Spectrum sensing method and device based on graph signal reconstruction and deep graph convolutional network

Abstract

In order to solve the spectrum sensing problem of cognitive radio network under the condition of irregular node distribution and local observation data missing, the present application provides a spectrum sensing method and device fusing graph signal reconstruction and deep graph convolution network (GCN). The method firstly models the distributed sensor network as an irregular weighted graph, and uses graph signal processing technology (such as kernel ridge regression, Laplacian interpolation or GSP low-pass reconstruction) to pre-reconstruct the signal of the unsampled nodes in the sensing blind area, and restores the signal features of the whole network. Subsequently, a multi-layer deep graph convolution network is designed and constructed, and the reconstructed whole network received signal strength (RSS) is used as the input feature, the spatial topological correlation is captured through the residual graph convolution layer, and the dynamic weighted balanced focal loss function is used for offline training. Thus, high-reliable spectrum sensing under low sampling overhead and strong shadow fading environment is realized, and the robustness of the system to irregular network topology and local data missing is improved.

Description

Technical Field

[0001] This invention belongs to the field of spectrum sensing technology, and in particular relates to a spectrum sensing method and apparatus based on graph signal reconstruction and depth graph convolutional networks. Background Technology

[0002] In cognitive radio (CR) networks, dynamic spectrum sharing by sensing spectrum holes in primary users (PUs) is a core method for improving spectrum utilization. Cooperative spectrum sensing (CSS) utilizes spatial diversity gain to effectively overcome the performance bottlenecks of single-point sensing under shadowing fading and multipath effects. However, with the expansion of wireless network scale and the increase in environmental complexity, achieving accurate inference of the spectrum status of all network nodes faces significant challenges.

[0003] Traditional cooperative spectrum sensing techniques are primarily based on the ideal assumptions of complete sensing data and regularly distributed nodes. Their core technologies often rely on energy detection, spatial interpolation (such as Kriging), or convolutional neural networks (CNNs) based on gridded images. While these methods perform well in ideal environments, they suffer from significant limitations when directly applied to real-world, irregular, and data-deficient sensor networks.

[0004] (1) Traditional deep learning perception schemes (such as CNN) usually require the input data to have a regular grid structure. However, in reality, sensor nodes are mostly randomly and irregularly distributed, which are typical non-Euclidean space data. Existing technologies often force sensor nodes to be mapped into a regular two-dimensional grid through coordinate transformation, resulting in the quantization loss of the original information and limiting the further improvement of perception accuracy.

[0005] (2) Existing sensing solutions for missing data mostly employ independent reconstruction and detection steps. In this model, the front-end reconstruction module uses methods such as linear interpolation to estimate the missing node data, which results in a certain reconstruction error. This error may be amplified in subsequent classification and decision-making stages, leading to a decrease in the robustness of the spectrum sensing system.

[0006] (3) Traditional graph signal reconstruction algorithms (such as pure Laplace smoothing) or statistical interpolation methods rely heavily on prior knowledge of linear smoothness. In complex urban or indoor wireless environments, signals are not only affected by macroscopic path loss, but also by strong nonlinear fading caused by obstacle obstruction and multipath effects. Traditional methods often suffer from severe model mismatch when facing these scenarios, resulting in a sharp decline in reconstruction performance as the sampling ratio decreases, which cannot meet the requirements of high-reliability dynamic spectrum access.

[0007] Therefore, considering the current challenges of handling irregular topologies and large-scale data gaps, there is an urgent need for an intelligent spectrum sensing technology that can deeply integrate physical topology priors with data-driven learning features and has environmental adaptability. Summary of the Invention

[0008] In view of this, the present invention aims to overcome the shortcomings of the above-mentioned problems in the prior art and proposes a spectrum sensing method and device based on graph signal reconstruction and deep graph convolutional network. The method utilizes graph signal processing technology to pre-reconstruct the sparsely sampled signals in the network and combines deep learning of the spatial topological features of all network nodes with deep graph convolutional network (GCN). Thus, reliable sensing performance can still be obtained under the conditions of no need for master user signal and channel prior knowledge and the existence of missing node data.

