of the signal sensed on the i-th receiving antenna; is the trajectory covariance matrix of the signal perceived on the i-th receiving antenna, and also represents the TCM modal information of the signal perceived on the i-th receiving antenna. The specific parameters for node attribute construction are set as follows: M = 8; τ = 8; ξ = N-τ.

[0114] Step B. Adjacency matrix construction: Use radial basis function to calculate the real-valued distance between modal information to construct the adjacency matrix of the layer intra-graph. The construction process of the adjacency matrix of each modal layer intra-graph can be specifically expressed as:

[0115] b1) Calculate RSS modal information using radial basis function The real-valued distance between them is used to construct the RSS modal layer internal graph G RSS The adjacency matrix of

[0116]

[0117] b2) Calculate PDV modal information using radial basis function The real-valued distance between PDV The adjacency matrix of

[0118]

[0119] b3) Use radial basis function to calculate PDK modal information The real-valued distance between them is used to construct the PDK modal layer internal graph G PDK The adjacency matrix of

[0120]

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

[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, and is composed of a coupling graph C and an identity matrix I. Since the multimodal information is constructed from different mapping transformations of the same received signal, the coupling graph C can be pre-defined as:

[0124]

[0125] Step C. Node Embedding module construction: The Node Embedding module uses a nonlinear mapping function to map node features to a common dimensional space to promote effective aggregation of multimodal information, which is implemented as follows:

[0126] c1) Use the fully connected function to map the RSS modal layer on the receiving antenna array to G RSS The node attribute T RSS Mapping to node embeddings

[0127] X RSS =σ(T RSS W RSS +b RSS )

[0128] c2) Use the fully connected function to map the PDV modal layer on the receiving antenna array to G PDV The node attribute T PDV Mapping to node embeddings

[0129] X PDV =σ(T PDV W PDV +b PDV )

[0130] c3) Use the fully connected function to map the PDK modal layer on the receiving antenna array to G PDK The node attribute T PDK Mapping to node embeddings

[0131] X PDK =σ(T PDK W PDK +b PDK )

[0132] c4) Use convolution and 3D convolution with size 2×τ×τ and stride 1 to map G inside the TCM modality layer on the receive antenna array. TCM The node attribute T TCM Mapping to node embeddings

[0133]

[0134] in, Represent the weight mapping matrices of the fully connected layer in RSS, PDV and PDK modes respectively; Respectively represent the bias parameters of the fully connected layer in RSS, PDV and PDK modes; Represents the convolution kernel of the 3D convolution layer, whose size is 2×τ×τ and the stride is 1; is the output feature of the pth output channel of the 3D convolution, is the node embedding that senses the signal on the i-th receiving antenna; Represent the embedding matrices of node attributes in the graph of RSS, PDV, PDK and TCM modality layers respectively.

[0135] Embodiment 3: The overall implementation steps of the spectrum sensing method proposed in this embodiment are the same as those in Embodiment 1. Figure 2 , the specific construction steps of the spectrum sensing framework based on heterogeneous graph convolutional network are further described in detail as follows:

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

[0137] (1.1) Hyper-Laplacian Spectral Decomposition: Decomposing the Hyper-adjacency Matrix of a Multiplexing Network is the intra-layer super-adjacency matrix and the inter-layer hyperadjacency matrix

[0138]

[0139] Hyper-Laplacian Matrix of Decomposition Multiplexing Network is the intra-layer hyper-Laplacian matrix and the inter-layer hyper-Laplacian matrix

[0140]

[0141] The hyper-Laplacian matrix is ​​defined as is a diagonal excess matrix and represents the adjacency matrix of the graph in the rth modal layer, is the adjacency matrix between the rth and kth modal layers; represents the Laplacian matrix of the graph in the rth modal layer; is the Laplacian matrix of the inter-layer graph between the rth and kth modes; the inter-layer super Laplacian matrix is ​​defined as represents the Laplace matrix of the coupling graph C, is the identity matrix;

[0142] Perform eigendecomposition on the hyper-Laplacian matrix to obtain the eigenvalue matrix and eigenvector matrix. The formula is as follows:

[0143]

[0144] Among them, U represents the eigenvector matrix in spectral decomposition, and Λ is the diagonal eigenvalue matrix in spectral decomposition. Therefore, the shared eigenbasis U and eigenspectrum Λ of the hyper-Laplacian matrix are used to construct heterogeneous graph convolution.

