A method and system for detecting concealed devices

By constructing a high-dimensional phase space reconstructed trajectory and a directional attention weight matrix, the problem of device detection under conditions of overlapping multi-source radio frequency signals and complex electromagnetic environments was solved, achieving efficient and accurate detection of concealed devices and reducing false alarm and missed alarm rates.

CN122885836BActive Publication Date: 2026-11-17GUANGZHOU BOCHENG ELECTRONICS CO LTD
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202611328087.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-08-31
Publication Date
2026-11-17
Estimated Expiration
2046-08-31

AI Technical Summary

Technical Problem

Existing technologies struggle to accurately distinguish between legitimate and concealed devices in environments with overlapping multi-source radio frequency signals and complex electromagnetic conditions, leading to high false alarm or false alarm rates. Furthermore, existing attention mechanisms do not adequately consider the asymmetric distribution differences between signal clusters and variations in environmental complexity.

Method used

By acquiring broadband radio frequency signals, extracting multidimensional time-series features, constructing a high-dimensional phase space to reconstruct the trajectory, using density clustering to separate signal clusters, calculating unidirectional KL divergence and kurtosis to construct a directional attention weight matrix, generating a symmetric positive definite tensor, and adaptively adjusting the judgment threshold, cross-cluster feature fusion and anomaly detection are achieved.

Benefits of technology

It improves the accuracy and robustness of detecting concealed devices in complex environments, reduces false alarm and false negative rates, and enhances the system's adaptability in complex spectrum environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122885836B_ABST
    Figure CN122885836B_ABST
Patent Text Reader

Abstract

This invention relates to the field of detection technology, and more particularly to a method and system for detecting concealed devices. The method includes: acquiring broadband radio frequency signals within the target space, extracting continuous temporal multidimensional features, and dividing different device signal clusters through high-dimensional phase space reconstruction and density clustering to obtain cluster center feature vectors and intra-cluster covariance matrices; constructing a Gaussian distribution based on the signal clusters, and combining one-way KL divergence, distribution kurtosis, and directional attention weights to complete neighborhood feature aggregation and cluster feature updating; then constructing a global symmetric positive definite tensor based on the updated cluster feature vectors, calculating its geodesic distance from the Frescher mean, and performing adaptive threshold determination based on historical normal sample distance distribution and environmental complexity factors. This invention improves the stability and adaptability of concealed device detection in complex electromagnetic environments.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of detection technology, and in particular to a method and system for detecting concealed devices. Background Technology

[0002] With the development of the Internet of Things (IoT) and wireless communication technologies, various wireless radio frequency (RF) devices are being used in office and living spaces. While improving the convenience of information exchange, this also brings risks of privacy leaks and the illegal collection of confidential information. Illegal devices such as miniature eavesdropping devices and hidden cameras typically exhibit characteristics such as low power, burst transmission, or low interception probability, and their RF signals are easily masked by legitimate device signals and environmental noise. Traditional detection methods often rely on single RF characteristics for identification. In scenarios with overlapping multi-source signals and significant background electromagnetic fluctuations, it is difficult to accurately distinguish between legitimate and hidden devices, leading to false alarms or missed detections, and failing to meet the detection needs of high-security scenarios.

[0003] Chinese patent document CN112946697B discloses a deep learning-based method for detecting and repairing cycle slips in satellite signals, comprising: 1. performing high-order differential processing on the carrier phase of the satellite signal to amplify cycle slip information while eliminating the influence of integer ambiguity and other errors; 2. converting the carrier phase time series into a vector set through phase space reconstruction, and constructing a cycle slip-free training sample set using cycle slip-free carrier phase values; 3. constructing and training a deep learning network model; 4. defining cycle slip detection statistics to determine whether a cycle slip has occurred in the satellite signal and to effectively repair the satellite signal.

[0004] As the above solutions demonstrate, existing technologies have attempted to combine nonlinear dynamics analysis and deep learning methods to separate signal clusters from different devices through phase space reconstruction and density clustering, and then use attention mechanisms to fuse cross-cluster features before determining anomalies based on the distance between the fused features and normal samples. However, existing attention mechanisms typically assume that the information transmission relationship between clusters is approximately symmetrical, failing to fully consider the unidirectional distribution differences and kurtosis characteristics between different signal clusters, resulting in a mismatch between weight allocation and the asymmetric perturbation relationship in the actual electromagnetic environment. Existing feature aggregation methods often employ simple splicing or weighted addition, which underutilizes higher-order statistical information such as intra-cluster covariance, easily leading to loss of identification information. Furthermore, existing anomaly determinations are mostly based on Euclidean distance and fixed thresholds, failing to fully utilize geodesic distances on symmetric positive definite matrix manifolds, and lacking a mechanism for adaptively adjusting thresholds according to environmental complexity, thus limiting the system's detection stability and adaptability in complex electromagnetic environments. Summary of the Invention

[0005] To address the problems of existing technologies, such as difficulty in separating overlapping multi-source radio frequency signals, distortion in cross-cluster attention weight allocation, and high false alarm and false negative rates due to the inability of Euclidean distance and fixed threshold to adapt to complex electromagnetic environment changes, this invention provides a method and system for detecting concealed devices.

[0006] In a first aspect, the present invention provides a method for detecting concealed devices, employing the following technical solution:

[0007] S1. Acquire broadband radio frequency signals in the space to be tested and extract multidimensional feature sequences of continuous time series; construct high-dimensional phase space reconstruction trajectories based on multidimensional feature sequences, and perform density-based spatial clustering on the phase space reconstruction trajectories to obtain multiple signal clusters representing signal sources of different devices, cluster center feature vectors of each signal cluster, and feature covariance matrices within the clusters. S2. Construct Gaussian distributions for different signal clusters and calculate the one-way KL divergence between the distributions. Based on the divergence and the kurtosis of the characteristic distribution of each signal cluster, construct a directional attention weight matrix and calculate the spectral norm of the matrix as an environmental complexity factor. After aggregating the features passed by the neighbor nodes with weights, multiply the matrix by the transpose of the orthogonal eigenvector matrix of the feature covariance matrix of the target node within the cluster, and add the residuals to the initial cluster center eigenvector of the target node to obtain the updated feature vector of each signal cluster. S3: Calculate the second-order global covariance matrix of all updated feature vectors to generate a symmetric positive definite tensor representing the global features of the scene; calculate the geodesic distance between the symmetric positive definite tensor and the Frescher mean; determine the baseline quantile threshold based on the distribution of geodesic distances of historical normal samples; construct a positively correlated monotonically increasing function using the environmental complexity factor to regularize and relax the baseline quantile threshold; determine the presence of a hidden device when the actual geodesic distance is greater than the relaxed threshold.

[0008] By acquiring broadband radio frequency signals and extracting multi-dimensional temporal features, a high-dimensional phase space reconstructed trajectory is constructed. Density clustering is used to separate different signal sources, obtaining signal clusters, cluster center feature vectors, and covariance matrices, effectively addressing environments with overlapping multi-source radio frequency signals. Furthermore, a directional attention weight matrix is ​​constructed using one-way KL divergence and distribution kurtosis, and the spectral norm is calculated as an environmental complexity factor. Combining orthogonal feature vector matrices and residual aggregation mechanisms to update cluster features enables efficient fusion of cross-cluster asymmetric correlation information, suppressing redundant interference. Finally, anomaly detection is performed based on symmetric positive definite tensors and geodesic distances, and the environmental complexity factor is used to adaptively relax the detection threshold, improving the accuracy and robustness of detecting concealed devices in complex environments and reducing false alarm and false negative rates.