[0009] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0010] In a first aspect, the present invention provides a spectrum sensing method based on graph signal reconstruction and depth graph convolutional networks, comprising the following steps:

[0011] Step 1: Constructing the graph model and data sampling, including:

[0012] The system acquires the geographic coordinates of each node in the distributed sensor network and constructs a weighted adjacency matrix that reflects the irregular spatial topology based on the preset nearest neighbor rules. At any sensing moment, the system extracts the raw received signal strength RSS data in real time through some observation nodes in the network and maps it into a sparse sampled signal vector according to the node index relationship.

[0013] Step 2: Perform full network signal pre-reconstruction, including:

[0014] The sparse sampled signal vector and the weighted adjacency matrix are input into the graph signal processing algorithm. By minimizing the graph gradient energy or performing low-frequency basis projection, the estimated signal intensity of the unsampled nodes is calculated, and a complete signal vector containing all nodes in the network is generated. The complete signal vector is then concatenated with the mask matrix indicating the observation state of the nodes to construct a multidimensional feature matrix of all nodes in the network.

[0015] Step 3: Training and feature learning based on the Deep Graph Convolutional Network (GCN), including:

[0016] The multi-dimensional node feature matrix and weighted adjacency matrix of the entire network are input into a preset depth graph convolutional network GCN. The spatial features are aggregated and extracted layer by layer using multi-layer stacked residual graph convolutional blocks. Using validation set data, the dynamic decision threshold that makes the F1-Score criterion reach its maximum value is determined by traversing and searching in the [0,1] probability space.

[0017] Step 4: Online inference and spectrum state determination, including:

[0018] The sparse RSS samples collected in real time are used to reconstruct the graph signal and generate a reconstructed feature matrix for the time to be detected. The trained GCN model performs forward inference on the reconstructed feature matrix and outputs a probability vector reflecting the probability that each spatial node in the network is occupied by the master user. The binary decision unit compares the output probability vector with the dynamic decision threshold element by element: if the probability value is greater than the threshold, it is determined that there is an authorized user signal at the corresponding node location; otherwise, it is determined that the spectrum at that location is idle.

[0019] Furthermore, the method for setting the detection threshold in step 3 includes: during the offline training phase, inputting validation set data containing known labels into the trained deep graph convolutional network to obtain the predicted value of the main user activity probability of each node; traversing candidate thresholds in the probability space [0,1] with a preset step value and calculating the F1-Score index under each candidate threshold; and selecting the candidate threshold that makes the F1-Score reach the maximum value as the final spectrum sensing dynamic decision threshold.

[0020] Furthermore, the method for determining the dynamic decision threshold in step 3 includes: the GCN model contains 5 convolutional layers and spatial features for completing the signal. Each graph convolutional layer is followed by a ReLU activation function for nonlinear mapping. The last layer maps the features to the main user occupancy probability vector of each node through a Sigmoid activation function.

[0021] Furthermore, in step 4, during online detection and inference, the sparse samples to be detected collected in real time are pre-reconstructed as described in step 2 and then input into the trained depth graph convolutional network model; the model extracts spatial features and outputs the occupancy probability vector of all network nodes; the probability vector is compared element-by-element with the dynamic decision threshold determined in step 3 to determine whether the authorized user corresponding to the node exists, thereby realizing rapid spectrum perception of irregularly distributed nodes in the entire network.

[0022] In a second aspect, the present invention provides a spectrum sensing device, including...

[0023] The data sampling and preprocessing module is used to construct the graph model and sample data, including:

[0024] The system acquires the geographic coordinates of each node in the distributed sensor network and constructs a weighted adjacency matrix that reflects the irregular spatial topology based on the preset nearest neighbor rules. At any sensing moment, the system extracts the raw received signal strength RSS data in real time through some observation nodes in the network and maps it into a sparse sampled signal vector according to the node index relationship.