[0145]

[0146] in Represent the spectral filter functions of intra-layer graph convolution and inter-layer graph convolution, respectively.

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

[0148]

[0149] In the formula, θ k is the approximation coefficient in the polynomial. The calculation formula of the intra-layer graph convolution based on polynomial approximation is:

[0150]

[0151] In-layer hyper-Laplacian matrix It is composed of the union of the Laplacian matrices of the graphs in all modal layers, and the graph convolution on the graph in each modal layer can be parameterized separately. Therefore, the graph convolution on the graph in each 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 (·) is the Chebyshev approximation polynomial, and its expression is; P k (x) = 2x·P k-1 (x)-P k-2 (x), P 1 =x,P 0 =1; and Represents the output node embedding of intra-layer graph convolution on the intra-layer graph of RSS, PDV, PDK and TCM modalities respectively. Represent the polynomial approximation coefficients of the intra-layer graph convolution on the intra-layer graph of RSS, PDV, PDK and TCM modalities, respectively.

[0154] (1.3) Inter-layer graph convolution: Inter-layer graph convolution is constructed using a graph attention network, which represents the weighted sum of node embeddings of graphs in different modal layers. The formula is as follows:

[0155]

[0156] In the formula, ⊙ represents the dot product operation; They represent the embedding weight of the i-th node in the graph of the RSS, PDV, PDK and TCM modal layers respectively, and their calculation expressions are:

[0157]

[0158] In the formula, Respectively represent the node weight mapping matrix of the graph in each modal layer; and Represent the attention vectors in the graph of each modal layer respectively; Represents the output node embedding of the inter-layer convolution on the RSS modality inter-layer graph.

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

[0160]

[0161] Step 2. Build global graph pooling: Use global summation graph pooling to extract the reused network structure information and generate a fixed-size global representation. The processing flow is as follows:

[0162]

[0163] in, Represents the global node embedding of the graph pooling output; They represent the i-th node embedding on the RSS, PDV, PDK and TCM modalities of the heterogeneous graph convolution output respectively.

[0164] Step 3. Build a classifier: The classifier is used to map the global representation of the multiplexing network into a decision statistic. The classifier is composed of two fully connected layers stacked together, and its processing formula can be expressed as:

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

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

[0167] Step 4. Construct decision statistics: Use the modal mapping function and radial basis function to map the received signal on the receiving antenna array to the multiplexing network; input the multiplexing network into the heterogeneous graph convolutional network to construct the decision vector

[0168]

[0169] in, Respectively represent the hypothesis H in the decision vector 1 and H 0 The class probability under G θ (·) represents the expression of the proposed model; according to the Neyman-Pearson criterion, the optimal decision statistic is set as:

[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 expressed as:

[0172]

[0173] The IEEE 802.22 protocol stipulates 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 hypothesis H 0 Building a multiplexing network based on down-sensing signals To construct a new dataset;

[0175]

[0176] (4.2) Use the trained heterogeneous graph convolutional network to calculate the decision statistic T of the new data set H0 ;

[0177] (4.3) Arrange the decision statistics in descending order and construct the decision statistics data set Ω;

[0178] (4.4) Use Get the constant false alarm rate detection threshold, where represents the false alarm rate;

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

[0180] The effect of the present invention will be further described below in conjunction with experiments.

