Time-frequency fusion Hookes dynamic wireless sensor network topology hierarchical inference method
By proposing a time-frequency fusion Hawkes dynamic wireless sensor network topology hierarchical inference method, the bottleneck of topology inference in non-cooperative scenarios of wireless sensor networks is solved, achieving efficient and accurate topology reconstruction and dynamic adaptation, which is suitable for situational awareness of large-scale wireless sensor networks.
Patent Information
- Application Number
- CN202511063900.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-31
- Publication Date
- 2025-11-18
AI Technical Summary
Existing wireless sensor network topology inference technologies suffer from limitations in static modeling, single perception dimension, high computational complexity, and insufficient dynamic adaptability in non-cooperative scenarios, resulting in limited network situational awareness and security vulnerabilities.
A time-frequency fusion Hawkes dynamic wireless sensor network topology hierarchical inference method is adopted. By constructing a dynamic time-frequency feature correlation model, a dynamic causal excitation model, and hierarchical subnet inference, multi-dimensional signal decoupling and dynamic topology reconstruction are achieved, reducing computational complexity and improving adaptability.
It achieves efficient and accurate topology reconstruction in non-cooperative environments, improves link detection accuracy, reduces false positive rate, and has excellent dynamic adaptability and engineering practical value.
Smart Images
Figure CN120979954A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of wireless sensor network topology inference technology, specifically relating to a time-frequency fusion Hawkes dynamic wireless sensor network topology hierarchical inference method, which is applicable to situational awareness of large-scale wireless sensor networks in non-cooperative, highly dynamic environments. Background Technology
[0002] Wireless sensor networks (WSNs) serve as core infrastructure in critical fields such as military reconnaissance and the Industrial Internet of Things (IIoT), and their topology inference technology is directly related to national and industrial security. In non-cooperative scenarios, due to the protection of sensitive information and network security, nodes commonly employ countermeasures such as information concealment, dynamic frequency hopping, and communication encryption. This renders traditional topology inference methods based on protocol handshakes or traffic pattern analysis completely ineffective. This technological bottleneck not only severely restricts network situational awareness capabilities but may also lead to significant vulnerabilities in the security protection of critical infrastructure. Therefore, developing topology inference technology adapted to non-cooperative environments has extremely important strategic significance and application value.
[0003] Existing technologies suffer from four major drawbacks: First, static modeling has limitations; methods based on graph theory or steady-state assumptions cannot adapt to topological abrupt changes caused by node movement or failures. Second, the perception dimension is singular; relying on a single signal feature makes it difficult to distinguish concurrent radiation sources in electromagnetic aliasing environments. Third, computational complexity is high; global topology synchronization inference algorithms are complex and cannot meet the real-time requirements of large-scale networks. Fourth, dynamic adaptability is insufficient; there is a lack of fine-grained tracking mechanisms for topology evolution triggered by local events. Therefore, a novel technology is urgently needed that meets the following core requirements: achieving multi-dimensional signal decoupling to separate the time-frequency fingerprint features of concurrent nodes; establishing dynamic causal modeling to capture the spatiotemporal correlation of events triggered between nodes; adopting a hierarchical computing architecture to reduce the global computational load through subnet collaborative inference; and constructing an incremental update mechanism to respond to node state changes in real time and maintain topology spatiotemporal consistency. Summary of the Invention
[0004] Purpose of the invention: To address the technical bottlenecks of radiation source aliasing, poor dynamic adaptability, and high computational complexity in topology inference of non-cooperative wireless sensor networks, this invention proposes a time-frequency fusion Hawkes dynamic wireless sensor network topology hierarchical inference method to achieve efficient and accurate reconstruction of dynamic topology.
