High-altitude unmanned aerial vehicle remote sensing power grid multi-disaster comprehensive identification method and system

Abstract

The invention provides a high-altitude unmanned aerial vehicle remote sensing power grid multi-disaster comprehensive identification method and system, and relates to the technical field of power system intelligent monitoring, and the method comprises the steps: fusing historical disaster data, geographic information and satellite images for a to-be-inspected power grid region, and constructing a digital twinborn model; performing risk clustering and quantitative evaluation on the transmission line corridor based on risk distribution output by the digital twinborn model to obtain a risk level quantitative matrix; obtaining a differential acquisition strategy parameter table according to the risk level quantization matrix; performing unmanned aerial vehicle group task parameterization optimization by using the differential acquisition strategy parameter table to obtain a collaborative observation scheme; and scheduling an unmanned aerial vehicle cluster to collect a multi-source remote sensing data stream according to the collaborative observation scheme. According to the invention, through constructing a full-process closed loop of digital twinning, risk-driven acquisition, multi-source data fusion, parallel disaster identification and intelligent decision response, efficient intelligent comprehensive identification and disposal of multiple disasters of the power grid are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent monitoring of power systems, in particular to a high-altitude unmanned aerial vehicle remote sensing power grid multi-disaster comprehensive identification method and system. BACKGROUND

[0002] With the intensification of global climate change, power grid systems are threatened by multiple disasters such as forest fires, icing, and geological disasters. Traditional manual inspection and single-point monitoring methods have been unable to meet the real-time and accurate identification requirements. Although unmanned aerial vehicle remote sensing technology can obtain multi-source high-resolution data, there are still systematic technical bottlenecks in building an automated and intelligent disaster comprehensive identification and disposal system, especially in the data processing process.

[0003] The existing technology mainly has the following defects: first, the unmanned aerial vehicle task planning is mostly static preset, without the dynamic risk model built by integrating meteorological, topographical, and historical disaster information, so it cannot realize the adaptive and differentiated deployment of detection resources (such as sensor types and sampling density) for high-risk areas, which may lead to uneven data source quality; second, the data is transmitted back to the remote cloud center through a single link, and then offline time and space registration and manual correction are performed, resulting in significant delay from collection to analysis, which may not be able to support the minute-level response requirements of disaster emergency; third, existing algorithms are mostly designed for single disaster, lacking multi-source feature linkage mechanism, for example, there is a lack of information triggering and cooperation between icing identification and associated feature analysis such as insulator contamination and tower deformation, which may restrict accurate identification and in-depth analysis of complex disasters; fourth, the disaster output results are often presented in simple reports, without deep integration with power grid topology, geographic information, transportation network, and resource scheduling system, which makes it difficult for decision-makers to quickly assess the impact and generate optimized disposal plans in a unified visualization scenario, which may cause information gap and response delay. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a high-altitude unmanned aerial vehicle remote sensing power grid multi-disaster comprehensive identification method and system, which realizes efficient and intelligent comprehensive identification and disposal of power grid multi-disasters by building a full-process closed loop of digital twin, risk-driven collection, multi-source data fusion, parallel disaster identification, and intelligent decision response.

[0005] To solve the above technical problems, the technical solution of the present application is as follows: In a first aspect, a high-altitude unmanned aerial vehicle remote sensing power grid multi-disaster comprehensive identification method is provided, the method comprising: For the power grid area to be inspected, a digital twin model is built by integrating historical disaster data, geographic information, and satellite images; Based on the risk distribution output by the digital twin model, risk clustering and quantitative assessment are performed on the transmission line corridor to obtain a risk level quantification matrix; a differentiated data acquisition strategy parameter table is obtained based on the risk level quantification matrix. The parameter table of the differentiated acquisition strategy is used to perform parameterized tuning of the UAV swarm mission to obtain a collaborative observation scheme; the UAV swarm is then scheduled to acquire multi-source remote sensing data streams based on the collaborative observation scheme. Multi-source remote sensing data streams are transmitted in real time to edge computing nodes via 5G private network and satellite relay communication links; Edge computing nodes receive multi-source remote sensing data streams and perform spatiotemporal registration to obtain spatiotemporal registration data; based on the spatiotemporal registration data, combined with differentiated acquisition strategy parameters, adaptive correction and feature-level fusion are performed to obtain multidimensional fused data; Multi-scale feature analysis is performed on multi-dimensional fused data to obtain a multi-scale feature set; parallel disaster feature parsing is performed based on the multi-scale feature set to obtain a disaster feature parameter set; comprehensive analysis of the disaster feature parameter set is performed to obtain a quantitative disaster assessment report; The quantitative disaster assessment report is mapped into a digital twin model to obtain the disaster scenario; based on the disaster scenario, the power grid topology and traffic network information are integrated to perform route planning and obtain a comprehensive response plan; based on the comprehensive response plan, coordinated response instructions are automatically generated and issued.

[0006] Secondly, a high-altitude unmanned aerial vehicle (UAV) remote sensing system for comprehensive identification of multiple disasters in power grids includes: The module is used to build a digital twin model for the power grid area to be inspected by integrating historical disaster data, geographic information and satellite imagery. The risk assessment module is used to perform risk clustering and quantitative assessment of transmission line corridors based on the risk distribution output by the digital twin model, and obtain a risk level quantification matrix; and obtain a differentiated data acquisition strategy parameter table based on the risk level quantification matrix. The collaborative acquisition module is used to perform parameterized optimization of UAV swarm missions using a differentiated acquisition strategy parameter table to obtain a collaborative observation scheme; and to schedule the UAV swarm to acquire multi-source remote sensing data streams based on the collaborative observation scheme. The transmission module is used to transmit multi-source remote sensing data streams to edge computing nodes in real time via 5G private network and satellite relay communication links; The fusion module is used by edge computing nodes to receive multi-source remote sensing data streams and perform spatiotemporal registration to obtain spatiotemporal registered data; based on the spatiotemporal registered data, combined with differentiated acquisition strategy parameters, adaptive correction and feature-level fusion are performed to obtain multi-dimensional fused data; The analysis module is used to perform multi-scale feature analysis on multi-dimensional fused data to obtain a multi-scale feature set; based on the multi-scale feature set, parallel disaster feature parsing is performed to obtain a disaster feature parameter set; and a comprehensive analysis of the disaster feature parameter set is performed to obtain a quantitative disaster assessment report. The response module maps the quantitative disaster assessment report to a digital twin model to obtain the disaster scenario; based on the disaster scenario, it integrates power grid topology and traffic network information to perform route planning and obtain a comprehensive response plan; based on the comprehensive response plan, it automatically generates and issues collaborative response instructions.

[0007] Thirdly, a computing device, comprising: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0008] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0009] The above-described solution of the present invention has at least the following beneficial effects: By integrating multiple types of data to construct a digital twin model, a unified spatiotemporal data platform is formed, enabling the correlation and integration of historical and current data. Quantitative assessments are completed based on model risk distribution, and precise risk classification is achieved through data clustering, providing a data-driven basis for collection strategies and supporting differentiated resource allocation. Parameter optimization is performed based on differentiated parameter tables to improve the collaborative efficiency of UAV swarms and ensure the targeted and collaborative nature of multi-source remote sensing data collection. A dual-link transmission architecture is adopted, leveraging the complementarity of 5G and satellite relay to ensure the real-time performance and stability of multi-source data stream transmission, avoiding data transmission interruptions. Spatiotemporal registration eliminates data spatiotemporal biases, adaptive correction optimizes data quality, and feature-level fusion strengthens the intrinsic correlation of data, improving data usability. Multi-scale feature analysis adapts to different disaster characteristic dimensions, parallel parsing improves processing efficiency, and comprehensive analysis forms a structured parameter set and quantitative report, enhancing the value of data interpretation. The assessment results are mapped to intuitive model construction scenarios, integrating multi-dimensional road network topology data to support path planning, and automatically generating instructions to achieve efficient response supported by data. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating a method for comprehensive identification of multiple disasters in a power grid using high-altitude unmanned aerial vehicle (UAV) remote sensing, provided by an embodiment of the present invention. Figure 2 This is a schematic diagram of a high-altitude unmanned aerial vehicle (UAV) remote sensing power grid multi-hazard integrated identification system provided by an embodiment of the present invention. Detailed Implementation

[0011] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0012] like Figure 1 As shown in the figure, an embodiment of the present invention proposes a method for comprehensive identification of multiple disasters in a power grid using high-altitude unmanned aerial vehicle (UAV) remote sensing. The method includes the following steps: Step 100: For the power grid area to be inspected, a digital twin model is constructed by integrating historical disaster data, geographic information and satellite imagery; Step 200: Based on the risk distribution output by the digital twin model, perform risk clustering and quantitative assessment on the transmission line corridor to obtain a risk level quantification matrix; and obtain a differentiated acquisition strategy parameter table based on the risk level quantification matrix. Step 300: Optimize the parameters of the UAV swarm mission using the differentiated acquisition strategy parameter table to obtain a collaborative observation scheme; and schedule the UAV swarm to acquire multi-source remote sensing data streams based on the collaborative observation scheme. Step 400: The multi-source remote sensing data stream is transmitted to the edge computing node in real time through the 5G private network and satellite relay communication link; Step 500: The edge computing node receives multi-source remote sensing data streams and performs spatiotemporal registration to obtain spatiotemporal registration data; based on the spatiotemporal registration data, combined with differentiated acquisition strategy parameters, adaptive correction and feature-level fusion are performed to obtain multidimensional fused data; Step 600: Perform multi-scale feature analysis on the multi-dimensional fused data to obtain a multi-scale feature set; perform parallel disaster feature parsing based on the multi-scale feature set to obtain a disaster feature parameter set; and conduct comprehensive analysis on the disaster feature parameter set to obtain a quantitative disaster assessment report. Step 700: Map the quantitative disaster assessment report to the digital twin model to obtain the disaster scenario; based on the disaster scenario, integrate power grid topology and traffic network information to perform path planning and obtain a comprehensive response plan; based on the comprehensive response plan, automatically generate and issue collaborative response instructions.

