A multi-network collaborative wireless radio frequency signal full-domain detection intelligent prevention and control system

The intelligent control system for full-domain detection of wireless radio frequency signals through multi-network collaboration solves the problems of multi-dimensional feature extraction and full-domain situational awareness in complex electromagnetic environments of existing systems. It realizes accurate positioning of target signal sources and threat assessment, and improves the detection accuracy and resource allocation efficiency of the system.

CN123123985APending Publication Date: 2026-11-17ZHENGZHOU YOUMEI INTELLIGENT TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611254365.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-18
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing wireless radio frequency signal detection systems struggle to extract multi-dimensional features in complex electromagnetic environments, leading to decreased detection accuracy and stability. Furthermore, they lack comprehensive situational awareness, hindering effective assessment of signal source threats and efficient resource allocation.

Method used

The intelligent prevention and control system for full-domain detection of wireless radio frequency signals through multi-network collaboration utilizes the computing module to extract the temporal scale spectrum, polarization state, and frequency drift matrix of I/Q data, constructs a joint feature tensor, generates a local situational syntax graph through a graph convolutional network, and establishes a full-domain network graph by combining the correlation between geographic and information entropy, thereby realizing the location of target signal sources and threat assessment, and generating resource allocation strategies.

Benefits of technology

It enables multi-dimensional deep feature recognition in complex electromagnetic environments, reduces perception blind spots, improves the collaborative detection and prevention capabilities of radio frequency signal threats, and can accurately assess security threats and optimize resource allocation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123985A_ABST
    Figure CN123123985A_ABST
Patent Text Reader

Abstract

This invention provides a multi-network collaborative wireless radio frequency signal detection and intelligent control system, comprising: detection nodes collecting radio frequency I / Q data, obtaining three types of matrices through wavelet transform, polarization decomposition, and second-order phase difference; constructing a joint feature tensor by dimension-aligned stacking; tensor blocks serving as topology nodes, networking based on time-frequency polarization similarity, outputting a local situational awareness syntax graph through hyper-network parameter tuning graph convolutional network and transmitting it back to the central unit; the central unit constructs a global network graph based on detection points, solves feature weights and collaborative strength by combining scale features, and fuses to generate a global situational awareness map to locate signal sources; after associating target signal sources with their positions relative to detection nodes and connecting them to the global network graph, calculating PageRank to determine threat levels, solving for the optimal suppression resource allocation scheme, and distributing it to control nodes to complete radio frequency security management.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application belongs to the field of wireless radio frequency signals, and in particular relates to a multi-network collaborative wireless radio frequency signal full-domain detection and control system. Background Technology

[0002] With the rapid development of wireless communication technology, the space electromagnetic environment is becoming increasingly complex and dynamic, with various legitimate communication signals intertwined with potential illegal interference and eavesdropping signals. In such radio frequency (RF) environments, relying solely on a single detection node or simple data analysis methods is insufficient for accurate detection of unknown RF signal sources. Existing detection systems typically focus on single-dimensional or shallow feature extraction when processing RF signals, making it difficult to fully identify the comprehensive attributes of signals across multiple dimensions such as time-frequency, polarization, and phase. These limitations lead to a decline in detection accuracy and stability when facing weak signals, complex interference, or multi-source aliasing scenarios. Furthermore, in distributed multi-node detection networks, the large amount of local situational information collected by each detection node often lacks deep collaborative fusion and structured processing. The central dispatching unit struggles to reconstruct the true situation of the entire electromagnetic signal domain from scattered local data, resulting in blind spots and lags in the overall network's perception of RF signals.

[0003] After detecting a target signal source, accurately assessing its threat level and implementing effective suppression is a crucial issue in radio security. Conventional defense systems, in the transition from signal detection to defense decision-making, often lack a deep understanding of the global network topology and signal propagation relationships. Faced with highly concealed radio frequency security threats, existing systems often struggle to calculate the actual impact of a target signal source on the entire network and to scientifically assess its threat level. This disconnect between perception and decision-making prevents the system from developing reasonable resource allocation schemes based on the true threat characteristics of the target source when performing collaborative defense tasks. This can lead to uneven resource allocation and insufficient collaborative efficiency among multiple defense nodes performing suppression tasks. Therefore, how to achieve joint extraction of multi-dimensional features of radio frequency signals in a distributed architecture and complete target source localization, threat assessment, and intelligent allocation of defense resources based on global situational awareness has become a pressing technical problem to be solved. Summary of the Invention

[0004] According to a first aspect of the present invention, a multi-network collaborative wireless radio frequency signal full-domain detection and control system is proposed, which is used to realize the joint extraction of multi-dimensional features of radio frequency signals under a distributed architecture, and includes the following modules: The calculation module is used to acquire I / Q data by collecting radio frequency signals from multiple detection nodes, perform wavelet transform on the I / Q data to obtain the time-scale spectrum matrix, generate the polarization state matrix by combining polarization data, calculate the second-order phase difference to obtain the frequency drift matrix, and stack the time-scale spectrum matrix, polarization state matrix and frequency drift matrix after dimension alignment processing to generate a joint feature tensor. The construction module is used to segment the joint feature tensor into local feature blocks and treat them as nodes, construct a topology graph based on time-frequency and polarization similarity, and generate a local situation grammar graph using a graph convolutional network. The parameters of the graph convolutional network are generated by the supernetwork based on the global signal situation of the previous period or the initial preset. Each detection node reports the local situation grammar graph, scale features and location to the central unit. The central unit uses the detection node as the vertex and establishes a global network graph based on the correlation between geography and information entropy. The generation module is used to calculate the importance of features within nodes and the strength of collaborative association between nodes based on scale features on the global network graph, fuse them to generate a global signal situation map, locate the target signal source based on the global signal situation map, connect the target signal source to the global network graph based on the positional association between the target signal source and the detection node, calculate the PageRank value to assess the security threat, generate a resource allocation strategy to maximize the suppression effectiveness of the target signal source and distribute it to the prevention and control nodes for execution.