[0005] Technical solution: The time-frequency fusion Hawkes dynamic wireless sensor network topology hierarchical inference method of the present invention specifically includes the following steps:
[0006] S1. Construct a dynamic time-frequency feature correlation model: Convert the radio frequency signals collected by the multi-sensor array into a time-frequency domain feature space, and establish a dynamic mapping relationship between the spatial location of nodes and electromagnetic features;
[0007] S2. Establish a dynamic causal excitation model: Construct event triggering relationships between nodes based on Hawkes processes, quantify spatiotemporal correlations through baseline intensity parameters and dynamic excitation coefficients, and set the decay rate of historical event influences;
[0008] S3, Hierarchical Subnet Reasoning: Initialize the subnet with boundary-aware nodes, calculate the subnet centroid based on the inverse weight of node degree, prioritize expanding the uncovered nodes that are closest to the centroid and have sufficient feature similarity, and control the node overlap rate between subnets to be no less than the preset value.
[0009] S4. Global Topology Fusion: The inference results of each subnet are weighted and fused with confidence, and time-frequency feature backtracking verification is performed on conflicting links to generate a dynamic adjacency matrix with timestamps.
[0010] Furthermore, the implementation process of step S1 is as follows:
[0011] S11: Deploy a symmetrical triangular sensor array to enhance directional resolution through spatial diversity;
[0012] S12: A sliding time window mechanism is used to achieve segmented signal processing;
[0013] S13: Perform Hamming window short-time Fourier transform to obtain spectral characteristics within a certain bandwidth;
[0014] S14: Extract the three-dimensional feature vector containing energy density, dominant frequency offset, and instantaneous bandwidth. Perform a short-time Fourier transform on the monitoring node to obtain the time-frequency matrix ||STFT. s (t,f)||;
[0015] S15: Based on a pre-established node location database, including node coordinates and feature fingerprints, an improved joint decision algorithm is used to detect aliasing signals and estimate the number of sources. First, a bandwidth abrupt change triggering mechanism is designed based on multipath effect theory. When the sensor detects an instantaneous bandwidth B... s A sudden increase, exceeding 80% of the historical average bandwidth of a single source, indicates the presence of signal aliasing. Secondly, the number of sources is determined through dominant frequency clustering, and the time-frequency matrix ||STFT is analyzed. s (t,f)|| Perform frequency domain peak search. If multiple energy peaks exist within a 3Hz bandwidth, it is determined to be multi-source aliasing. The number of radiation sources is equal to the number of independent frequency peaks.
[0016] S16: Separate the different signal sources and reconstruct the signal sequentially by constructing a binary mask matrix:
[0017]
[0018] Signal s iThe signal s is obtained by multiplying the corresponding mask matrix with the short-time leaf transform result. i The short-time Fourier transform results are as follows:
[0019]
[0020] Performing a short-time inverse Fourier transform on the transformed result yields the original radiation source signal after separation and reconstruction.
[0021] Furthermore, step S2 is implemented as follows:
[0022] S21: Initialize the intensity function λ(t) parameters and construct an interactive event model with spatiotemporal decay characteristics:
[0023] λ(t)=μ+∑α·exp(-β(tt i ))
[0024] Where μ is the baseline strength reflecting the inherent interaction tendency of nodes, α is the dynamic excitation coefficient, characterizing the influence of historical events on the current trigger strength, and β is the time decay parameter. The baseline strength parameter μ is initialized to 0.1 times the identity matrix, with a matrix dimension of N×N, where N is the number of nodes. The dynamic excitation coefficient α is initialized to 0.5 times the identity matrix. The initial value of the time decay parameter β is set to 0.8.
[0025] S22: Establish a time decay model, optimize the decay rate through maximum likelihood estimation, and use an improved EM algorithm for parameter estimation; E-step: Based on the current parameter μ k ,α k ,β k Calculate the conditional probability γ ij =α ij ·exp(-β(t j -t i )) / λ(t j M-step: Maximize the log-likelihood function using the gradient ascent method, and update the formula as follows: ε is the learning rate; optimization terminates when the difference in the likelihood function between two consecutive iterations is less than 1e-6.
[0026] S23: Dynamic update mechanism: When a node moves more than σ / 4, where σ is the signal wavelength, it triggers a local parameter update.
