Regional electromagnetic inspection method and system based on vehicle-mounted electromagnetic detection equipment

Abstract

The invention provides a regional electromagnetic patrol method and system based on vehicle-mounted electromagnetic detection equipment, and the method comprises the following steps: recognizing a target monitoring node based on historical data and an electromagnetic propagation model, and constructing an electromagnetic environment diagram network; a network is divided into sub-graphs through multi-scale clustering, a vehicle-mounted electromagnetic inspection vehicle and an edge server are deployed in each sub-graph area, and a central server is deployed in a central area. And the vehicle-mounted equipment performs electromagnetic detection according to the optimized path and acquires data. And the edge server updates the sub-graph by using edge calculation and federated learning technologies, and performs local anomaly analysis. And the central server aggregates all sub-graph update information to complete global update and anomaly analysis. And finally, local and global analysis results are fused to form a comprehensive electromagnetic anomaly analysis result, and the comprehensive electromagnetic anomaly analysis result is visually displayed on a regional map. According to the invention, real-time and accurate large-range electromagnetic environment monitoring and analysis are realized, and the electromagnetic management capability is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of electromagnetic inspection, and in particular relates to a regional electromagnetic inspection method and system based on vehicle-mounted electromagnetic detection equipment. Background Art

[0002] With the rapid development of the information society, electromagnetic environment management faces unprecedented challenges and opportunities. On the one hand, the widespread application of wireless communication technologies, such as mobile communications, wireless networks, and the Internet of Things, has made the electromagnetic environment increasingly complex. The popularity of various electronic devices and wireless devices has led to a surge in the number of electromagnetic interference sources and a diversification of interference types, which has brought tremendous pressure to the monitoring and management of the electromagnetic environment. On the other hand, the development of emerging technologies such as 5G, smart cities, and autonomous driving has put forward higher requirements on the quality of the electromagnetic environment. The normal operation of these technologies requires the support of a stable and reliable electromagnetic environment, and any interference may lead to serious consequences. At the same time, the public's attention to the health effects of electromagnetic radiation has continued to increase, and higher expectations have been put forward for the transparency and effectiveness of electromagnetic environment management.

[0003] In modern urban environments, there are a large number of electromagnetic interference sources, including various communication equipment, power facilities, industrial equipment, etc. The location, intensity and characteristics of these interference sources may change at any time, forming an extremely complex electromagnetic scene. Therefore, the traditional fixed monitoring point layout is difficult to capture this dynamic change, resulting in the monitoring data being difficult to truly reflect the current electromagnetic environment. In addition, changes in the electromagnetic environment often have temporal and spatial correlations, and monitoring of a single point or local area is difficult to reveal the changing trends and potential problems of the overall electromagnetic environment. Traditional methods lack the ability to conduct systematic and comprehensive analysis of large areas, and it is difficult to accurately identify and locate the source of abnormal electromagnetic events. The above problems seriously restrict the efficiency and accuracy of electromagnetic environment management, and cannot meet the high requirements of modern society for electromagnetic environment safety and quality. New technical solutions are urgently needed to solve this challenge. Summary of the invention

[0004] The present invention provides a regional electromagnetic inspection method and system based on a vehicle-mounted electromagnetic detection device to solve the technical problems mentioned in the above background technology.

[0005] In a first aspect, the present invention provides a method for regional electromagnetic inspection based on a vehicle-mounted electromagnetic detection device, the method comprising the following steps:

[0006] Obtain regional maps and historical electromagnetic survey data for target areas;

[0007] Based on the historical electromagnetic detection data and according to the electromagnetic propagation model, identifying historical electromagnetic radiation points as target monitoring nodes on the regional map;

[0008] Using the road network in the regional map as the node edge, and combining all the target monitoring nodes to construct an electromagnetic environment map network of the target area;

[0009] Using a hierarchical clustering algorithm to perform multi-scale clustering on the target monitoring nodes in the electromagnetic environment map network, and dividing the electromagnetic environment map network into multiple regional sub-maps based on the preliminary clustering results of the multi-scale clustering and using a spectral clustering algorithm;

[0010] Deploy an on-board electromagnetic inspection vehicle equipped with an edge server in the area where each of the regional sub-maps is located, and deploy a central server in the central area of ​​the target area, import the regional sub-maps into the corresponding edge servers, import the electromagnetic environment map network into the central server, and all the edge servers are communicatively connected with the central server;

[0011] For any of the regional sub-graphs, the optimal electromagnetic inspection path of the vehicle-mounted electromagnetic inspection vehicle in the region to which the regional sub-graph belongs is intelligently planned based on the regional sub-graph, and the vehicle-mounted electromagnetic inspection vehicle is controlled to perform electromagnetic inspection along the optimal electromagnetic inspection path, and electromagnetic inspection data of the region to which the regional sub-graph belongs is collected;

[0012] Using edge computing and federated learning technology and based on the electromagnetic inspection data, the regional sub-map is updated in the edge server, and local electromagnetic anomaly analysis is performed on the updated regional sub-map by the edge server to obtain a local anomaly analysis result;

[0013] When all the regional sub-maps are updated, the update information of all the regional sub-maps is aggregated through the central server and the global update of the electromagnetic environment map network is completed in the central server, and the global electromagnetic anomaly analysis is performed on the updated electromagnetic environment map network through the central server to obtain a global anomaly analysis result;

[0014] The local anomaly analysis result and the global anomaly analysis result are merged into a comprehensive electromagnetic anomaly analysis result of the target area, and the comprehensive electromagnetic anomaly analysis result is imported into the regional map for visual display.

[0015] Optionally, the identifying historical electromagnetic radiation points as target monitoring nodes on the regional map based on the historical electromagnetic detection data and according to an electromagnetic propagation model comprises the following steps:

[0016] Using a free space propagation model as an electromagnetic propagation model for the target area;

[0017] Based on the electromagnetic propagation model, the radiation points of the target area are reversely estimated according to the historical electromagnetic detection data and the least square method;

[0018] In the reverse calculation process, the objective function of the least square method is constructed by combining the historical electromagnetic detection data and the output value of the electromagnetic propagation model;

[0019] The reverse calculation is completed by solving the objective function through an optimization algorithm, and all potential historical electromagnetic radiation points are marked on the regional map;

[0020] Clustering all the potential historical electromagnetic radiation points using a density clustering algorithm, and obtaining all the historical electromagnetic radiation points in the regional map after clustering;

[0021] Extracting radiation point features of each of the historical electromagnetic radiation points from the historical electromagnetic detection data, wherein the radiation point features include the historical electromagnetic field intensity, the historical abnormal frequency and the spatial impact range of the historical electromagnetic radiation point;

[0022] The historical electromagnetic radiation point is used as a target monitoring node, and the node weight of the target monitoring node is calculated according to the corresponding radiation point characteristics.

[0023] Optionally, the step of performing multi-scale clustering on the target monitoring nodes in the electromagnetic environment map network by using a hierarchical clustering algorithm, and dividing the electromagnetic environment map network into a plurality of regional sub-maps by using a spectral clustering algorithm based on the preliminary clustering results of the multi-scale clustering comprises the following steps:

[0024] Constructing a multidimensional feature vector of the target monitoring node according to the radiation point features;

[0025] Using principal component analysis to reduce the multidimensional feature vector to a low-dimensional feature vector;

[0026] Clustering the low-dimensional feature vector using a hierarchical clustering algorithm based on a minimum spanning tree to obtain a feature clustering tree;

[0027] Calculating a plurality of different cutoff thresholds according to the average node distance between the root node and the leaf nodes in the feature clustering tree;

[0028] Truncating the feature clustering tree based on each of the truncation thresholds to obtain feature clustering results of multiple different scales;

[0029] Constructing an affinity matrix based on the feature clustering results at multiple different scales;

[0030] The affinity matrix is ​​used as a constraint condition of a spectral clustering algorithm, and the electromagnetic environment map network is divided into a plurality of regional sub-maps based on the constraint condition and by using the spectral clustering algorithm.

[0031] Optionally, the step of using the affinity matrix as a constraint of a spectral clustering algorithm and dividing the electromagnetic environment map network into a plurality of regional sub-maps based on the constraint and using the spectral clustering algorithm comprises the following steps:

[0032] Calculate and obtain the degree matrix of the electromagnetic environment map network, and calculate and obtain the normalized Laplace matrix of the electromagnetic environment map network according to the degree matrix;

[0033] Calculating and obtaining the matrix eigenvalues ​​and matrix eigenvectors of the normalized Laplace matrix;

[0034] Select the matrix eigenvectors corresponding to the k smallest non-zero matrix eigenvalues ​​to form a spectral clustering matrix;

[0035] Construct the target optimization function of the spectral clustering algorithm, the target optimization function is:

[0036] min tr(Y T LY)+λ[tr(Y T (DA)Y)]

[0037] Wherein: tr represents the trace of the matrix, Y represents the spectral clustering matrix, L represents the normalized Laplace matrix, λ represents the balance parameter, T represents the transposed matrix, D represents the degree matrix, and A represents the affinity matrix;

[0038] Solving the target optimization function by generalized eigenvalues, and forming a subgraph matrix with K target matrix eigenvectors obtained by solving;

[0039] The row vectors of the sub-graph matrix are clustered using a K-means algorithm, and the electromagnetic environment graph network is divided into a plurality of regional sub-graphs according to the clustering result.

[0040] Optionally, the intelligent planning of the optimal electromagnetic inspection path of the vehicle-mounted electromagnetic inspection vehicle in the area in which it belongs based on the area sub-map includes the following steps:

[0041] Acquire historical traffic flow data of the road network in the area to which the regional subgraph belongs, and assign node edge weights to the node edges of the regional subgraph according to the historical traffic flow data;

[0042] Taking the target monitoring node with the highest node weight as the starting node;

[0043] An optimal path selection step is performed based on the starting node, and the optimal path selection step is as follows:

[0044] Calculate the node distances between the starting node and all adjacent nodes;

[0045] Constructing a multi-objective reward function by combining the node distance, the node weights of the adjacent nodes, and the node edge weights of the node edges between the starting node and the adjacent nodes;

[0046] Based on the multi-objective reward function and using a Monte Carlo tree search method, an optimal adjacent node is selected from all the adjacent nodes as an optimal path node;

[0047] The optimal path selection step is repeatedly performed with the optimal path node as the starting node until the farthest node distance between the optimal path node and the starting node exceeds a preset distance threshold, thereby obtaining the optimal electromagnetic inspection path of the vehicle-mounted inspection vehicle.