[0005] Preferably, the plurality of detection nodes acquire radio frequency signals to obtain I / Q data, and perform wavelet transform on the I / Q data to obtain a time-scale spectrum matrix, including: The I / Q data are concatenated into a complex signal, and a complex Morlet wavelet is selected as the transformation basis. A scale sequence containing 256 scales was established to cover the analysis frequency band in a logarithmically equidistant manner; The complex signal is convolved and integrated with the complex Morlet wavelet basis functions corresponding to each scale by a sliding time window, and the absolute modulus of the result is extracted to generate a two-dimensional time-scale spectrum matrix.

[0006] Preferably, the step of generating a polarization state matrix by combining polarization data includes: Receives the horizontally polarized signal and the vertically polarized signal output from the dual-polarized antenna; Calculate the instantaneous amplitude, phase difference, and cross-correlation coefficient of the two polarized signals; Based on the calculated parameters, the Stokes parameters are solved, and the polarization state matrix is ​​constructed to identify the instantaneous polarization characteristics of the signal source.

[0007] Preferably, the step of segmenting the joint feature tensor into local feature blocks and treating them as nodes, and constructing a topological graph based on time-frequency and polarization similarity, includes: The joint feature tensor is segmented using a fixed-size three-dimensional sliding window to obtain the local feature blocks; After dimensionless processing of frequency interval, time evolution difference and polarization parameter difference, the weighted Euclidean distance between any two feature blocks is calculated. When the weighted Euclidean distance between two feature blocks is less than a preset threshold, a connection edge is established between the corresponding nodes, and a local topology graph is constructed using the reciprocal of the distance as the weight.

[0008] Preferably, the parameters of the graph convolutional network are generated by the supernetwork based on the previous cycle or an initially preset global signal state, including: Construct a hypernetwork model containing three fully connected layers; The global signal situation vector of the previous cycle is input into the super network model. If it is the initial cycle, the initial preset global signal situation vector is used, and a one-dimensional feature vector is output through the activation function and linear mapping. The one-dimensional feature vector is reshaped into a weight matrix, replacing the aggregation function parameters and node feature transformation matrices in the graph convolutional network.

[0009] Preferably, the central unit establishes a global network graph based on the correlation between geographic location and information entropy, with the detection node as the vertex, including: The actual geographical distance between nodes is calculated using the spherical trigonometry formula, and the reciprocal of the distance is taken as the spatial weight. Calculate the KL divergence of the feature distribution of the local situation syntax graph of two nodes, and take the reciprocal of the smoothed KL divergence as the information entropy similarity. The spatial weights and information entropy similarity are weighted and summed after being processed to be dimensionless. If the result exceeds a set threshold, a directed connection is established between the nodes and the corresponding edge weight is assigned.

[0010] Preferably, the step of calculating the importance of features within nodes and the strength of collaborative associations between nodes based on scale characteristics on the global network graph includes: One-dimensional convolution and the Softmax function are used to normalize the multi-scale features within the node, and the feature importance weight matrix is ​​output. Calculate the cross-covariance matrix of the corresponding scale features between nodes, perform singular value decomposition on the cross-covariance matrix, extract the ratio of the largest singular value to the sum of all singular values ​​as the cooperative association strength, and remove redundant connections with association strength below a set threshold.

[0011] Preferably, the step of generating a resource allocation strategy that maximizes the suppression effectiveness against the target signal source and distributing it to the control nodes for execution includes: An adversarial optimization model is established with the objective function of maximizing the suppression effectiveness against the target signal source. Using the maximum transmit power and available frequency bands of each control node as constraints, a mapping relationship between interference suppression parameters and suppression effectiveness is established. The adversarial optimization model is solved using a heuristic algorithm, and the output suppression frequency, interference power, and beam pointing parameters of the target signal source are used as resource allocation strategies.

[0012] According to a second aspect of the present invention, a multi-network collaborative wireless radio frequency signal full-domain detection and intelligent control method is proposed, comprising the following steps: Multiple detection nodes acquire radio frequency signals to obtain I / Q data, perform wavelet transform on the I / Q data to obtain a time-scale spectrum matrix, combine it with polarization data to generate a polarization state matrix, calculate the second-order phase difference to obtain a frequency drift matrix, and then stack the time-scale spectrum matrix, polarization state matrix and frequency drift matrix after dimension alignment to generate a joint feature tensor. The joint feature tensor is segmented into local feature blocks and regarded as nodes. A topology graph is constructed based on time-frequency and polarization similarity. A local situation grammar graph is generated using a graph convolutional network. The parameters of the graph convolutional network are generated by the supernetwork based on the global signal situation of the previous period or the initial preset. Each detection node reports the local situation grammar graph, scale features and location to the central unit. The central unit establishes a global network graph based on the correlation between geography and information entropy, with the detection node as the vertex. On the global network graph, the importance of features within nodes and the strength of collaborative association between nodes are calculated based on scale characteristics, and a global signal situation map is generated by fusion. The target signal source is located based on the global signal situation map. The target signal source is connected to the global network graph based on the positional association between the target signal source and the detection node. The PageRank value is calculated to assess the security threat. A resource allocation strategy that maximizes the suppression effectiveness of the target signal source is generated and distributed to the prevention and control nodes for execution.

[0013] Optionally, the plurality of detection nodes acquire radio frequency signals to obtain I / Q data, and perform wavelet transform on the I / Q data to obtain a time-scale spectrum matrix, including: The I / Q data are concatenated into a complex signal, and a complex Morlet wavelet is selected as the transformation basis. A scale sequence containing 256 scales was established to cover the analysis frequency band in a logarithmically equidistant manner; The complex signal is convolved and integrated with the complex Morlet wavelet basis functions corresponding to each scale by a sliding time window, and the absolute modulus of the result is extracted to generate a two-dimensional time-scale spectrum matrix.

[0014] Optionally, generating the polarization state matrix by combining polarization data includes: Receives the horizontally polarized signal and the vertically polarized signal output from the dual-polarized antenna; Calculate the instantaneous amplitude, phase difference, and cross-correlation coefficient of the two polarized signals; Based on the calculated parameters, the Stokes parameters are solved, and the polarization state matrix is ​​constructed to identify the instantaneous polarization characteristics of the signal source.