[0027] Furthermore, step S3 is implemented as follows:
[0028] S31: The initial node selection of the subnet is to select sensing nodes that satisfy the boundary signal strength difference greater than 3dB;
[0029] S32: Calculate the centroid coordinates of the current subnet based on the centroid advancement algorithm, and prioritize the expansion of the node closest to the centroid among the uncovered nodes with a time-frequency feature similarity greater than 85%;
[0030] S33: Control the subnet overlap rate to be greater than 65%, divide the large-scale network into small-scale networks for decision-making, and ensure that there are enough common nodes between subnets when dividing the network topology;
[0031] S34: Perform time-frequency feature backtracking verification through conflict resolution mechanism to check the consistency of signal features of nodes in overlapping areas; when multiple subnets are found to have conflicting attribution judgments of the same node, the historical time-frequency feature data of the node within the sliding time window will be retrieved; by calculating the time-frequency feature correlation coefficient and setting the threshold to 0.7, the feature matching degree with the core nodes of each candidate subnet is verified to ensure the high consistency between the topology inference results and the time-frequency observation data.
[0032] Furthermore, step S4 is implemented as follows:
[0033] The topology inference results of each subnet are weighted and fused to obtain the global topology. The adjacency matrices of each subnet are linearly combined according to their confidence weights using a weighted voting algorithm to generate a global dynamic adjacency matrix.
[0034] Beneficial Effects: Compared with existing technologies, the beneficial effects of this invention are as follows: This invention achieves higher link detection accuracy and reduces false positive rate through time-frequency feature fusion and Hawkes dynamic modeling; This invention has excellent dynamic adaptability and can achieve incremental updates of local parameters in node movement or fault scenarios; This invention adopts a hierarchical inference strategy to reduce computational complexity and significantly improve computational efficiency; This invention innovatively realizes topology inference in non-cooperative environments, and can be applied to various encrypted frequency hopping communication scenarios without protocol interaction, demonstrating outstanding engineering practical value. Attached Figure Description
[0035] Figure 1 This is a flowchart of the present invention;
[0036] Figure 2 This is a schematic diagram of the simulation node deployment;
[0037] Figure 3 This is a schematic diagram of the time-frequency domain analysis of the signal received by the measurement node;
[0038] Figure 4 This is a schematic diagram of a network adjacency matrix. Detailed Implementation
[0039] The present invention will now be described in further detail with reference to the accompanying drawings.
[0040] like Figure 1As shown, this invention proposes a time-frequency fusion Hawkes dynamic wireless sensor network topology hierarchical inference method, the specific implementation process of which is as follows:
[0041] S1, Construct a dynamic time-frequency feature extraction model.
[0042] The radio frequency signals acquired by the multi-sensor array are converted into a time-frequency domain feature space, and a dynamic mapping relationship between the spatial location of nodes and electromagnetic features is established.
[0043] S11: Deploy a symmetrical triangular sensor array, such as Figure 2 As shown (sensor A / B / C layout), spatial diversity enhances directional resolution.
[0044] S12: A sliding time window mechanism (100ms window length / 50ms step size) is used to implement signal segmentation processing, as shown in the attached diagram. Figure 2 The horizontal axis shows the time scale.
[0045] S13: Perform a 256-point Hamming window short-time Fourier transform to obtain the spectral characteristics within an 800Hz bandwidth, corresponding to... Figure 3 Vertical axis frequency resolution.
[0046] S14: Extract the three-dimensional feature vector and perform a short-time Fourier transform on the monitoring nodes to obtain the time-frequency matrix ||STFT. s (t,f)||, including: energy density: average energy of the time-domain signal, calculate the energy value (modulus squared of the complex number) at each time frequency point (t,f), sum the energy values at all time points t, and then divide by the time window length T to obtain the energy density;
[0047]
[0048] Main frequency offset: such as Figure 3 Find the frequency peak location of the -50dB / Hz high intensity region in the direction of Node2; for each frequency f, accumulate the time-frequency amplitude along the time axis t, and find the frequency point with the largest accumulated amplitude.
[0049]
[0050] Instantaneous bandwidth: Based on the calculation of the second-order moment of the spectrum (characterizing the signal modulation characteristics), for each frequency f, calculate its bandwidth relative to the dominant frequency f. p,s The square root of the deviation is normalized to obtain the instantaneous bandwidth.