[0013] In this embodiment of the invention, scattered historical disaster data, geographic information, and satellite imagery are integrated to form a complete data view of the power grid area. A digital twin model provides an intuitive and comprehensive data carrier for subsequent risk analysis, achieving effective correlation and unified presentation of multi-source basic data. Risk feature clustering identifies the commonalities and differences in risks across different regions, and a risk level quantification matrix provides stronger data support for risk assessment results. A differentiated data collection strategy parameter table clarifies data collection needs, providing targeted basis for collection and improving the relevance and rationality of data collection. Parameter optimization ensures that the drone swarm task configuration is highly adapted to the characteristics of the risk area, and a collaborative observation scheme ensures the orderly operation of the drone swarm, guaranteeing that multi-source remote sensing data streams can cover the core characteristics of high-risk areas, reducing redundant data collection, and improving data collection efficiency and quality. A dual guarantee of 5G private network and satellite relay communication links avoids single-link transmission interruptions, achieving stable multi-source remote sensing data streams. Real-time transmission enables edge computing nodes to quickly acquire data, saving time for subsequent data processing; spatiotemporal registration eliminates deviations in time and space between multi-source data, providing a unified analytical benchmark; adaptive correction combined with differentiated acquisition parameters corrects data distortion; feature-level fusion extracts core and effective information from multi-source data, strengthening the intrinsic correlation of data and improving its usability and accuracy; multi-scale feature analysis mines disaster-related information in different dimensions of data; parallel disaster feature parsing improves data processing efficiency; key disaster parameters are quickly extracted; and the resulting quantitative disaster assessment report makes disaster information more organized and data-driven; disaster scenario mapping makes the connection between disaster conditions and power grid entities more intuitive; the integration of power grid topology and transportation network information makes route planning more aligned with actual operation and maintenance needs; comprehensive response plans provide guidance for subsequent actions; and the automatic generation and issuance of collaborative response instructions improves the timeliness and coordination of operation and maintenance responses.

[0014] In a preferred embodiment of the present invention, step 100, for the power grid area to be inspected, integrates historical disaster data, geographic information, and satellite imagery to construct a digital twin model, including: Step 101: Collect multi-source basic geographic data covering the power grid area to be inspected. This multi-source basic geographic data includes: high-resolution visible light and synthetic aperture radar remote sensing images, high-precision digital elevation model data, historically archived disaster point distribution data, and regional meteorological background field data containing temperature, humidity, and wind speed. Specifically, the data is collected using a collaborative approach combining satellite remote sensing, ground-based supplementary measurements, and system retrieval. All data are correlated with the WGS-84 coordinate system, and the collection time and source are labeled. The high-resolution visible light remote sensing images are acquired by the Gaofen-6 satellite, equipped with a panchromatic / multispectral camera, with a spatial resolution of no less than 0.8m, covering the blue, green, red, and near-infrared bands. Images are taken during clear, cloudless midday hours to reduce shadow interference. The synthetic aperture radar images are acquired by the Sentinel-1 satellite, using C-band, wide-swath interferometric mode, and VV and VH dual polarization, with a spatial resolution of 10m. The system simultaneously records radar transmission and reception timestamps and orbital parameters; high-precision digital elevation model data is combined with UAV lidar supplementary measurements and geospatial database retrieval, with supplementary measurements focusing on mountainous areas with weak satellite data coverage, lidar point cloud density not less than 50 points / square meter, and elevation accuracy error controlled within 0.1m; historical disaster point distribution data is retrieved from the power grid operation and maintenance management system and local emergency management department databases, covering the occurrence time, type, impact range, level, and damage to power grid facilities of disasters in the past 10 years; regional meteorological background field data is obtained through real-time collection by automatic weather stations along the power grid and API calls from meteorological departments, with a collection frequency of 15 minutes / time, and core parameters including air temperature (accuracy ±0.2℃), relative humidity (accuracy ±3%), average wind speed (accuracy ±0.3m / s), and instantaneous wind speed extreme values, with data coverage extending 5 kilometers beyond the boundary of the power grid area to be inspected.

[0015] Step 102 involves performing radiometric calibration, atmospheric correction, and geometric fine correction on the high-resolution visible light and synthetic aperture radar remote sensing images to obtain a standardized remote sensing image base. The high-precision digital elevation model data and regional meteorological background field data are then subjected to format standardization and spatial interpolation to obtain a topographic and meteorological environment dataset spatially aligned with the standardized remote sensing image base. Specifically, this includes performing specialized preprocessing on the collected multi-source data, with the core being specialized processing of the high-resolution visible light and synthetic aperture radar remote sensing images, and standardization and interpolation processing of the high-precision digital elevation model and regional meteorological background field data. Ultimately, this achieves pixel-level matching between each data layer and the standardized remote sensing image base.

[0016] Visible light remote sensing images are processed in the following order: radiometric calibration, atmospheric correction, and geometric fine correction. Radiometric calibration, based on the satellite sensor response function and combined with the satellite's built-in calibration coefficients, converts the image's digital quantization values ​​into surface reflectance. Atmospheric correction uses the FLAASH model, inputting atmospheric profile data such as air pressure, temperature, and water vapor content at the time of capture to eliminate the effects of atmospheric scattering and absorption. Geometric fine correction uses prominent features such as the tops of power grid towers and road intersections as control points, with no fewer than 15 control points per image and a planar error not exceeding 1 pixel. Coordinate correction is completed through quadratic polynomial fitting to generate a standardized remote sensing image base. High-precision digital elevation model data is uniformly converted to GeoTIFF format, with the resolution adjusted to match the standardized remote sensing image base. Data gaps are filled using Kriging interpolation, with interpolation errors controlled within 0.2m. Regional meteorological background field data is uniformly converted to NetCDF format, and discrete station data is transformed into continuous areal data using inverse distance weighted interpolation. The interpolation grid size is consistent with the image base pixels to ensure complete alignment of the topographic and meteorological datasets with the spatial location of the image base.

[0017] Step 103: Based on the standardized remote sensing image base, a multi-scale semantic feature tensor is obtained through forward propagation calculation using a pre-trained deep convolutional neural network; based on the multi-scale semantic feature tensor, region proposal and target classification are performed to complete the initial localization of key power facilities; pixel-level instance segmentation is performed on the initial localization to obtain a contour mask; the contour mask is vectorized and spatial coordinate transformed to obtain a vector dataset of the spatial distribution of power facilities. Specifically, this includes: using the standardized remote sensing image base as input data, the core of which is to complete feature extraction and target processing through a pre-trained deep convolutional neural network. This neural network uses Re... The model is built on the sNet-50 architecture, which includes multiple layers of alternating convolutional and pooling layers, a tail-connected fully connected layer, and a softmax activation function for the output layer to achieve multi-class probability output. The model is trained on a dataset containing 100,000 images of power grid scenes. The images in the dataset cover power grid facilities under different terrains and climates. Each image is labeled with the category and location information of power facilities such as towers, conductors, insulators, and fittings. The training process optimizes the network parameters through the backpropagation algorithm, calculates the error between the predicted and labeled values ​​using the cross-entropy loss function, and iterates until the model's classification accuracy on the validation set is stable and meets the target.

[0018] When applying the model, the standardized remote sensing image base is first scaled to 224×224 pixels and input into the network. Forward propagation is performed through convolutional and pooling layers. The convolutional layers extract features using a sliding window, and the pooling layers reduce the dimensionality of the feature map while retaining key information. This process sequentially acquires shallow texture features, mid-level structural features, and deep semantic features of the image, ultimately outputting a multi-scale semantic feature tensor with dimensions of 1024×14×14. Based on this feature tensor, a region proposal network is invoked to generate candidate target regions. The candidate region area threshold is set to 10 to 1000 pixels. Duplicate candidate regions are eliminated using the intersection-over-union (IoU) ratio, with an IoU threshold of 0.7. Then, combined with predefined power facility categories, target classification is completed, achieving the initial localization of key power facilities. Masking is used... The R-CNN algorithm performs pixel-level instance segmentation on the initially located power facility area. It outputs a binary contour mask for each facility through the generated mask branch, with a mask pixel accuracy of no less than 95%. The contour mask is then vectorized, and the Douglas-Puk algorithm is used to simplify the contour curve while retaining key feature points. Finally, the pixel coordinates are converted to WGS-84 geographic coordinates through coordinate transformation, resulting in a vector dataset of the spatial distribution of power facilities containing information such as facility type, geographic coordinates, contour size, etc.

[0019] Step 104: Based on the standardized remote sensing image base, extract and fuse multi-scale contextual features, perform pixel-by-pixel semantic segmentation to obtain the road pixel-level classification confidence distribution; perform topology optimization and denoising processing on the road pixel-level classification confidence distribution to obtain a binarized road mask; perform skeleton extraction and vectorization processing on the binarized road mask to obtain a road spatial distribution vector dataset. Specifically, this includes: using the standardized remote sensing image base as the processing object, the core process involves multi-scale feature extraction, semantic segmentation, and vectorization processing to generate a road spatial distribution vector dataset; employing a pyramid hierarchical strategy to extract features, first downsampling the images at 1x, 2x, and 4x scales respectively, then extracting road features at each scale using a convolutional neural network, and superimposing them through a feature fusion module to enhance feature recognition; and based on the fused feature map, performing segmentation using a U-Net architecture segmentation model. For pixel-by-pixel semantic segmentation, the model uses 5000 remote sensing images of labeled roads as training samples and outputs a pixel-level classification confidence distribution of roads (between 0 and 1). A confidence threshold of 0.8 is set, and pixels above the threshold are classified as road pixels, generating a preliminary road mask. The mask undergoes topology optimization and denoising: morphological opening operations are used to eliminate noise spots smaller than 50 pixels, and connected component analysis is used to repair road breaks with a width of 3 pixels or less, resulting in a complete binary road mask. The Zhang-Suen thinning algorithm is used to extract the skeleton from the mask, refining the road region into single-pixel-width skeleton lines while preserving connectivity and direction. The skeleton lines are vectorized, and feature points such as start points, end points, and inflection points are extracted and their geographic coordinates are calculated. Line fitting is used to connect the feature points, ultimately generating a vector dataset containing information such as road centerline coordinates, width, and grade.