[0181] 1. Experimental conditions:

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

[0183] 2. Experimental content:

[0184] The performance of the spectrum sensing algorithm of the present invention is tested: the test data is input into the optimized heterogeneous graph convolutional network, and the PD-SNR curve under the signal-to-noise ratio of -30dB to -12dB and the ROC curve under the false alarm rate of 0.1 are plotted according to the detection results. The detection probability and the area under the ROC curve (AUC) are used as evaluation indicators to measure the accuracy, robustness and generalization performance of the detection algorithm in various non-ideal channel environments.

[0185] This experiment evaluates the performance of the method of the present invention in scenarios such as AWGN, CGGN, FWGN, large-scale fading, robustness and scalability, cooperative spectrum sensing based on OFDM mechanism and imperfect reporting channel. In each channel environment, the detection effect of the method of the present invention is compared with the mainstream method from the two perspectives of ROC curve and PD-SNR curve. The mainstream methods mainly include ME-GM [Paper: Performance of statistical tests for single-source detection using random matrix theory], PDVD [Paper: Phase difference variancebased 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], 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 The comparison chart of the detection results of the method of the present invention and the reference method in a noisy channel environment is shown in FIG. Figure 3 (a) and Figure 4 (a) ROC curves corresponding to AWGN and CGGN channels, respectively. Figure 3 (b) and Figure 4 (b) PD-SNR curves corresponding to AWGN and CGGN channels, respectively. By comparing the existing spectrum sensing method and the method of the present invention on the two evaluation curves, it can be seen that the method of the present invention has the best effect and is significantly better than the mainstream benchmark method in the prior art. Figure 4"HGCN-SS-AWGN" is the ROC curve and PD-SNR curve of the detection performance in the CGGN channel environment after pre-training in the AWGN channel. The spectrum sensing algorithm proposed in the present invention does not need to be retrained or fine-tuned, and can directly predict the spectrum occupancy state in the CGGN environment, showing robustness to the CGGN channel fading environment.

[0188] Figure 5 The figure shows the detection performance of the method of the present invention and the mainstream spectrum sensing method in a large-scale fading environment, where (a) represents the PD-SNR curve and (b) represents the ROC curve. By comparing the mainstream method with the method of the present invention, the method of the present invention effectively adapts to the large-scale fading scenario and alleviates the impact of path loss and shadow fading on the detection performance. Figure 5 "HGCN-SS-AWGN" is the ROC curve and PD-SNR curve of the detection performance of the pre-trained model under AWGN channel in a large-scale channel environment. The spectrum sensing algorithm proposed in the present invention does not need to be retrained or fine-tuned, and shows detection performance comparable to that of the model trained under large-scale fading scenarios, which shows that the method of the present invention has good generalization ability.

[0189] Figure 6 The ROC curves and PD-SNR curves of the pre-trained model in three scenarios with different numbers of modulation symbols and three different numbers of receiving antennas are shown. The algorithm proposed in the present invention can directly detect the spectrum state without retraining or fine-tuning, and the detection performance increases with the increase of the number of modulation symbols and the number of receiving antennas.

[0190] Figure 7 It is expressed as the detection performance of the method of the present invention and the mainstream method in the collaborative spectrum sensing based on the OFDM mechanism. By comparing the mainstream method with the method of the present invention, there is a significant performance improvement between the method of the present invention and other benchmark methods. The method of the present invention utilizes the multipath effect of the transmission channel in the collaborative spectrum sensing and incorporates multiple secondary users into the decision-making, thereby improving the detection performance.

[0191] Figure 8 This is a comparison chart of the detection results of the method of the present invention and the mainstream method in the imperfect channel environment of the collaborative spectrum sensing scenario. By comparing the existing collaborative spectrum sensing method and the method of the present invention on two evaluation curves, it can be seen that the method of the present invention has the highest area under the ROC curve and the highest detection gain.

[0192] Parts of the present invention that are not described in detail belong to common knowledge among those skilled in the art.