[0009] Preferably, the method for extracting multidimensional feature sequences from continuous time series is as follows: The broadband radio frequency signal received in the space under test is subjected to spectrum scanning, channelization separation, or peak band tracking to obtain at least one candidate narrowband signal component; the candidate narrowband signal component is divided into time windows of equal length according to a fixed time sequence; instantaneous envelope value, instantaneous phase value, carrier frequency offset value, and received signal strength indication value are extracted within each time window; the instantaneous phase value is subjected to phase unwrapping processing to obtain continuous phase value; the instantaneous envelope value, continuous phase value, carrier frequency offset value, and received signal strength indication value are subjected to dimensionless normalization processing; all normalized values ​​within the same time window are concatenated to form a one-dimensional feature vector; according to the time series order, the one-dimensional feature vectors generated by all continuous time windows are stacked accordingly to form a feature matrix, and the feature matrix as a whole is determined as a multidimensional feature sequence.

[0010] By performing spectrum scanning, channelization separation, or peak band tracking on broadband radio frequency signals, multiple narrowband signal components are extracted. Then, multi-dimensional features such as instantaneous envelope, continuous phase, carrier frequency offset, and RSSI are extracted using a fixed time window. After normalization and column concatenation, a one-dimensional feature vector is formed, and finally, a time-series multi-dimensional feature sequence is constructed. This can capture the time-varying characteristics and modulation details of radio frequency signals, providing high-quality input data for subsequent high-dimensional phase space reconstruction and cluster analysis, and enhancing the method's adaptability to different types of covert devices.

[0011] Preferably, constructing a high-dimensional phase space reconstruction trajectory based on multidimensional feature sequences includes: in the historical normal sample modeling stage, calculating the delay time parameter of each dimension feature in the multidimensional feature sequence using the mutual information method, determining the embedding dimension parameter of each dimension feature using the false nearest neighbor method, and using the determined delay time parameter and embedding dimension parameter as fixed phase space reconstruction parameters for the phase space reconstruction of historical normal samples and current detection samples; combining the delay time parameter and embedding dimension parameter, converting each dimension feature in the original feature sequence into a corresponding sub-state space matrix using coordinate delay processing, and pruning and aligning each sub-state space matrix according to the common effective time index or the shortest effective sample length, and then splicing them along the column direction to form a high-dimensional state space matrix as the phase space reconstruction trajectory.

[0012] In the historical normal sample modeling stage, the delay time and embedding dimension of each feature dimension are determined by mutual information and spurious nearest neighbor methods, respectively, and used as fixed parameters for phase space reconstruction, ensuring the comparability of historical and current samples in the same feature space. By processing coordinate delay and aligning and concatenating the sub-state space matrices, a high-dimensional state space matrix is ​​generated, which effectively preserves the nonlinear dynamic characteristics of the original signal and improves the physical rationality of subsequent clustering and feature updates.

[0013] Preferably, density-based spatial clustering is performed on the phase space reconstructed trajectory to obtain multiple signal clusters representing signal sources from different devices, the cluster center feature vectors of each signal cluster, and the intra-cluster feature covariance matrix, including: A density-based spatial clustering algorithm with limitations on the neighborhood search radius and the minimum number of core sample points is applied to perform local density operations on the state space matrix to obtain multiple disjoint, non-noise data point clusters as signal clusters. The arithmetic mean of the coordinate features of all member vectors within a single signal cluster is calculated as the corresponding cluster center feature vector. The second-order central moments of all sample vectors within the signal cluster relative to this cluster center feature vector are calculated and combined to generate an associated covariance matrix. The covariance matrix is ​​then subjected to a small diagonal loading regularization process to ensure that the covariance matrix satisfies the symmetric positive definite property, which serves as the corresponding intra-cluster feature covariance matrix.

[0014] A density-based spatial clustering algorithm is used to cluster the reconstructed trajectories in phase space. This algorithm can automatically separate different signal sources and identify outlier noise points without presetting the number of clusters. By calculating the arithmetic mean and second-order central moments of the eigenvectors within each cluster and performing small diagonal loading regularization, a symmetric positive definite intra-cluster feature covariance matrix is ​​obtained, providing a stable statistical basis for subsequent Gaussian modeling, KL divergence calculation, and feature updates.

[0015] Preferably, a Gaussian distribution is constructed for different signal clusters, and the one-way KL divergence between the distributions is calculated, including: using the independent cluster center eigenvectors of each signal cluster as the distribution mean, and the corresponding intra-cluster feature covariance matrix as the distribution covariance, a matching multidimensional Gaussian probability density mathematical distribution is established for all independent signal clusters respectively; traversing any two different signal clusters, taking one as the first examined signal cluster and the other as the second examined signal cluster, the one-way KL divergence from the first examined signal cluster to the second examined signal cluster is calculated.

[0016] A multidimensional Gaussian distribution is constructed using the cluster center eigenvectors and intra-cluster feature covariance matrices of each signal cluster, and the one-way KL divergence between different signal clusters is calculated, which can quantify the asymmetric distribution differences between signal sources. This one-way metric better reflects the asymmetric interference relationship between signal sources in the actual electromagnetic environment compared to traditional symmetric metrics, providing a theoretical basis for constructing a directional attention weight matrix and improving the physical realism and effectiveness of cross-cluster feature fusion.

[0017] Preferably, the method for calculating the environmental complexity factor includes: calculating the average of the standardized fourth-order central distances in all higher-order feature dimensions covered by a single signal cluster, as the kurtosis of the intra-cluster feature distribution; obtaining the one-way KL divergence value from the source signal cluster to the target signal cluster, converting the one-way KL divergence value into an inter-cluster similarity component through negative correlation mapping, and multiplying the inter-cluster similarity component with the intra-cluster feature distribution kurtosis adjustment factor corresponding to the source signal cluster to obtain the directed communication initial component value on the corresponding network connection edge; converting all input initial components of the same target node into communication weights using a normalization function; writing the weights into the corresponding structure of the asymmetric attention graph to form a directed attention weight matrix; calculating the maximum eigenvalue of the product of the directed attention weight matrix and the conjugate transpose of the directed attention weight matrix, performing positive square root processing on the maximum eigenvalue to obtain the spectral norm, and setting the spectral norm as the environmental complexity factor.

[0018] By calculating the kurtosis and one-way KL divergence of the feature distribution within a cluster, an asymmetric directed attention weight matrix is ​​constructed, and its spectral norm is calculated as an environmental complexity factor, which can dynamically reflect the interference intensity and asymmetric disturbance level of the current electromagnetic environment. This mechanism enables subsequent cross-cluster feature aggregation and threshold adjustment to adapt to environmental changes in real time, enhancing the system's adaptability and detection stability in complex spectral environments.

[0019] Preferably, the method for obtaining the updated feature vectors of each signal cluster includes: performing eigenvalue decomposition on the intra-cluster feature covariance matrix of a single signal cluster representing the target node to obtain the orthogonal eigenvector matrix of the intra-cluster feature covariance matrix; extracting the initial cluster center feature vectors of each initially connected neighbor node in the transmission direction, multiplying them by the weights of the corresponding directional attention weight matrix to obtain individual weighted vectors, and summing the individual weighted vectors of all corresponding neighbors to obtain the target transmission aggregation vector; multiplying the transpose of the orthogonal eigenvector matrix on the left by the target transmission aggregation vector to obtain the residual offset vector after feature space projection alignment; and accumulating the residual offset vector with the original initial cluster center feature vector of the target node to obtain the corresponding output updated feature vector.