[0025] The image signal pre-reconstruction module is used for pre-reconstruction of the entire network image signal, including:

[0026] The sparse sampled signal vector and the weighted adjacency matrix are input into the graph signal processing algorithm. By minimizing the graph gradient energy or performing low-frequency basis projection, the estimated signal intensity of the unsampled nodes is calculated, and a complete signal vector containing all nodes in the network is generated. The complete signal vector is then concatenated with the mask matrix indicating the observation state of the nodes to construct a multidimensional feature matrix of all nodes in the network.

[0027] The offline model training and threshold optimization module is used for training and feature learning based on the Deep Graph Convolutional Network (GCN), including:

[0028] The multi-dimensional node feature matrix and weighted adjacency matrix of the entire network are input into a preset depth graph convolutional network GCN. The spatial features are aggregated and extracted layer by layer using multi-layer stacked residual graph convolutional blocks. Using validation set data, the dynamic decision threshold that makes the F1-Score criterion reach its maximum value is determined by traversing and searching in the [0,1] probability space.

[0029] The online inference and state decision module is used for online inference and spectrum state decision, including:

[0030] The sparse RSS samples collected in real time are used to reconstruct the graph signal and generate a reconstructed feature matrix for the time to be detected. The trained GCN model performs forward inference on the reconstructed feature matrix and outputs a probability vector reflecting the probability that each spatial node in the network is occupied by the master user. The binary decision unit compares the output probability vector with the dynamic decision threshold element by element: if the probability value is greater than the threshold, it is determined that there is an authorized user signal at the corresponding node location; otherwise, it is determined that the spectrum at that location is idle.

[0031] Thirdly, the present invention provides an electronic device, including at least one processor and at least one memory communicatively connected to the processor, wherein: the memory stores program instructions executable by the processor, and the processor can execute a spectrum sensing method based on graph signal reconstruction and depth graph convolutional network by calling the program instructions.

[0032] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements a spectrum sensing method based on graph signal reconstruction and depth graph convolutional networks.

[0033] Compared with existing technologies, the spectrum sensing method and apparatus based on graph signal reconstruction and depth graph convolutional networks described in this invention have the following advantages:

[0034] This invention overcomes the spatial quantization limitations of traditional gridded models under irregular topologies: traditional cooperative spectrum sensing schemes (such as CNN-based models) typically require sensor data to be forcibly mapped into a regular two-dimensional grid, which results in severe loss of physical spatial feature quantization. This invention directly utilizes Graph Convolutional Networks (GCNs) to process the received irregular graph structure data, accurately representing arbitrarily distributed sensor topologies through adjacency matrices, fully preserving the spatial fidelity of the original data, and effectively overcoming the performance bottleneck of traditional methods in spatial modeling.

[0035] This invention addresses the challenge of sensing in large-scale sensor networks when node data is missing: existing sensing methods largely rely on the ideal assumption of complete sensing data, leading to a drastic degradation in accuracy when data loss occurs due to node dormancy or link failures. This invention introduces graph signal reconstruction technology, utilizing the prior smoothness of wireless signals in physical space to repair features in sensing blind spots. This achieves accurate recovery from local sparse observations to global features across the entire network, significantly improving the system's performance with low sampling overhead.

[0036] This invention suppresses the cascading effect of reconstruction errors in deep neural networks: traditional collaborative perception processes often neglect the impact of interpolation errors on subsequent decisions. This invention corrects reconstruction errors by constructing a deep graph convolutional network (GCN) with multiple layers of residual connections and combining it with a dynamically weighted balanced focal loss function. This ensures the stable transmission of original spatial features in the deep network, avoids the problem of over-smoothing features, and enhances the robustness of perception in environments with strong shadow fading.