[0193] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Obviously, for professionals in this field, after understanding the content and principles of the present invention, they may make various modifications and changes in form and details without departing from the principles and structures of the present invention. However, these modifications and changes based on the ideas 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 network, characterized in that Integrate multimodal information, multiplexing networks, and heterogeneous graph convolutional networks into a unified spectrum sensing framework to achieve spectrum occupancy status detection in an end-to-end manner; the specific steps are as follows: (1) Building a multiplexing network: (1.1) Using different modal mapping functions, the received signal is mapped into multiple modal information, and the modal information is used to construct the node attributes of the inner graph in the multiplexing network; (1.2) Using radial basis functions and multiple modal information to construct the adjacency matrix of the intra-layer graph in the multiplexing network, and using coupling graphs and identity matrices to construct the adjacency matrix of the inter-layer graph in the multiplexing network; (1.3) The reuse 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 nonlinear mapping functions to map different modal information into a common dimensional space and construct a NodeEmbedding module; (2.2) Heterogeneous graph convolution is composed of intra-layer graph convolution and inter-layer graph convolution. The intra-layer graph convolution uses graph spectral convolution to aggregate the same modality information on the receiving antenna array from the spectral domain perspective of node attributes; the inter-layer graph convolution uses the graph attention mechanism to assign corresponding weights to the information of different modalities on the same receiving antenna according to the inter-layer adjacency matrix and aggregate multi-modal information; (2.3) Use global sum pooling to extract abstract features of the reused network and obtain a 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 into a decision statistic; (2.5) The Node Embedding module is cascaded with the heterogeneous graph convolution, the global pooling layer and the classifier in sequence to obtain a heterogeneous graph convolutional network; (3) Setting the initial hyperparameters, inputting the reused network into the heterogeneous graph convolutional network, iteratively training it until convergence, and obtaining the optimized heterogeneous graph convolutional network; (4) The trained heterogeneous graph convolutional network is used to map the multiplexing network into a decision statistic. The Monte Carlo method is then used to construct a detection threshold with a fixed false alarm rate based on the decision statistic. The final spectrum occupancy state is then determined based on the binary hypothesis to achieve spectrum sensing.

2. The method according to claim 1, characterized in that: The multiple modal information described in step (1.1) includes at least any two or more of the four modal information of 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 modal mapping functions are used to map the received signal into multiple modal information, including using the mapping function F RSS (·),F PDV (·),F PDK (·) and F TCM (·) Map the received signal into received signal strength modal information Phase difference variance modal information Phase difference kurtosis modal information and trajectory covariance modal information These four modal information are used to construct the node attributes of the graphs in multiple modal layers in the reuse network, which are expressed as follows: (1.1.1) Using F RSS (·) Map the received signal into RSS modal information Use all RSS modal information to build the in-layer graph G RSS Node properties (1.1.2) Use F PDV (·) Map the received signal into PDV modal information Construct the intra-layer graph G using all PDV modality information PDV Node properties Among them, Re(y i (t)) and Im(y i (t)) represent the real part and imaginary part of the sampled value of the received signal at sampling time t respectively; is the instantaneous phase of the sampled value of the perceived signal at the tth moment on the i-th receiving antenna; the mod2π function normalized phase θ t i In the range [0,2π]; is the phase difference Θ i The mean of (1.1.3) Use F PDK (·) Map the received signal into PDK modal information Construct the layer internal graph G using all PDK modal information PDK Node properties (1.1.4) Using F TCM (·) Map the received signal into TCM modal information Use all TCM modality information to construct the intra-layer graph G TCM Node properties in, is a vector sampled by a rectangular window function with a length of ξ and a step size of 1; τ represents the number of rectangular window function sampling vectors; 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 graph, including the adjacency matrix A of the RSS modal layer graph. RSS , the adjacency matrix A of the graph in the PDV modal layer PDV , the adjacency matrix A of the graph in the PDK modal layer PDK and the adjacency matrix A of the graph in the TCM modal layer TCM ; The coupling graph and the identity matrix are used to construct the inter-layer adjacency matrix in the multiplexing network It is expressed as follows: (1.2.1) Calculate the modal information T using radial basis functions r i and T r j The real-valued distance between them is used to construct the intra-layer graph G r The adjacency matrix of 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 nonlinear mapping function to map node attributes to a common dimensional space, which is implemented as follows: (2.1.1) Use the fully connected function to transform the node attributes T of the RSS, PDV, and PDK modal layers into RSS 、T PDV 、T PDK Mapping to node embedding T RSS ,T PDV , 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 step size of 1 to transform the node attributes T of the graph in the TCM modal layer TCM Mapping to node embeddings Among them, W RSS ,W PDV , Respectively represent the weight mapping matrix of the fully connected layer in RSS, PDV and PDK modes; b RSS ,b PDK , is the bias parameter of the fully connected layer; Represents the convolution kernel of the 3D convolution layer; is the output feature of the pth output channel of the 3D convolution.