[0015] Optionally, the step of segmenting the joint feature tensor into local feature blocks and treating them as nodes, and constructing a topology graph based on time-frequency and polarization similarity, includes: The joint feature tensor is segmented using a fixed-size three-dimensional sliding window to obtain the local feature blocks; After dimensionless processing of frequency interval, time evolution difference and polarization parameter difference, the weighted Euclidean distance between any two feature blocks is calculated. When the weighted Euclidean distance between two feature blocks is less than a preset threshold, a connection edge is established between the corresponding nodes, and a local topology graph is constructed using the reciprocal of the distance as the weight.

[0016] Optionally, the parameters of the graph convolutional network are generated by the supernetwork based on the previous cycle or an initially preset global signal state, including: Construct a hypernetwork model containing three fully connected layers; The global signal situation vector of the previous cycle is input into the super network model. If it is the initial cycle, the initial preset global signal situation vector is used, and a one-dimensional feature vector is output through the activation function and linear mapping. The one-dimensional feature vector is reshaped into a weight matrix, replacing the aggregation function parameters and node feature transformation matrices in the graph convolutional network.

[0017] Optionally, the central unit establishes a global network graph based on the correlation between geographic location and information entropy, with the detection node as the vertex, including: The actual geographical distance between nodes is calculated using the spherical trigonometry formula, and the reciprocal of the distance is taken as the spatial weight. Calculate the KL divergence of the feature distribution of the local situation syntax graph of two nodes, and take the reciprocal of the smoothed KL divergence as the information entropy similarity. The spatial weights and information entropy similarity are weighted and summed after being processed to be dimensionless. If the result exceeds a set threshold, a directed connection is established between the nodes and the corresponding edge weight is assigned.

[0018] Optionally, the step of calculating the importance of features within nodes and the strength of collaborative associations between nodes based on scale characteristics on the global network graph includes: One-dimensional convolution and the Softmax function are used to normalize the multi-scale features within the node, and the feature importance weight matrix is ​​output. Calculate the cross-covariance matrix of the corresponding scale features between nodes, perform singular value decomposition on the cross-covariance matrix, extract the ratio of the largest singular value to the sum of all singular values ​​as the cooperative association strength, and remove redundant connections with association strength below a set threshold.

[0019] Optionally, generating a resource allocation strategy that maximizes the suppression effectiveness against the target signal source and distributing it to the control nodes for execution includes: An adversarial optimization model is established with the objective function of maximizing the suppression effectiveness against the target signal source. Using the maximum transmit power and available frequency bands of each control node as constraints, a mapping relationship between interference suppression parameters and suppression effectiveness is established. The adversarial optimization model is solved using a heuristic algorithm, and the output suppression frequency, interference power, and beam pointing parameters of the target signal source are used as resource allocation strategies.

[0020] This invention extracts the time-scale spectrum, polarization state, and frequency drift matrix of radio frequency (RF) signals and constructs a joint feature tensor, enabling the characterization of multi-dimensional deep features of RF signals in complex electromagnetic environments. By converting the joint features into a topological graph and generating a local situational syntax graph using a graph convolutional network, combined with a hypernetwork mechanism to adaptively configure network parameters based on the global situation, it can identify implicit correlations between local signals. Furthermore, by fusing geographical correlation and information entropy correlation through central units, a global network graph is constructed, enabling collaborative processing and global fusion of data from multiple detection nodes, reducing the perception blind spots caused by isolated information points. After generating a global signal situational graph based on feature importance and the strength of collaborative correlation between nodes, target signal sources can be located. Based on the positional correlation between the target signal source and detection nodes, the target signal source is integrated into the global network graph. The PageRank algorithm is used to calculate and assess the degree of security threat, thereby formulating a resource allocation strategy to maximize suppression effectiveness and improving the collaborative detection, situational assessment, and prevention and control capabilities for RF signal threats. Attached Figure Description

[0021] Figure 1 This is a schematic diagram illustrating the evolution trajectory of instantaneous polarization characteristics; Figure 2 This is a schematic diagram of the distribution curve of the importance of scale features; Figure 3 This is a schematic diagram of the resource allocation strategy for the target signal source. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] In the first embodiment, the present invention proposes a multi-network collaborative wireless radio frequency signal full-domain detection intelligent prevention and control system, comprising the following modules: The calculation module is used to acquire I / Q data by collecting radio frequency signals from multiple detection nodes, perform wavelet transform on the I / Q data to obtain the time-scale spectrum matrix, generate the polarization state matrix by combining polarization data, calculate the second-order phase difference to obtain the frequency drift matrix, and stack the time-scale spectrum matrix, polarization state matrix and frequency drift matrix after dimension alignment processing to generate a joint feature tensor.

[0024] Radio frequency (RF) signal sampling is performed. Before entering the baseband acquisition module, the RF signal undergoes low-noise amplification, filtering, and local oscillator mixing at the RF front-end. This down-converts the RF subband to be analyzed to the baseband or intermediate frequency range that meets the sampling bandwidth requirements, acquiring in-phase I and quadrature Q digital sequences. The complex I and Q sequences are calculated to obtain the absolute modulus of the wavelet coefficients at each scale and time point, constructing a two-dimensional time-scale spectrum matrix. Horizontally and vertically polarized I and Q signals are acquired, Stokes parameters are calculated, and these parameters are normalized and smoothed, then concatenated according to the time step to form a polarization state matrix. The Stokes parameters are parameters characterizing the polarization state of electromagnetic waves or optical signals using four real components: total intensity, horizontal / vertical linear polarization difference, 45° linear polarization component, and circular polarization component. The complex phases of the I and Q data are unpacked to avoid phase jumps, and two consecutive calculations are performed to obtain the second-order phase difference. This second-order phase difference is then time-normalized based on the sampling time interval to obtain frequency drift characteristic values ​​representing the frequency change trend. This result is truncated and arranged at a fixed length within a time window to generate a frequency drift matrix. The time-scale spectral matrix and the frequency drift matrix are interpolated and scaled to the same number of time frames and feature dimensions. For the polarization state matrix, it is first aligned along the time dimension, and the normalized Stokes features are mapped to a polarization embedding matrix with the same feature dimension through a shared linear projection layer or one-dimensional convolution, achieving dimension alignment. The three aligned matrices are merged along the newly added channel dimension to generate a three-dimensional joint feature tensor containing the time step, feature dimension, and number of channels. Before stacking, the time-scale spectral matrix, polarization embedding matrix, and frequency drift matrix are normalized or standardized to ensure that the three types of matrices are on a unified numerical scale.