[0048] Optionally, controlling the vehicle-mounted electromagnetic inspection vehicle to perform electromagnetic inspection according to the optimal electromagnetic inspection path and collecting electromagnetic inspection data of the area to which the area sub-map belongs includes the following steps:

[0049] Before taking the optimal path node as the starting node in each round of the optimal path selection step, obtaining real-time traffic flow data of the road network in the area to which the regional subgraph belongs, and updating the edge weights of all the nodes in the regional subgraph according to the real-time traffic flow data;

[0050] After the optimal path node is determined in each round in the optimal path selection step, the vehicle-mounted electromagnetic inspection vehicle is controlled to travel from the starting node to the optimal path node, and electromagnetic inspection data around the vehicle-mounted electromagnetic inspection vehicle is continuously collected during the driving process.

[0051] Optionally, the step of updating the regional submap in the edge server based on the electromagnetic inspection data by using edge computing and federated learning technology, and performing local electromagnetic anomaly analysis on the updated regional submap by the edge server to obtain the local anomaly analysis result comprises the following steps:

[0052] Deploy the same graph neural network model on all of the edge servers;

[0053] For any of the edge servers, the preprocessed electromagnetic inspection data is input into the graph neural network model, the regional subgraph on the edge server is updated using a graph structure learning method, and the model parameters of the graph neural network model are updated synchronously;

[0054] Using a federated averaging algorithm to exchange the updated model parameters with all other edge servers;

[0055] Update the local graph neural network model based on the exchanged model parameters;

[0056] The updated multi-scale anomaly score of the regional subgraph is calculated by using the multi-layer graph attention network structure in the graph neural network model, and the local electromagnetic anomaly analysis result of the regional subgraph is calculated by using an integrated learning method.

[0057] Optionally, aggregating the update information of all the regional sub-maps through the central server and completing the global update of the electromagnetic environment map network in the central server, and performing global electromagnetic anomaly analysis on the updated electromagnetic environment map network through the central server to obtain the global anomaly analysis result includes the following steps:

[0058] Aggregating update information of all the regional sub-graphs from all the edge servers through the central server and using an incremental update strategy;

[0059] Based on all the update information and using a graph difference algorithm, a global update of the electromagnetic environment map network is completed in the central server;

[0060] Extracting abnormal features of the time dimension from the updated electromagnetic environment map network using an autoregressive integrated moving average method;

[0061] Extracting spatial dimension abnormal features from the updated electromagnetic environment map network using a spatial autocorrelation analysis method;

[0062] The spatial dimension abnormality features and the temporal dimension abnormality features are combined and an integrated learning method is used to analyze and obtain a global abnormality analysis result.

[0063] Optionally, after fusing the local anomaly analysis result and the global anomaly analysis result into the comprehensive electromagnetic anomaly analysis result of the target area, the following steps are further included:

[0064] Build a spatiotemporal graph convolutional network model;

[0065] Inputting the electromagnetic environment graph network and the comprehensive electromagnetic anomaly analysis results as input data into the spatiotemporal graph convolutional network model;

[0066] Using the spatiotemporal graph convolutional network model to predict the spatiotemporal evolution trend of electromagnetic anomalies in the target area, and outputting the predicted spatiotemporal evolution results of electromagnetic anomalies;

[0067] The graph network structure of the electromagnetic environment graph network and the node weights of the target monitoring nodes are adjusted according to the spatiotemporal evolution results of the electromagnetic anomaly.

[0068] In a second aspect, the present invention also provides a regional electromagnetic patrol system based on vehicle-mounted electromagnetic detection equipment, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the computer program, the regional electromagnetic patrol method based on vehicle-mounted electromagnetic detection equipment as described in the first aspect is implemented.

[0069] The beneficial effects of the present invention are:

[0070] The present invention realizes the refined division and management of large-scale complex environments by constructing an electromagnetic environment graph network and combining a multi-scale clustering algorithm, overcoming the limitation that the traditional fixed monitoring point layout is difficult to adapt to the dynamic electromagnetic environment. The on-board electromagnetic inspection vehicle is used for flexible and mobile monitoring, which greatly improves the coverage and real-time performance of monitoring and can capture the instantaneous changes of the electromagnetic environment in a timely manner. The present invention innovatively introduces edge computing and federated learning technologies to realize real-time processing and analysis of data on the edge server, significantly improving the response speed and efficiency of the system. Through the global analysis of the central server, the present invention realizes the systematic cognition of the overall electromagnetic environment and overcomes the problem of fragmented data analysis in traditional methods. In addition, the multi-level anomaly analysis mechanism of the present invention, combined with local and global analysis results, greatly improves the recognition accuracy and positioning accuracy of electromagnetic anomaly events. The present invention also uses visualization display technology to intuitively present the electromagnetic environment status, which is convenient for managers to make quick decisions. In general, the present invention constructs an intelligent and adaptive electromagnetic environment monitoring and management system, which can not only cope with complex and changeable electromagnetic environments, but also automatically adjust monitoring strategies according to environmental changes, greatly improving the efficiency and accuracy of electromagnetic environment management. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a flowchart of a regional electromagnetic inspection method based on a vehicle-mounted electromagnetic detection device in one embodiment of the present application.

[0072] Figure 2 This is a schematic diagram of the system architecture of a frequency selector for a vehicle-mounted electromagnetic detection device in one embodiment of the present application.

[0073] Figure 3 This is a schematic diagram of the system architecture of PC data collection software in one embodiment of the present application.

[0074] Figure 4 This is a visual display diagram of a path in one embodiment of the present application that combines the inspection path, inspection results and area map.

[0075] Figure 5 This is a grid-format visualization diagram combining inspection results and area maps in one embodiment of the present application. DETAILED DESCRIPTION

[0076] The following will be combined with the drawings in the embodiments of the present application to clearly describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments in the present application belong to the scope of protection of this application.

[0077] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first", "second", etc. are generally of one type, and the number of objects is not limited. For example, the first object can be one or more. In addition, "and / or" in the specification and claims represents at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0078] Figure 1 FIG. 1 is a flow chart of a method for regional electromagnetic inspection based on a vehicle-mounted electromagnetic detection device in one embodiment. It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps in the above method may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps. Figure 1 As shown, the regional electromagnetic inspection method based on the vehicle-mounted electromagnetic detection equipment disclosed in the present invention specifically includes the following steps:

[0079] S101. Obtain a regional map and historical electromagnetic detection data of the target area.

[0080] Among them, regional maps are usually obtained in digital format, including two main forms: vector maps and raster maps. Vector maps use geometric elements such as points, lines, and surfaces to represent geographic entities, which are suitable for expressing linear features such as road networks; raster maps divide the area into regular grids, each grid stores corresponding geographic information, and are suitable for expressing continuously changing geographical phenomena. Map data can be obtained from a geographic information system (GIS) database or remote sensing images, and usually includes information such as topography, landforms, land use, and building distribution. The resolution and accuracy of map data directly affect the accuracy of subsequent analysis, so it is necessary to select an appropriate spatial resolution, usually between 1 and 10 meters.

[0081] Historical electromagnetic detection data includes electromagnetic field strength measurements made in the target area over a period of time in the past. These data usually include information such as measurement time, geographic location (latitude and longitude coordinates), frequency range, electric field strength, magnetic field strength, etc. The data may come from fixed monitoring stations, mobile measurement equipment, or satellite remote sensing. The data format may include CSV, JSON, or proprietary formats, which need to be standardized. In order to ensure the representativeness and reliability of the data, the data needs to be preprocessed, including outlier detection, missing value processing, and spatiotemporal consistency checks. The processed data needs to be spatially aligned with the map to ensure that the electromagnetic detection data accurately corresponds to the geographic location. This can be achieved through coordinate transformation and projection alignment. Commonly used projection systems include UTM and Gauss-Krüger projections. Finally, the processed data is stored in a spatial database such as PostGIS or SpatiaLite for subsequent rapid retrieval and analysis.

[0082] S102. Based on historical electromagnetic detection data and in accordance with an electromagnetic propagation model, historical electromagnetic radiation points are identified on a regional map as target monitoring nodes.

[0083] Among them, first of all, it is necessary to select a suitable electromagnetic propagation model to simulate the propagation characteristics of electromagnetic waves in space. Commonly used models include free space propagation model, Okumura-Hata model and COST-231 model. Taking the free space propagation model as an example, its path loss can be expressed as: L = 32.45 + 20log (f) + 20log (d), where f is the frequency (MHz) and d is the distance (km). For complex terrain, ray tracing or finite difference time domain method (FDTD) can be used for more accurate simulation. When selecting the model, factors such as the terrain characteristics of the area, building density and electromagnetic wave frequency need to be considered. Next, the selected model is used to reversely calculate the historical electromagnetic detection data to determine the possible location of the radiation source. This can be achieved by least squares method or maximum likelihood estimation. By solving this minimization problem through an optimization algorithm (such as Levenberg-Marquardt algorithm), the most likely location and intensity of the radiation source can be obtained. For the identified radiation points, cluster analysis is required to remove redundant points. Specifically, the density-based spatial clustering algorithm DBSCAN can be used: for each point p, if there are at least MinPts points in its ε neighborhood, a cluster is formed. The center point after clustering is used as the final target monitoring node. Finally, the identified target monitoring nodes are superimposed on the regional map to form a preliminary electromagnetic environment monitoring network.

[0084] S103. Using the road network in the regional map as node edges, and combining all target monitoring nodes, construct an electromagnetic environment map network of the target area.

[0085] First, it is necessary to extract road network information from the regional map. This can be obtained directly from vector map data or extracted from raster maps using image processing technology. For raster maps, edge detection algorithms (such as the Canny algorithm) and morphological operations can be used to extract road lines. The main steps of the Canny algorithm include: Gaussian filtering, calculating gradients, non-maximum suppression, and double threshold detection. The extracted road network needs to be topologically processed to ensure the connectivity and integrity of the roads. This can be achieved through topological cleaning algorithms, such as the Douglas-Peucker algorithm for simplifying road lines and the Snap algorithm for repairing broken connections. Next, the road network is converted into a graph structure, where road intersections and endpoints are used as nodes of the graph and road segments are used as edges of the graph. Then, the target monitoring node (i.e., the identified historical electromagnetic radiation point) is added to the graph. For monitoring nodes that are not on the road, the nearest road point needs to be found and connected. This can be achieved through a nearest neighbor search algorithm, such as a KD tree: a KD tree containing all road points is constructed, and then a nearest neighbor query is performed on each monitoring node.