[0037] This invention enhances the system's adaptive decision-making and optimization capabilities in complex electromagnetic environments: traditional sensing algorithms often employ static decision thresholds, which struggle to adapt to rapidly changing channel environments and asymmetric sample distributions. This invention innovatively applies a dynamic threshold optimization mechanism based on the F1-Score maximization criterion. By traversing and searching for the optimal segmentation point during offline training, the system can autonomously balance the false alarm probability and the detection probability. This reduces the false alarm rate in complex fading scenarios and improves the utilization efficiency of spectrum resources while ensuring the same level of interference protection. Attached Figure Description

[0038] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0039] Figure 1 This is a schematic diagram of the spectrum sensing algorithm framework of the present invention;

[0040] Figure 2This is a structural diagram of the GCN model of the present invention. Detailed Implementation

[0041] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0042] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0043] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0044] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0045] Example 1

[0046] Combination Figure 1 The schematic diagram of the method framework shows that the spectrum sensing method in this embodiment mainly includes four processing stages that are executed sequentially and collaboratively:

[0047] First, through the data acquisition and preprocessing stage, a weighted adjacency matrix reflecting the physical spatial connectivity of the network is constructed, and the sparse received signal strength (RSS) data observed in real time is extracted.

[0048] Secondly, in the graph signal pre-reconstruction stage, the system uses physical space smoothness or frequency domain prior knowledge to estimate the signal of missing nodes, thereby completing the initial repair and data completion of the network features.

[0049] Subsequently, during the offline training phase, the system inputs the completed multidimensional feature matrix into the depth map convolutional network (GCN), uses historical complete data for error backpropagation and parameter optimization, and determines the dynamic detection threshold that optimizes the overall decision performance of the system.

[0050] Finally, during the online detection phase, the system feeds the feature data collected and reconstructed in real time into the trained model for end-to-end binary probability inference, ultimately determining whether the main user exists at each spatial location.

[0051] The solution of the present invention will be described in detail below with reference to the specific implementation steps of the present invention:

[0052] Step 1: Constructing the graph model and data sampling

[0053] Step 1.1: Digital Modeling of Network Nodes

[0054] The system obtains the data deployed in the target area. The geographical coordinates of each sensor node are abstracted into a set of graph vertices in non-Euclidean space, establishing a physical spatial mapping relationship for the cognitive radio network.

[0055] Step 1.2: Construct a weighted adjacency matrix

[0056] The processor calculates the spatial connectivity between sensor nodes based on the geographical distance between them and preset nearest neighbor rules (such as the KNN algorithm), generating a weighted adjacency matrix that reflects the network's spatial topology. It is used to quantify the correlation between different sensor nodes in physical space.

[0057] Step 1.3: Sparse Data Acquisition and Mask Mapping

[0058] At any given moment, the data acquisition module extracts raw received signal strength (RSS) data from some active observation nodes (not dormant or without communication failures) in the network in real time, forming a locally sparse sampled signal vector. At the same time, the processor generates a value related to the total number of nodes. The corresponding binary mask matrix is ​​used to accurately identify which nodes in the network are in an observed state and which nodes are in a missing data region.

[0059] Step 2: Perform full network signal pre-reconstruction

[0060] Step 2.1: Signal Reconstruction Path Selection and Execution

[0061] The pre-reconstruction module receives sparse sampled signal vectors. and weighted adjacency matrix Based on the fading characteristics of the current wireless channel environment, the processor calls one of the following preset algorithms to calculate the estimated signal of unsampled nodes and generate a complete signal vector for the entire network. .

[0062] A. Laplace Interpolation: Laplace interpolation is based on the smoothness assumption of graph signals: the signal values ​​of neighboring nodes should be similar. Its core idea is to minimize the graph gradient energy of the signal, and the steps are as follows:

[0063] Signal With Laplace matrix Blocking by observation / missing data:

[0064]

[0065] in: For the submatrix between observation nodes; The submatrix between missing nodes; For cross submatrices;

[0066] Substitute the block matrix into the smoothing term:

[0067]

[0068] This optimization problem is equivalent to:

[0069]

[0070] right Differentiating, we get:

[0071]

[0072] The final solution is obtained, and the complete interpolated signal is obtained by splicing the solutions together.