[0025] In an optional embodiment, the plurality of detection nodes acquire radio frequency signals to obtain I / Q data, and perform wavelet transform on the I / Q data to obtain a time-scale spectrum matrix, including: The I / Q data are concatenated into a complex signal, and a complex Morlet wavelet is selected as the transformation basis. A scale sequence containing 256 scales was established to cover the analysis frequency band in a logarithmically equidistant manner; The complex signal is convolved and integrated with the complex Morlet wavelet basis functions corresponding to each scale by a sliding time window, and the absolute modulus of the result is extracted to generate a two-dimensional time-scale spectrum matrix.

[0026] The system acquires data with a front-end RF sampling rate set to 100MHz, where 100MHz is the baseband or intermediate frequency sampling rate after down-conversion. The effective analysis bandwidth corresponding to a single acquisition does not exceed the Nyquist bandwidth allowed by this sampling rate. The system performs segmented scanning or sub-band coverage of the target RF frequency band within the range of 10MHz to 10GHz through local oscillator frequency switching, RF sub-band selection, or multi-channel parallel down-conversion. The in-phase component I(t) and quadrature component Q(t) are concatenated in the complex form s(t) = I(t) + jQ(t) to form an original complex signal sequence with a duration of 10ms. In the characteristic transform domain, the complex Morlet wavelet center frequency parameter and bandwidth parameter determined according to the sampling rate and wavelet library calibration rules are selected as the basic transformation basis. Subsequently, based on the correspondence between pseudo-frequency and scale of the complex Morlet wavelet, a logarithmically equidistant scale sequence containing 256 scales is pre-calibrated so that each scale corresponds to the target analysis frequency in the baseband or intermediate frequency range after downconversion; the so-called analysis frequency band range covering 10MHz to 10GHz refers to the coverage range of the RF center frequency band after segmented downconversion by the RF front end, rather than directly covering the 10GHz instantaneous bandwidth with a single 100MHz baseband sampling.

[0027] A sliding time window with a length of 1024 sampling points and a step size of 512 sampling points is constructed to slice the complex signal sequence. Within each truncated window, the time-domain complex signal slice is convolved and integrated pointwise with the complex Morlet wavelet basis functions after translation and scaling under the above 256 different scale parameters. After integration, the absolute modulus of the complex response is extracted and normalized to the [0,1] interval. A two-dimensional time-scale spectrum matrix with a size of 256 scale dimensions multiplied by the total number of time steps is generated in memory to represent the time-frequency domain amplitude response.

[0028] In an optional embodiment, generating the polarization state matrix by combining polarization data includes: Receives the horizontally polarized signal and the vertically polarized signal output from the dual-polarized antenna; Calculate the instantaneous amplitude, phase difference, and cross-correlation coefficient of the two polarized signals; Based on the calculated parameters, the Stokes parameters are solved, and the polarization state matrix is ​​constructed to identify the instantaneous polarization characteristics of the signal source.

[0029] The front-end RF module synchronously receives the horizontal polarization channel time-series signal output from the orthogonal dual-polarized antenna. and vertical polarization channel time series signal The instantaneous analytical envelopes of the two signals are extracted separately to obtain the instantaneous amplitude. and And calculate the instantaneous phase of the two to obtain the phase difference. Using 2048 sampling points, corresponding to a 20.48 microsecond time window as a fixed sliding window length, for... and Mean removal and energy normalization are performed within the corresponding time window. The normalized cross-correlation function is calculated, and the absolute value of the zero-delay cross-correlation coefficient or the cross-correlation peak is extracted as the polarization channel correlation. The polarization channel correlation with a value range of [0,1] is obtained.

[0030] The four-dimensional Stokes parameter vector is solved by combining the acquired instantaneous parameters: the parameters representing the total radiated power are calculated separately. , representing the difference between horizontal and vertical polarization components , representing the 45-degree orthogonal polarization component and representing circular polarization components The obtained four-dimensional sequence is then subjected to noise reduction and smoothing using a moving average filter with a length of 100 sample points. To avoid mixing the total power term and the polarization normalization component, the following steps are first performed... Normalized according to the maximum power or preset power calibration value within the sampling window to obtain Calculate separately , and , where ε is a preset minimum positive number; by , , and Construct a normalized Stokes feature sequence, where Located in the interval [0,1], , and Located in the interval [-1, 1]. 1024 consecutive time sampling steps are taken along the time axis. The four denoised and normalized parameter row vectors are stacked layer by layer, and then... Located in the interval [0,1] , and Constrained within the range [-1, 1], a real-valued polarization state matrix with dimensions of 4 rows and 1024 columns is constructed, outputting the multidimensional evolution trajectory of the instantaneous polarization characteristics of the spatial target. The instantaneous polarization characteristic evolution trajectory is as follows: Figure 1 As shown.

[0031] The construction module is used to segment the joint feature tensor into local feature blocks and treat them as nodes, construct a topology graph based on time-frequency and polarization similarity, and generate a local situation grammar graph using a graph convolutional network. The parameters of the graph convolutional network are generated by the supernetwork based on the global signal situation of the previous period or an initial preset. Each detection node reports the local situation grammar graph, scale features, and location to the central unit. The central unit uses the detection nodes as vertices and establishes a global network graph based on the correlation between geography and information entropy.

[0032] A fixed step size and kernel size are used to slide and prune along the joint feature tensor along the time and feature dimensions to extract local feature blocks, each of which is treated as an independent node in the graph network structure. After dimensionless processing of the frequency interval, temporal evolution difference, and polarization parameter difference of corresponding local feature blocks between nodes, the weighted Euclidean distance between any two nodes is calculated. If the weighted Euclidean distance is less than a preset threshold, a connection edge is established between the two nodes to obtain a topology graph. Feature aggregation and dimensionality reduction are performed on the topology graph and node feature blocks to update the node state sequence and generate a local situation syntax graph. To achieve adaptive parameter tuning, a multilayer perceptron supernetwork model containing multiple fully connected layers is constructed. The model takes the global signal situation vector from the previous cycle as input during the forward propagation process, or an initially preset global signal situation vector as input, and outputs a set of weight tensors, which are then assigned to the weight parameter matrix of the graph convolutional network layer module running at the current detection node. Each detection node sends the serialized local situation syntax map data packet, scale features, and latitude and longitude location information obtained by parsing the NMEA0183 protocol using a UDP-based network communication socket to the central unit's listening port. The central unit receives the data from each node, initializes the global network graph, adds each detection node as a vertex, calculates the spherical geographic distance between any two nodes based on latitude and longitude, calculates the information entropy difference between the node's local situation syntax map, calculates the weighted sum of the reciprocals of the distance and information entropy difference to obtain the edge weight, and calls the add weighted edge function to add weighted edges, thus establishing the global network graph.