[0086] Furthermore, in order to deal with the time-varying characteristics of the electromagnetic environment, the time dimension can be introduced to expand the graph into a space-time graph: G(V, E, T), where T represents the time series. Each time point has a corresponding graph snapshot, and the attributes of nodes and edges change over time. In order to improve the expressiveness of the graph, a multi-layer graph structure can be introduced, with different layers representing different frequency bands or different types of electromagnetic sources. The relationship between layers can be represented by cross edges. Considering the continuity of the electromagnetic environment, interpolation methods can be used to fill in the electromagnetic properties of unknown areas in the graph. Commonly used methods include inverse distance weighting (IDW) and Kriging. This electromagnetic environment graph network constructed based on road networks and monitoring nodes can not only effectively express the spatial distribution of the electromagnetic environment, but also reflect the impact of the road network on electromagnetic propagation.

[0087] S104. A hierarchical clustering algorithm is used to perform multi-scale clustering on the target monitoring nodes in the electromagnetic environment map network. Based on the preliminary clustering results of the multi-scale clustering and using a spectral clustering algorithm, the electromagnetic environment map network is divided into multiple regional sub-maps.

[0088] The main steps of hierarchical clustering include: 1) treating each node as a cluster; 2) calculating the distance between all cluster pairs; 3) merging the two closest clusters; 4) repeating steps 2 and 3 until the preset number of clusters or distance threshold is reached. Distance calculation can be simply based on the spatial location and electromagnetic properties of the node, for example:

[0089] d = sqrt((x1-x2) 2 +(y1-y2) 2 )+λ*|E1-E2|

[0090] Where (x, y) is the spatial coordinate, E is the electromagnetic intensity, and λ is the equilibrium parameter. In order to achieve multi-scale clustering, the clustering tree can be truncated at different distance thresholds to obtain clustering results of different granularities. This can be achieved through a dynamic tree cutting algorithm, such as the dynamic hybrid tree cutting method: the cutting position is dynamically determined according to the ratio of the intra-cluster variance and the inter-cluster distance. Next, based on the multi-scale clustering results, an affinity matrix A is constructed, where Aij represents the frequency that nodes i and j belong to the same cluster at different scales. Then, the entire graph network is partitioned using a spectral clustering algorithm. The main steps of spectral clustering include: 1) calculating the Laplace matrix L = DA, where D is the degree matrix; 2) calculating the eigenvalues ​​and eigenvectors of L; 3) selecting the eigenvectors corresponding to the k smallest non-zero eigenvalues ​​to form a matrix U; 4) treating the rows of U as points in k-dimensional space and clustering them using the K-means algorithm. In the partitioning process, the continuity of the electromagnetic environment and the integrity of the road network need to be considered. Constraints need to be introduced based on the affinity matrix. This electromagnetic environment graph network partitioning method based on multi-scale clustering and spectral clustering can not only effectively capture the spatial structural characteristics of the electromagnetic environment, but also adapt to monitoring needs at different scales.

[0091] S105. Deploy an on-board electromagnetic inspection vehicle equipped with an edge server in the area where each regional sub-map is located, and deploy a central server in the central area of ​​the target area, import the regional sub-map into the corresponding edge server, import the electromagnetic environment map network into the central server, and all edge servers are connected to the central server in communication.

[0092] Among them, the on-board electromagnetic inspection vehicle includes an intelligent unmanned inspection vehicle, an on-board frequency selector, an on-board measuring probe and a satellite positioning device. The on-board frequency selector and the on-board measuring probe are installed on the roof of the inspection vehicle. No vehicle modification is required, and the installation and use are convenient, and the connection reliability is guaranteed. The on-board measuring probe is not less than 0.2m away from the roof, and is equipped with a non-vehicle-modified on-board bracket for easy installation and disassembly. In order to meet the minimum speed limit of 60km / h on expressways, the equipment sampling time interval is no more than 300 milliseconds to meet the requirement that the distance between survey points should not exceed 5m. By optional mobile power supply, 24-hour fixed-point long-term continuous monitoring is supported. Reference Figure 2 The vehicle-mounted frequency selector includes a multi-layer architecture including application layer, communication layer, hardware layer and storage layer. The vehicle-mounted measurement probe can be a medium- and short-wave frequency selector probe, a radio frequency electromagnetic field probe or a low-frequency electromagnetic field probe. Different probes can be replaced according to the actual environment. The following are the relevant parameters of each probe:

[0093] 1. The parameters of medium and short wave frequency selective probe are as follows:

[0094] 1) Frequency range: 100kHz~30MHz;

[0095] 2) Detection limit: electric field 0.05V / m-500V / m, magnetic field 0.001A / m-10A / m;

[0096] 3) Frequency response: ≤±1.5dB;

[0097] 4) Isotropy: ≤1dB;

[0098] 5) Dynamic range: ≥80dB;

[0099] 6) Frequency error: <±0.1%;

[0100] 7) Linearity: ≤±1dB;

[0101] 8) Meet the technical requirements of the standards "Method for Monitoring Electromagnetic Radiation Environment of Medium Wave Broadcasting Transmitters" (HJ1136-2020) and "Method for Monitoring Electromagnetic Radiation Environment of Short Wave Broadcasting Transmitters" (HJ1199-2021).

[0102] 9) Frequency resolution: optional, supports 5kHz, 10kHz, 30kHz, 50kHz, 100kHz, etc.

[0103] 10) Data collection sampling rate: supports 1 time / second.

[0104] 11) Measurement results: electric field spectrum and frequency-selective field strength value, magnetic field spectrum and frequency-selective field strength value, real-time value and 6-minute average value.

[0105] 2. The parameters of the RF electromagnetic field probe are as follows:

[0106] 1) Frequency range: 20MHz~6GHz;

[0107] 2) Detection limit: Lower detection limit ≤7×10-6W / m2 (0.05V / m),

[0108] Upper detection limit ≥663W / m2 (500V / m);

[0109] 3) Frequency response: ≤±1.5dB (900MHz~3GHz); ≤±3dB (<900MHz or >3GHz);

[0110] 4) Isotropy: <1dB (30MHz~3GHz), ≤2dB (3GHz~6GHz);

[0111] 5) Dynamic range: ≥80dB;

[0112] 6) Frequency error: <±0.1%;

[0113] 7) Linearity: ≤±1.5dB.

[0114] 3. The parameters of low-frequency electromagnetic field probe are as follows:

[0115] 1) Frequency range: 1Hz~100kHz; support 50Hz point frequency monitoring;

[0116] 2) Detection limit: ≤0.05V / m under electric field, ≤10nT under magnetic field;

[0117] 3) Isotropy: electric field ≤1dB, magnetic field ≤1dB;

[0118] 4) Dynamic range: ≥60dB;

[0119] 5) Linearity: ≤±1dB.

[0120] In this embodiment, both the edge server and the central server can use industrial-grade computers, and the server configuration requires high-performance computing capabilities. When importing the regional sub-map into the edge server, data compression and format conversion are required. Graph compression algorithms, such as k-core decomposition or edge sampling methods, can be used to reduce the amount of data. The data format should choose a lightweight format suitable for edge computing, such as Protocol Buffers or MessagePack. When the electromagnetic environment map network is imported into the central server, in addition to the complete graph structure, metadata such as creation time, update frequency, coordinate system information, etc. need to be imported. Distributed storage systems, such as Apache Cassandra, can be used to improve data reading and writing performance. The import process should design a fault-tolerant mechanism, such as using a checksum to verify data integrity and setting an automatic retry mechanism to handle transmission failures. In order to ensure data security, strong encryption measures should be implemented, such as using the AES-256 encryption algorithm, and setting a strict access control policy. The communication connection between the edge server and the central server needs to consider security, reliability and efficiency. VPN technology can be used to build a secure channel, such as using the IPSec protocol. This distributed electromagnetic environment monitoring system based on vehicle-mounted edge servers and central servers can not only realize large-scale, high-precision electromagnetic environment monitoring, but also improve data processing efficiency and response speed through edge computing.

[0121] In another implementation, in order to save costs, the edge server configuration can be cancelled, and only the data collection terminal is installed on the inspection vehicle, and the PC data collection software is deployed in the data collection terminal. Figure 3 The PC data collection software can configure the UI layer, interface layer, service layer and storage layer to achieve simple inspection data storage, data processing and data display process.

[0122] S106. For any regional sub-map, the optimal electromagnetic inspection path of the vehicle-mounted electromagnetic inspection vehicle in the area to which it belongs is intelligently planned based on the regional sub-map, and the vehicle-mounted electromagnetic inspection vehicle is controlled to perform electromagnetic inspection along the optimal electromagnetic inspection path to collect electromagnetic inspection data of the area to which the regional sub-map belongs.

[0123] In one embodiment, a path planning model can be constructed based on the regional subgraph, and the problem is specifically modeled as a vehicle routing problem with a time window (VRPTW): V = {v1, v2, ..., vn} represents the set of nodes to be visited, E = {(vi, vj) | vi, vj∈V, i≠j} represents the set of feasible paths, c(vi, vj) represents the cost (such as distance or time) of the path (vi, vj), and [ai, bi] represents the time window of the node vi. The goal is to minimize the total cost: minΣc(vi, vj)x(vi, vj), where x(vi, vj) is a 0-1 variable indicating whether the path (vi, vj) is selected. Constraints include: each node can only be visited once, the vehicle must arrive within the time window, and the vehicle capacity limit is not exceeded. Considering the dynamics of the electromagnetic environment, a dynamic weight can be introduced: w(vi, t) = f(Hi(t), Si(t)), where Hi(t) is the historical abnormal frequency of node vi at time t, and Si(t) is the current electromagnetic signal strength. The path planning algorithm can use the improved ant colony algorithm (ACO). In order to improve the efficiency of the algorithm, local search optimization, such as the 2-opt algorithm, can be used. Taking into account the real-time traffic conditions, real-time traffic data can be integrated to dynamically adjust the path cost. For example, the Dijkstra algorithm is used to calculate the shortest path in real time, and the cost function can be defined as: c(vi,vj,t) = d(vi,vj) / v(t)+w(vj,t), where d is the distance, v(t) is the current average speed, and w is the node weight. In order to handle emergencies, a dynamic re-planning mechanism can be designed: define the trigger condition, and re-execute the path planning algorithm when triggered.

[0124] Electromagnetic detection during inspections needs to consider multiple factors. The detection frequency can be dynamically adjusted according to the vehicle speed and the rate of change of the environment: f = max (f_min, v / d + k*ΔE / Δt), where f_min is the minimum detection frequency, v is the vehicle speed, d is the spatial resolution, ΔE / Δt is the rate of change of the electromagnetic environment, and k is the adjustment coefficient. Detection parameters (such as frequency band, bandwidth) should be dynamically optimized based on historical data and real-time conditions. During data collection, real-time quality control needs to be implemented, and the Kalman filter is used to remove noise: x(k) = Ax(k-1) + Bu(k) + w(k), z(k) = Hx(k) + v(k), where x is the state vector, u is the control vector, z is the observation vector, and w and v are process noise and observation noise. Considering the energy constraints of on-board equipment, an adaptive sampling strategy can be designed: reduce the sampling frequency in areas with stable electromagnetic environments and increase the sampling frequency in areas with drastic changes. By continuously optimizing algorithms and parameters, it can adapt to different types of electromagnetic environments and monitoring requirements, providing efficient and flexible data collection capabilities for electromagnetic environment monitoring systems.