[0051]
[0052] S15: Based on a pre-established node location database (containing the coordinates and feature fingerprints of each node), an improved joint decision algorithm is first used to detect aliasing signals and estimate the number of sources. After determining that the signal received by the monitoring node is an aliasing signal and the number of radiation sources, the aliasing signal is separated by combining energy density, dominant frequency offset, and instantaneous bandwidth indicators: when the dominant frequency difference detected by observation nodes A / B / C is less than 3Hz and the bandwidth ratio is within the range of [0.8, 1.2], such as Figure 3 Independent signal sources from different directions are identified as signals from the same radiation source. For concurrent signals with overlapping main frequencies but bandwidth differences greater than 25%, the time-frequency covariance matrix between sensors is constructed, and the aliasing components are separated by eigenvalue decomposition. Finally, a mapping relationship database is established to achieve high-precision positioning of the radiation source.
[0053] S16: Next, the signals from the separated signal sources are reconstructed sequentially by constructing a binary mask matrix:
[0054]
[0055] Signal s i The signal s is obtained by multiplying the corresponding mask matrix with the short-time leaf transform result. i The short-time Fourier transform results are as follows:
[0056]
[0057] Performing a short-time inverse Fourier transform on the above changes yields the original radiation source signal after separation and reconstruction.
[0058] S2, Hawkes Dynamic Excitation Causal Modeling: Using the 3D feature vector generated in S1 as input, a spatiotemporal correlation model of event triggering between nodes is constructed. The event triggering relationship between nodes is built based on the Hawkes process, and the spatiotemporal correlation is quantified through baseline intensity parameters and dynamic excitation coefficients, while setting the decay rate of historical event influence.
[0059] S21: Initialize the intensity function λ(t) parameters; construct an interactive event model with spatiotemporal decay characteristics:
[0060] λ(t)=μ+∑α·exp(-β(tt i ))
[0061] Among them, the baseline intensity parameter μ is initialized to 0.1 times the identity matrix, reflecting the inherent interaction tendency of the nodes, with a matrix dimension of N×N (N is the number of nodes); the dynamic excitation coefficient α is initialized to 0.5 times the identity matrix, representing the influence of historical events on the current trigger intensity; the time decay parameter β is initially set to 0.8, and subsequently optimized through maximum likelihood estimation.
[0062] S22: Establish a time decay model, optimize the decay rate through maximum likelihood estimation, and use an improved EM algorithm for parameter estimation. E-step: Calculate the conditional probability γ of the communication relationship between sensor nodes i and j. k =α k ·exp(-β k (t j -t i )) / (μ+∑α k ·exp(-β k (tt i ))), where μ k ,α k ,β k These represent the specific values of the baseline intensity parameter μ, dynamic excitation coefficient α, and time decay parameter β in the k-th optimization step; M-step: Maximize the log-likelihood function using the gradient ascent method, with the update formula as follows: The learning rate ε = 0.01; the optimization is terminated when the difference in the likelihood function between two consecutive iterations is less than 1e-6.
[0063] S23: Dynamic update mechanism: When a node moves more than σ / 4 (σ is the signal wavelength), a local parameter update is triggered.
[0064] S3, Hierarchical Subnet Topology Inference: Based on the causal probability matrix of S2, topology hierarchical resolution is achieved through progressive spatial expansion.
[0065] S31: Initialize the boundary node, where the signal strength difference is greater than 3dB.
[0066] S32: The centroid advancement algorithm calculates the centroid coordinates of the current subnet and prioritizes expanding the node closest to the centroid among the uncovered nodes with a time-frequency feature similarity greater than 85%. This expansion strategy based on dual constraints of spatial location and signal features can effectively maintain the topological coherence and signal feature consistency of nodes within the subnet, thereby improving the accuracy of network topology inference.
[0067] S33: Control subnet overlap rate greater than 65%, divide large-scale networks into subnets such as... Figure 2 The small-scale network shown is used for decision-making. During network topology partitioning, sufficient common nodes exist between subnets to provide a crucial data association foundation for subsequent global topology fusion, while avoiding topology fragmentation due to excessively low overlap. This threshold setting ensures both the independence of subnet partitioning (each subnet contains at least 35% unique nodes) and the necessary continuity of the topology structure (reliable connections between subnets are achieved through 65% shared nodes), thus achieving an optimal balance between topology inference accuracy and computational complexity.