[0033] In an optional embodiment, the step of segmenting the joint feature tensor into local feature blocks and treating them as nodes, and constructing a topology graph based on time-frequency and polarization similarity, includes: The joint feature tensor is segmented using a fixed-size three-dimensional sliding window to obtain the local feature blocks; After dimensionless processing of frequency interval, time evolution difference and polarization parameter difference, the weighted Euclidean distance between any two feature blocks is calculated. When the weighted Euclidean distance between two feature blocks is less than a preset threshold, a connection edge is established between the corresponding nodes, and a local topology graph is constructed using the reciprocal of the distance as the weight.

[0034] A 3D joint feature tensor of size 1024×256×3, generated by alignment and stitching, is used, representing the time step, number of scales, and total number of feature channels, respectively. The three feature channels correspond to the normalized time-scale spectrum matrix, polarization embedding matrix, and frequency drift matrix. The polarization embedding matrix is ​​obtained by time alignment and feature mapping of multiple Stokes components in the polarization state matrix, without increasing the number of channels in the joint feature tensor. A 3D sliding scan window with a spatial size of 16×16×3 is set, and overlapping scans are performed along the time and scale dimensions with a fixed step value of 8. Each extracted 3D sub-tensor is set as an independent local feature block and abstracted as a node to be built in an undirected graph network. For any two extracted nodes i and j, the frequency interval at the center point is calculated. Difference in time evolution The root mean square error of the Stokes parameters was extracted as the polarization parameter difference. The three absolute difference parameters mentioned above are mapped to the [0,1] interval by using the maximum-minimum range standardization method, thereby achieving dimensionless unification of the feature scale.

[0035] Input a preset feature preference weight vector [0.4, 0.3, 0.3], and calculate the weighted Euclidean distance between two feature block nodes according to the normalized results of frequency interval, time evolution difference, and polarization parameter difference. ,in The connection determination threshold is set to 0.25, and the distance is evaluated using a numerical comparator. If the calculated result is less than 0.25, feature block i and feature block j are determined to have spatiotemporal and polarization correlation, and an undirected edge is inserted into the node pair in the memory adjacency list; the reciprocal of the weighted Euclidean distance is extracted, and a protection constant is added. To prevent division by zero anomalies, the edge is assigned as a topological weight, with values ​​ranging from 4 to... Within the interval, complete the creation of the edge set and weight set of the bottom layer of the local topological association structure graph.

[0036] In an optional embodiment, the parameters of the graph convolutional network are generated by the supernetwork based on the previous cycle or an initially preset global signal state, including: Construct a hypernetwork model containing three fully connected layers; The global signal situation vector of the previous cycle is input into the super network model. If it is the initial cycle, the initial preset global signal situation vector is used, and a one-dimensional feature vector is output through the activation function and linear mapping. The one-dimensional feature vector is reshaped into a weight matrix, replacing the aggregation function parameters and node feature transformation matrices in the graph convolutional network.

[0037] The hypernetwork model's network structure includes one input layer, two hidden layers, and one output layer. The input of the hypernetwork model is a global signal situation vector, and the output is a continuous one-dimensional feature vector. The graph convolutional network's network structure includes a node feature linear transformation layer and a neighbor node aggregation layer. The input of the graph convolutional network is a local topological association graph and the feature matrix of each node, and the output is a local situation syntax graph updated with features. Before inputting into the graph convolutional network, each 16×16×3 local feature block is flattened to obtain a 768-dimensional original local feature vector. This vector is then mapped to a 64-dimensional node feature vector through a shared linear projection layer, an average pooling layer, or a one-dimensional convolutional dimensionality reduction layer. This 64-dimensional feature vector serves as the node input feature matrix for the graph convolutional network, thus matching the node feature dimension with the graph convolutional weight matrix generated by the subsequent hypernetwork. A hypernetwork architecture consisting of three multilayer perceptrons is instantiated, with the input layer dimension set to 256, the number of hidden layer neurons set to 512 and 1024 respectively, and the output layer dimension set to 2080. At the start of the first execution cycle of the system, i.e., at T=0, a 256-dimensional tensor consisting of random numbers following a Gaussian distribution with a mean of 0 and a variance of 0.1 is generated as the initial preset global signal situation vector input to the system. In subsequent execution cycles, i.e., T≥1 second, the central unit performs global pooling and linear mapping on the global signal situation map generated in the previous cycle to obtain a 256-dimensional global signal situation vector, and transmits or broadcasts this 256-dimensional global signal situation vector back to or to each detection node as the input data source of the hypernetwork.

[0038] The input vector propagates forward within the supernetwork. A LeakyReLU nonlinear activation function with a negative slope of 0.2 is used to prevent gradient vanishing, and a Dropout rate of 0.3 is injected into the hidden layers for regularization. At the top-level output, no activation function is applied; instead, a pure linear fully connected mapping is used to generate a continuous one-dimensional feature vector of length 2080. Tensor reshaping is performed: the first 2048 values ​​are truncated and reshaped into a principal feature projection matrix with dimensions of 64 rows and 32 columns. The remaining 32 values ​​are truncated to generate a bias constant array of length 32. This generated high-dimensional matrix object is used to replace the original feature linear transformation kernel weight parameters and neighbor node aggregation bias parameters of the target graph convolutional network model, completing the update of the cross-period situational awareness in the network hierarchy. The graph convolutional network outputs a local situational awareness syntax graph composed of 32-dimensional node situational features and their connections.