[0020] Step 105: Integrate the aforementioned power facility spatial distribution vector dataset, road spatial distribution vector dataset, terrain and meteorological environment dataset, and historically archived disaster point distribution data to construct a multi-dimensional spatiotemporal background database that integrates terrain elevation, facility location, road network connectivity, meteorological conditions, and historical disaster information. Specifically, this includes: using GIS as the data carrier platform, integrating multiple types of data through layered storage and associated indexes, employing PostgreSQL and PostGIS management systems to ensure data security and access efficiency, and ultimately constructing a spatiotemporal background database that supports multi-condition queries; the power facility vector data from Step 103 is stored as the core layer in the power facility layer, including facility identifiers, The data includes attributes such as category and coordinates; the road vector data from step 104 is stored in the traffic network layer, recording information such as road signs and grades; the terrain and meteorological data from step 102 are stored in the terrain elevation and meteorological background raster layers respectively (time series storage of temperature, humidity, and other data); historical disaster point data is stored in the historical disaster vector layer, associated with disaster type, occurrence time, and facility damage information; a joint index is established using geographic coordinates and timestamps to achieve data association: power facilities and roads are topologically associated through a 10m radius buffer analysis; terrain and meteorological data are pixel-level matched with the former two through spatial coordinates; historical disaster data is associated with meteorological data through timestamps, and with power facilities and roads through the scope of impact.

[0021] Step 106: Based on the multi-dimensional spatiotemporal background database, data organization and spatial correlation are performed to obtain integrated scene data; based on the integrated scene data, the electrical topology relationships of the power grid are integrated to construct a data-topology association structure; based on the data-topology association structure, three-dimensional geometric reconstruction and dynamic rendering are performed to obtain a visual simulation environment; simulation and risk inference rules are configured in the visual simulation environment to complete the construction of an integrated space-air-ground digital twin model. Specifically, this includes: based on the multi-dimensional spatiotemporal background database, following a step-by-step process of data integration, topology construction, three-dimensional reconstruction, and rule configuration, a digital twin model is constructed on the GIS platform. The model supports core functions such as data updates, scene simulation, and risk inference; using an object-oriented modeling method, entities such as power facilities, roads, and terrain in the database are abstracted into data objects, and the attributes, methods, and relationships of each object are defined to generate logically consistent integrated scene data, laying the foundation for subsequent modeling; the electrical topology data of the power grid in the area to be inspected, including line connection relationships, is retrieved from the power grid dispatch management system. The system includes information such as the electrical number of the poles and towers, conductor type, and impedance parameters. Electrical topology data is linked to power facility objects in GIS using unique entity identifiers, constructing a two-way association structure between geospatial data and power grid electrical topology, enabling linked queries of geographical location and electrical attributes. Based on this association structure, ContextCapture software is used for 3D geometric reconstruction, using standardized remote sensing imagery as the texture data source and digital elevation models as the terrain foundation to create detailed models of entities such as power poles, conductors, and roads with centimeter-level accuracy. Dynamic rendering is performed using the Unity3D engine, configuring real-time lighting, atmospheric scattering, and other environmental effects to generate a visual simulation environment supporting 360-degree panoramic browsing and multi-view switching. Based on historical disaster data and physical models, simulation and risk projection rules are configured in the simulation environment for wildfire spread speed, conductor icing and its correlation with temperature and humidity, and the impact of geological disasters on pole foundations. These rules are then linked with real-time data from a spatiotemporal background database to achieve dynamic projection of disaster scenarios, ultimately completing the construction of an integrated space-air-ground digital twin model.

[0022] In a preferred embodiment of the present invention, step 200 involves performing risk clustering and quantitative assessment on the transmission line corridor based on the risk distribution output by the digital twin model, obtaining a risk level quantification matrix; and obtaining a differentiated acquisition strategy parameter table based on the risk level quantification matrix, including: Step 201: Based on the risk distribution output by the digital twin model, calculate the spatial proximity of each risk point within the transmission line corridor and construct a risk spatial density distribution field. Specifically, this includes: using the power grid risk distribution data output by the digital twin model as the core input. This risk distribution data includes the geographical coordinates, risk type labels, and preliminary risk probability values ​​of each risk point within the transmission line corridor to be inspected. The geographical coordinates are in the WGS-84 coordinate system, and the risk type labels include: wildfire hazard, icing risk, geological instability, etc.; first, determine the spatial extent of the transmission line corridor, extending 50m to both sides of the line centerline to form a rectangular area as the analysis boundary, and excluding areas outside this boundary. Invalid risk point data; then, the spatial proximity of each risk point is calculated, and the straight-line distance between any two risk points is calculated using the Euclidean distance formula, with the distance calculation accuracy retained to 0.1m; a risk spatial density distribution field is constructed based on the spatial proximity, using the kernel density estimation method, setting the search radius to 50m, matching the width of one side of the line corridor, and dividing the line corridor area into 10m×10m grid cells, calculating the risk point density in each grid cell, with the density value being the ratio of the number of risk points in the cell to the grid area, and finally outputting a risk spatial density distribution field containing density information in grid units, with the data format being raster data, consistent with the spatial benchmark of the digital twin model.

[0023] Step 202: Based on the risk spatial density distribution field, identify and mark the core risk point set by setting a neighborhood radius and a minimum point threshold. Specifically, this includes: setting two core judgment indicators based on the risk spatial density distribution field: neighborhood radius and minimum point threshold; the neighborhood radius is set differently based on the risk type characteristics, with a neighborhood radius of 30m for risks such as wildfires and geological disasters, and a neighborhood radius of 20m for risks such as icing. This parameter is calibrated using historical disaster spread range statistics; the minimum point threshold is related to the grid density. When the risk point density of the grid cell is greater than or equal to 0.05 points / square meter, the minimum point threshold is set to 8; when... When the density is between 0.02 and 0.05 points / square meter, the threshold is set to 5; when the density is less than 0.02 points / square meter, the threshold is set to 3 to ensure that the core point identification accuracy is balanced in different density areas. For each risk point in the risk spatial density distribution field, a neighborhood analysis is performed. Taking the point as the center, the number of other risk points included in the neighborhood radius of the corresponding risk type is counted. If the number is greater than or equal to the corresponding minimum point threshold, the point is marked as a core risk point. At the same time, the correlation relationship of all risk points in its neighborhood is recorded, and finally a core risk point set containing the coordinates of the core risk point, the risk type, and the number of associated points is formed, which is stored in vector data format.

[0024] Step 203: Based on the core risk point set, clusters are expanded and merged through connectivity analysis and density reachability judgment to generate an initial risk cluster partition. The initial risk cluster partition includes both clusters and noise points. Specifically, it includes: using the core risk point set as the starting data, cluster expansion and merging are completed through connectivity analysis and density reachability judgment; density reachability is defined as: if risk point A is a core risk point, and risk point B is within the neighborhood of A, then B and A are density reachable; if B is a core risk point, and risk point C is within the neighborhood of B, then C and A are indirectly density reachable; firstly, a core risk point in an unassigned cluster is selected as the starting point, and all density reachable risk points in its neighborhood are included in the same initial cluster. This process is repeated until all core risk points have been assigned to clusters, with each cluster assigned a unique cluster identifier. Then, a cluster merging operation is performed. The distance between the cluster centers of any two clusters is calculated, where the cluster center is the average of the coordinates of all risk points within the cluster. If two clusters belong to the same risk type and their cluster center distance is less than or equal to 1.5 times the neighborhood radius of the corresponding risk type, the two clusters are merged into a new cluster, and the cluster identifier and the set of risk points within the cluster are updated. Risk points not covered by the neighborhood of any core risk point are marked as noise points and not included in any cluster. Finally, the initial risk clustering result, containing the clusters and the set of noise points, is output. Each cluster includes a cluster identifier, risk type, set of risk points within the cluster, and cluster center coordinates.

[0025] Step 204 involves performing a statistical consistency test on the internal risk feature vectors of each cluster in the initial risk clustering partition, and re-discriminating the noisy points at the boundaries to obtain the final risk clustering partition set. Specifically, this includes: performing a statistical consistency test on the internal risk feature vectors of each cluster in the initial risk clustering partition. The risk feature vectors include three core indicators: risk probability value, risk development trend coefficient, and environmental impact factor. A chi-square test is used to calculate the deviation between the feature vectors of all risk points within a cluster and the mean feature vector within the cluster. A significance level of 0.05 is set. If the test result shows that the deviation is within the allowable range, the cluster is considered to have consistent internal features, and the cluster structure is retained. If the deviation exceeds the allowable range... The cluster is then split into 2 to 3 sub-clusters based on the clustering results of the feature vectors, and the cluster identifiers are reassigned. For noise points located at the boundaries of clusters, the feature similarity between them and each neighboring cluster is calculated using the cosine similarity algorithm, with similarity values ​​ranging from 0 to 1. If the maximum similarity value is greater than or equal to 0.75, the noise point is assigned to the corresponding cluster. If the maximum similarity value is less than 0.75, the noise point attribute is maintained, but its spatial location and feature information are recorded for subsequent analysis. The spatial boundaries of each optimized cluster are determined, and the minimum bounding polygon of the cluster is generated using the convex hull algorithm as the cluster partition boundary. Finally, a risk cluster partition set containing partition boundaries, cluster identifiers, risk types, and internal risk point statistics is formed.