[0073] .

[0074] B. Graph Signal Processing Low-Pass Reconstruction: The GSP low-pass method interprets the smoothness of graph signals from a frequency domain perspective. It assumes that the real signal is mainly concentrated in the low-frequency subspace of the Laplace. The steps are as follows:

[0075] Graph Laplacian eigenvalue decomposition, then sorted by eigenvalues ​​from smallest to largest, and the first K low-frequency bases are selected:

[0076] ,

[0077] Only the rows corresponding to the observed nodes are retained:

[0078]

[0079] Searching for coefficients ,make As close as possible to the observation :

[0080]

[0081] The final smoothed signal is obtained by combining the low-frequency bases with the optimal coefficients. This method can effectively suppress high-frequency noise, and is especially suitable for signals with strong spatial smoothness, such as RSS fields.

[0082] C. Kernel-based graph signal reconstruction: The kernel method reconstructs signals from the perspectives of regularization and kernel regression. Firstly, it uses the graph Laplacian matrix... and topology smoothing regularization parameters Define the kernel matrix on the graph:

[0083]

[0084] Extract the row and column corresponding to the observation node from the kernel matrix:

[0085]

[0086] Solve for the kernel coefficients, then cross the kernel matrix with the graph. Perform multiplication to obtain the full image reconstruction signal:

[0087] ,

[0088] in This represents the preset ridge regression penalty parameter, used to prevent overfitting during the local matrix inversion process.

[0089] The advantage of this method is that it can balance local smoothness with global structural relationships and has strong robustness to noise and local graph irregularities.

[0090] Step 2.2: Fusion of physical characteristics of the entire network signal.

[0091] After the processor calculates the signal estimate of the missing node through any of the above paths, it compares it with the sparse sampled signal vector known in step 1.3. Spatial concatenation is performed according to the node index sequence to recover a complete one-dimensional column vector reflecting the received signal strength (RSS) of the entire network.

[0092] Step 2.3: Construct the multidimensional node input feature matrix.

[0093] The processor performs dimensionality expansion and concatenation on the full-network completion signal vector generated in step 2.2 and the binary mask matrix generated in step 1.3, constructing a multi-dimensional input feature matrix. (dimension is) This is used as the standard input format for downstream depth graph convolutional networks (GCNs).

[0094] Step 3: Training and feature learning based on a deep graph convolutional network (GCN)

[0095] Step 3.1: Layer-by-layer aggregation of deep spatial features

[0096] In this embodiment, the depth map convolutional network employs the following method: Figure 2 The structure shown indicates that the model will complete the graph signal feature matrix (dimension 1). The input data consists of 1824 learnable parameters. The input feature projection layer contains 96 hidden units, projecting the original 16-dimensional node features onto a 96-dimensional hidden feature space. The model has 5 graph convolutional layers, each with 96 hidden units. The second to fifth graph convolutional layers each contain 9504 learnable parameters. Residual connections are introduced between adjacent graph convolutional layers, adding the inputs and outputs element-wise to ensure feature fidelity in the deep network and suppress oversmoothing. Finally, a fully connected layer with a sigmoid activation function is used for classification; this layer contains 96 input units and 97 learnable parameters. Since this invention models the spectrum perception of all nodes in the network as a binary probability discrimination problem, the size of the last output layer is set to 1, directly outputting the main user occupancy probability vector (dimension 1) for each node. The specific parameter configurations are shown in Table 1.