[0039] In an optional embodiment, the central unit establishes a global network graph based on the geographic and information entropy correlation, using the detection node as the vertex, including: The actual geographical distance between nodes is calculated using the spherical trigonometry formula, and the reciprocal of the distance is taken as the spatial weight. Calculate the KL divergence of the feature distribution of the local situation syntax graph of two nodes, and take the reciprocal of the smoothed KL divergence as the information entropy similarity. The spatial weights and information entropy similarity are weighted and summed after being processed to be dimensionless. If the result exceeds a set threshold, a directed connection is established between the nodes and the corresponding edge weight is assigned.

[0040] The GPS coordinate parameters of each underlying detection node converged to the central unit were extracted and converted into longitude λ and latitude ϕ in radians under the WGS-84 coordinate system. Using the semi-versine spherical trigonometric formula and substituting the Earth's standard radius constant of 6371 km, the great circle distance between any two communication nodes u and v in space was calculated. After adding a displacement compensation constant of 0.1 km, the reciprocal is calculated to obtain the inter-node distance attenuation coefficient, which is the initial spatial weight. Extract the 32-dimensional local situation syntax graph probability distribution feature vectors P and Q from nodes u and v, which are output from the underlying layer and normalized by Softmax exponential mapping. Calculate the KL divergence between their probability distributions element by element. Input a constant in the denominator To prevent the denominator from being zero, a smoothing factor of 0.01 is added to the KL divergence values, and then the reciprocal is used to generate the corresponding information entropy similarity. .

[0041] For all candidate detection node pairs, the corresponding spatial weight set Similarity set with information entropy Normalized to a unified distribution benchmark with a mean of 0 and a standard deviation of 1, and linearly weighted scalar summations are calculated using fusion parameter tuning coefficients of 0.6 and 0.4 respectively. The output hybrid evaluation value serves as the comprehensive connectivity confidence of the nodes. A global threshold judgment logic is input, with a constant of 0.65. If and only if the connectivity confidence from node u to node v is greater than this threshold, a directed structural connection originating from u and terminating at v is appended to the edge set of the global graph storage medium. The fusion connectivity confidence parameter is assigned to the edge data object as a topology transfer weight, resulting in a global directed communication graph containing dual evaluation indicators of geometric distance and high-level semantic overlap.

[0042] The generation module is used to calculate the importance of features within nodes and the strength of collaborative association between nodes based on scale features on the global network graph, fuse them to generate a global signal situation map, locate the target signal source based on the global signal situation map, connect the target signal source to the global network graph based on the positional association between the target signal source and the detection node, calculate the PageRank value to assess the security threat, generate a resource allocation strategy to maximize the suppression effectiveness of the target signal source and distribute it to the prevention and control nodes for execution.

[0043] The node scale features are normalized to obtain the feature importance score within each node. For adjacent nodes with connectivity in the global network graph, the cross-covariance matrix of their corresponding scale features is calculated, and singular value decomposition is performed on the cross-covariance matrix. The ratio of the maximum singular value to the sum of all singular values ​​is extracted as the inter-node cooperative association strength. Using the inter-node cooperative association strength as the edge weight, the local situation features of each detection node in the global network graph are fused using global aggregation pooling to generate a global signal situation map tensor representing the entire network perception. The global signal situation map tensor is forward-inferred to extract interference signal features, and the arrival timestamp of the same target signal is extracted from at least four detection nodes synchronized with a unified clock. The unified clock synchronization is achieved through GNSS time synchronization, PPS synchronization signal, or network time synchronization calibration. Based on the time difference of arrival (TDOA) between multiple detection nodes, a multi-station hyperboloid positioning equation system is constructed, and the three-dimensional spatial coordinates of the target signal source are estimated to achieve signal source positioning.

[0044] The estimated coordinates of the target signal source are registered as a new node. The weights of the target association edges are calculated based on the spatial Euclidean distance between the new node and existing detection nodes. Two-way target association edges are established between detection nodes and the target signal source node that meet a preset distance threshold. The target signal source node is then connected to the global network graph through these two-way target association edges. The PageRank centrality index of each node in the graph is calculated to obtain the target signal source node's score, which is then mapped to a normalized security threat level index. An adversarial optimization model is constructed, using the target signal source's security threat level, target operating frequency band, and node link gain as inputs, and suppression frequency, interference power, and beam pointing as variables to be optimized. The optimal resource allocation strategy is solved in the central unit. This allocation strategy is encapsulated as a lightweight data exchange format JSON command and sent to the corresponding UAV control node. The control node receives the command and calls the hardware driver application programming interface to control the phase shifter weights of the phased array antenna, adjust the power amplifier output level, and complete the electromagnetic suppression action against the target signal source.

[0045] In an optional embodiment, the step of calculating the intra-node feature importance and inter-node collaborative association strength based on scale characteristics on the global network graph includes: One-dimensional convolution and the Softmax function are used to normalize the multi-scale features within the node, and the feature importance weight matrix is ​​output. Calculate the cross-covariance matrix of the corresponding scale features between nodes, perform singular value decomposition on the cross-covariance matrix, extract the ratio of the largest singular value to the sum of all singular values ​​as the cooperative association strength, and remove redundant connections with association strength below a set threshold.

[0046] The data matrix obtained at a single communication node contains 256 depth scales and has a single-scale vector dimension of 64. A one-dimensional temporal convolution operator with a size of 1×3 and 1 channel is set and used to extract local correlations with a stride of 1. Zero padding is applied at both ends to maintain alignment of the scale dimension, and the resulting dimensionality-reduced one-dimensional internal scale response feature sequence is extracted. Apply a temperature-controlled Softmax activation function with a preset annealing temperature coefficient τ=0.5 to the feature sequence. Scaling is performed to transform all scale-based excitation values ​​into probability distribution intervals summing to 1, generating a 256×1 feature importance weight column matrix to allocate feature weights across different resolution scales. The scale-based feature importance distribution is as follows: Figure 2 As shown.