[0020] By performing eigenvalue decomposition on the intra-cluster feature covariance matrix of the target node to obtain an orthogonal feature vector matrix, and using directional attention weights to weighted aggregate the cluster center feature vectors of neighboring nodes, and then updating the feature vector of the target node by adding the orthogonal projection and residual, the interference information introduced by feature space misalignment can be effectively reduced, the structural information of the original features can be preserved, and the accuracy and robustness of cross-cluster feature fusion can be improved.

[0021] Preferably, the calculation of the second-order global covariance matrix of all updated feature vectors and the generation of a symmetric positive definite tensor representing the global features of the scene include: stacking all corresponding updated feature vectors in columns according to the order of signal cluster category labels to form a global updated feature matrix; calculating the demeaning matrix of the global updated feature matrix relative to the mean vector of the global features of the matrix; performing matrix multiplication on the demeaning matrix and the transpose of the matrix itself, and dividing by a scalar of degrees of freedom equal to the total number of signal clusters minus 1 to generate a second-order global covariance matrix; performing small diagonal loading regularization on the second-order global covariance matrix to ensure that the second-order global covariance matrix satisfies the symmetric positive definite property, and using the processed symmetric positive definite matrix as the symmetric positive definite tensor of the global features of the scene.

[0022] All updated feature vectors are concatenated column-wise to form a global updated feature matrix. A second-order global covariance matrix is ​​generated through mean reduction and covariance operations, followed by diagonal loading regularization to obtain a symmetric positive definite tensor. This tensor has a good geometric structure on a Riemannian manifold and can serve as a high-order representation of global scene features, providing mathematically stable input for subsequent geodesic distance calculations and anomaly detection.

[0023] Preferably, the baseline quantile threshold is determined based on the geodesic distance distribution of historical normal samples. A positively correlated monotonically increasing function is constructed using the environmental complexity factor to regularize and relax the baseline quantile threshold. This includes: obtaining all historical geodesic distance offset distribution models on the pre-installed projection stream of historical attack-free, clean environmental record data; extracting values ​​with a cumulative distribution ratio reaching 95% within the distribution model and using these as the baseline quantile threshold; generating a regulation factor positively correlated with the environmental clutter level by applying a monotonically increasing smoothing adjustment operator to the environmental complexity factor; and multiplying the baseline quantile threshold by the regulation factor to obtain a judgment threshold that increases with increasing environmental complexity, thus achieving regularization and relaxation of the baseline quantile threshold.

[0024] The baseline quantile threshold is extracted based on the geodesic distance distribution of historical normal samples. The threshold is adaptively relaxed by constructing a positively correlated monotonically increasing function using the environmental complexity factor. This allows the judgment criteria to be dynamically adjusted with environmental complexity, avoiding false alarms or missed alarms caused by fixed thresholds in complex electromagnetic environments.

[0025] Secondly, the present invention provides a concealed device detection system, which adopts the following technical solution: A covert device detection system includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, a covert device detection method according to the above is implemented.

[0026] The present invention has the following technical effects: This invention extracts multidimensional temporal features from radio frequency signals, constructs a high-dimensional phase space reconstructed trajectory, and performs density clustering to separate signal clusters corresponding to different signal sources and characterize their statistical properties. It utilizes one-way divergence and distribution kurtosis to construct a directional attention weight matrix, combining it with orthogonal feature vector matrices and residual aggregation mechanisms to achieve the fusion of cross-cluster asymmetric correlation information, reducing the impact of redundant interference on feature updates. The updated global features are constructed as symmetric positive definite tensors, and their geodesic distance to the Fréchez mean is calculated in manifold space to enhance the ability to identify abnormal radio frequency structural changes. Simultaneously, the baseline quantile threshold is adaptively relaxed based on an environmental complexity factor, enabling the judgment criteria to adapt to changes in background electromagnetic fluctuations, thereby reducing the risk of false alarms in complex environments and improving the reliability of concealed device detection. Attached Figure Description

[0027] Figure 1 This is a flowchart of a concealed device detection method according to the present invention.

[0028] Figure 2 This is a schematic diagram of the instantaneous envelope feature extraction of the present invention.

[0029] Figure 3 This is a schematic diagram comparing the ablation experimental effects of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] This invention discloses a method for detecting concealed devices, referring to... Figure 1 This includes the following steps: S1, obtain the signal cluster, cluster center eigenvector and covariance matrix.

[0032] Broadband radio frequency signals within the test space are acquired, and multidimensional feature sequences of continuous time series are extracted. Based on the multidimensional feature sequences, high-dimensional phase space reconstruction trajectories are constructed, and density-based spatial clustering is performed on the phase space reconstruction trajectories to obtain multiple signal clusters representing signal sources from different devices, cluster center feature vectors of each signal cluster, and intra-cluster feature covariance matrices.

[0033] Baseband in-phase and quadrature data samples of broadband radio frequency signals are acquired using general-purpose software radio peripherals. The time-domain signal is converted to the frequency domain using a Fast Fourier Transform (FFT). Spectrum scanning, channelization separation, or peak band tracking are performed on the frequency-domain signal to obtain at least one candidate narrowband signal component. Features such as received signal strength indication, instantaneous envelope, continuous phase, and carrier frequency offset corresponding to the candidate narrowband signal component are calculated. After dimensionless normalization of features of different dimensions, a multidimensional feature sequence of continuous time series is obtained. In the historical normal sample modeling stage, the mutual information algorithm is used to calculate the time series delay time, and the spurious nearest neighbor algorithm is used to determine the optimal embedding dimension. The determined delay time and embedding dimension are used as fixed phase space reconstruction parameters for historical normal samples and current detection samples. Delay coordinate mapping is performed on the multidimensional feature sequence to reconstruct the phase space, generating a high-dimensional phase space reconstruction trajectory matrix. The density-based spatial clustering algorithm, DBSCAN, is used to perform unsupervised clustering of the row vectors of the reconstructed trajectory matrix in high-dimensional phase space. After removing outlier noise points, index labels of multiple signal clusters representing independent signal sources are output. For each segmented signal cluster, the arithmetic mean of the feature vectors along the sample dimension is calculated as the cluster center feature vector, and the covariance function is called to calculate the covariance matrix of all sample feature vectors in the signal cluster as the intra-cluster feature covariance matrix.

[0034] In an optional embodiment, acquiring a broadband radio frequency signal within the test space and extracting a multidimensional feature sequence of continuous time series includes: performing spectrum scanning, channelization separation, or peak band tracking on the broadband radio frequency signal received within the test space to obtain at least one candidate narrowband signal component; dividing the candidate narrowband signal component into non-overlapping time windows of equal length according to a fixed time sequence; extracting instantaneous envelope values, instantaneous phase values, carrier frequency offset values, and received signal strength indication values ​​within each time window; performing phase unwrapping processing on the instantaneous phase values ​​to obtain continuous phase values; performing dimensionless normalization processing on the instantaneous envelope values, continuous phase values, carrier frequency offset values, and received signal strength indication values ​​respectively; concatenating the normalized instantaneous envelope values, continuous phase values, carrier frequency offset values, and received signal strength indication values ​​within the same time window to form a one-dimensional feature vector; stacking the one-dimensional feature vectors generated by all continuous time windows in the order of time series to form a feature matrix, and determining the feature matrix as a whole as a multidimensional feature sequence representing continuous time series.