[0026] Step 205: For each risk clustering unit in the risk clustering partition set, extract the terrain, vegetation, and historical disaster features within its spatial range based on the digital twin model to form a multi-dimensional risk feature vector. Specifically, this includes: for each risk clustering unit in the risk clustering partition set, extracting feature information based on the multi-dimensional spatiotemporal background database of the digital twin model to construct a multi-dimensional risk feature vector; regarding terrain features, extracting the average elevation, maximum slope, and terrain complexity index within the partition, where the terrain complexity index is calculated as the ratio of the elevation standard deviation within the partition to the area of ​​the region; regarding vegetation features, extracting the vegetation coverage and dominant vegetation types within the partition. The data includes the average vegetation type and height, with dominant vegetation types including coniferous forests, broad-leaved forests, and shrubs. The data comes from the semantic segmentation results of remote sensing images from a digital twin model. Regarding historical disaster characteristics, the data extracts the number of historical disasters, the proportion of disaster types, and the maximum disaster intensity level within a 1-kilometer radius of the zone over the past 10 years. The intensity level is divided into 1 to 5 levels according to industry standards. Each feature indicator corresponds to a specific value or code. For example, dominant vegetation types are represented by numerical codes, such as 1 for coniferous forests, 2 for broad-leaved forests, etc. Finally, each cluster unit forms a multi-dimensional risk feature vector containing 8 feature indicators. The dimensions of the vector are arranged in the order of topography, vegetation, and historical disasters.

[0027] Step 206 involves normalizing the multi-dimensional risk feature vector to generate a standardized risk feature vector. Specifically, this includes: normalizing the multi-dimensional risk feature vector to eliminate dimensional differences between different feature indicators; for numerical features, such as average elevation, vegetation coverage, and historical disaster frequency, using a min-max normalization method to map feature values ​​to a range of 0 to 1, with the mapping process based on the maximum and minimum values ​​of the feature across all clustering units to ensure that the normalized data reflects the relative level of each unit within the whole; for categorical coding features, such as dominant vegetation types and the proportion of historical disaster types, using one-hot coding to convert a single code into multiple 0-1 variables, for example, converting vegetation type code 1 (coniferous forest) into a vector form of [1, 0, 0, ...] to ensure that categorical features can participate in subsequent quantitative assessment calculations; after completing the standardization of all features, reorganizing them according to the original feature order to obtain a standardized risk feature vector, with each vector having a unified dimension, providing standardized input for subsequent weighted assessment.

[0028] Step 207: Based on the standardized risk feature vector, a weighted assessment is performed to generate a quantitative risk level value for each unit. Specifically, this includes: a weighted assessment based on the standardized risk feature vector, with weights determined using an analytic hierarchy process (AHP) combined with expert scoring to construct a hierarchical model. The target layer is the risk level assessment; the criteria layer comprises three major feature categories: terrain, vegetation, and historical disasters; and the indicator layer comprises specific feature indicators. Ten experts in power grid operation and maintenance and disaster prevention are invited to score the relative importance of each criterion and indicator layer using a 1-9 scale. The rationality of the scoring results is verified through a consistency test (consistency ratio CR less than 0.1). Finally, the weight values ​​of each feature indicator are calculated, with relatively high weights for the number of historical disasters and the maximum disaster intensity level (0.25 and 0.2 respectively), while the terrain complexity index and vegetation coverage each have a weight of 0.15, and the remaining indicators have a combined weight of 0.25. The quantitative risk level value for each cluster unit is obtained by multiplying the values ​​of each dimension of the standardized risk feature vector by their corresponding weights and summing the results. The quantitative value ranges from 0 to 10, with a higher value indicating a higher risk level.

[0029] Step 208 involves organizing the quantified values ​​of all units into a risk level quantification matrix based on spatial topological relationships. Specifically, this includes: organizing the quantified risk level values ​​of all risk clustering units into a risk level quantification matrix based on spatial topological relationships; firstly, determining the row and column indexing rules of the matrix, using the tower number of the transmission line as the vertical index (row index), with each tower corresponding to 1 to 3 clustering units under its jurisdiction; and using the segment number of the line as the horizontal index (column index), with each segment containing 5 to 8 consecutive towers corresponding to clustering units; each element of the matrix contains three core pieces of information: a unique identifier for the clustering unit, a quantified risk level value, and spatial range coordinates, where the spatial range coordinates are represented by the vertex coordinates of the clustering partition boundary; the matrix is ​​stored in a two-dimensional array format and is associated with the spatial topological relationships of the digital twin model, supporting rapid location of the corresponding clustering partition in the digital twin model through matrix elements.

[0030] Step 209: Based on the risk level quantization matrix, and combined with preset UAV platform performance parameters and sensor physical constraints, the quantization values ​​of each unit are transformed and boundary-checked using a preset risk-strategy mapping function to obtain an optimized strategy parameter set for each unit. The optimized strategy parameter set includes at least spatial sampling density, sensor operating frequency, and flight altitude. Specifically, it includes: generating an optimized strategy parameter set based on the risk level quantization matrix, combined with preset UAV platform performance parameters and sensor physical constraints. The UAV platform performance parameters include a maximum flight altitude of 1500m, a minimum flight altitude of 50m, a maximum speed of 20m / s, and a maximum of 4 sensors. The sensor physical constraints include a maximum visible light camera resolution of 8000×6000 pixels, an infrared thermal imager temperature detection range of -40℃ to 150℃, and a maximum lidar point cloud density of 500 points / square meter. The preset risk-strategy mapping function... The data is divided into 5 intervals based on quantized values: 0 to 2 (low risk), 2 to 4 (relatively low risk), 4 to 6 (medium risk), 6 to 8 (relatively high risk), and 8 to 10 (high risk). Each interval corresponds to a set of basic strategy parameters. The spatial sampling density of the high-risk interval is set to 10 points / square meter, the sensor operating frequency is 10 Hz, and the flight altitude is 80 m. The relatively high-risk interval is adjusted to 8 points / square meter, 8 Hz, and 100 m respectively; the medium-risk interval is 5 points / square meter, 5 Hz, and 150 m; the relatively low-risk interval is 3 points / square meter, 3 Hz, and 200 m; and the low-risk interval is 1 point / square meter, 1 Hz, and 300 m. Boundary checks are performed on the mapped basic parameters to ensure that the flight altitude is within the range of 50 to 1500 m and the sampling density does not exceed the maximum performance of the sensor. If the basic parameters exceed the constraint range, they are adjusted to the nearest constraint threshold, thus forming the optimized strategy parameter set for each cluster unit.

[0031] Step 210: Organize the optimization strategy parameter sets of all units into a differentiated acquisition strategy parameter table according to their spatial indexes. Specifically, this includes: organizing the optimization strategy parameter sets of all risk clustering units into a differentiated acquisition strategy parameter table according to their spatial indexes. The spatial index adopts a three-level coding format of line number, tower number, and clustering unit number, encoded as a 10-character string. The first two characters are the line number, the middle three characters are the tower number, and the last five characters are the clustering unit number, ensuring that the index of each unit is unique. The parameter table is organized in tabular form and contains seven fields: spatial index code, clustering unit spatial range, risk level quantification value, and spatial index. The system includes sampling density, sensor operating frequency, flight altitude, and recommended sensor types. Recommended sensor types are matched based on risk type: multispectral sensors and infrared thermal imagers are prioritized for wildfire hazards; lidar and infrared thermal imagers are prioritized for icing risks; and synthetic aperture radar and visible light cameras are prioritized for geological disasters. The parameter table supports filtering by risk level, route segmentation, spatial location, and other conditions. Data is saved in both Excel and XML formats. Excel format is used for manual review, while XML format is used for automatic parsing and parameter loading by the UAV mission planning system, enabling multi-scenario data adaptation.

[0032] In a preferred embodiment of the present invention, step 300 above, which involves using a differentiated acquisition strategy parameter table to perform parameterized tuning of the UAV swarm mission to obtain a collaborative observation scheme, and scheduling the UAV swarm to acquire multi-source remote sensing data streams according to the collaborative observation scheme, includes: Step 301: Based on the differentiated acquisition strategy parameter table, perform range estimation to obtain an initial range estimation matrix. Specifically, this includes: using the differentiated acquisition strategy parameter table as the core input, first extracting the spatial coordinates of each risk cluster unit; using the minimum bounding rectangle method to determine the latitude and longitude of the unit boundary; calculating the coordinates of the unit center point as the core reference point for UAV operations; combining the UAV platform performance parameters (endurance 240 minutes, cruising speed 20 m / s, hovering energy consumption 1.5 times that of cruising) to perform range estimation. The range calculation covers three parts: the round-trip range from the take-off / landing point to the first cluster unit, the operational range within the cluster unit (i.e., calculating the flight path density based on the sampling density), and the high... The risk zone uses a grid-like path, the low-risk zone uses a polygonal path, and the transfer range between units is calculated as a straight-line distance, with an additional 5% airspace avoidance redundancy. At the same time, the impact of the sensor's working mode on the range is considered. For example, when the lidar is turned on, the UAV speed needs to be reduced to 15m / s, and the range calculation coefficient is adjusted accordingly. The estimated range, estimated energy consumption, and operation time of each UAV (distinguished by numbers 1 to N) for each cluster unit are organized into an initial range estimation matrix. The matrix row index is the UAV number, the column index is the cluster unit spatial index, and the matrix elements include the range in kilometers, energy consumption in kilowatt-hours, and operation time in minutes. The data precision is retained to 0.1 kilometers, 0.1 kilowatt-hours, and 1 minute, respectively.