[0047] During cross-node processing, adjacent node pairs u and v with connections are traversed in the global graph. Multi-scale feature sequences of u and v are extracted after importance-weighted modulation. A 64×64 cross-channel cross-covariance matrix is ​​output using a centralized calculation method. Singular values ​​of the cross-covariance matrix are extracted and arranged in descending order of value to obtain a sequence of non-negative real singular values. , ... Perform a division operation to determine the largest singular value. Sum of all singular values The ratio is used to obtain the proportion of singular values. This ratio is used as a representation of the collaborative association strength that defines the data consistency convergence efficiency of nodes in the (0,1] space. When this ratio is less than a set threshold of 0.15, the network connection is removed from the adjacency list, redundant edges are cut off, and a connected backbone network is constructed.

[0048] In an optional embodiment, generating a resource allocation strategy that maximizes the suppression effectiveness against the target signal source and distributing it to the control nodes for execution includes: An adversarial optimization model is established with the objective function of maximizing the suppression effectiveness against the target signal source. Using the maximum transmit power and available frequency bands of each control node as constraints, a mapping relationship between interference suppression parameters and suppression effectiveness is established. The adversarial optimization model is solved using a heuristic algorithm, and the output suppression frequency, interference power, and beam pointing parameters of the target signal source are used as resource allocation strategies.

[0049] An optimization model is constructed for the target signal source node, and the global optimization objective function is set as follows: ,in , This represents the free-space path loss factor after conversion to the linear domain, or the linear propagation loss factor obtained after calibration. To preset a very small positive number, This represents the equivalent noise power obtained by integrating the signal at the target signal receiving location within the target bandwidth. Indicates the target signal received power. This represents the spectral matching factor between the suppression frequency, the target signal's center frequency, and the target's occupied bandwidth, with a value ranging from 0 to 1. When the suppression frequency falls within the target's occupied bandwidth or has a high degree of overlap with the target signal's spectrum... Take the larger value when the suppression frequency deviates from the target occupied bandwidth. The power is reduced to near zero, so that both the denominator and numerator correspond to power dimensions and form a dimensionless suppression efficiency ratio, thereby maximizing the sum of the equivalent suppression efficiency scores projected by the five coordinated control nodes to the target signal source receiving position. Input constraint: Control the allocated transmit power variable of the i-th radio frequency node. The maximum power output should be less than 50W of the transmitter's linear saturation power, and the total power output for suppression should not exceed 200W; the transmission signal frequency should be limited. Within the designated target operating frequency band of 900MHz to 1.5GHz, and prioritizing the target signal's occupied frequency band or its adjacent suppressable frequency band, the beam control angle... The feasible search region is limited to a continuous interval of [0, 360°]; a nonlinear mathematical mapping relationship is constructed between the above three types of parameters and the terminal signal-to-noise ratio attenuation performance, so that the link gain corresponds to the greater the propagation loss. The smaller the value, the less likely the path loss will be mistakenly treated as a positive gain and added to the suppression performance.

[0050] For this constrained system with a high-dimensional non-convex solution space, the data is converted into a real-number encoding format and input into a genetic algorithm for solution. The configuration parameters are set as follows: initial population size of 100, maximum evolutionary iterations of 500 generations, and a fixed crossover operator probability of 0.8. The mutation factor is allowed to expand from 0.01 to an upper limit of 0.1 as the population fitness variance decreases. After reaching the preset convergence accuracy, the optimal solution inverse decoding operation is performed, outputting a specific control configuration list; for example, resolving parameters such as the azimuth angle of antenna #2 being 125°, the center frequency being 1025MHz, and injecting 45W of power. The control form containing all parameters is encapsulated into a standard low-level message and pushed to the distributed control node radio frequency components for execution, completing real-time resource configuration. The resource allocation strategy for the target signal source is as follows: Figure 3 As shown.

[0051] In the second embodiment, the present invention also proposes a multi-network collaborative wireless radio frequency signal full-domain detection and intelligent control method, comprising the following steps: Multiple detection nodes acquire radio frequency signals to obtain I / Q data, perform wavelet transform on the I / Q data to obtain a time-scale spectrum matrix, combine it with polarization data to generate a polarization state matrix, calculate the second-order phase difference to obtain a frequency drift matrix, and then stack the time-scale spectrum matrix, polarization state matrix and frequency drift matrix after dimension alignment to generate a joint feature tensor. The joint feature tensor is segmented into local feature blocks and regarded as nodes. A topology graph is constructed based on time-frequency and polarization similarity. A local situation grammar graph is generated using a graph convolutional network. The parameters of the graph convolutional network are generated by the supernetwork based on the global signal situation of the previous period or the initial preset. Each detection node reports the local situation grammar graph, scale features and location to the central unit. The central unit establishes a global network graph based on the correlation between geography and information entropy, with the detection node as the vertex. On the global network graph, the importance of features within nodes and the strength of collaborative association between nodes are calculated based on scale characteristics, and a global signal situation map is generated by fusion. The target signal source is located based on the global signal situation map. The target signal source is connected to the global network graph based on the positional association between the target signal source and the detection node. The PageRank value is calculated to assess the security threat. A resource allocation strategy that maximizes the suppression effectiveness of the target signal source is generated and distributed to the prevention and control nodes for execution.

[0052] The above description represents the preferred embodiments of the present invention. It should be noted that, for those skilled in the art, various improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A multi-network collaborative wireless radio frequency signal full-domain detection and intelligent control system, characterized in that, include: The calculation module is used to acquire I / Q data by collecting radio frequency signals from multiple detection nodes, perform wavelet transform on the I / Q data to obtain the time-scale spectrum matrix, generate the polarization state matrix by combining polarization data, calculate the second-order phase difference to obtain the frequency drift matrix, and stack the time-scale spectrum matrix, polarization state matrix and frequency drift matrix after dimension alignment processing to generate a joint feature tensor. The construction module is used to segment the joint feature tensor into local feature blocks and treat them as nodes, construct a topology graph based on time-frequency and polarization similarity, and generate a local situation grammar graph using a graph convolutional network. The parameters of the graph convolutional network are generated by the supernetwork based on the global signal situation of the previous period or the initial preset. Each detection node reports the local situation grammar graph, scale features and location to the central unit. The central unit uses the detection node as the vertex and establishes a global network graph based on the correlation between geography and information entropy. The generation module is used to calculate the importance of features within nodes and the strength of collaborative association between nodes based on scale features on the global network graph, fuse them to generate a global signal situation map, locate the target signal source based on the global signal situation map, connect the target signal source to the global network graph based on the positional association between the target signal source and the detection node, calculate the PageRank value to assess the security threat, generate a resource allocation strategy to maximize the suppression effectiveness of the target signal source and distribute it to the prevention and control nodes for execution.