[0035] Using radio frequency receiving front-end equipment, such as an SDR software radio platform, continuous baseband I / Q signals are acquired at a relatively high sampling rate of 10MS / s to 40MS / s, preferably 20MS / s, within a set frequency band. A fast Fourier transform is performed on the wideband I / Q signals to obtain the spectral amplitude distribution. Based on the local peak detection results, the center frequency and bandwidth range of candidate frequency bands are determined. Bandpass filtering and digital down-conversion are performed on the candidate frequency bands to isolate the corresponding frequency band signals and convert them into baseband narrowband signal components. When multiple peak frequency bands are detected, one or more candidate narrowband signal components are selected based on spectral energy intensity, duration, or frequency band stability as feature extraction targets for subsequent instantaneous envelope, continuous phase, carrier frequency offset, and received signal strength indication values.

[0036] Using a rectangular window function, the data stream is divided into segments of equal length. For example, a 20-microsecond time window with non-overlapping beginnings and endings. Within each time window containing 400 sampling points, the average complex modulus of the sampling points is calculated as the instantaneous envelope value, typically ranging from 0.1mV to 5mV. The instantaneous phase at each point is calculated using the arctangent function, and a phase unwrapping algorithm is performed using a π-radian threshold-based transition detection to compensate for the 2π offset, outputting a continuous phase average. The carrier frequency offset within the window is calculated using phase slope estimation or an FFT peak search algorithm, for example, obtaining a frequency offset estimate on the order of kHz. The received signal strength indication value is obtained by taking the logarithmic average of the sampled power over the entire window, for example, -65dBm. An example RF signal instantaneous envelope versus time curve is shown below. Figure 2 As shown. After obtaining the single-window parameters, Min-Max extremum normalization is performed to map the four types of feature values ​​to comparable numerical ranges. These features are then vertically concatenated in a fixed order as column vectors, constructing a 4×1 one-dimensional feature column vector. Based on the progression of the time windows, a total of N time windows are extracted and saved in chronological order of reception, with lengths ranging from 1000 to 5000 (e.g., N=2000). These 2000 four-dimensional feature column vectors are then sequentially stacked horizontally to form a 4×2000-dimensional temporal two-dimensional feature matrix.

[0037] In an optional embodiment, a high-dimensional phase space reconstruction trajectory is constructed based on a multi-dimensional feature sequence, and density-based spatial clustering is performed on the phase space reconstruction trajectory to obtain multiple signal clusters representing signal sources from different devices, cluster center feature vectors of each signal cluster, and intra-cluster feature covariance matrices, including: In the historical normal sample modeling stage, the delay time parameter of each dimension of the multidimensional feature sequence is calculated by mutual information method, and the embedding dimension parameter of each dimension of the feature is determined by false nearest neighbor method. The determined delay time parameter and embedding dimension parameter are used as fixed phase space reconstruction parameters for the phase space reconstruction of historical normal samples and current detection samples. Combining the delay time parameter and the embedding dimension parameter, the features of each dimension in the original feature sequence are converted into corresponding sub-state space matrices by coordinate delay processing. After pruning and aligning each sub-state space matrix according to the common effective time index or the shortest effective sample length, they are spliced ​​along the column direction to form a high-dimensional state space matrix as the phase space reconstruction trajectory. A density-based spatial clustering algorithm with a set neighborhood search radius and a minimum number of core sample points is applied to perform local density operation on the state space matrix to obtain multiple non-overlapping non-noise data point clusters as signal clusters. Calculate the arithmetic mean of the coordinate features of all member vectors within a single signal cluster as the corresponding cluster center feature vector; calculate the second-order central moments of all sample vectors within the signal cluster relative to this cluster center feature vector, combine them to generate the associated covariance matrix, and perform small diagonal loading regularization on the covariance matrix to ensure that the covariance matrix satisfies the symmetric positive definite property, which serves as the corresponding cluster feature covariance matrix.

[0038] For each one-dimensional sequence in the previously extracted multi-dimensional feature matrix, such as a continuous phase feature sequence, the mutual information algorithm is independently called during the historical normal sample modeling stage to solve for the number of hysteresis steps corresponding to the decay of the mutual information function to 1 / e of its original value, thus establishing independent delay time parameters for each dimension. For example, envelope =5, phase =8. The spurious nearest neighbor method is used to monitor the separation degree of neighboring points as the phase space dimension increases incrementally. When the proportion of spurious neighbors decreases to a set threshold, the increment is stopped and the optimal embedding dimension parameter for each feature dimension is output. Preferably 3 to 8 dimensions, for example =4-dimensional. This was determined during the modeling phase of the historical normal sample. and The parameter set is fixed and reused in subsequent detection samples to ensure that the symmetric positive definite matrix generated by the current sample and the historical normal samples are in the same dimensional space. The obtained independent parameter set is then applied. , Each original feature sequence of length L is expanded by coordinate delay mapping to construct a sub-state space matrix of its respective dimension. All feature sub-state spaces are then clipped and aligned according to their shortest row boundaries, followed by column-wise concatenation and merging. The output is reconstructed as follows: A multidimensional state space matrix of ×M dimensions, where the total embedding dimension is... For example, in matrix 18, each row represents an M-dimensional phase space state point. Clustering is performed on the reconstructed phase space data using the DBSCAN algorithm. The neighborhood search radius Eps is adjusted based on the current dimension and the discrete variance of the distribution, preferably ranging from 0.5 to 2.0, for example, 1.2, and the minimum number of core sample points MinPts is limited, preferably from 10 to 50 points, for example, 25. By connecting high-density connected regions and isolating free noise, disjoint high-dimensional data clusters are segmented as signal clusters representing different device signal sources. For the k-th independent signal cluster, the mean value of all vector nodes within the independent signal cluster is calculated, resulting in an M×1 column vector, which is established as the cluster center feature vector. The product of the entire demeaned sample vector set within the signal cluster and its transpose is calculated to obtain the second-order sample covariance central moments. To address the singularity problem caused by matrix underrank, a small array of artificial bias constants λI is superimposed on all principal eigenelements along the diagonal of the resulting square matrix. The regularization parameter λ is typically between... arrive For example, it is preferred to set as After regularization, an M×M intra-cluster feature covariance matrix is ​​generated.

[0039] S2, constructs directional attention based on KL divergence and kurtosis, and updates the feature vectors of each cluster.

[0040] Gaussian distributions are constructed for different signal clusters, and the one-way KL divergence between the distributions is calculated. Based on the divergence and the kurtosis of the characteristic distribution of each signal cluster, a directional attention weight matrix representing the asymmetric flow of information between clusters is constructed, and the spectral norm of the matrix is ​​calculated as an environmental complexity factor. The features passed by the neighbor nodes are aggregated and then multiplied by the transpose of the orthogonal eigenvector matrix of the feature covariance matrix of the target node within the cluster. The residuals are added to the initial cluster center eigenvector of the target node to obtain the updated feature vector of each signal cluster.