[0125] S107. Use edge computing and federated learning technology to update the regional submap in the edge server based on the electromagnetic inspection data, and perform local electromagnetic anomaly analysis on the updated regional submap through the edge server to obtain local anomaly analysis results.

[0126] S108. After all regional sub-maps are updated, the update information of all regional sub-maps is aggregated through the central server and the global update of the electromagnetic environment map network is completed in the central server. The central server performs a global electromagnetic anomaly analysis on the updated electromagnetic environment map network to obtain a global anomaly analysis result.

[0127] Among them, first the central server needs to collect update information from each edge server. In order to reduce communication overhead, an incremental update strategy can be adopted to transmit only the changed parts. When aggregating update information, possible conflicts need to be resolved. Version-based concurrency control (MVCC) can be used: assign a version number to each update, and when a conflict occurs, keep the latest version or use a predefined merge strategy. For updating the graph structure, a graph difference algorithm can be used: calculate the difference between the original graph and the updated graph, and only apply the difference. For example, use the A* algorithm to find the minimum edit distance path between two graph versions: f(n) = g(n) + h(n), where g(n) is the actual cost from the starting state to the current state, and h(n) is the estimated cost from the current state to the target state.

[0128] After completing the global update, the central server needs to perform global anomaly analysis on the entire electromagnetic environment map network. Due to the large scale of data, a distributed computing framework such as Apache Spark can be used. Use GraphX ​​for graph calculation: valgraph = GraphLoader.edgeListFile(sc,"edges.txt"), and then use the Pregel API for iterative calculation: val result = graph.pregel(initialMsg)(vprog,sendMsg,mergeMsg). Global anomaly analysis can be performed from multiple dimensions. First, anomaly detection in the spatial dimension can be performed using the spatial autocorrelation analysis method. Anomaly detection in the time dimension can use the time series analysis method. For example, the ARIMA (autoregressive integrated moving average) model is used. Finally, in order to integrate the results of multiple anomaly detection methods, ensemble learning techniques can be used. For example, multiple basic models are trained using Stacking, and then the outputs of these models are combined using a meta-learner. This global electromagnetic anomaly analysis method based on a central server can not only integrate large-scale distributed data, but also discover global patterns and trends that may be overlooked by local analysis.

[0129] S109. The local anomaly analysis results and the global anomaly analysis results are integrated into the comprehensive electromagnetic anomaly analysis results of the target area, and the comprehensive electromagnetic anomaly analysis results are imported into the regional map for visual display.

[0130] In order to fully integrate local and global information, a multi-task learning framework can be used. For example, a multi-task neural network with hard parameter sharing: the bottom layer shares parameters, and the top layer sets a specific output layer for each task (local anomaly detection and global anomaly detection). The model can be expressed as:

[0131] y l =softmax(W l *h+b l ), y g =softmax(W g *h+b g )

[0132] Where h is the output of the shared hidden layer. The loss function can be defined as: L = α*L + β*L g +λ*R(W), where L l and L g are the losses for local and global tasks respectively, R(W) is the regularization term, and α, β and λ are balancing parameters.

[0133] Considering the diversity of electromagnetic anomalies, a multi-label learning method can be used. For example, a label correlation-aware multi-label learning algorithm ML-kNN is used: for each instance x, its k nearest neighbors are found, and then multiple labels of x are predicted based on the label distribution of these neighbors. The predicted probability can be expressed as: P(Hj|Cx) = (s+cj) / (s*m+c), where Hj represents the event with the jth label 1, Cx represents that there are c instances with label j among the k nearest neighbors of x, s is the smoothing parameter, and m is the number of labels. Based on the above method, the local anomaly analysis results and the global anomaly analysis results can be fused into the comprehensive electromagnetic anomaly analysis results of the target area.

[0134] Finally, the regional map is digitized using Geographic Information System (GIS) software, and the comprehensive electromagnetic anomaly analysis results are converted into a geographic coordinate system and imported into the regional map, usually using latitude and longitude or UTM coordinates. Figure 4 Each abnormal point can be represented by a mark of different color or size. The color can reflect the degree of abnormality, and the size can represent the scope of impact. The inspection path can be represented by continuous line segments. Line segments of different colors can represent different inspection periods, forming a visual display diagram in the form of a path.

[0135] The regional map can also be divided into uniform grids, and the grid size can be determined according to the monitoring accuracy and regional characteristics. Each grid cell corresponds to a specific geographical location and area. Next, the comprehensive electromagnetic anomaly analysis results are mapped to these grid cells. Interpolation algorithms such as Kriging or Inverse Distance Weighted (IDW) can be used to expand the discrete measurement point data to the entire area. The degree of anomaly of each grid cell can be represented by different colors to form a heat map effect, such as Figure 5 As shown. The specific value of the color can be obtained through normalization: Color_value = (Anomaly_value-Min_value) / (Max_value-Min_value). The legend should clearly explain the corresponding relationship between color and degree of anomaly. This grid-like visualization can fully reflect the electromagnetic environment of the entire area, help identify abnormal hotspots, assess the scope of impact, and provide a basis for formulating regional control measures.

[0136] In one embodiment, identifying historical electromagnetic radiation points as target monitoring nodes on a regional map based on historical electromagnetic detection data and according to an electromagnetic propagation model includes the following steps:

[0137] The free space propagation model is used as the electromagnetic propagation model of the target area;

[0138] Based on the electromagnetic propagation model, the radiation points in the target area are reversely calculated according to the historical electromagnetic detection data and the least square method;

[0139] In the reverse calculation process, the objective function of the least squares method is constructed by combining the historical electromagnetic detection data and the output values ​​of the electromagnetic propagation model;

[0140] The objective function is solved by an optimization algorithm to complete the reverse calculation, and all potential historical electromagnetic radiation points are marked on the regional map;

[0141] All potential historical electromagnetic radiation points are clustered using a density clustering algorithm. After clustering, all historical electromagnetic radiation points in the regional map are obtained;

[0142] Extract radiation point characteristics of each historical electromagnetic radiation point from the historical electromagnetic detection data, the radiation point characteristics include the historical electromagnetic field intensity, historical abnormal frequency and spatial impact range of the historical electromagnetic radiation point;

[0143] The historical electromagnetic radiation points are taken as target monitoring nodes, and the node weights of the target monitoring nodes are calculated according to the corresponding radiation point characteristics.

[0144] In this embodiment, the basic formula of the free space propagation model is: Pr = Pt*(λ / (4πd)) 2*Gt*Gr, where Pr is the received power, Pt is the transmitted power, λ is the wavelength, d is the propagation distance, Gt and Gr are the transmit and receive antenna gains respectively. In practical applications, the model can be adjusted by introducing additional parameters, such as the path loss exponent n: Pr = Pt*(λ / (4πd)) n *Gt*Gr. This adjustment can better adapt to different environmental conditions. For example, in open areas n is close to 2, while in urban environments n may reach 4 or higher. The advantage of using the free space propagation model is its simplicity and universality, which can quickly provide a preliminary estimate of the electromagnetic field strength for a large area. Next, based on the free space propagation model, the least squares method is used to reversely infer the radiation points in the target area. The large amount of historical electromagnetic detection data obtained includes field strength measurements at different locations. Establish an optimization problem whose goal is to find the most likely location and power of the radiation source so that the difference between the predicted value and the actual measured value is minimized. Specifically, the error function can be defined: E=∑(Pi m -Pi p ) 2 , where Pi_m is the actual field strength at the i-th measurement point, and Pi_p is the predicted field strength calculated based on the assumed radiation source location and power. By minimizing this error function, the optimal radiation source parameters can be obtained. In practice, an iterative algorithm such as the gradient descent method can be used to solve this optimization problem. In addition, considering that there may be multiple radiation sources, a step-by-step approximation method can be used: first locate the main radiation source, and then gradually consider the contribution of the secondary source. This reverse calculation method can effectively utilize historical data and provide important radiation source information for electromagnetic environment monitoring and management.

[0145] In the process of reverse calculation, the key step is to construct the objective function of the least square method. This function needs to comprehensively consider the output values ​​of the historical electromagnetic detection data and the electromagnetic propagation model. The objective function is usually defined as the sum of the squares of the differences between the actual measured values ​​and the model predicted values: F = ∑wi*(Mi-Pi(θ)) 2 , where Mi is the actual field strength at the i-th measurement point, Pi(θ) is the predicted field strength calculated based on the current parameters θ (including the location and power of the radiation source), and wi is the weight factor. The weight factor can be set according to the reliability or importance of the measurement point. For example, recent data or high signal-to-noise ratio measurements can be given a higher weight. The reverse calculation is then completed by solving the objective function through an optimization algorithm. Commonly used optimization algorithms include gradient descent, Newton's method, Levenberg-Marquardt algorithm, etc. Taking the gradient descent method as an example, its iterative formula is: Where η is the learning rate, is the gradient of the objective function under the current parameters. In order to accelerate convergence and avoid falling into local optimality, momentum method or adaptive learning rate method such as Adam algorithm can be used. In the solution process, it is necessary to set appropriate initial values ​​and stop conditions. For example, a grid search method can be used to find a better initial value, and decide when to stop based on the change in the objective function value or the number of iterations. After the solution is completed, the optimal parameter θ* contains the inferred radiation source location and power information. Next, these potential historical electromagnetic radiation points need to be marked on the regional map. Geographic Information System (GIS) tools can be used to convert longitude and latitude coordinates into specific locations on the map.

[0146] Next, a density clustering algorithm is used to cluster all potential historical electromagnetic radiation points, thereby identifying the true radiation source location and eliminating the influence of noise points. Common density clustering algorithms include DBSCAN (density-based spatial clustering with applied noise) and OPTICS (sorting points to identify clustering structures). Taking DBSCAN as an example, its core idea is that an area containing at least MinPts points within a given radius ε is considered a dense area. The main steps of the algorithm include: 1) For each unvisited point p, calculate the number of points in its ε-neighborhood; 2) If the number of points is greater than or equal to MinPts, create a new cluster and mark p as a core point; 3) Recursively add all points in the ε-neighborhood of p to the cluster; 4) Repeat the above steps until all points are visited. A key advantage of DBSCAN is that it can automatically identify noise points without pre-specifying the number of clusters. In practical applications, the ε and MinPts parameters need to be reasonably set according to the characteristics of the electromagnetic environment. For example, these parameters can be determined based on the expected density and distribution range of the radiation source. After clustering is completed, the center of each cluster can be regarded as a historical electromagnetic radiation point. For each cluster, its centroid can be calculated as the final radiation point position: (x, y) = (∑xi / n, ∑yi / n), where (xi, yi) is the coordinate of the i-th point in the cluster, and n is the total number of points in the cluster. This method can effectively merge densely distributed potential radiation points, while filtering out isolated points that may be caused by measurement errors or environmental interference, thereby obtaining a more reliable historical electromagnetic radiation point distribution.