[0097] Table 1

[0098]

[0099] The processor inputs the feature matrix of all network nodes after dimension concatenation into a pre-defined residual depth map convolutional network. This network consists of an input projection layer, five residual map convolutional blocks, and a fully connected output layer. In any given... In the layer residual map convolutional block, the processor performs the following forward propagation and feature aggregation operations:

[0100]

[0101] in, To add self-connected adjacency matrices, and These represent the training weight matrix and bias vector, respectively. Indicates correspondence The degree matrix, Layer normalization is used to stabilize the distribution of hidden features during training. Furthermore, residual connections are introduced between consecutive layers.

[0102]

[0103] After multi-hop feature aggregation in deep GCN, the model outputs a classification probability matrix after sigmoid normalization. .

[0104] Step 3.2: Backpropagation of network weight errors:

[0105] After five layers of spatial feature aggregation, the processor outputs the predicted occupancy probability of each node through a fully connected layer and a sigmoid function. The processor uses a balanced focus loss function (Focal Loss) to calculate the deviation between the predicted probability and the true label, and calls the Adam optimizer to perform gradient calculation and parameter updates. The formula for this loss function is:

[0106]

[0107] in, This represents the node occupancy probability predicted by the model. The system (or processor) dynamically adjusts this probability by calculating the ratio of positive to negative samples in the current batch. This allows the model to focus more on sparse active samples and difficult-to-classify samples with large interpolation errors during backpropagation.

[0108] During backpropagation, this invention uses the Adam optimization algorithm to dynamically calculate gradients and update weights until the model converges.

[0109] Step 3.3: Optimization calculation of dynamic decision threshold

[0110] After model convergence, the processor inputs validation set data with known labels into the model to obtain predicted probabilities. The processor traverses candidate thresholds within the probability space [0,1] with a preset step value, calculating the harmonic mean (F1-Score) of precision and recall at different thresholds. The system selects a specific candidate threshold that maximizes the F1-Score globally. This is used as the optimal dynamic decision threshold under this network topology and stored in memory.

[0111] Step 4: Online Inference and Spectrum State Decision

[0112] Step 4.1: Real-time Feature Reconstruction and Inference

[0113] The system acquires real-time sensing data from locally sparse nodes and calls the graph signal pre-reconstruction module to generate a reconstructed feature matrix for the current detection time. Subsequently, this reconstructed feature matrix is ​​directly input into the trained Deep Graph Convolutional Network (GCN) model to perform one forward propagation computation.

[0114] Step 4.2: Binary Spectral State Decision

[0115] After the preceding steps (the GSP module reconstructs the missing signals, and the GCN model extracts features), the network output layer generates a continuous vector rate reflecting the probability that each spatial node location in the entire network is occupied by the main user. The binary decision maker reads each element of the probability vector one by one and compares it with the optimal dynamic decision threshold pre-stored in memory. A hard decision comparison is performed, following the logic:

[0116]

[0117] in, To detect the decision threshold, specifically, the optimal decision threshold that maximizes the F1-Score based on the probability distribution of the observed nodes.

[0118] Through the above closed-loop process, this invention can directly map sparsely collected signals into high-confidence global spectrum state decision results without requiring prior knowledge from the main user.

[0119] Example 2

[0120] This invention provides a spectrum sensing device, including

[0121] The data sampling and preprocessing module is used to construct the graph model and sample data, including:

[0122] The system acquires the geographic coordinates of each node in the distributed sensor network and constructs a weighted adjacency matrix that reflects the irregular spatial topology based on the preset nearest neighbor rules. At any sensing moment, the system extracts the raw received signal strength RSS data in real time through some observation nodes in the network and maps it into a sparse sampled signal vector according to the node index relationship.

[0123] The image signal pre-reconstruction module is used for pre-reconstruction of the entire network image signal, including:

[0124] The sparse sampled signal vector and the weighted adjacency matrix are input into the graph signal processing algorithm. By minimizing the graph gradient energy or performing low-frequency basis projection, the estimated signal intensity of the unsampled nodes is calculated, and a complete signal vector containing all nodes in the network is generated. The complete signal vector is then concatenated with the mask matrix indicating the observation state of the nodes to construct a multidimensional feature matrix of all nodes in the network.