[0041] Using the cluster center eigenvectors and intra-cluster feature covariance matrices of each signal cluster obtained in the above steps, the feature distribution of each signal cluster is modeled as a multivariate Gaussian distribution. Based on the closed-form solution formula of the KL divergence of the multivariate Gaussian distribution, the one-way KL divergence value between the Gaussian distributions corresponding to any two different signal clusters is calculated through matrix inversion, determinant calculation, and matrix trace operation. The kurtosis value of each signal cluster in each feature dimension is calculated, and the arithmetic mean is taken as the kurtosis of the feature distribution. The one-way KL divergence is converted into an inter-cluster similarity component through negative correlation mapping, and then multiplied by the kurtosis adjustment factor of the feature distribution corresponding to the source signal cluster. Softmax normalization is performed on the input components of the same target node to generate a directed attention weight matrix. The singular value decomposition function, i.e., the SVD algorithm, is called to calculate the maximum singular value of the directed attention weight matrix, and the maximum singular value is used as the environment complexity factor. For each target signal cluster, the target signal cluster is taken as the target node, and other signal clusters in the space are taken as neighbor nodes. Feature aggregation is achieved by linearly weighting the cluster center eigenvectors of all neighbor nodes using the corresponding weight coefficients in the directed attention weight matrix. The orthogonal eigenvector matrix is ​​obtained by orthogonally decomposing the intra-cluster feature covariance matrix of the target node using the eigenvalue decomposition function. The aggregated feature column vectors are then left-multiplied by the transpose of the orthogonal eigenvector matrix for feature space alignment and principal component projection. The projection result is then added to the initial cluster center feature vector of the target node to achieve residual addition. The updated feature vectors of each signal cluster that incorporate global context information are output.

[0042] In an optional embodiment, a Gaussian distribution is constructed for different signal clusters, and the one-way KL divergence between the distributions is calculated, including: using the independent cluster center eigenvectors of each signal cluster as the distribution mean and the corresponding intra-cluster feature covariance matrix as the distribution covariance, and establishing a matching multidimensional Gaussian probability density mathematical distribution for all independent signal clusters respectively; traversing any two different signal clusters, taking one as the first examined signal cluster and the other as the second examined signal cluster, calculating the one-way KL divergence from the first examined signal cluster to the second examined signal cluster, and using it as an asymmetric measure of the difference in directed distributions between clusters.

[0043] The M×1 dimensional independent cluster center eigenvectors and the M×M dimensional intra-cluster feature covariance matrices of each previously generated signal cluster are retrieved and respectively assigned to the distribution mean vector μ and probability distribution covariance matrix Σ in a multivariate normal distribution. Using this analytical correspondence, a parameterized Gaussian probability density function model is fitted to each signal cluster. Based on a multi-node directed graph analysis architecture, the signal source topology network is traversed. In each iteration, the first cluster of signal sources responsible for initiating environmental disturbances is extracted, following a probability distribution. The second target signal cluster node, which is responsible for receiving the impact, follows a probability distribution. The relative entropy is evaluated for two independent multidimensional Gaussian functions. During computation, the bi-Gaussian analytical expression is called to calculate the logarithmic difference of the determinants. Inverse matrix cross-multiplication trace and the weighted mean displacement quadratic Mahalanobis distance of the covariance inverse solution Subtracting the corresponding spatial feature dimension constant M from the sum of these three terms and then multiplying by the constant 0.5 yields the original theoretical relative entropy value, representing the intensity of cross-node distribution differences when the first observation device points to the second observation target. This value is then used to construct the one-way KL divergence. When constructing the complete directed attention weight matrix, all different signal clusters are traversed according to the direction from the source signal cluster to the target signal cluster, and the one-way KL divergence of each directed cluster pair is calculated.

[0044] In an optional embodiment, a directed attention weight matrix representing the asymmetric flow of information between clusters is constructed based on the divergence and the kurtosis of the characteristic distribution of each signal cluster, and the spectral norm of the matrix is ​​calculated as an environmental complexity factor. This includes: calculating the average of the normalized fourth-order center distances in all higher-order feature dimensions covered by a single signal cluster as the intra-cluster feature distribution kurtosis describing the kurtosis morphology; obtaining the one-way KL divergence values ​​from the source signal cluster to the target signal cluster, converting the one-way KL divergence values ​​into inter-cluster similarity components through negative correlation mapping, and multiplying the inter-cluster similarity components with the intra-cluster feature distribution kurtosis adjustment factor corresponding to the source signal cluster to obtain the directed communication initial component values ​​on the corresponding network connection edge; converting all input initial components of the same target node into communication weights using the Softmax normalization function; writing the weights into the corresponding structure of the asymmetric attention graph to form the directed attention weight matrix; calculating the maximum eigenvalue of the product of the directed attention weight matrix and the conjugate transpose of the directed attention weight matrix, performing positive square root processing on the maximum eigenvalue to obtain the spectral norm, and setting the spectral norm as an environmental complexity factor representing the intensity of the influence of the perturbation environment in the entire space.

[0045] Retrieve the column vectors of all phase space reconstructed feature data for the i-th single independent signal source cluster, and perform mean and variance shifting and standardization operations on each M-dimensional data component in this higher-order supermanifold. Solve for the standardized fourth-order central moment, and subtract the natural kurtosis constant 3 of the Gaussian distribution to obtain the kurtosis metric. Sum the kurtosis parameters of all M data dimensions of the cluster and divide by the number of feature dimensions to perform arithmetic mean dimensionality reduction, obtaining a scalar parameter representing the level of concentration of signal fluctuations, for example... =1.85, established as the characteristic distribution kurtosis representing steep signal distortion within a cluster. To assess the influence projection between cascaded nodes, the real number of the one-way KL divergence from the source signal node to the destination signal node in the network structure is extracted, for example, measured... =4.2, and this one-way KL divergence is converted into inter-cluster similarity components through negative correlation mapping, for example... Where σ is a positive scaling parameter. The kurtosis of the source-end node feature distribution is used to generate a kurtosis adjustment factor through a bounded smooth mapping, for example... , where β is a non-negative adjustment coefficient. This is used as the initial value for directed communication components. The calculated scalar is assigned to the current communication connection path attribute parameter table to obtain the initial value for directed communication components. Softmax normalization is performed on the K directed input initial components flowing into the entire network from the same target cluster to obtain the communication weights. This ensures that the input weights of the same target node are in the range of 0 to 1 and that the sum of the weights is 1. Based on the initiator and receiver indices in the directed graph (i.e., row i, column j), the generated weights are filled into an empty matrix of size K×K, where K is the number of cluster source points, constructing a directed attention weight matrix A. The perturbation level features are extracted from the obtained weight matrix, and the current dense matrix A and its conjugate transpose are then processed. If the element is a real number, then it is Multiplying them together yields Square matrix; using numerical analysis to decompose the eigenvalues ​​of the square matrix and select the largest positive real eigenvalue. Solving for the scalar of this characteristic root The obtained matrix spectral norm is used to characterize the current topological perturbation expansion level, and the spectral norm is set as the environmental complexity factor ρ that is updated in real time.

[0046] In an optional embodiment, the features of the neighbor nodes are aggregated by weighted transfer and then multiplied by the transpose of the orthogonal eigenvector matrix of the intra-cluster feature covariance matrix corresponding to the target node. The residual is added to the initial cluster center eigenvector of the target node to obtain the updated feature vector of each signal cluster. This includes: performing eigenvalue decomposition calculation on the intra-cluster feature covariance matrix of a single signal cluster representing the target node to obtain the orthogonal eigenvector matrix of the intra-cluster feature covariance matrix. Extract the initial cluster center feature vectors of each initially connected neighbor node in the propagation direction, multiply them by the corresponding weights in the directional attention weight matrix to obtain individual weighted vectors, and sum the individual weighted vectors of all corresponding neighbors to obtain the target propagation aggregation vector; multiply the target propagation aggregation vector by the transpose of the orthogonal feature vector matrix to obtain the residual offset vector after feature space projection alignment; accumulate the residual offset vector with the original initial cluster center feature vector of the target node to obtain the corresponding output updated feature vector.