[0147] Next, the characteristics of each historical electromagnetic radiation point are extracted from the historical electromagnetic detection data. The radiation point characteristics usually include the historical electromagnetic field strength, the historical abnormal frequency and the spatial impact range. For the historical electromagnetic field strength, the average, maximum, minimum and standard deviation over a period of time can be calculated. The historical abnormal frequency can be defined as the frequency at which the field strength exceeds a certain threshold, and the threshold can be set as the average value plus a number of times the standard deviation. The spatial impact range can be determined by analyzing the attenuation of the field strength with distance. The field strength distribution map can be generated using an interpolation method (such as Kriging interpolation), and then the distance when the field strength drops to a certain threshold (such as 10% of the maximum value) is determined as the impact range. After feature extraction, the historical electromagnetic radiation point is used as the target monitoring node, and the node weight is calculated based on the corresponding radiation point characteristics. The calculation of the node weight needs to take into account multiple factors, and the weighted summation method can be used: W = w1*F1+w2*F2+w3*F3, where F1, F2, and F3 represent the normalized historical electromagnetic field strength, historical abnormal frequency and spatial impact range, respectively, and w1, w2, and w3 are the corresponding weight coefficients. Normalization can use the min-max method. The choice of weight coefficient can be based on the needs of specific application scenarios. For example, if you pay more attention to abnormal situations, you can increase the weight of historical abnormal frequency. . In order to make the weight more dynamic, a time decay factor can be introduced: W(t) = W*exp(-λ*Δt), where λ is the decay coefficient and Δt is the time interval. This allows the most recent observation to have a greater impact on the weight. The weight information can be used to construct an electromagnetic environment graph network, where the edge weight between nodes can be defined as a function of the weights of the two nodes. This weight calculation method based on historical data and feature analysis can effectively capture the important characteristics of the electromagnetic environment and provide an important basis for subsequent monitoring strategy formulation and anomaly detection.

[0148] In one embodiment, a hierarchical clustering algorithm is used to perform multi-scale clustering on target monitoring nodes in an electromagnetic environment map network, and the electromagnetic environment map network is divided into multiple regional sub-maps based on the preliminary clustering results of the multi-scale clustering and using a spectral clustering algorithm, including the following steps:

[0149] Construct a multi-dimensional feature vector of the target monitoring node based on the characteristics of the radiation point;

[0150] The principal component analysis method is used to reduce the multi-dimensional feature vector to a low-dimensional feature vector;

[0151] The low-dimensional feature vectors are clustered using a hierarchical clustering algorithm based on a minimum spanning tree to obtain a feature clustering tree;

[0152] Multiple different cutoff thresholds are calculated based on the average node distance between the root node and the leaf node in the feature clustering tree;

[0153] The feature clustering tree is truncated based on each truncation threshold to obtain feature clustering results of multiple different scales;

[0154] Construct an affinity matrix based on the feature clustering results at multiple different scales;

[0155] The affinity matrix is ​​used as the constraint condition of the spectral clustering algorithm. Based on the constraint condition and the spectral clustering algorithm, the electromagnetic environment map network is divided into multiple regional sub-graphs.

[0156] In this embodiment, the radiation point characteristics include historical electromagnetic field strength, frequency distribution, spatial influence range, etc. For each monitoring node, a feature vector V = [E_avg, E_max, E_std, f_main, BW, R] is constructed. In order to process features of different dimensions, standardization processing, such as Z-score standardization, is usually required. The feature vector processed in this way can better reflect the electromagnetic environment characteristics of each monitoring node, and provide a comparable data basis for subsequent analysis and clustering. The principal component analysis (PCA) is then used to reduce the multidimensional feature vector to a low-dimensional feature vector. The core idea of ​​PCA is to find the main direction of change of the data, that is, the principal component. The specific steps include: 1) calculating the covariance matrix; 2) solving the eigenvalues ​​and eigenvectors of the covariance matrix; 3) selecting the eigenvectors corresponding to the largest k eigenvalues ​​as the principal components; 4) projecting the original data onto these k principal components. The eigenvector after dimensionality reduction can be expressed as: V new =P T *V, where P is a matrix consisting of k principal components and V is the original eigenvector. Usually the number of principal components that can explain 80% to 90% of the variance is selected. PCA not only reduces the data dimension, but also removes noise, highlights the main features, and provides a more concise and effective data representation for subsequent cluster analysis.

[0157] Next, we use the hierarchical clustering algorithm based on the minimum spanning tree to cluster the low-dimensional feature vectors and obtain the feature clustering tree. First, we construct a complete graph, and the distance between nodes can be calculated using the Euclidean distance: Then, the Kruskal algorithm is used to construct the minimum spanning tree: 1) sort all edges by weight; 2) select the edge with the minimum weight in turn, and add it to the tree if it does not form a cycle; 3) repeat until n-1 edges are selected. Finally, hierarchical clustering is performed based on the minimum spanning tree: starting from the longest edge, it is gradually deleted, and a new cluster is formed each time an edge is deleted. This method can effectively capture the hierarchical structure of the data and is suitable for processing data with non-spherical distribution. The truncation threshold is calculated according to the characteristic clustering tree, and the tree is truncated at each threshold to obtain the multi-scale clustering result, and then the affinity matrix is ​​constructed. Specifically, multiple percentiles can be selected as thresholds, such as 25%, 50%, and 75% of the node distance. The element aij of the affinity matrix A represents the number of times nodes i and j are clustered into one class at different scales. Finally, the affinity matrix is ​​used as the constraint condition of spectral clustering, and the electromagnetic environment map network is divided into multiple regional subgraphs based on the constraint conditions and through the spectral clustering algorithm. This multi-scale constrained spectral clustering method can better capture the spatial structure of the electromagnetic environment and obtain a more reasonable regional division.

[0158] In one implementation, the affinity matrix is ​​used as a constraint condition of the spectral clustering algorithm, and the electromagnetic environment map network is divided into multiple regional sub-maps based on the constraint condition and by the spectral clustering algorithm, including the following steps:

[0159] The degree matrix of the electromagnetic environment map network is calculated, and the normalized Laplace matrix of the electromagnetic environment map network is calculated according to the degree matrix;

[0160] Calculate the matrix eigenvalues ​​and matrix eigenvectors of the normalized Laplace matrix;

[0161] Select the matrix eigenvectors corresponding to the k smallest non-zero matrix eigenvalues ​​to form a spectral clustering matrix;

[0162] Construct the objective optimization function of the spectral clustering algorithm;

[0163] The target optimization function is solved by generalized eigenvalues, and the K target matrix eigenvectors obtained by solving the problem are combined into a subgraph matrix;

[0164] The K-means algorithm is used to cluster the row vectors of the subgraph matrix, and the electromagnetic environment map network is divided into multiple regional subgraphs according to the clustering results.

[0165] In this embodiment, the adjacency matrix W of the electromagnetic environment graph network is first constructed, where Wij represents the similarity or connection strength between nodes i and j. The degree matrix D is a diagonal matrix whose diagonal elements Dii are equal to the degree of node i. The normalized Laplace matrix L is defined as: L = ID (-1 / 2) WD (-1 / 2), where I is the identity matrix. This normalization can reduce the impact of node degree and make the clustering results more balanced. Next, we need to calculate the matrix eigenvalues ​​and matrix eigenvectors of the normalized Laplacian matrix. This step uses the eigendecomposition theory in linear algebra. For the n×n symmetric matrix L, solve the characteristic equation Lv=λv, where λ is the eigenvalue and v is the corresponding eigenvector. The QR algorithm or Lanczos algorithm is usually used for numerical calculation. As a result, n eigenvalues ​​λ1≤λ2≤...≤λn and their corresponding eigenvectors v1, v2,..., vn are obtained. The eigenvalues ​​reflect the important structural information of the graph, and the eigenvectors corresponding to the smaller eigenvalues ​​contain the key information of the graph partitioning. Then select the matrix eigenvectors corresponding to the k smallest non-zero matrix eigenvalues ​​to form the spectral clustering matrix. Here k is usually equal to the desired number of clusters. The smallest non-zero eigenvalues ​​are selected because they contain the most important information of the graph structure. These k eigenvectors are arranged in columns to form an n×k spectral clustering matrix Y. This matrix Y actually maps the original n-dimensional data points to the k-dimensional space. In this new space, the distribution of data points is easier to cluster. This step not only achieves dimensionality reduction, but also retains the key information of the graph structure, laying the foundation for subsequent clustering.

[0166] Next, we construct the objective optimization function of the spectral clustering algorithm in order to formalize the problem into a mathematical optimization problem. The objective function of standard spectral clustering is usually defined as: min tr(Y T LY), stY T Y = I, this optimization problem is equivalent to minimizing the normalized cut. However, in practical applications, it is necessary to introduce additional constraints based on the affinity matrix to improve the clustering effect. The final target optimization function is:

[0167] min tr(Y T LY)+λ[tr(Y T (DA)Y)]

[0168] Where: tr represents the trace of the matrix, Y represents the spectral clustering matrix, L represents the normalized Laplace matrix, λ represents the balance parameter, T represents the transposed matrix, D represents the degree matrix, and A represents the affinity matrix.

[0169] Next, the objective optimization function is solved by the generalized eigenvalue problem. Considering the aforementioned constraints, this optimization problem can be transformed into solving the generalized eigenvalue equation: (L+αC)y=λDy, where D is the degree matrix. Numerical methods such as the Lanczos algorithm or the Arnoldi iteration can be used to solve this equation. Specifically, it is necessary to find the eigenvectors corresponding to the K smallest generalized eigenvalues. These eigenvectors form an n×K subgraph matrix, where each column represents an eigenvector and each row corresponds to the representation of a node in the new K-dimensional space. This step is actually looking for an optimal low-dimensional representation so that in the new space, similar nodes are closer and dissimilar nodes are farther away. The subgraph matrix contains the essential information of the graph structure and provides an ideal data representation for the final clustering.