[0125] The offline model training and threshold optimization module is used for training and feature learning based on the Deep Graph Convolutional Network (GCN), including:

[0126] The multi-dimensional node feature matrix and weighted adjacency matrix of the entire network are input into a preset depth graph convolutional network GCN. The spatial features are aggregated and extracted layer by layer using multi-layer stacked residual graph convolutional blocks. Using validation set data, the dynamic decision threshold that makes the F1-Score criterion reach its maximum value is determined by traversing and searching in the [0,1] probability space.

[0127] The online inference and state decision module is used for online inference and spectrum state decision, including:

[0128] The sparse RSS samples collected in real time are used to reconstruct the graph signal and generate a reconstructed feature matrix for the time to be detected. The trained GCN model performs forward inference on the reconstructed feature matrix and outputs a probability vector reflecting the probability that each spatial node in the network is occupied by the master user. The binary decision unit compares the output probability vector with the dynamic decision threshold element by element: if the probability value is greater than the threshold, it is determined that there is an authorized user signal at the corresponding node location; otherwise, it is determined that the spectrum at that location is idle.

[0129] Example 3

[0130] To enable the deployment of the aforementioned spectrum sensing system in a real physical environment, embodiments of the present invention also provide an electronic device (e.g., an edge computing server or an advanced cognitive radio terminal). This electronic device includes:

[0131] At least one processor and a memory communicatively connected to the at least one processor.

[0132] The memory stores program instructions that can be executed by the at least one processor. When the processor calls the program instructions, it can execute the spectrum sensing method based on graph signal reconstruction and depth graph convolutional network described in any of the above method embodiments (i.e., steps 1 to 4).

[0133] Specifically, the processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0134] Example 4

[0135] Furthermore, embodiments of the present invention also provide a non-volatile computer-readable storage medium storing a computer program, which, when executed by the processor of the aforementioned electronic device, implements the steps of the spectrum sensing method described in any of the foregoing embodiments.

[0136] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A spectrum sensing method based on graph signal reconstruction and depth graph convolutional networks, characterized in that: Includes the following steps: Step 1: Constructing the graph model and data sampling, including: The system acquires the geographic coordinates of each node in the distributed sensor network and constructs a weighted adjacency matrix that reflects the irregular spatial topology based on the preset nearest neighbor rules. At any sensing moment, the system extracts the raw received signal strength RSS data in real time through some observation nodes in the network and maps it into a sparse sampled signal vector according to the node index relationship. Step 2: Perform full network signal pre-reconstruction, including: The sparse sampled signal vector and the weighted adjacency matrix are input into the graph signal processing algorithm. By minimizing the graph gradient energy or performing low-frequency basis projection, the estimated signal intensity of the unsampled nodes is calculated, and a complete signal vector containing all nodes in the network is generated. The complete signal vector is then concatenated with the mask matrix indicating the observation state of the nodes to construct a multidimensional feature matrix of all nodes in the network. Step 3: Training and feature learning based on the Deep Graph Convolutional Network (GCN), including: The multi-dimensional node feature matrix and weighted adjacency matrix of the entire network are input into a preset depth graph convolutional network GCN. The spatial features are aggregated and extracted layer by layer using multi-layer stacked residual graph convolutional blocks. Using validation set data, the dynamic decision threshold that makes the F1-Score criterion reach its maximum value is determined by traversing and searching in the [0,1] probability space. Step 4: Online inference and spectrum state determination, including: The sparse RSS samples collected in real time are used to reconstruct the graph signal and generate a reconstructed feature matrix for the time to be detected. The trained GCN model performs forward inference on the reconstructed feature matrix and outputs a probability vector reflecting the probability that each spatial node in the network is occupied by the master user. The binary decision unit compares the output probability vector with the dynamic decision threshold element by element: if the probability value is greater than the threshold, it is determined that there is an authorized user signal at the corresponding node location; otherwise, it is determined that the spectrum at that location is idle.