[0047] For a specific target network node that needs iterative updating, i.e., the j-th signal cluster, extract the intra-cluster feature covariance matrix of the object in the form of M×M. Eigenvalue decomposition is used to generate an M×M dimensional matrix of mutually perpendicular orthogonal eigenvectors. A local mapping projection basis adapted to the dominant dispersal direction of the target is established. In the fusion process, each associated source exerting interference on the target is extracted, including itself, totaling K source points, and the initial cluster center feature vector of each source point in M×1 dimensions is extracted. ; Call the corresponding decay scalar value between 0 and 1 in the targeted attention weight matrix. For example, 0.55. Using pointwise multiplication to establish association... Transform it into a single-item weighted column vector.

[0048] Execute formula By summing the M-dimensional weighted branches obtained from all associated tributary nodes, an M×1-dimensional target transfer aggregation vector representing the interferometric projection is obtained. To reduce the interference loss caused by spatial geometric misalignment, the eigenvector orthogonal matrix is... Perform matrix transpose For aggregated transfer body Perform vector projection mapping, that is, perform Mathematical multiplication. The target propagation aggregation vector is projected and aligned in the feature space using an orthogonal eigenvector matrix to generate an M×1 dimensional residual feature offset vector that matches the local statistical structure of the target signal cluster. The terminal aggregation network deploys a residual quick-jump structure model, which shifts the feature residual vector. Together with the target's own M×1 initial cluster center feature vector Perform column-wise addition; the calculation formula is as follows: A full-dimensional network convergence result is constructed that can take into account both inherent characteristics and background pressure. The residual update result serves as the update feature vector for the corresponding signal cluster, which is used for subsequent global symmetric positive definite tensor construction and anomaly detection.

[0049] The above process is achieved through an attention-based iterative convergence network. The input to the attention-based iterative convergence network consists of independent initial cluster center feature vectors for each signal cluster and the corresponding intra-cluster feature covariance matrix. The network structure includes a graph attention computation layer, a feature space projection alignment layer, and a residual-like quick-jump structure layer. The graph attention computation layer constructs communication components using the inter-cluster similarity component obtained by negative correlation mapping of the one-way Coulomb divergence and a feature distribution kurtosis adjustment factor, and obtains the directional attention weight matrix through Softmax normalization. The feature space projection alignment layer uses the transposed orthogonal feature vector matrix decomposed from the intra-cluster feature covariance matrix to map and project the target transfer aggregation vector obtained by weighted summation, outputting a residual offset vector. The residual-like quick-jump structure layer calculates the residual offset vector by adding the column-wise residuals of the initial cluster center feature vector. The output of the attention-based iterative convergence network is an updated feature vector that fuses the device's own transmission features and inter-cluster background correlation information.

[0050] S3 calculates the geodesic distance of the global covariance tensor and uses adaptive thresholding to detect concealed devices.

[0051] Calculate the second-order global covariance matrix of all updated feature vectors to generate a symmetric positive definite tensor representing the global features of the scene; calculate the geodesic distance between the symmetric positive definite tensor and the Frescher mean; determine the baseline quantile threshold based on the distribution of geodesic distances of historical normal samples; construct a positively correlated monotonically increasing function using the environmental complexity factor to regularize and relax the baseline quantile threshold; determine the presence of a hidden device when the actual geodesic distance is greater than the relaxed threshold.

[0052] Concatenate the updated eigenvectors of all signal clusters column-wise to form a global feature matrix. Calculate the second-order global covariance matrix of this global feature matrix, and add a value of 10 to the diagonal of this second-order global covariance matrix. -6 A diagonal regularization term is loaded to ensure that the matrix satisfies the symmetric positive definite property, generating a symmetric positive definite tensor representing the global features of the current spatial scene.

[0053] The geodesic distance between the symmetric positive definite tensor and the Frescher mean is calculated. Specifically, the Frescher mean matrix, calculated in the manifold space of the symmetric positive definite matrix from historical normal samples, is retrieved. The symmetric positive definite tensor and the Frescher mean matrix corresponding to the current scene are input into the affine invariant Riemannian metric model. The matrix product of the inverse square root of the Frescher mean matrix, the current symmetric positive definite tensor, and the inverse square root of the Frescher mean matrix is ​​calculated. The logarithm of the resulting matrix is ​​taken, and the Frobenius norm of the logarithm is calculated. This norm value is used as the geodesic distance of the current scene relative to the historical normal state. The formula is as follows: Where C is the current symmetric positive definite tensor. The Fraser mean of the historical normal sample. It is the Frobenius norm.

[0054] The set of geodesic distances from historical normal samples to the Fraser mean is extracted, and the 95th percentile value is calculated as the baseline quantile threshold. A positively correlated, monotonically increasing adjustment factor of not less than 1 is constructed using the environmental complexity factor. This adjustment factor is multiplied by the baseline quantile threshold to obtain the relaxed judgment threshold. The actual geodesic distance calculated at the current moment is compared with the relaxed threshold. If the actual geodesic distance is greater than the relaxed threshold, the system's anomaly alarm module is triggered to determine the presence of an unknown radio frequency signal emitted by a concealed device in the current test space.

[0055] Through the above processing, appropriately expanding the allowable range of normal fluctuations when environmental complexity increases helps to reduce the risk of false alarms caused by normal disturbances in complex electromagnetic environments. At the same time, the preceding directed attention fusion is still used to enhance the expression of abnormal cross-cluster relationships, thereby maintaining the ability to identify abnormal structural changes in concealed devices.

[0056] In an optional embodiment, the second-order global covariance matrix of all updated feature vectors is calculated to generate a symmetric positive definite tensor representing the global features of the scene. This includes: stacking all corresponding updated feature vectors in columns according to the order of signal cluster category labels to form a global updated feature matrix; calculating the demeaning matrix of the global updated feature matrix relative to the mean vector of the matrix's global features; performing matrix multiplication on the demeaning matrix and its own transpose, and dividing by a scalar of degrees of freedom equal to the total number of signal clusters minus 1 to generate a second-order global covariance matrix; performing small diagonal loading regularization on the second-order global covariance matrix to ensure that the second-order global covariance matrix satisfies the symmetric positive definite property, and using the processed symmetric positive definite matrix as the scene's global symmetric positive definite tensor.

[0057] The feature vectors of all active signal clusters obtained after attention fusion are updated, sorted in a fixed order according to a preset cluster index, and then aggregated to form a global feature representation. Following this sorting mechanism, the k-th feature node of size M×1, representing the convergence of the entire network's interactions, is continuously extracted to update the feature vector. K independent one-dimensional column vectors are aligned and concatenated in parallel to generate a global update feature matrix of dimension M×K. Traverse along the column directions of the matrix to generate an M×1 dimensional global feature mean vector. By subtracting the origin translation parameter from the phase of each of the K columns of the globally fused large matrix, the amount of invariant stationary DC drift error is reduced, resulting in a relatively unbiased demeaned matrix with dimensions always M×K. When the number of effective non-noise signal clusters K is not less than 2, the mean-reduced matrix with an M×K size structure on the left is... Perform cross matrix dot product inner multiplication with the transpose of the matrix on the right with a K×M configuration, and then divide by the degree of freedom coefficient K-1 for proportional compression. When the number of effective non-noise signal clusters K is less than 2, do not perform global covariance calculation by dividing by K-1. Instead, use the intra-cluster characteristic covariance matrix of the current effective signal cluster, or use a preset positive definite unit matrix after diagonal loading as the scene's global symmetric positive definite tensor to ensure that subsequent geodesic distance calculations can continue to be performed. Establish the square matrix restored to a fixed dimension M×M scale as a second-order global covariance square matrix.