[0068] S34: The conflict resolution mechanism performs time-frequency feature backtracking verification to check the consistency of signal features of nodes in overlapping areas. Specifically, when multiple subnets are found to have conflicting classifications of the same node, the system retrieves the historical time-frequency feature data of that node within the sliding time window (including three-dimensional feature vectors such as energy density, dominant frequency offset, and instantaneous bandwidth). By calculating the correlation coefficient (with a threshold set to be greater than 0.7), the system verifies the feature matching degree between the node and the core nodes of each candidate subnet, thereby ensuring a high degree of consistency between the topology inference results and the time-frequency observation data, and effectively solving the subnet division conflict problem caused by electromagnetic environmental interference or node movement.
[0069] S4, Global Topology Fusion Verification. The inference results of each subnet are weighted and fused based on confidence levels. Time-frequency feature backtracking verification is performed on conflicting links to generate a dynamic adjacency matrix with timestamps.
[0070] S41: Output a visualized topology map, as shown below. Figure 4 As shown, the connectivity between node 1 and nodes 2, 3, and 4 was successfully deduced in each successfully partitioned subnet.
[0071] S42: The topology inference results of each subnet are weighted and fused to obtain the global topology. A weighted voting algorithm is then used to linearly combine the adjacency matrices of each subnet according to their confidence weights, generating a global dynamic adjacency matrix. This fusion method effectively suppresses misjudgment interference from individual subnets, improving the overall topology inference accuracy in experimental data while meeting real-time requirements.
[0072] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.
Claims
1. A time-frequency fusion Hawkes dynamic wireless sensor network topology hierarchical inference method, characterized in that, Includes the following steps: S1. Construct a dynamic time-frequency feature correlation model: Convert the radio frequency signals collected by the multi-sensor array into a time-frequency domain feature space, and establish a dynamic mapping relationship between the spatial location of nodes and electromagnetic features; S2. Establish a dynamic causal excitation model: Construct event triggering relationships between nodes based on Hawkes processes, quantify spatiotemporal correlations through baseline intensity parameters and dynamic excitation coefficients, and set the decay rate of historical event influences; S3, Hierarchical Subnet Reasoning: Initialize the subnet with boundary-aware nodes, calculate the subnet centroid based on the inverse weight of node degree, prioritize expanding the uncovered nodes that are closest to the centroid and have sufficient feature similarity, and control the node overlap rate between subnets to be no less than the preset value. S4. Global Topology Fusion: The inference results of each subnet are weighted and fused with confidence, and time-frequency feature backtracking verification is performed on conflicting links to generate a dynamic adjacency matrix with timestamps.
2. The time-frequency fusion Hawkes dynamic wireless sensor network topology hierarchical inference method according to claim 1, characterized in that, The implementation process of step S1 is as follows: S11: Deploy a symmetrical triangular sensor array to enhance directional resolution through spatial diversity; S12: A sliding time window mechanism is used to achieve segmented signal processing; S13: Perform Hamming window short-time Fourier transform to obtain spectral characteristics within a certain bandwidth; S14: Extract the three-dimensional feature vector containing energy density, dominant frequency offset, and instantaneous bandwidth. Perform a short-time Fourier transform on the monitoring node to obtain the time-frequency matrix ||STFT. s (t,f)||; S15: Based on a pre-established node location database, an improved joint decision algorithm is used to detect aliased signals and estimate the number of sources. S16: Separate the different signal sources and reconstruct the signal sequentially by constructing a binary mask matrix: Signal s i The signal s is obtained by multiplying the corresponding mask matrix with the short-time leaf transform result. i The short-time Fourier transform results are as follows: Performing a short-time inverse Fourier transform on the transformed result yields the original radiation source signal after separation and reconstruction.
3. The time-frequency fusion Hawkes dynamic wireless sensor network topology hierarchical inference method according to claim 2, characterized in that, The node location database includes the coordinates and feature fingerprints of each node.