[0170] Finally, the K-means algorithm is used to cluster the row vectors of the subgraph matrix. This process is also the key to converting continuous spectral embedding into discrete clustering results. The basic idea of ​​the K-means algorithm is to divide n data points into K clusters so that each data point belongs to the cluster center closest to it. The specific steps are as follows: 1) Randomly select K initial cluster centers; 2) Assign each data point to the nearest cluster center; 3) Recalculate the center of each cluster; 4) Repeat steps 2 and 3 until convergence or the maximum number of iterations is reached. Here, the data point is the row vector of the subgraph matrix Y, and each row vector represents the coordinates of a node in the K-dimensional space. After clustering is completed, each node is assigned to a cluster, which results in the regional subgraph division of the electromagnetic environment map network.

[0171] In one embodiment, intelligently planning the optimal electromagnetic inspection path of the vehicle-mounted electromagnetic inspection vehicle in the region based on the regional sub-graph includes the following steps:

[0172] Obtain the historical traffic flow data of the road network in the area to which the regional subgraph belongs, and assign node edge weights to the node edges of the regional subgraph according to the historical traffic flow data;

[0173] The target monitoring node with the highest node weight is taken as the starting node;

[0174] The optimal path selection steps are performed based on the starting node. The optimal path selection steps are as follows:

[0175] Calculate the node distance between the starting node and all adjacent nodes;

[0176] A multi-objective reward function is constructed by combining the node distance, the node weights of the adjacent nodes, and the node edge weights of the node edges between the starting node and the adjacent nodes;

[0177] Based on the multi-objective reward function and using the Monte Carlo tree search method, the best adjacent node is selected from all adjacent nodes as the optimal path node;

[0178] The optimal path node is used as the starting node to repeatedly perform the optimal path selection step until the farthest node distance between the optimal path node and the starting node exceeds a preset distance threshold, thereby obtaining the optimal electromagnetic inspection path of the vehicle-mounted inspection vehicle.

[0179] In this implementation, historical traffic flow data can be obtained from the traffic management department or intelligent transportation system. These data usually include information such as traffic flow and average vehicle speed at different time periods. Then, these data are mapped to the road network of the regional subgraph. For example, the weighted average method can be used to calculate the traffic flow weight of each road: W = ∑(w i *f i ) / ∑w i , where f_i is the traffic volume in the i-th time period, and w_i is the weight of the time period (which can be adjusted according to the inspection time). This method not only takes into account the importance of the road, but also reflects the traffic characteristics of different time periods, providing an important reference for subsequent path planning. Then traverse all target monitoring nodes and compare their weight values. A simple linear search or a more efficient data structure such as a maximum heap can be used to find the maximum weight node. For example, if there are nodes A (weight 10), B (weight 15), and C (weight 8), node B is selected as the starting node. There are several advantages to selecting a high-weight node as the starting point: 1) the most important or most complex areas can be monitored first; 2) it is conducive to covering the maximum monitoring value within a limited time; 3) it may be easier to detect potential electromagnetic anomalies.

[0180] Next, the optimal path selection step is performed based on the starting node. The optimal path selection step involves the following sub-steps:

[0181] 1) Calculate the node distance between the starting node and all adjacent nodes. The distance here can be physical distance, travel time, or weighted distance that takes into account multiple factors. For example, the straight-line distance between two nodes can be calculated using Euclidean distance, or the actual road length can be used. This step provides basic spatial information for subsequent path selection, helping to determine the possible range and direction of movement.

[0182] 2) Construct a multi-objective reward function by combining node distance, node weights of adjacent nodes, and node edge weights. This reward function needs to balance multiple objectives, including minimizing driving distance, maximizing monitoring value, optimizing traffic flow, etc. The form of the reward function can be: R = α*(1 / d) + β*W n +γ*W e, where d is the node distance, W_n is the neighboring node weight, W_e is the edge weight, and α, β, and γ are balancing parameters. This function encourages the selection of paths with moderate distance, high node importance, and good road conditions. The parameters can be adjusted according to specific needs. For example, if more attention is paid to monitoring coverage, the value of β can be increased.

[0183] 3) Based on the multi-objective reward function, the Monte Carlo Tree Search (MCTS) method is used to select the optimal adjacent nodes. MCTS evaluates the long-term value of each choice by simulating a large number of possible paths. Its basic steps include: selection, expansion, simulation, and backtracking. In the selection stage, the UCT (Upper Confidence Bound for Trees) formula is used to balance exploration and exploitation. The advantage of MCTS is that it can find approximate optimal solutions in complex decision spaces, which is particularly suitable for dynamically changing environments. Through a large number of simulations, MCTS can predict the long-term impact of different choices and make more informed decisions.

[0184] Next, the optimal path node is used as the new starting node to repeat the optimal path selection step until the preset distance threshold is reached. This process realizes the gradual construction of the entire inspection path. Each iteration selects a new optimal node to join the path. In each iteration, the list of visited nodes needs to be updated to avoid repeated visits. In addition, the parameters of the reward function can be adjusted dynamically. For example, as the path extends, the weight of returning to the starting point can be gradually increased. The end of this process may be in a variety of situations: reaching the distance threshold, visiting all important nodes, or reaching the inspection time limit. The final path not only takes into account spatial coverage, but also balances the monitoring value and traffic conditions, providing an optimized electromagnetic monitoring path solution for vehicle-mounted inspections.

[0185] In one embodiment, controlling the vehicle-mounted electromagnetic inspection vehicle to perform electromagnetic inspection according to the optimal electromagnetic inspection path and collecting electromagnetic inspection data of the area to which the regional sub-map belongs includes the following steps:

[0186] Before taking the optimal path node as the starting node in each round of the optimal path selection step, obtain the real-time traffic flow data of the road network in the area to which the regional subgraph belongs, and update the edge weights of all nodes in the regional subgraph according to the real-time traffic flow data;

[0187] After the optimal path node is determined in each round of the optimal path selection step, the on-board electromagnetic inspection vehicle is controlled to travel from the starting node to the optimal path node, and electromagnetic inspection data around the on-board electromagnetic inspection vehicle is continuously collected during the driving process.

[0188] In this embodiment, in each round of the optimal path selection step, the real-time traffic flow data of the road network in the area to which the regional subgraph belongs is updated and the node edge weight is updated accordingly. This step first requires real-time acquisition of traffic flow data from the intelligent transportation system or road condition monitoring equipment. The data includes vehicle density, average speed, congestion index, etc. The traffic flow density D=N / L (N is the number of vehicles, L is the road length) can be used to represent the degree of road congestion. Then, these real-time data are combined with historical data to update the weight of each edge. The exponential moving average method can be used for smooth update. This method takes into account both the current road conditions and the historical trend information. The updated weight directly affects the next path selection, allowing the inspection vehicle to avoid sudden congested sections and choose a more unobstructed path. This dynamic update mechanism can make the inspection path planning more flexible, adapt to real-time traffic changes, and improve inspection efficiency and coverage. After the optimal path node is determined in the optimal path selection step, the on-board electromagnetic inspection vehicle is controlled to travel from the starting node to the optimal path node, and the surrounding electromagnetic inspection data is continuously collected during the driving process.

[0189] In one embodiment, edge computing and federated learning technologies are used to update a regional submap in an edge server based on electromagnetic inspection data, and local electromagnetic anomaly analysis is performed on the updated regional submap by the edge server to obtain a local anomaly analysis result, including the following steps:

[0190] Deploy the same graph neural network model on all edge servers;

[0191] For any edge server, the preprocessed electromagnetic inspection data is input into the graph neural network model, and the regional subgraph on the edge server is updated using the graph structure learning method, and the model parameters of the graph neural network model are updated synchronously;

[0192] Use the federated averaging algorithm to exchange updated model parameters with all other edge servers;

[0193] Update the local graph neural network model based on the exchanged model parameters;

[0194] The multi-scale anomaly score of the updated regional subgraph is calculated through the multi-layer graph attention network structure in the updated graph neural network model, and the local electromagnetic anomaly analysis result of the regional subgraph is calculated by the ensemble learning method.

[0195] In this implementation, we first need to design a graph neural network model suitable for electromagnetic environment analysis, which usually includes multiple layers of graph convolution or graph attention layers. For example, we can use Graph Attention Network (GAT) as the basic structure: Where h_i is the feature of node i, α_ij is the attention coefficient, and W is the weight matrix. The input layer of the model can receive node features (such as location, historical abnormal frequency) and edge features (such as signal strength changes). The middle layer performs feature extraction and information transmission, and the output layer is used for anomaly detection or classification. After the model structure is determined, it needs to be initialized and deployed to each edge server. Ensure that all servers obtain the same initial model. The deployment process also includes configuring the operating environment, setting up data interfaces, etc.

[0196] For any edge server, the preprocessed electromagnetic inspection data is input into the graph neural network model, and the graph structure learning method is used to update the regional subgraph and model parameters. Specifically, preprocessing can include data standardization, outlier processing and feature extraction. The graph structure learning method can dynamically adjust the topological structure of the graph, such as using the attention mechanism to calculate the weight of the edge. The model parameter update usually uses the back propagation algorithm. This process not only updates the model parameters, but also adjusts the structure of the regional subgraph according to the new data, so that the model can better capture the dynamic characteristics of the local electromagnetic environment. The updated model parameters are then exchanged with all other edge servers using the federated average algorithm to achieve distributed learning. The federated average algorithm allows the learning results of all edge servers to be utilized while protecting data privacy. The specific process is as follows: First, each edge server i trains the model on local data to obtain the updated parameters θ_i. Then, all servers share the parameters in a point-to-point network. Each edge server calculates the average parameter: θ avg =(1 / N)*∑ i θ i , where N is the number of edge servers. To protect privacy, a secure aggregation protocol such as homomorphic encryption can be used. Finally, the exchanged model parameters are used to update the local graph neural network model to complete the federated learning of the model in the edge server.

[0197] Next, the multi-layer graph attention network structure in the updated graph neural network model is used to calculate the multi-scale anomaly score of the updated regional subgraph, and the local electromagnetic anomaly analysis results of the regional subgraph are calculated using an ensemble learning method. The multi-layer graph attention network can capture features at different scales. The attention mechanism of each layer can be expressed as:

[0198]

[0199] Multi-scale anomaly scores can be calculated by aggregating the outputs of different layers: score_i = f(h_i^1, h_i^2, ..., h_i^L), where f can be a simple average or a weighted sum. Ensemble learning methods can combine the outputs of multiple models, such as using random forests or gradient boosting trees. The final anomaly analysis result is a probability distribution or a binary classification (normal / abnormal).