[0033] Step 302: Based on the initial range estimation matrix, perform task bundling and allocation to obtain a conflict-free task allocation scheme. Specifically, this includes: generating a scheme based on the initial range estimation matrix using a combination of task bundling and intelligent allocation. Task bundling prioritizes bundling spatially adjacent risk clustering units (unit center point distance less than or equal to 1 km) into task packages. Within the same task package, the risk type of units must match the sensor configuration of the same UAV. For example, units containing wildfire hazards are bundled and allocated to UAVs equipped with multispectral and infrared sensors. Task allocation uses an improved genetic algorithm, with the optimization objectives being the shortest total range, the most balanced energy consumption, and the earliest task completion time. The algorithm parameters are set to a population size of 50 and iteration... The algorithm has 100 iterations, a crossover probability of 0.7, and a mutation probability of 0.05. A dual conflict detection mechanism is introduced during the allocation process: airspace conflict detection is achieved by setting a buffer zone for UAV operations (horizontal radius 50m, vertical height 20m) to avoid overlapping operations by different UAVs in the same time and space segment; time conflict detection is achieved by planning the execution time periods of each task package to ensure that the same clustering unit is not repeatedly operated by multiple UAVs, with a time interval set to 5 minutes to allow for data verification time; for allocation results with conflicts, corrections are made by adjusting the task execution order or changing the executing UAV, ultimately outputting a conflict-free task allocation scheme. The scheme includes a task package list for each UAV, execution order, estimated start and end times, and energy consumption budget.

[0034] Step 303: Based on the conflict-free task allocation scheme, the final track point sequence and sensor operating timing of each UAV are obtained through track smoothing and cooperative rendezvous planning. Specifically, this includes: determining core parameters based on the conflict-free task allocation scheme using track smoothing and cooperative rendezvous planning techniques; optimizing the initially planned polygonal track using cubic B-spline curves, selecting track inflection points as curve control points, and adjusting the spacing between control vertices according to flight altitude (20m at 80m altitude, 50m at 300m altitude) to ensure the track curvature change rate is less than or equal to 0.05 radians / meter, avoiding drastic attitude adjustments by the UAVs; and for scenarios where multiple UAVs operate on the same line segment, setting the rendezvous point 10m directly above the top of the line tower during rendezvous. The time window is accurate to the second. By adjusting the takeoff time and flight speed of each UAV, it is ensured that the horizontal distance between UAVs is greater than or equal to 50m and the vertical distance is greater than or equal to 10m when they meet. The final waypoint sequence is represented by WGS-84 coordinates. Each waypoint includes longitude, latitude, altitude and arrival time. The point interval is determined according to the sampling density, with an interval of 5m in high-risk areas and 20m in low-risk areas. The sensor working sequence is bound to the waypoint sequence. When a UAV arrives at a waypoint, the sensor working command is triggered. The working frequency is executed according to the strategy parameters. For example, a 10Hz frequency in high-risk areas corresponds to collecting data once every 0.1 seconds. At the same time, the sensor exposure time, gain value and other parameters are recorded to ensure that the sensor detection timing of different UAVs in the same area is synchronized and the time deviation is controlled within 100 milliseconds.

[0035] Step 304: Integrate the conflict-free task allocation scheme, the final waypoint sequence, and the sensor operating timeline to construct a collaborative observation scheme. Specifically, this includes integrating three core modules: the conflict-free task allocation scheme, the final waypoint sequence, and the sensor operating timeline, to build a structurally complete collaborative observation scheme. The scheme adopts a layered architecture design. The first layer is the task overview layer, including task number, execution date, UAV cluster size, total area of ​​the risk zone covered, and estimated total operation time. The second layer is the individual UAV task layer, corresponding to each UAV number, including the UAV's task package details, waypoint sequence file path, sensor configuration list, and energy consumption budget. The third layer is the collaborative control layer, including the communication frequency between UAVs, status feedback cycle, and emergency response. The mechanism and data transmission protocol are defined, with the communication frequency between UAVs (i.e., the 5G private network communication frequency) set to 2.4GHz, and the status feedback cycle being 1 second / time. The solution embeds spatiotemporal reference calibration information, uniformly adopting UTC time and the WGS-84 coordinate system to ensure data benchmark consistency across modules. Simultaneously, solution verification fields are added, including trajectory feasibility verification results and sensor compatibility verification results. The trajectory feasibility verification result checks for flight attitudes exceeding UAV performance, while the sensor compatibility verification result checks for sensor operating frequency and flight speed matching. If verification fails, the previous steps are returned for re-optimization. The final solution is stored in XML format, supporting direct parsing by the UAV ground control station system, and also generates a visual version for manual review.

[0036] Step 305: Based on the aforementioned collaborative observation scheme, analyze and generate collaborative control commands for the UAV. Specifically, this includes: first, extracting the core parameters of the individual UAV task layer, converting the waypoint sequence into flight control parameters recognizable by the UAV, including heading angle, pitch angle, roll angle, and flight speed, where the heading angle accuracy is retained to 0.1 degrees and the speed accuracy to 0.1 m / s; sensor control commands are generated according to the working sequence, including sensor startup time, working mode, acquisition frequency, and data storage path. For example, the infrared thermal imager startup command needs to include temperature range setting (-40℃ to 150℃) and frame rate parameters. The frequency (10Hz) is used for data synchronization. The coordinated commands include cluster coordination parameters, such as the formation distance between UAVs, data synchronization trigger signals, and rendezvous point rendezvous time, to ensure coordinated actions of multiple UAVs. The commands are encapsulated using the MavLink communication protocol. Each command frame contains the command type, target UAV number, parameter data, and checksum. The checksum is generated using a cyclic redundancy check algorithm to avoid command transmission errors. The commands are divided into pre-takeoff commands, flight control commands, sensor control commands, and return-to-home commands according to the execution order. The pre-takeoff commands are generated and issued 10 minutes in advance, and the remaining commands are dynamically generated according to the real-time flight path progress to ensure the timeliness of the commands.

[0037] Step 306: Based on the cooperative control command, control the UAV cluster to perform cooperative flight and synchronize sensor detection to obtain raw data. Specifically, this includes: based on the cooperative control command, the ground control station sends control signals to the UAV cluster through a dual-link communication system of 5G private network and satellite relay. The dual links adopt a load balancing mechanism. Under normal circumstances, the 5G private network undertakes the main communication tasks. When the 5G signal strength is lower than -85dBm, it automatically switches to the satellite relay link to ensure communication continuity. After receiving the command, the UAVs perform cooperative flight. During the takeoff phase, they take off sequentially at 30-second intervals according to a preset order to avoid airspace congestion. During the flight phase, the UAVs provide real-time feedback on their status to the ground control station. Data, including current location, flight attitude, battery level, and sensor operating status, is fed back every 1 second. The ground control station monitors the cluster's operational status in real time through a situational awareness display interface. Synchronous sensor detection is performed according to the working sequence. When the UAV reaches the target waypoint and its attitude is stable (attitude error less than or equal to 0.5 degrees), the sensor acquisition action is triggered. The collected raw data includes visible light images, infrared thermal images, lidar point clouds, and ultraviolet imaging data. The raw data is stored in a dual-mode system of local caching and real-time transmission. Local storage uses a high-speed SD card with a read / write speed of greater than or equal to 100MB / s. Real-time transmission uses a compression algorithm (JPEG 2000 compression format, compression ratio adjusted as needed, 10:1 compression ratio in high-risk areas, 20:1 compression ratio in low-risk areas) to reduce the data volume, ensure smooth transmission, and avoid loss of raw data.

[0038] Step 307 involves performing spatiotemporal synchronization and format standardization on the raw data to obtain a multi-source remote sensing data stream. Specifically, this includes: performing spatiotemporal synchronization and format standardization on the raw data; spatiotemporal synchronization using the UAV's GPS timestamp and IMU inertial measurement unit data as a reference to associate the data collected by different sensors with the same spatiotemporal coordinates; time synchronization using GPS pulse signal calibration to uniformly correct the timestamps of each sensor's data to UTC time, with time accuracy controlled within 1ms; and spatial synchronization using the UAV's real-time position and attitude data to convert the pixel coordinates or point cloud data collected by the sensors into WGS-84 geographic coordinates, ensuring that the multi-source data for the same observation target are spatially synchronized. The data is positioned consistently; format standardization converts different types of data into a unified format, with visible light and infrared images converted to TIFF format and metadata such as geographic coordinates, acquisition time, and sensor parameters added; LiDAR point clouds are converted to LAS format, including the three-dimensional coordinates, reflectance intensity, and classification information of the points; ultraviolet imaging data are converted to PNG format; the standardized data undergoes quality screening to remove invalid data such as blurry images (resolution below 300 dpi) and noisy point clouds (reflectance intensity below the threshold), and finally the data is organized according to the naming rules of UAV number, acquisition time, and risk unit index to form a structured multi-source remote sensing data stream, supporting real-time processing and analysis by subsequent edge computing nodes.