[0058] Since highly homogeneous signal interactions may result in linear collinearity, leading to dimensional degradation and causing the covariance matrix to tend towards singular states, thus disrupting the inverse capability of subsequent differential manifold operations, a fixed diagonal constant positron matrix λI with an ultra-small scale parameter is added parallel to the main diagonal inside the joint initial high-order variance matrix as a structural correction term. For example, a term with a value of λ=1× Since λ is a positive number, the eigenvalues ​​of the matrix after adding λI are all greater than 0, thus satisfying the requirements of a symmetric positive definite matrix. This symmetric matrix, which contains information on the overall macroscopic mapping and changes in the environment, is established as the scene's global symmetric positive definite tensor, serving as the core reference for subsequent geodesic tracing and algorithm recognition verification.

[0059] In an optional embodiment, a baseline quantile threshold is determined based on the geodesic distance distribution of historical normal samples. A positively correlated monotonically increasing function is constructed using the environmental complexity factor to regularize and relax the baseline quantile threshold. This includes: obtaining all historical geodesic distance offset distribution models on the pre-installed projection stream of historical attack-free, clean environment record data; extracting values ​​with a cumulative distribution ratio reaching 95% within the distribution model as the baseline quantile threshold; generating an adjustment factor positively correlated with the environmental clutter level by applying a monotonically increasing smoothing adjustment operator to the environmental complexity factor; and multiplying the baseline quantile threshold by the adjustment factor to obtain a judgment threshold that increases with increasing environmental complexity, thereby achieving regularization and relaxation of the baseline quantile threshold.

[0060] During the offline model calibration phase, a secure, static, full-band spectrum channel with strict control and legally authorized communication, free from security risks, is used as the foundation. Tens of thousands of complete, fault-free, and stable historical records without malicious intent are detected and archived. These records are extracted and converted into a sequence of secure covariate tensors to construct a secure space database. Based on the characteristics of non-Euclidean space, a Frescher characteristic mean template is calculated on a symmetric positive definite manifold. Affine invariant distance is used to calculate the absolute multi-order curved geodesic distance from all historical secure tensor models to this secure template. Using the statistical kernel density smoothing estimation method, with an empirical bandwidth such as h=0.05, a probability density statistical profile curve covering the full range of distance fluctuations is fitted to establish a distribution model containing all historical geodesic distance offsets. Based on the distribution curve, an integral surface accumulation is performed to find the tolerance boundary line where the accumulated distribution accounts for 95% of the total capacity events. The extreme distance values ​​at the truncated boundary line are used as the base for determining whether the event is insurmountable. This base is extracted as a static judgment benchmark threshold, i.e., the benchmark quantile threshold. For example, the constant value is 4.62. Input the environmental complexity factor ρ based on a fully positive scalar, and use a monotonically increasing smooth adjustment function not less than 1 to generate the adjustment multiple, for example, using h(ρ) = 1 + αln(1 + ρ), where α is the environmental sensitivity adjustment constant, preferably set within the range of 0.1 to 0.5, for example, using a configuration value of α = 0.2. Perform the calculation; when the input measured environmental complexity factor ρ = 3.15, we obtain h(ρ) = 1 + 0.2ln(4.15) ≈ 1.28. This operation generates an adjustment multiple not less than 1 that increases with increasing environmental complexity. The obtained offline baseline quantile threshold is then used. Multiplying the constant by the online real-valued adjustment factor, the relaxed decision threshold is calculated to be approximately 4.62 × 1.28 ≈ 5.91. This output is confirmed as the adaptive decision threshold. By changing the fixed static attribute and switching to a dynamic decision method that relaxes according to the degree of external signal interference, threshold regularization relaxation based on environmental complexity is achieved. This method helps reduce false alarms caused by normal background fluctuations in complex electromagnetic environments. Furthermore, combined with the preceding cross-cluster directional attention feature update process, it improves the ability to distinguish abnormal radio frequency structural changes.

[0061] Ablation experiments were conducted under experimental conditions combining a radio frequency microwave anechoic chamber and a multipath fading channel simulator. Five legitimate broadband radio frequency transceiver nodes and two illegitimate interference source nodes were deployed in the test environment to simulate spoofing attacks. The environmental signal-to-noise ratio (SNR) varied randomly within the range of -15dB to 20dB. A total of 3000 sets of continuous time-series radio frequency data samples were collected as an independent validation set. Three experimental groups were set up for comparison: the first group was a complete scheme model including all technical features; the second group was a first ablation model that removed the attention iterative convergence network module and directly constructed a tensor using the initial feature vector; the third group was a second ablation model that removed the environmental complexity threshold adaptive relaxation module and relied solely on a fixed static quantile threshold for attack detection.

[0062] The experimental results obtained through system computation and detection evaluation on the full validation set samples are as follows: The anomaly recognition accuracy of the complete scheme model is 98.4%, the overall false alarm rate is 0.9%, and the average computation latency per frame is 12ms; the anomaly recognition accuracy of the first ablation model decreases to 91.6%, the overall false alarm rate increases to 4.8%, and the average computation latency per frame is 8ms; the anomaly recognition accuracy of the second ablation model decreases to 87.2%, the overall false alarm rate increases to 9.5%, and the average computation latency per frame is 11ms. Figure 3 As shown.

[0063] Based on the above experimental results, it can be seen that the core module significantly improves detection performance. Compared with the first ablation model, the complete scheme, by introducing a graph attention network and a feature residual affine alignment mapping mechanism, can reduce the impact of high-dimensional feature misalignment caused by multi-source radio frequency concurrency and inter-node interference clutter, thereby improving the anomaly identification accuracy by 6.8 percentage points and reducing the false alarm rate by 3.9 percentage points. Compared with the second ablation model, the complete scheme utilizes a threshold adaptive relaxation mechanism driven by environmental complexity factors to provide a more reasonable judgment tolerance for normal background disturbances in spectral environments with drastic signal-to-noise ratio fluctuations, reducing false alarms caused by fixed static threshold strategies, and improving the detection stability of the system in complex electromagnetic environments.

[0064] This invention also discloses a covert device detection system, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement a covert device detection method according to the present invention.

[0065] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0066] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting concealed devices, characterized in that, Including the following steps: S1. Acquire broadband radio frequency signals in the space to be tested and extract multidimensional feature sequences of continuous time series; construct high-dimensional phase space reconstruction trajectories based on multidimensional feature sequences, and perform density-based spatial clustering on the phase space reconstruction trajectories to obtain multiple signal clusters representing signal sources of different devices, cluster center feature vectors of each signal cluster, and feature covariance matrices within the clusters. S2. Construct Gaussian distributions for different signal clusters and calculate the one-way KL divergence between the distributions. Based on the divergence and the kurtosis of the characteristic distribution of each signal cluster, construct a directional attention weight matrix and calculate the spectral norm of the matrix as an environmental complexity factor. After aggregating the features passed by the neighbor nodes with weights, multiply the matrix by the transpose of the orthogonal eigenvector matrix of the feature covariance matrix of the target node within the cluster, and add the residuals to the initial cluster center eigenvector of the target node to obtain the updated feature vector of each signal cluster. S3: Calculate the second-order global covariance matrix of all updated feature vectors to generate a symmetric positive definite tensor representing the global features of the scene; calculate the geodesic distance between the symmetric positive definite tensor and the Frescher mean; determine the baseline quantile threshold based on the distribution of geodesic distances of historical normal samples; construct a positively correlated monotonically increasing function using the environmental complexity factor to regularize and relax the baseline quantile threshold; determine the presence of a hidden device when the actual geodesic distance is greater than the relaxed threshold.