[0200] In one embodiment, the update information of all regional sub-maps is aggregated by a central server and the global update of the electromagnetic environment map network is completed in the central server. The global electromagnetic anomaly analysis is performed on the updated electromagnetic environment map network by the central server to obtain the global anomaly analysis result, which includes the following steps:

[0201] Aggregate the updated information of all regional subgraphs from all edge servers through the central server and adopt an incremental update strategy;

[0202] Based on all the updated information and using the graph difference algorithm, the global update of the electromagnetic environment map network is completed in the central server;

[0203] The autoregressive integrated moving average method is used to extract the abnormal features of the time dimension from the updated electromagnetic environment map network;

[0204] The spatial dimension anomaly features are extracted from the updated electromagnetic environment map network using the spatial autocorrelation analysis method;

[0205] The abnormal features of spatial dimension and time dimension are combined and the integrated learning method is used to obtain the global abnormal analysis results.

[0206] In this embodiment, the incremental update strategy only transmits the changed part, rather than the entire subgraph, which greatly reduces the amount of data transmission. In specific implementation, each edge server maintains a version number and update log. When the local subgraph changes, the server records the changes (such as adding nodes, deleting edges, modifying attributes, etc.) and the corresponding timestamp. The central server requests updates from the edge server regularly or on demand, and the edge server only sends the changes since the last synchronization. After receiving these update information, the central server applies the changes in timestamp order. To handle possible conflicts, version-based concurrency control can be used: if the received update is based on an old version, the central server may need to request the complete subgraph data. This method not only reduces the network load, but also improves the real-time performance of the update, enabling the global graph network to quickly reflect local changes.

[0207] Based on all the update information and using the graph difference algorithm, the global update of the electromagnetic environment map network is completed in the central server. Specifically, the graph difference algorithm can efficiently handle the dynamic changes of large-scale graphs. The basic idea is to represent the changes of the graph as a series of basic operations (such as adding / deleting nodes or edges, modifying attributes, etc.). The algorithm first constructs a difference graph ΔG to represent all changes. Then, this difference graph is applied to the original graph G to obtain the updated graph G'. In specific implementation, the graph can be represented by an adjacency list or an adjacency matrix, and the difference operation can be represented as the modification of these data structures.

[0208] After completing the global update, the autoregressive integrated moving average (ARIMA) method is used to extract abnormal features in the time dimension from the updated electromagnetic environment map network. The ARIMA model combines three components: autoregression (AR), difference (I), and moving average (MA), and can effectively process time series data. For each node or key feature in the electromagnetic environment map network, an ARIMA (p, d, q) model can be constructed, where p is the number of autoregressive terms, d is the number of difference terms, and q is the number of moving average terms. The model can be expressed as: φ(B)(1-V) d X t =θ(B)ε t , where B is the lag operator, φ and θ are the AR and MA polynomials respectively. When implemented, the time series is first tested for stationarity (such as ADF test) and then differentiated if non-stationary. The values ​​of p and q are then determined using the autocorrelation function (ACF) and partial autocorrelation function (PACF). After fitting the model, it can be used to predict future values ​​and calculate prediction intervals. Anomalies can be defined as situations where the actual value is outside the prediction interval. For example, if the actual value is outside the 95% prediction interval, it may indicate an anomaly.

[0209] On the other hand, spatial autocorrelation analysis is used to extract spatial dimension anomaly features from the updated electromagnetic environment map network. Spatial autocorrelation analysis is based on the first law of geography: things that are close are more related than things that are far away. In the electromagnetic environment map network, this is manifested as the electromagnetic characteristics of adjacent areas may be more similar. Commonly used spatial autocorrelation indicators include the Global Moran's I and the Local Moran's I. The calculation formula for the Global Moran's I is:

[0210]

[0211] Where N is the number of nodes, W is the sum of weights, and w_ij is the spatial weight. I values ​​close to 1 indicate positive correlation, close to -1 indicate negative correlation, and close to 0 indicate random distribution. The local Moran index is used to identify local clusters, and hot and cold spots can be found. When implementing, you first need to define the spatial weight matrix, which can be based on distance or adjacency. Then calculate the global and local indices and perform significance tests. Anomalies can be defined as significant high-value clusters (hot spots) or low-value clusters (cold spots), or outliers that are significantly different from the surrounding environment. This method can effectively identify spatial anomaly patterns in the electromagnetic environment, such as the location of interference sources or the propagation range of abnormal signals.

[0212] Finally, the global anomaly analysis results are obtained by combining the abnormal features of the spatial dimension and the abnormal features of the temporal dimension and using an ensemble learning method. Ensemble learning improves the overall performance and robustness by combining the predictions of multiple models. In this scenario, methods such as random forests, gradient boosting trees (such as XGBoost), or stacked ensembles can be used. First, the temporal anomaly features obtained from the ARIMA model and the spatial anomaly features obtained from the spatial autocorrelation analysis are combined into a feature vector. Then, multiple basic models are trained, each of which may focus on a different subset of features or use a different algorithm. For example, a decision tree can be trained to focus on temporal patterns, and a support vector machine can be trained to focus on spatial patterns. Finally, voting, weighted averaging, or a meta-learner is used to integrate the predictions of these basic models. The above steps can not only comprehensively consider the spatiotemporal features, but also improve the accuracy and generalization of detection by integrating multiple models, providing reliable results for the global anomaly analysis of the electromagnetic environment.

[0213] In one embodiment, after the local anomaly analysis result and the global anomaly analysis result are merged into the comprehensive electromagnetic anomaly analysis result of the target area, the following steps are also included:

[0214] Build a spatiotemporal graph convolutional network model;

[0215] The electromagnetic environment map network and the comprehensive electromagnetic anomaly analysis results are fed as input data into the spatiotemporal graph convolutional network model;

[0216] The spatiotemporal graph convolutional network model is used to predict the spatiotemporal evolution trend of electromagnetic anomalies in the target area, and the predicted spatiotemporal evolution results of electromagnetic anomalies are output;

[0217] The graph network structure of the electromagnetic environment graph network and the node weights of the target monitoring nodes are adjusted according to the spatiotemporal evolution results of electromagnetic anomalies.

[0218] In this embodiment, constructing a spatiotemporal graph convolutional network model is a key step in capturing the spatiotemporal dynamic characteristics of the electromagnetic environment. The model combines the advantages of graph convolutional networks (GCN) and temporal convolutional networks (TCN) and can process information in both spatial and temporal dimensions. The model structure usually includes multiple spatiotemporal convolutional blocks, each of which consists of a graph convolutional layer and a temporal convolutional layer. The graph convolutional layer captures spatial dependencies and can be expressed as: in is the normalized adjacency matrix, H^(l) is the node feature of the lth layer, and W^(l) is the learnable weight matrix. The temporal convolution layer uses causal convolution to process temporal information, which can be expressed as: t =f(∑ i W i *I (t-i) ), where I_t is the input sequence and W_i is the convolution kernel. The model can also include attention mechanisms, such as self-attention layers: To capture long-distance dependencies. Finally, the prediction results are output through a fully connected layer and an activation function (such as ReLU). This structure allows the model to simultaneously learn the spatial correlation and temporal evolution of the electromagnetic environment, providing a strong foundation for subsequent anomaly prediction.

[0219] Next, the electromagnetic environment map network and the comprehensive electromagnetic anomaly analysis results are input as input data into the space-time graph convolutional network model. The space-time graph convolutional network model is used to predict the space-time evolution trend of electromagnetic anomalies in the target area, and the predicted space-time evolution results of electromagnetic anomalies are output. During the prediction process, the model forward propagates through multiple space-time convolution blocks, and each block captures space-time dependencies at different scales. For example, the first block may focus on local short-term patterns, while subsequent blocks may capture larger and longer-term patterns. The predicted output can be the anomaly probability or degree of anomaly for each node in the next H time steps. Specifically, if a softmax output layer is used, the model can output a probability distribution for each node in each future time step: P(y i , t|X,A,T)=softmax(f θ (X, A, T)), where f_θ represents the function of the entire network. To evaluate the prediction performance, indicators such as mean square error (MSE), mean absolute error (MAE) or Area Under the ROC Curve (AUC) can be used.

[0220] Based on the predicted anomaly evolution trend, the edge weights of the graph network can be adjusted. For example, if the anomaly correlation between two nodes increases, the weight of the edge between them can be increased: ij =w ij *(1+α*corr(y i ,y j )), where Corr(y i ,y j ) is the correlation coefficient of the predicted anomalies of nodes i and j, and α is the adjustment factor. For node weight adjustment, it can be based on the degree of abnormality and uncertainty of the prediction. For example, information entropy is used to quantify the uncertainty of the prediction: H i =-∑ k p ik log(p ik ), where p ik is the predicted probability of node i in category k. The new node weight can be defined as: w′ i =w i *(1+β*H i +γ*y i), where y_i is the predicted degree of anomaly, and β and γ are balance parameters. This adjustment strategy enables the network structure to adapt to the dynamic changes of the electromagnetic environment, focusing on high-risk and high-uncertainty areas. This adaptive method can improve monitoring efficiency and better allocate limited monitoring resources, thereby improving the overall electromagnetic environment anomaly detection capability.

[0221] The present invention also discloses a regional electromagnetic patrol system based on a vehicle-mounted electromagnetic detection device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the regional electromagnetic patrol method based on the vehicle-mounted electromagnetic detection device as described in any one of the above embodiments.

[0222] Among them, the processor can adopt a central processing unit (CPU). Of course, according to actual usage, other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. can also be adopted. The general-purpose processor can adopt a microprocessor or any conventional processor, etc., and this application does not impose any restrictions on this.

[0223] Among them, the memory can be an internal storage unit of a computer device, such as a hard disk or memory of a computer device, or an external storage device of a computer device, such as a plug-in hard disk, a smart memory card (SMC), a secure digital card (SD) or a flash memory card (FC) equipped on the computer device, etc., and the memory can also be a combination of an internal storage unit and an external storage device of a computer device. The memory is used to store computer programs and other programs and data required by the computer device. The memory can also be used to temporarily store data that has been output or is to be output, and this application does not impose any restrictions on this.

[0224] A person skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of protection of the present application is limited to these examples. In line with the concept of the present application, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of different aspects of one or more embodiments of the present application as above, which are not provided in detail for the sake of simplicity.

[0225] One or more embodiments of the present application are intended to cover all such substitutions, modifications and variations that fall within the broad scope of the present application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of one or more embodiments of the present application should be included in the protection scope of the present application.