4. The time-frequency fusion Hawkes dynamic wireless sensor network topology hierarchical inference method according to claim 2, characterized in that, The implementation process of S15 is as follows: First, a bandwidth mutation triggering mechanism is designed based on multipath effect theory. When the sensor detects an instantaneous bandwidth B... s A sudden increase, exceeding 80% of the historical average bandwidth of a single source, indicates the presence of signal aliasing. Secondly, the number of sources is determined through dominant frequency clustering, and the time-frequency matrix ||STFT is analyzed. s (t,f)|| Perform frequency domain peak search. If multiple energy peaks exist within a 3Hz bandwidth, it is determined to be multi-source aliasing, and the number of radiation sources is equal to the number of independent frequency peaks.
5. The time-frequency fusion Hawkes dynamic wireless sensor network topology hierarchical inference method according to claim 1, characterized in that, The implementation process of step S2 is as follows: S21: Initialize the intensity function λ(t) parameters and construct an interactive event model with spatiotemporal decay characteristics: λ(t)=μ+∑α·exp(-β(tt i )) Where μ is the baseline strength reflecting the inherent interaction tendency of the node, α is the dynamic excitation coefficient, characterizing the influence of historical events on the current trigger strength, and β is the time decay parameter; S22: Establish a time decay model, optimize the decay rate through maximum likelihood estimation, and use an improved EM algorithm for parameter estimation; E-step: Based on the current parameter μ k ,α k ,β k Calculate the conditional probability γ ij =α ij ·exp(-β(t j -t i )) / λ(t j M-step: Maximize the log-likelihood function using the gradient ascent method, and update the formula as follows: ε is the learning rate; optimization terminates when the difference in the likelihood function between two consecutive iterations is less than 1e-6. S23: Dynamic update mechanism: When a node moves more than σ / 4, where σ is the signal wavelength, it triggers a local parameter update.
6. The time-frequency fusion Hawkes dynamic wireless sensor network topology hierarchical inference method according to claim 5, characterized in that, The baseline intensity parameter μ is initialized to a 0.1 times identity matrix with a matrix dimension of N×N, where N is the number of nodes.
7. The time-frequency fusion Hawkes dynamic wireless sensor network topology hierarchical inference method according to claim 5, characterized in that, The dynamic excitation coefficient α is initialized to 0.5 times the identity matrix; The initial value of the time decay parameter β is set to 0.
8.
8. The time-frequency fusion Hawkes dynamic wireless sensor network topology hierarchical inference method according to claim 1, characterized in that, The implementation process of step S3 is as follows: S31: The initial node selection of the subnet is to select sensing nodes that satisfy the boundary signal strength difference greater than 3dB; S32: Calculate the centroid coordinates of the current subnet based on the centroid advancement algorithm, and prioritize the expansion of the node closest to the centroid among the uncovered nodes with a time-frequency feature similarity greater than 85%; S33: Control the subnet overlap rate to be greater than 65%, divide the large-scale network into small-scale networks for decision-making, and ensure that there are enough common nodes between subnets when dividing the network topology; S34: Perform time-frequency feature backtracking verification through conflict resolution mechanism to check the consistency of signal features of nodes in overlapping areas; when multiple subnets are found to have conflicting attribution judgments of the same node, the historical time-frequency feature data of the node within the sliding time window will be retrieved; by calculating the time-frequency feature correlation coefficient and setting the corresponding threshold, the feature matching degree with the core nodes of each candidate subnet is verified to ensure the high consistency between the topology inference results and the time-frequency observation data.
9. The time-frequency fusion Hawkes dynamic wireless sensor network topology hierarchical inference method according to claim 8, characterized in that, The threshold in step S34 is 0.
7.
10. The time-frequency fusion Hawkes dynamic wireless sensor network topology hierarchical inference method according to claim 1, characterized in that, The implementation process of step S4 is as follows: The topology inference results of each subnet are weighted and fused to obtain the global topology. The adjacency matrices of each subnet are linearly combined according to their confidence weights using a weighted voting algorithm to generate a global dynamic adjacency matrix.