[0039] In a preferred embodiment of the present invention, step 400, which transmits the multi-source remote sensing data stream to the edge computing node in real time via a 5G private network and a satellite relay communication link, includes: Step 401 involves parsing, classifying, and prioritizing the multi-source remote sensing data stream to obtain a priority data sequence. This priority data sequence is then encapsulated, protocol-converted, and time-series reassembled to obtain a standardized data packet queue. Specifically, this includes: taking the multi-source remote sensing data stream as input, which contains various types of data such as visible light images, infrared thermal maps, and lidar point clouds, with each data entry associated with a risk clustering unit index, acquisition timestamp, and sensor parameters; firstly, data parsing and classification are performed, distinguishing data categories through the type identifier field in the data header, specifically dividing them into four categories: high-risk area critical facility data, high-risk area environmental data, medium- and low-risk area facility data, and medium- and low-risk area environmental data. High-risk area critical facility data includes, for example, infrared data of insulators in risk areas of levels 8 to 10, while high-risk area environmental data includes multispectral data of wildfires in the same region. Priority sorting is then performed based on risk level and data timeliness requirements, using a 1-4 level priority mechanism, with level 1 being the highest priority and level 4 the lowest. High-risk area critical facility data includes, for example, infrared data of insulators in risk areas of levels 8 to 10, and multispectral data of wildfires in the same region. Key facility data is set to Level 1, requiring a transmission latency of less than or equal to 500ms; high-risk area environmental data is set to Level 2, with a latency of less than or equal to 1s; medium- and low-risk area facility data is set to Level 3, with a latency of less than or equal to 3s; and medium- and low-risk area environmental data is set to Level 4, with a latency of less than or equal to 5s. During the sorting process, if data with the same priority appears, it is arranged in ascending order of collection timestamp to ensure that data collected earlier is transmitted first. TLV type, length, and value format encapsulation is used, with the header containing priority identifier, data type, timestamp, sequence number, and checksum fields. The data body contains the original data content, and a 4-byte cyclic redundancy checksum is added at the end. The proprietary transmission protocols of different sensors are uniformly converted to the IP network protocol, with Level 1 and 2 data using UDP to ensure real-time performance and Level 3 and 4 data using TCP to ensure reliability. Time-series reassembly is performed based on timestamps and sequence numbers, eliminating abnormal data with timestamp deviations exceeding 1s to ensure that multi-source data from the same scene is arranged in the order of collection, ultimately forming a standardized data packet queue sorted by priority.

[0040] Step 402: Based on preset link status and transmission strategies, the standardized data packet queue is allocated to the 5G private network link and the satellite relay communication link to obtain the data packets allocated to the 5G private network link and the satellite relay communication link. Specifically, this includes: completing data packet allocation based on preset link status judgment indicators and transmission strategies. The preset indicators include real-time status parameters of the 5G private network link and the satellite relay communication link: 5G link indicators include signal strength (dBm), packet loss rate (%), and bandwidth (Mbps); satellite link indicators include signal-to-noise ratio (dB), link latency (ms), and bit error rate (%). These indicators are collected and updated every 100ms by the link status monitoring module. The preset transmission strategy dynamically adapts to the link status according to priority: Level 1 priority data is preferentially allocated to the 5G private network link, while backup transmission is performed on the satellite link. The system is configured to form a primary / backup dual-transmission mode. Level 2 priority data is allocated based on the 5G link signal strength: data with a signal strength greater than or equal to -85dBm is assigned to the 5G link, while data with a signal strength less than -85dBm is switched to the satellite link. Level 3 and 4 priority data are allocated using a load balancing method: data with a 5G link bandwidth utilization rate less than 70% is prioritized for allocation to the 5G link, while data with a utilization rate greater than or equal to 70% is allocated to the satellite link. During link allocation, a conflict avoidance mechanism is implemented: data of the same type with the same timestamp will not simultaneously occupy the core bandwidth of two links. 30% of the 5G link core bandwidth is reserved for Level 1 data burst transmission, and 20% of the satellite link core bandwidth is reserved for backup data transmission. After allocation, a link identifier field is added to the data packets of each link: G for 5G links and S for satellite links, facilitating subsequent cross-link fusion identification.

[0041] Step 403: For the data packets allocated to the 5G private network link, perform forward error correction coding and flow control based on the transmission control protocol to obtain the 5G link transport stream. Specifically, this includes: performing targeted processing on the data packets allocated to the 5G private network link; firstly, performing forward error correction coding using RS (Reed-Solomon) coding, with the coding parameters set to RS(255, 239), i.e., adding 16 bytes of checksum for every 239 bytes of data, which can correct consecutive errors within 8 bytes. This parameter is adapted to the typical error characteristics of the 5G link; secondly, performing flow control based on the transmission control protocol, using a sliding window mechanism, with the window size dynamically adjusted according to the 5G link bandwidth: the window size is greater than or equal to 50Mbps. The transmission rate is set to 8192 bytes, 4096 bytes when the bandwidth is between 20 and 50 Mbps, and 2048 bytes when the bandwidth is less than 20 Mbps. A congestion control strategy is also set: when a link packet loss rate of 2% or higher is detected, a slow start mechanism is triggered, reducing the transmission rate to 50% of the current rate. The rate is gradually restored once the packet loss rate is less than 0.5%. Processed data packets are arranged into 5G link transport streams according to priority and timestamp order. A transport stream header is added, containing the stream identifier, total number of data packets, and start timestamp. The data is transmitted through the edge node access network of the 5G private network. During transmission, ACK confirmation signals are fed back in real time. Data packets that are not confirmed are automatically retransmitted after 100ms to ensure transmission reliability.

[0042] Step 404: For the data packets allocated to the satellite relay communication link, perform efficient compression, anti-interference channel coding, and adaptive modulation based on link quality to obtain the satellite link transmission stream. Specifically, this includes: Addressing the high latency and susceptibility to interference inherent in satellite relay communication links, the allocated data packets undergo triple processing. First, efficient compression is performed, employing differentiated compression algorithms based on data type: visible light and infrared images use the JPEG 2000 compression standard, with a compression ratio of 10:1 for high-risk areas to preserve details and 20:1 for medium- and low-risk areas to improve efficiency; lidar point clouds use an improved LZ77 algorithm, reducing data volume by deleting duplicate point cloud coordinates, achieving a compression ratio of 8:1; text-based sensor parameters use Huffman coding, with a compression ratio of approximately 3:1; anti-interference channel coding is performed using LDPC (Low-Density Parity-Check) coding at a code rate of 1 / 2, constructing a sparse parity-check matrix to achieve dual correction for random and burst errors, adapting to the complex channel environment of the satellite link; adaptive modulation is then performed based on link quality, through real-time monitoring... Signal-to-noise ratio (SNR) modulation method: 16QAM modulation is used when the SNR is greater than or equal to 15dB to improve the transmission rate; QPSK modulation is used when the SNR is between 8 and 15dB to balance the rate and reliability; BPSK modulation is used when the SNR is less than 8dB to ensure basic transmission capability; the processed data packets are organized into a satellite link transmission stream according to the frame structure. Each frame contains a frame synchronization word, frame sequence number, data length, and check field. The frame synchronization word uses a fixed sequence 1A2B3C4D to ensure accurate frame synchronization at the receiving end. It is sent to the relay satellite through the uplink of the satellite ground station, and then transmitted to the satellite receiving module of the edge computing node through the satellite downlink.

[0043] Step 405: The 5G link transport stream and the satellite link transport stream are parsed, decoded, and their integrity verified to obtain a set of valid data packets. Based on the timestamps and sequence numbers of the data packets in the set of valid data packets, cross-link alignment and deduplication fusion are performed to obtain a fused data packet stream. Specifically, this includes: the dual-link receiving module of the edge computing node receives the 5G link transport stream and the satellite link transport stream respectively; first, parsing and decoding are performed: the transport stream header and frame structure are stripped; RS decoding is performed on the 5G link data to remove error correction codes; LDPC decoding and decompression are performed on the satellite link data sequentially to restore it to a standardized data packet format; data packet integrity verification is performed by verifying the trailing cyclic redundancy check code to determine whether the data is complete; data packets that fail the verification are marked as... Invalid and discarded data packets are recorded along with their sequence number and link identifier. If the data is of priority level 1, a retransmission request is triggered. Valid data packets that pass verification are categorized and organized by timestamp and sequence number to form a set of valid data packets indexed by collection timestamp and data type. Coarse alignment is performed based on timestamp, grouping data packets with timestamp deviations of less than or equal to 100ms into the same data group. Fine alignment is performed based on sequence number to ensure that data packets within the same data group are arranged in the order of transmission. The deduplication rule is to retain the earliest arriving and complete data packet. If both the 5G link and the satellite link transmit the same data packet (by matching the sequence number), the 5G link data is retained and the satellite link backup data is discarded. After fusion, the data packets are arranged in ascending order by timestamp to form a continuous and complete fused data packet stream.

[0044] Step 406 involves restoring the fused data packet stream to generate a reconstructed data stream consistent with the multi-source remote sensing data stream, and pushing the reconstructed data stream to the edge computing node. Specifically, this includes: performing format restoration processing on the fused data packet stream, stripping the TLV encapsulation header and link identifier field, and converting the IP network protocol back to the original data protocol that the edge computing node can directly process. Image data is restored to TIFF / LAS format, and sensor parameters are restored to JSON format, ensuring that the data format is completely consistent with the multi-source remote sensing data stream output in step 307. During the restoration process, data consistency verification is performed, comparing key features of the reconstructed data with the original data stream, such as image resolution, number of point clouds, and parameters. If the deviation exceeds 5%, a data verification process is triggered, and the data packet with the corresponding sequence number is retrieved again for secondary restoration. After confirmation of consistency, a reconstructed data stream is generated, and a data transmission completion marker is added, which includes statistical information such as transmission success rate and average latency. The reconstructed data stream is pushed to the processing module of the edge computing node through the internal high-speed bus. The push adopts the memory-mapped file method to avoid the latency caused by data copying, ensuring that the total latency of Level 1 priority data from collection to push to the processing module is less than or equal to 1 second, meeting the minute-level response requirements of disaster emergency. At the same time, the transmission statistics are synchronized to the transmission monitoring interface of the digital twin model, so that the operation and maintenance personnel can grasp the data transmission status in real time.