2. The spectrum sensing method based on graph signal reconstruction and depth graph convolutional network according to claim 1, characterized in that: The method for setting the detection threshold in step 3 includes: during the offline training phase, inputting validation set data containing known labels into the trained deep graph convolutional network to obtain the predicted value of the main user activity probability of each node; traversing the candidate thresholds in the probability space [0,1] with a preset step value and calculating the F1-Score index under each candidate threshold; and selecting the candidate threshold that makes the F1-Score reach the maximum value as the final spectrum sensing dynamic decision threshold.

3. The spectrum sensing method based on graph signal reconstruction and depth graph convolutional network according to claim 1, characterized in that: The method for determining the dynamic decision threshold in step 3 includes: the GCN model contains 5 convolutional layers and spatial features for taking the complement signal. Each graph convolutional layer is followed by a ReLU activation function for nonlinear mapping. The last layer maps the features to the main user occupancy probability vector of each node through a Sigmoid activation function.

4. The spectrum sensing method based on graph signal reconstruction and depth graph convolutional network according to claim 1, characterized in that: In step 4, during online detection and inference, the sparse samples to be detected collected in real time are processed by the pre-reconstruction process described in step 2 and then input into the trained depth graph convolutional network model. The model extracts spatial features and outputs the occupancy probability vector of all network nodes. The probability vector is compared element-by-element with the dynamic decision threshold determined in step 3 to determine whether the authorized user corresponding to the node exists, thereby realizing rapid spectrum perception of irregularly distributed nodes in the entire network.

5. A spectrum sensing device, characterized in that: include The data sampling and preprocessing module is used to construct the graph model and sample data, including: The system acquires the geographic coordinates of each node in the distributed sensor network and constructs a weighted adjacency matrix that reflects the irregular spatial topology based on the preset nearest neighbor rules. At any sensing moment, the system extracts the raw received signal strength RSS data in real time through some observation nodes in the network and maps it into a sparse sampled signal vector according to the node index relationship. The image signal pre-reconstruction module is used for pre-reconstruction of the entire network image signal, including: The sparse sampled signal vector and the weighted adjacency matrix are input into the graph signal processing algorithm. By minimizing the graph gradient energy or performing low-frequency basis projection, the estimated signal intensity of the unsampled nodes is calculated, and a complete signal vector containing all nodes in the network is generated. The complete signal vector is then concatenated with the mask matrix indicating the observation state of the nodes to construct a multidimensional feature matrix of all nodes in the network. The offline model training and threshold optimization module is used for training and feature learning based on the Deep Graph Convolutional Network (GCN), including: The multi-dimensional node feature matrix and weighted adjacency matrix of the entire network are input into a preset depth graph convolutional network GCN. The spatial features are aggregated and extracted layer by layer using multi-layer stacked residual graph convolutional blocks. Using validation set data, the dynamic decision threshold that makes the F1-Score criterion reach its maximum value is determined by traversing and searching in the [0,1] probability space. The online inference and state decision module is used for online inference and spectrum state decision, including: The sparse RSS samples collected in real time are used to reconstruct the graph signal and generate a reconstructed feature matrix for the time to be detected. The trained GCN model performs forward inference on the reconstructed feature matrix and outputs a probability vector reflecting the probability that each spatial node in the network is occupied by the master user. The binary decision unit compares the output probability vector with the dynamic decision threshold element by element: if the probability value is greater than the threshold, it is determined that there is an authorized user signal at the corresponding node location; otherwise, it is determined that the spectrum at that location is idle.

6. An electronic device, characterized in that: It includes at least one processor and at least one memory communicatively connected to the processor, wherein: the memory stores program instructions executable by the processor, and the processor can execute the method as described in any one of claims 1-4 by invoking the program instructions.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1-4.

Images