Claims

1. A regional electromagnetic inspection method based on vehicle-mounted electromagnetic detection equipment, characterized in that: The steps include: Obtain regional maps and historical electromagnetic survey data for the target area; Based on the historical electromagnetic detection data and according to the electromagnetic propagation model, identifying historical electromagnetic radiation points as target monitoring nodes on the regional map; Using the road network in the regional map as the node edge, and combining all the target monitoring nodes to construct an electromagnetic environment map network of the target area; Using a hierarchical clustering algorithm to perform multi-scale clustering on the target monitoring nodes in the electromagnetic environment map network, and dividing the electromagnetic environment map network into multiple regional sub-maps based on the preliminary clustering results of the multi-scale clustering and using a spectral clustering algorithm; Deploy an on-board electromagnetic inspection vehicle equipped with an edge server in the area where each of the regional sub-maps is located, and deploy a central server in the central area of ​​the target area, import the regional sub-maps into the corresponding edge servers, import the electromagnetic environment map network into the central server, and all the edge servers are communicatively connected with the central server; For any of the regional sub-graphs, the optimal electromagnetic inspection path of the vehicle-mounted electromagnetic inspection vehicle in the region to which the regional sub-graph belongs is intelligently planned based on the regional sub-graph, and the vehicle-mounted electromagnetic inspection vehicle is controlled to perform electromagnetic inspection along the optimal electromagnetic inspection path, and electromagnetic inspection data of the region to which the regional sub-graph belongs is collected; Using edge computing and federated learning technology and based on the electromagnetic inspection data, the regional sub-map is updated in the edge server, and local electromagnetic anomaly analysis is performed on the updated regional sub-map by the edge server to obtain a local anomaly analysis result; When all the regional sub-maps are updated, the update information of all the regional sub-maps is aggregated through the central server and the global update of the electromagnetic environment map network is completed in the central server, and the global electromagnetic anomaly analysis is performed on the updated electromagnetic environment map network through the central server to obtain a global anomaly analysis result; The local anomaly analysis result and the global anomaly analysis result are merged into a comprehensive electromagnetic anomaly analysis result of the target area, and the comprehensive electromagnetic anomaly analysis result is imported into the regional map for visual display.

2. The regional electromagnetic inspection method based on vehicle-mounted electromagnetic detection equipment according to claim 1 is characterized in that: The identifying of historical electromagnetic radiation points as target monitoring nodes on the regional map based on the historical electromagnetic detection data and according to the electromagnetic propagation model comprises the following steps: Using a free space propagation model as an electromagnetic propagation model for the target area; Based on the electromagnetic propagation model, the radiation points of the target area are reversely estimated according to the historical electromagnetic detection data and the least square method; In the reverse calculation process, the objective function of the least square method is constructed by combining the historical electromagnetic detection data and the output value of the electromagnetic propagation model; The reverse calculation is completed by solving the objective function through an optimization algorithm, and all potential historical electromagnetic radiation points are marked on the regional map; Clustering all the potential historical electromagnetic radiation points using a density clustering algorithm, and obtaining all the historical electromagnetic radiation points in the regional map after clustering; Extracting radiation point features of each of the historical electromagnetic radiation points from the historical electromagnetic detection data, the radiation point features including the historical electromagnetic field intensity, historical abnormal frequency and spatial impact range of the historical electromagnetic radiation point; The historical electromagnetic radiation point is used as a target monitoring node, and the node weight of the target monitoring node is calculated according to the corresponding radiation point characteristics.

3. The regional electromagnetic inspection method based on vehicle-mounted electromagnetic detection equipment according to claim 2 is characterized in that: The method of using a hierarchical clustering algorithm to perform multi-scale clustering on the target monitoring nodes in the electromagnetic environment map network, and dividing the electromagnetic environment map network into multiple regional sub-maps based on the preliminary clustering results of the multi-scale clustering and using a spectral clustering algorithm, comprises the following steps: Constructing a multidimensional feature vector of the target monitoring node according to the radiation point features; Using principal component analysis to reduce the multidimensional feature vector to a low-dimensional feature vector; Clustering the low-dimensional feature vector using a hierarchical clustering algorithm based on a minimum spanning tree to obtain a feature clustering tree; Calculating a plurality of different cutoff thresholds according to the average node distance between the root node and the leaf nodes in the feature clustering tree; Truncating the feature clustering tree based on each of the truncation thresholds to obtain feature clustering results of multiple different scales; Constructing an affinity matrix based on the feature clustering results at multiple different scales; The affinity matrix is ​​used as a constraint condition of a spectral clustering algorithm, and the electromagnetic environment map network is divided into a plurality of regional sub-maps based on the constraint condition and by using the spectral clustering algorithm.

4. The regional electromagnetic inspection method based on vehicle-mounted electromagnetic detection equipment according to claim 3 is characterized in that: The method of using the affinity matrix as a constraint condition of the spectral clustering algorithm and dividing the electromagnetic environment map network into a plurality of regional sub-maps based on the constraint condition and using the spectral clustering algorithm comprises the following steps: Calculate and obtain the degree matrix of the electromagnetic environment map network, and calculate and obtain the normalized Laplace matrix of the electromagnetic environment map network according to the degree matrix; Calculating and obtaining the matrix eigenvalues ​​and matrix eigenvectors of the normalized Laplace matrix; Select the matrix eigenvectors corresponding to the k smallest non-zero matrix eigenvalues ​​to form a spectral clustering matrix; Construct the target optimization function of the spectral clustering algorithm, the target optimization function is: min tr(Y T LY)+λ[tr(Y T (DA)Y)] Wherein: tr represents the trace of the matrix, Y represents the spectral clustering matrix, L represents the normalized Laplace matrix, λ represents the balance parameter, T represents the transposed matrix, D represents the degree matrix, and A represents the affinity matrix; Solving the target optimization function by generalized eigenvalues, and forming a subgraph matrix with K target matrix eigenvectors obtained by solving; The row vectors of the sub-graph matrix are clustered using a K-means algorithm, and the electromagnetic environment graph network is divided into a plurality of regional sub-graphs according to the clustering result.

5. The regional electromagnetic inspection method based on vehicle-mounted electromagnetic detection equipment according to claim 2 is characterized in that: The intelligent planning of the optimal electromagnetic inspection path of the vehicle-mounted electromagnetic inspection vehicle in the area based on the area sub-map comprises the following steps: Acquire historical traffic flow data of the road network in the area to which the regional subgraph belongs, and assign node edge weights to the node edges of the regional subgraph according to the historical traffic flow data; Taking the target monitoring node with the highest node weight as the starting node; An optimal path selection step is performed based on the starting node, and the optimal path selection step is as follows: Calculate the node distances between the starting node and all adjacent nodes; Constructing a multi-objective reward function by combining the node distance, the node weights of the adjacent nodes, and the node edge weights of the node edges between the starting node and the adjacent nodes; Based on the multi-objective reward function and using a Monte Carlo tree search method, an optimal adjacent node is selected from all the adjacent nodes as an optimal path node; The optimal path selection step is repeatedly performed with the optimal path node as the starting node until the farthest node distance between the optimal path node and the starting node exceeds a preset distance threshold, thereby obtaining the optimal electromagnetic inspection path of the vehicle-mounted inspection vehicle.

6. The regional electromagnetic inspection method based on vehicle-mounted electromagnetic detection equipment according to claim 5 is characterized in that: The controlling the vehicle-mounted electromagnetic inspection vehicle to perform electromagnetic inspection according to the optimal electromagnetic inspection path and collecting electromagnetic inspection data of the area to which the area sub-map belongs comprises the following steps: Before taking the optimal path node as the starting node in each round of the optimal path selection step, obtaining real-time traffic flow data of the road network in the area to which the regional subgraph belongs, and updating the edge weights of all the nodes in the regional subgraph according to the real-time traffic flow data; After the optimal path node is determined in each round in the optimal path selection step, the vehicle-mounted electromagnetic inspection vehicle is controlled to travel from the starting node to the optimal path node, and electromagnetic inspection data around the vehicle-mounted electromagnetic inspection vehicle is continuously collected during the driving process.

7. The regional electromagnetic inspection method based on vehicle-mounted electromagnetic detection equipment according to claim 1 is characterized in that: The method of updating the regional sub-map in the edge server based on the electromagnetic inspection data by using edge computing and federated learning technology, and performing local electromagnetic anomaly analysis on the updated regional sub-map by the edge server to obtain the local anomaly analysis result comprises the following steps: Deploy the same graph neural network model on all of the edge servers; For any of the edge servers, the preprocessed electromagnetic inspection data is input into the graph neural network model, the regional subgraph on the edge server is updated using a graph structure learning method, and the model parameters of the graph neural network model are updated synchronously; Using a federated averaging algorithm to exchange the updated model parameters with all other edge servers; Update the local graph neural network model based on the exchanged model parameters; The updated multi-scale anomaly score of the regional subgraph is calculated by using the multi-layer graph attention network structure in the graph neural network model, and the local electromagnetic anomaly analysis result of the regional subgraph is calculated by using an integrated learning method.

8. The regional electromagnetic inspection method based on vehicle-mounted electromagnetic detection equipment according to claim 7 is characterized in that: Aggregating the update information of all the regional sub-maps through the central server and completing the global update of the electromagnetic environment map network in the central server, performing global electromagnetic anomaly analysis on the updated electromagnetic environment map network through the central server, and obtaining the global anomaly analysis result comprises the following steps: Aggregating update information of all the regional sub-graphs from all the edge servers through the central server and using an incremental update strategy; Based on all the update information and using a graph difference algorithm, a global update of the electromagnetic environment map network is completed in the central server; Extracting abnormal features of the time dimension from the updated electromagnetic environment map network using an autoregressive integrated moving average method; Extracting spatial dimension abnormal features from the updated electromagnetic environment map network using a spatial autocorrelation analysis method; The spatial dimension abnormality features and the temporal dimension abnormality features are combined and an integrated learning method is used to analyze and obtain a global abnormality analysis result.

9. The regional electromagnetic inspection method based on vehicle-mounted electromagnetic detection equipment according to claim 1 is characterized in that: After the local anomaly analysis result and the global anomaly analysis result are merged into the comprehensive electromagnetic anomaly analysis result of the target area, the following steps are also included: Build a spatiotemporal graph convolutional network model; Inputting the electromagnetic environment graph network and the comprehensive electromagnetic anomaly analysis results as input data into the spatiotemporal graph convolutional network model; Using the spatiotemporal graph convolutional network model to predict the spatiotemporal evolution trend of electromagnetic anomalies in the target area, and outputting the predicted spatiotemporal evolution results of electromagnetic anomalies; The graph network structure of the electromagnetic environment graph network and the node weights of the target monitoring nodes are adjusted according to the spatiotemporal evolution results of the electromagnetic anomaly.

10. A regional electromagnetic inspection system based on a vehicle-mounted electromagnetic detection device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the regional electromagnetic patrol method based on the vehicle-mounted electromagnetic detection equipment as described in any one of claims 1 to 9 is implemented.

Images