[0045] In a preferred embodiment of the present invention, in step 500 above, the edge computing node receives multi-source remote sensing data streams and performs spatiotemporal registration to obtain spatiotemporal registration data; based on the spatiotemporal registration data, combined with differentiated acquisition strategy parameters, adaptive correction and feature-level fusion are performed to obtain multidimensional fused data, including: Step 501: The edge computing node receives the reconstructed data stream as a multi-source remote sensing data stream. Specifically, the edge computing node acts as the core carrier for data processing, and its hardware configuration is adapted to the requirements of parallel processing of multi-source data. Specifically, it is equipped with an Intel Xeon Gold 6330 processor (28 cores and 56 threads), 128GB DDR4 memory, and a 2TB NVMe high-speed solid-state drive. It integrates a 5G private network receiving module and a satellite data receiving card, supporting simultaneous access to dual-link data. The edge computing node establishes a connection with the dual-link receiving module through an internal high-speed data bus and receives the reconstructed data stream using an interrupt-driven method, identifying it as the multi-source remote sensing data stream to be processed. During the reception process, a data integrity pre-verification is initiated. The data integrity is determined by verifying the transmission completion flag field in the data stream header. If the flag is 1, the reception is confirmed and stored in the high-speed cache. If the flag is 0, a retransmission request is immediately sent to the transmission module to ensure that the received data is complete, laying a complete data foundation for subsequent processing. This reception mechanism keeps the data reception latency within 50ms, meeting the real-time requirements of disaster emergency response.

[0046] Step 502: The multi-source remote sensing data stream is parsed and its types separated to obtain visible light imagery, infrared thermal imagery, and lidar point cloud data streams. Spatiotemporal reference information is extracted from each of the visible light imagery, infrared thermal imagery, and lidar point cloud data streams to obtain reference data with original spatiotemporal stamps. Specifically, the edge computing node calls a preset multi-source data parsing engine to perform parsing and type separation operations on the multi-source remote sensing data streams. The parsing engine distinguishes data categories by identifying the data type identifier field in the data header, where identifier V corresponds to the visible light imagery data stream, I corresponds to the infrared thermal imagery data stream, and L corresponds to the lidar point cloud data stream. During the parsing process, the encapsulation header and verification fields from the data transmission stage are stripped to restore the original file formats of each type of data. Visible light imagery is restored to TIFF format, infrared thermal imagery to RAW format, and lidar point cloud data stream to L. The point cloud was restored to LAS format, and invalid data blocks found during the parsing process were removed. Spatiotemporal reference information was extracted from the three types of data streams after separation to form reference data with original spatiotemporal stamps. The extraction of spatiotemporal reference information was based on the metadata fields of the data itself: GPS coordinates (longitude and latitude accurate to 0.00001 degrees), shooting timestamp (UTC time, accurate to ms), sensor focal length and exposure parameters were extracted from the EXIF ​​metadata of visible light and infrared images; the acquisition timestamp (with the same accuracy as the image timestamp), point cloud coordinate system (WGS-84) and scanning angle information were extracted from the LAS file header of the lidar point cloud. After extraction, a unique identifier code was added to each data stream. The encoding rules were data type, acquisition timestamp, and UAV number to ensure data traceability in subsequent processing and avoid confusion of multi-source data.

[0047] Step 503: Based on the reference data with the original spatiotemporal stamp and the unified reference of the digital twin model, spatiotemporal registration data is obtained through coordinate transformation and time synchronization. Specifically, this includes: using the reference data with the original spatiotemporal stamp as the processing object, performing coordinate transformation and time synchronization based on the unified spatiotemporal reference of the digital twin model (WGS-84 geographic coordinate system and UTC time system); the coordinate transformation is differentiated for the coordinate characteristics of different data types: the original coordinates of the lidar point cloud are relative to the aircraft body coordinates, and are converted into absolute geographic coordinates through the real-time attitude data of the UAV (pitch angle, roll angle, and heading angle output by the IMU inertial measurement unit) and GPS position data. The transformation process uses coordinate transformation matrix calculation to ensure that the spatial deviation between the converted point cloud coordinates and the terrain data of the digital twin model is less than 0.5m; visible light and infrared images are used through image geographic registration algorithms, with feature points such as tower apexes and road intersections in the digital twin model as ground control points, and coordinate correction is completed by using quadratic polynomial fitting to make the image pixel coordinates correspond accurately to the geographic coordinates, and the registration error is controlled within 1 pixel.

[0048] Time synchronization employs Network Time Protocol (NTP) combined with hardware clock calibration to align the original timestamps of the three types of data streams with the high-precision clocks (error less than or equal to 1ms) of the edge computing nodes. For data with timestamp deviations exceeding 100ms, corrections are made using track point time information from the UAV flight logs to ensure that the timestamp deviations of visible light images, infrared thermal images, and lidar point clouds under the same observation scenario are less than 100ms. After coordinate transformation and time synchronization are completed, the three types of data are organized according to the correlation between timestamps and spatial locations to form spatiotemporal registration data containing a unified spatiotemporal reference. The data format adopts a custom fusion data format for easy subsequent rapid access and processing.

[0049] Step 504: Based on the spatiotemporal registration data, and combined with the spatial sampling density, sensor operating frequency, and flight altitude parameters of each partition in the differentiated acquisition strategy parameter table, adaptive correction is performed on the geometric deformation and radiation deviation caused by terrain undulation and atmospheric conditions in the data to obtain the corrected data. Specifically, this includes: based on the spatiotemporal registration data, retrieving the parameter information of the corresponding risk clustering unit in the differentiated acquisition strategy parameter table, including spatial sampling density, sensor operating frequency, and flight altitude. The spatial sampling density is 1 to 10 points / square meter, the sensor operating frequency is 1 to 10 Hz, and the flight altitude is 80 to 300 m. These parameters serve as the dynamic adjustment basis for adaptive correction, achieving parameter and data matching. The correction process addresses the geometric deformation caused by terrain undulation and the radiation deviation caused by atmospheric conditions separately. Geometric deformation correction uses a digital surface model (DSM)-assisted method, extracting the DSM data of the corresponding area from the digital twin model, calculating the terrain slope in conjunction with the UAV flight altitude parameters, and performing several corrections on the tilted areas in the lidar point cloud and image data. Correction methods: For areas with a slope greater than 30°, a block correction method is used, dividing the data into 10m×10m sub-blocks, with each sub-block's coordinates adjusted based on the DSM elevation difference; for areas with a slope less than or equal to 30°, a holistic correction method is used, completing the correction through a single coordinate transformation to ensure that the corrected data accurately reflects the spatial morphology of power facilities, reducing geometric deformation errors to within 0.1m; Radiation deviation correction combines atmospheric conditions and sensor operating frequency parameters, using the Atmospheric Radiative Transfer Model (MODTRAN) for adaptive adjustment: for high-risk areas (sampling density greater than or equal to 8 points / square meter), the model iteration count is increased based on the sensor operating frequency (10 iterations) to accurately eliminate the influence of atmospheric scattering and absorption on image radiation values; for medium- and low-risk areas (sampling density less than or equal to 5 points / square meter), a simplified model (5 iterations) is used to balance correction accuracy and processing efficiency; during the correction process, the grayscale mean and standard deviation of the data before and after correction are compared in real time to ensure that the radiation deviation is reduced by more than 30%, ultimately obtaining corrected data with accurate geometric morphology and radiation characteristics.

[0050] Step 505 involves extracting features from the corrected data to obtain corresponding spectral feature sets, thermal infrared feature sets, and three-dimensional structural feature sets. Specifically, this includes: performing feature extraction on the corrected data according to its type, combining the disaster identification focus of each data type with targeted algorithms to ensure feature effectiveness; extracting spectral feature sets from visible light image data using an improved ResNet-18 convolutional neural network, pre-trained on 100,000 power grid images, focusing on extracting grayscale anomalies in insulator pollution, texture changes on tower surfaces, and color features of wildfires, outputting a spectral feature set containing 256-dimensional feature vectors, with each dimension of the feature vector corresponding to the response value of different spectral bands; and extracting thermal infrared feature sets from infrared thermal image data using a combination of adaptive threshold segmentation and texture analysis. The maximum inter-class variance method is used to determine the temperature threshold and segment high-temperature anomaly regions, such as the temperature difference region between iced conductors and normal conductors. Then, feature parameters such as average temperature, temperature standard deviation, thermal gradient, and texture entropy of the region are extracted to form a 128-dimensional thermal infrared feature set. Among them, the average temperature and thermal gradient features are mainly used for the identification of disasters such as icing and equipment overheating. The lidar point cloud data is used to extract a three-dimensional structural feature set through point cloud segmentation and feature descriptor calculation. The region growing algorithm is used to segment the point cloud of power facilities such as towers and conductors. Then, the three-dimensional dimensions, morphological parameters, and spatial positional relationship features of the facilities are calculated. Among them, the three-dimensional dimensions are such as tower height and conductor diameter, and the morphological parameters are such as tower tilt angle and conductor sag, generating a 512-dimensional three-dimensional structural feature set. This feature set provides data support for the calculation of tower deformation and conductor icing thickness.