2. The method for detecting concealed devices according to claim 1, characterized in that, The method for extracting multidimensional feature sequences from continuous time series is as follows: The received broadband radio frequency signal in the test space is subjected to spectrum scanning, channelization separation, or peak band tracking to obtain at least one candidate narrowband signal component; the candidate narrowband signal component is divided into time windows of equal length according to a fixed time sequence; the instantaneous envelope value, instantaneous phase value, carrier frequency offset value, and received signal strength indication value are extracted within each time window; the instantaneous phase value is subjected to phase unwrapping processing to obtain continuous phase value; the instantaneous envelope value, continuous phase value, carrier frequency offset value, and received signal strength indication value are subjected to dimensionless normalization processing; all normalized values ​​within the same time window are concatenated to form a one-dimensional feature vector. Following the time series order, the one-dimensional feature vectors generated by all consecutive time windows are stacked to form a feature matrix, and the feature matrix as a whole is determined as a multi-dimensional feature sequence.

3. The method for detecting concealed devices according to claim 1, characterized in that, The high-dimensional phase space reconstruction trajectory is constructed based on multidimensional feature sequences, including: in the historical normal sample modeling stage, the delay time parameter of each dimension feature in the multidimensional feature sequence is calculated by mutual information method, and the embedding dimension parameter of each dimension feature is determined by false nearest neighbor method. The determined delay time parameter and embedding dimension parameter are used as fixed phase space reconstruction parameters for the phase space reconstruction of historical normal samples and current detection samples. Combining the delay time parameter and embedding dimension parameter, the features in each dimension of the original feature sequence are converted into corresponding sub-state space matrices by coordinate delay processing. After pruning and aligning each sub-state space matrix according to the common effective time index or the shortest effective sample length, they are spliced ​​along the column direction to form a high-dimensional state space matrix as the phase space reconstruction trajectory.

4. The method for detecting concealed devices according to claim 1, characterized in that, Density-based spatial clustering is performed on the phase space reconstructed trajectory to obtain multiple signal clusters representing signal sources from different devices, the cluster center feature vectors of each signal cluster, and the intra-cluster feature covariance matrix, including: A density-based spatial clustering algorithm with limitations on the neighborhood search radius and the minimum number of core sample points is applied to perform local density operations on the state space matrix to obtain multiple disjoint, non-noise data point clusters as signal clusters. The arithmetic mean of the coordinate features of all member vectors within a single signal cluster is calculated as the corresponding cluster center feature vector. The second-order central moments of all sample vectors within the signal cluster relative to this cluster center feature vector are calculated and combined to generate an associated covariance matrix. The covariance matrix is ​​then subjected to a small diagonal loading regularization process to ensure that the covariance matrix satisfies the symmetric positive definite property, which serves as the corresponding intra-cluster feature covariance matrix.

5. The method for detecting concealed devices according to claim 1, characterized in that, For different signal clusters, construct Gaussian distributions and calculate the one-way KL divergence between distributions. This includes: using the independent cluster center eigenvectors of each signal cluster as the distribution mean and the corresponding intra-cluster feature covariance matrix as the distribution covariance, establishing matching multidimensional Gaussian probability density mathematical distributions for all independent signal clusters; traversing any two different signal clusters, taking one as the first examined signal cluster and the other as the second examined signal cluster, and calculating the one-way KL divergence from the first examined signal cluster to the second examined signal cluster.

6. The method for detecting concealed devices according to claim 1, characterized in that, The calculation method for the environmental complexity factor includes: calculating the average of the standardized fourth-order central distances in all higher-order feature dimensions covered by a single signal cluster, which is used as the kurtosis of the intra-cluster feature distribution; obtaining the one-way KL divergence value from the source signal cluster to the target signal cluster, converting the one-way KL divergence value into an inter-cluster similarity component through negative correlation mapping, and multiplying the inter-cluster similarity component with the intra-cluster feature distribution kurtosis adjustment factor corresponding to the source signal cluster to obtain the directed communication initial component value on the corresponding network connection edge; converting all input initial components of the same target node into communication weights using a normalization function; writing the weights into the corresponding structure of the asymmetric attention graph to form a directed attention weight matrix; calculating the maximum eigenvalue of the product of the directed attention weight matrix and the conjugate transpose of the directed attention weight matrix, performing positive square root processing on the maximum eigenvalue to obtain the spectral norm, and setting the spectral norm as the environmental complexity factor.

7. The method for detecting concealed devices according to claim 1, characterized in that, The method for obtaining the updated feature vectors of each signal cluster includes: performing eigenvalue decomposition on the intra-cluster feature covariance matrix of a single signal cluster representing the target node to obtain the orthogonal eigenvector matrix of the intra-cluster feature covariance matrix; extracting the initial cluster center feature vectors of each initially connected neighbor node in the transmission direction, multiplying them by the weights of the corresponding directional attention weight matrix to obtain individual weighted vectors, and summing the individual weighted vectors of all corresponding neighbors to obtain the target transmission aggregation vector; multiplying the transpose of the orthogonal eigenvector matrix on the left by the target transmission aggregation vector to obtain the residual offset vector after feature space projection alignment; and accumulating the residual offset vector with the original initial cluster center feature vector of the target node to obtain the corresponding output updated feature vector.

8. The method for detecting concealed devices according to claim 1, characterized in that, The process involves calculating the second-order global covariance matrix of all updated feature vectors and generating a symmetric positive definite tensor representing the global features of the scene. This includes: stacking all corresponding updated feature vectors column-wise according to the signal cluster category label order to form a global updated feature matrix; calculating the demeaning matrix of the global updated feature matrix relative to the mean vector of the matrix's global features; performing matrix multiplication on the demeaning matrix and its own transpose, and dividing by a scalar of degrees of freedom equal to the total number of signal clusters minus 1 to generate a second-order global covariance matrix; performing a small diagonal loading regularization on the second-order global covariance matrix to ensure that it satisfies the symmetric positive definite property, and using the processed symmetric positive definite matrix as the symmetric positive definite tensor of the scene's global features.

9. The method for detecting concealed devices according to claim 1, characterized in that, The baseline quantile threshold is determined based on the geodesic distance distribution of historical normal samples. A positively correlated monotonically increasing function is constructed using the environmental complexity factor to regularize and relax the baseline quantile threshold. This process includes: obtaining the distribution model of all historical geodesic distance offsets on the pre-installed projection stream of historical, attack-free, and clean environmental records; extracting values ​​with a cumulative distribution ratio reaching 95% within the distribution model and using these values ​​as the baseline quantile threshold; generating a regulation factor positively correlated with the level of environmental clutter by applying a monotonically increasing smoothing adjustment operator to the environmental complexity factor; and multiplying the baseline quantile threshold by the regulation factor to obtain a judgment threshold that increases with increasing environmental complexity, thus achieving regularization and relaxation of the baseline quantile threshold.

10. A concealed device detection system, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement a covert device detection method according to any one of claims 1-9.

Citation Information

Patent Citations

  • A Deep Learning-Based Method for Satellite Signal Cycle Slip Detection and Repair

    CN112946697B

  • Broadband spectrum compression detection method based on digital twinning and signal noise

    CN121864229A

  • Multipath clustering and tracking method for dynamic wireless channel

    CN122073510A