2. The system according to claim 1, characterized in that, The multiple detection nodes acquire radio frequency signals to obtain I / Q data, and perform wavelet transform on the I / Q data to obtain a time-scale spectrum matrix, including: The I / Q data are concatenated into a complex signal, and a complex Morlet wavelet is selected as the transformation basis. A scale sequence containing 256 scales was established to cover the analysis frequency band in a logarithmically equidistant manner; The complex signal is convolved and integrated with the complex Morlet wavelet basis functions corresponding to each scale by a sliding time window, and the absolute modulus of the result is extracted to generate a two-dimensional time-scale spectrum matrix.

3. The system according to claim 2, characterized in that, The step of generating a polarization state matrix by combining polarization data includes: Receives the horizontally polarized signal and the vertically polarized signal output from the dual-polarized antenna; Calculate the instantaneous amplitude, phase difference, and cross-correlation coefficient of the two polarized signals; Based on the calculated parameters, the Stokes parameters are solved, and the polarization state matrix is ​​constructed to identify the instantaneous polarization characteristics of the signal source.

4. The system according to claim 1, characterized in that, The step of segmenting the joint feature tensor into local feature blocks and treating them as nodes, and constructing a topological graph based on time-frequency and polarization similarity, includes: The joint feature tensor is segmented using a fixed-size three-dimensional sliding window to obtain the local feature blocks; After dimensionless processing of frequency interval, time evolution difference and polarization parameter difference, the weighted Euclidean distance between any two feature blocks is calculated. When the weighted Euclidean distance between two feature blocks is less than a preset threshold, a connection edge is established between the corresponding nodes, and a local topology graph is constructed using the reciprocal of the distance as the weight.

5. The system according to claim 1, characterized in that, The parameters of the graph convolutional network are generated by the supernetwork based on the global signal state of the previous cycle or an initial preset, including: Construct a hypernetwork model containing three fully connected layers; The global signal situation vector of the previous cycle is input into the super network model. If it is the initial cycle, the initial preset global signal situation vector is used, and a one-dimensional feature vector is output through the activation function and linear mapping. The one-dimensional feature vector is reshaped into a weight matrix, replacing the aggregation function parameters and node feature transformation matrices in the graph convolutional network.

6. The system according to claim 1, characterized in that, The central unit, with the detection node as the vertex, establishes a global network graph based on the correlation between geography and information entropy, including: The actual geographical distance between nodes is calculated using the spherical trigonometry formula, and the reciprocal of the distance is taken as the spatial weight. Calculate the KL divergence of the feature distribution of the local situation syntax graph of two nodes, and take the reciprocal of the smoothed KL divergence as the information entropy similarity. The spatial weights and information entropy similarity are weighted and summed after being processed to be dimensionless. If the result exceeds a set threshold, a directed connection is established between the nodes and the corresponding edge weight is assigned.

7. The system according to claim 1 or 6, characterized in that, The step of calculating the intra-node feature importance and inter-node collaborative association strength based on scale characteristics on the global network graph includes: One-dimensional convolution and the Softmax function are used to normalize the multi-scale features within the node, and the feature importance weight matrix is ​​output. Calculate the cross-covariance matrix of the corresponding scale features between nodes, perform singular value decomposition on the cross-covariance matrix, extract the ratio of the largest singular value to the sum of all singular values ​​as the cooperative association strength, and remove redundant connections with association strength below a set threshold.

8. The system according to claim 1, characterized in that, The process of generating a resource allocation strategy that maximizes the suppression effectiveness against the target signal source and distributing it to the control nodes for execution includes: An adversarial optimization model is established with the objective function of maximizing the suppression effectiveness against the target signal source. Using the maximum transmit power and available frequency bands of each control node as constraints, a mapping relationship between interference suppression parameters and suppression effectiveness is established. The adversarial optimization model is solved using a heuristic algorithm, and the output suppression frequency, interference power, and beam pointing parameters of the target signal source are used as resource allocation strategies.

9. A multi-network collaborative wireless radio frequency signal full-domain detection and intelligent control method, characterized in that, include: Multiple detection nodes acquire radio frequency signals to obtain I / Q data, perform wavelet transform on the I / Q data to obtain a time-scale spectrum matrix, combine it with polarization data to generate a polarization state matrix, calculate the second-order phase difference to obtain a frequency drift matrix, and then stack the time-scale spectrum matrix, polarization state matrix and frequency drift matrix after dimension alignment to generate a joint feature tensor. The joint feature tensor is segmented into local feature blocks and regarded as nodes. A topology graph is constructed based on time-frequency and polarization similarity. A local situation grammar graph is generated using a graph convolutional network. The parameters of the graph convolutional network are generated by the supernetwork based on the global signal situation of the previous period or the initial preset. Each detection node reports the local situation grammar graph, scale features and location to the central unit. The central unit establishes a global network graph based on the correlation between geography and information entropy, with the detection node as the vertex. On the global network graph, the importance of features within nodes and the strength of collaborative association between nodes are calculated based on scale characteristics, and a global signal situation map is generated by fusion. The target signal source is located based on the global signal situation map. The target signal source is connected to the global network graph based on the positional association between the target signal source and the detection node. The PageRank value is calculated to assess the security threat. A resource allocation strategy that maximizes the suppression effectiveness of the target signal source is generated and distributed to the prevention and control nodes for execution.

10. The method according to claim 9, characterized in that, The multiple detection nodes acquire radio frequency signals to obtain I / Q data, and perform wavelet transform on the I / Q data to obtain a time-scale spectrum matrix, including: The I / Q data are concatenated into a complex signal, and a complex Morlet wavelet is selected as the transformation basis. A scale sequence containing 256 scales was established to cover the analysis frequency band in a logarithmically equidistant manner; The complex signal is convolved and integrated with the complex Morlet wavelet basis functions corresponding to each scale by a sliding time window, and the absolute modulus of the result is extracted to generate a two-dimensional time-scale spectrum matrix.