6. The method according to claim 1, characterized in that: The heterogeneous graph convolution described in step (2.2) is composed of intra-layer graph convolution and inter-layer graph convolution. The specific implementation steps are as follows: (2.2.1) Hyper-Laplacian matrix decomposition: Hyper-Laplacian Matrix of Decomposition Multiplexing Network is the intra-layer hyper-Laplacian matrix and the inter-layer hyper-Laplacian matrix The formula is as follows: in, is a diagonal excess matrix and The Laplacian matrix representing the graph within the r modal layer; is the inter-layer hyper-Laplacian matrix; (2.2.2) Heterogeneous graph convolution construction: The eigendecomposition of the hyper-Laplacian matrix constructs the eigenvalue matrix and the eigenvector matrix. The formula is as follows: Where U represents the eigenvector matrix in spectral decomposition, and Λ is the diagonal eigenvalue matrix in spectral decomposition; The heterogeneous graph convolution is constructed using the shared feature basis U and feature spectrum Λ of the hyper-Laplacian matrix. The formula is as follows: in and Represent the spectral filter functions of intra-layer graph convolution and inter-layer graph convolution respectively; (2.2.3) In-layer graph convolution construction: Approximating the in-layer graph convolutional spectral filter function using polynomial approximation techniques The formula is as follows: In-layer hyper-Laplacian matrix It is composed of the union of the Laplacian matrices of the graphs in all modal layers. The graph convolution on the graphs in each modal layer can be parameterized separately, as follows: In the formula, κ j =cos((j+0.5)π / (K+1)) represents the Chebyshev interpolation node; P k (·) is the Chebyshev approximation polynomial, and its expression is: P k (x) = 2x·P k-1 (x)-P k-2 (x), P1 = x, P0 = 1; They represent the node embeddings of the intra-layer graph convolution output on the intra-layer graph of the r-modal layer respectively; They represent the polynomial approximation coefficients of the convolution of the inner graph of the r-modal layer respectively; (2.2.4) Inter-layer graph convolution construction: The inter-layer graph convolution is constructed using a graph attention network, which is a weighted summation of the node embeddings of the graphs in different modal layers. The formula is as follows: In the formula, ⊙ represents the dot product operation; They represent the embedding weight of the i-th node in the graph of the r-th modal layer, and their calculation expression is: In the formula, Represents the r-modal node weight mapping matrix; represents the r-th modal attention vector; Represents the i-th node embedding of the inter-layer convolution output; (2.2.5) The output of the heterogeneous graph convolution layer consists of intra-layer node embedding and inter-layer node embedding, expressed 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, Represents the global node embedding of the graph pooling output; and They represent the i-th node embedding on the RSS, PDV, PDK and TCM modalities of the heterogeneous graph convolution output respectively.

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

9. The method according to claim 1, characterized in that: The optimization training described in step (3) specifically uses the cross entropy function to train the heterogeneous graph convolutional network, and uses the Adam optimizer to iteratively update the network parameters until the convolutional network converges.

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

Images