[0051] Step 506: Spatial alignment and scale normalization are performed on the spectral feature set, thermal infrared feature set, and three-dimensional structural feature set to obtain an aligned multi-source feature set. Information from the aligned multi-source feature set is then integrated to obtain multi-dimensional fused data. Specifically, this includes: First, spatial alignment and scale normalization are performed on the spectral feature set, thermal infrared feature set, and three-dimensional structural feature set. Spatial alignment uses the three-dimensional coordinates of the lidar point cloud as a reference, associating the pixel features of visible light images and infrared thermal images with their corresponding three-dimensional positions through coordinate mapping, ensuring that the multi-source features of the same power facility correspond one-to-one in space. For issues of inconsistent feature dimensions, feature interpolation is used to supplement low-dimensional feature dimensions, unifying all three feature sets into 512-dimensional feature vectors. Scale normalization uses the min-max normalization method, mapping the values ​​of all feature vectors to 0. To eliminate dimensional differences between different features, temperature values ​​(-40℃ to 150℃) and point cloud coordinates (latitude, longitude, and elevation) are uniformly converted into feature values ​​of the same scale. After normalization, a feature fusion algorithm based on an attention mechanism is used to integrate multi-source features. The algorithm learns the feature importance weights in historical disaster data and assigns higher weights (0.3 to 0.5) to key features for disaster identification in high-risk areas (such as the temperature features of icing and the three-dimensional features of tower deformation) and lower weights (0.1 to 0.2) to auxiliary features. During the integration process, feature splicing and nonlinear transformation are used to fuse the three types of feature sets into a single 1024-dimensional feature vector. Each feature vector is associated with the corresponding power facility identifier, spatial coordinates, and collection timestamp, ultimately forming multi-dimensional fused data containing spatial location information, multi-source feature information, and attribute information.

[0052] In a preferred embodiment of the present invention, step 600 involves performing multi-scale feature analysis on the multi-dimensional fused data to obtain a multi-scale feature set; performing parallel disaster feature parsing based on the multi-scale feature set to obtain a disaster feature parameter set; and performing comprehensive analysis on the disaster feature parameter set to obtain a quantitative disaster assessment report, including: Step 601: Perform multi-scale convolutional neural network feature extraction on the multi-dimensional fused data to obtain a multi-scale feature set. Specifically, this includes: using 1024-dimensional fused data as the processing object, the data contains metadata such as power grid equipment identification, WGS-84 coordinates, and timestamps; a multi-scale convolutional neural network based on U-Net is used to perform feature extraction; this network contains 5 scale-progressive feature extraction modules and 4 cross-scale fusion modules, which can accurately adapt to different scale differences such as the microscopic features of insulator pollution and the macroscopic features of wildfire spread; microscopic features are extracted using 16 3×3 convolutional kernels (stride 4) at the 1 / 4 scale; mesoscopic features are captured using 32 3×3 convolutional kernels (stride 2) at the 1 / 2 scale; and 1-scale... At a scale of 1x, 64 3×3 convolutional kernels (stride 1) are used to extract conventional features; at a scale of 2x, 128 5×5 convolutional kernels (stride 0.5) are used to capture macroscopic features; at a scale of 4x, 256 7×7 convolutional kernels (stride 0.25) are used to extract ultra-macroscopic features. The fusion module introduces a channel attention mechanism, assigning weights of 0.3 to 0.4 to key feature channels such as icing and surface discharge, and weights of 0.05 to 0.1 to auxiliary features such as terrain and background. The network is pre-trained on 50,000 sets of labeled data and optimized with a cross-entropy loss function. After convergence, the feature extraction accuracy exceeds 92%. The final output is a multi-scale feature set with five dimensions, each dimension corresponding to a 256-dimensional feature vector. All features are associated with a unified spatiotemporal index.

[0053] Step 602: Based on the multi-scale feature set, analyze the 3D point cloud and visible light image to extract the icing feature parameters of the conductor and ground wire. Specifically, this includes: based on the multi-scale feature set, calling 1 / 2 scale and 1x scale feature data to construct a dual-source fusion system of 3D point cloud geometric modeling and visible light image radiation characteristic analysis to extract the icing features of the conductor and ground wire; in the 3D point cloud analysis, the RANSAC algorithm is used to separate the conductor and ground wire point cloud, removing interference from towers, vegetation, etc., with a filtering accuracy of over 98%; the conductor and ground wire point cloud is sliced ​​at 0.5m intervals, and the slice contour is obtained by fitting the minimum circumcircle. The diameter of the conductor was compared with the standard diameter of the conductor in the digital twin model to calculate the initial icing thickness. After truncating the icing area, the Otsu algorithm was used to segment the icing and the conductor. The grayscale difference between the two was distinguished based on the threshold of the 8-bit grayscale image of 40. Combined with the scale determined by the UAV flight altitude and the sensor focal length, the pixel features were converted into actual dimensions to obtain the equivalent area of ​​the icing cross section. The icing thickness was obtained by weighting the point cloud (0.6) and the image (0.4) to achieve an accuracy of ±0.2 mm. The volume was calculated by combining the equivalent area with the length of the continuous icing section (accuracy ±0.5 m), and the density was inverted (accuracy ±0.05 g / cm³). 3 Based on density and average gray value, it is divided into three categories: rain rime, hoarfrost, and mixed rime. All parameters are associated with the conductor segment number and spatial coordinate range.

[0054] Step 603: Based on the spatial location marked by the icing feature parameters, perform a linked analysis of the hyperspectral and ultraviolet data of the corresponding region to obtain surface discharge intensity feature parameters. Specifically, this includes: based on the spatial coordinates of the icing features, quickly associating the 1 / 4 scale features of the corresponding region in the multi-scale feature set using an R-tree index structure (retrieval response time less than or equal to 50ms); calling hyperspectral and ultraviolet data to perform linked analysis of icing hazards and discharge risks; specifically, generating a circular analysis area with a radius of 5m centered on the center coordinates of the icing area, and filtering hyperspectral images in the 300-1000nm band (spectral resolution 5nm) and ultraviolet images in the 200-400nm band within this area; extracting reflectance data in the 300-400nm ultraviolet band and the 700-800nm ​​near-infrared band, and calculating ultraviolet band reflectance anomalies (which are consistent with the regional background). The difference in values ​​is used to identify potential discharge precursors. When an outlier is greater than or equal to 20%, it is considered a precursor to discharge. Simultaneously, abrupt reflectivity changes in the red-edge band (700-750nm) are detected to identify material property changes caused by contamination accumulation on the insulator surface, serving as an auxiliary criterion for discharge risk. Ultraviolet data analysis employs a spot recognition algorithm: based on the ultraviolet radiation characteristics of discharge, spot areas with grayscale values ​​greater than or equal to 200 (8-bit grayscale image) are detected in the image. Morphological opening operations of 3×3 rectangular structural elements are used to filter interfering spots such as direct sunlight, and the number and area of ​​effective discharge spots are counted. Finally, discharge frequency, intensity level, duration, and associated icing thickness threshold are extracted to form a linked dataset of icing and discharge. The discharge frequency is the number of effective discharge points occurring per unit time, with a statistical period of 10s and an accuracy of ±1 time / minute. The intensity level is level 1, less than or equal to 0.1cm. 2 Level 5, greater than or equal to 1.0 cm 2 The accuracy of the duration is ±1s.

[0055] Step 604 involves simultaneously processing the multispectral and thermal infrared data from the multi-scale feature set to obtain wildfire characteristic parameters. Specifically, this includes: simultaneously calling the 2x and 4x scale features of the multi-scale feature set; using collaborative processing of multispectral and thermal infrared data to resolve the problem of misjudgment based on single data; calculating the Normalized Difference Vegetation Index (NDVI) by combining multispectral analysis with vegetation index and fire point spectral characteristics; marking suspected fire points when the NDVI is less than or equal to 0.1 and the near-infrared reflectance is greater than or equal to 0.8; extracting the ratio of the red band (630 to 690 nm) to the near-infrared band (760 to 900 nm) in the suspected area; confirming the area as a fire point when it is greater than or equal to 1.2, thus excluding high-temperature interference from areas without vegetation cover; and setting thermal infrared data according to environmental temperature differences. Thresholds: The threshold is 60℃ when the ambient temperature is less than or equal to 25℃ and 80℃ when it is greater than 25℃. High-temperature fire points are marked and distinguished as open flame, smoldering, and high-temperature hazards according to the extreme temperature values. Among them, open flame is greater than or equal to 300℃, smoldering is 100 to 300℃, and high-temperature hazards are 50 to 100℃. The integrated wildfire characteristic parameters are: the number of open flame and smoldering points is counted according to the extreme temperature values, the fire point coordinate accuracy is ±1m, the spread speed is calculated at 30-second intervals with an accuracy of ±0.1m / minute, the shortest distance to the power line (accuracy ±0.5m), and the burning vegetation type is identified through multispectral analysis. Fire points are clustered according to the spread area, and each cluster is assigned a unique identifier (such as Fire-20251207-001) and associated with the power grid line protection level.

[0056] Step 605: Merge the structural point cloud time series data associated with the icing feature parameters and wildfire feature parameters to analyze and obtain structural damage feature parameters. Specifically, this includes: retrieving the icing feature parameters and wildfire feature parameters, extracting the structural point cloud time series data within the affected area through equipment identification association, including current data and 3 sets of historical data from the previous hour (sampling interval of 15 minutes, taken from the edge computing node cache), and using time series comparison and deformation modeling methods to analyze structural damage and establish the causal relationship between disaster and damage. Tower damage analysis: ICP registration is performed between the temporal point cloud and the digital twin standard tower point cloud (error ≤ 0.3cm). Areas with a deviation ≥ 5cm are marked as suspected damage. If the point cloud missing rate of the suspected area is ≥ 10% or the coordinate abrupt change is ≥ 10cm, tower damage is confirmed. Conductor damage analysis: The sag change is calculated. If the change is ≥ 20% compared with the standard sag, it is marked as abnormal. If the point cloud fracture length is ≥ 1m and associated with icing (≥ 15mm) or wildfire (distance ≤ 5m), it is determined as conductor damage. The damage type is clearly defined as tower or conductor damage. The location coordinate accuracy is ±10cm for tower and ±0.5m for conductor. The degree of conductor damage is classified into mild (≤ 5%), moderate (5% to 20%), and severe (≥ 20%) according to the cross-sectional damage ratio. The associated disaster type is classified as icing-dominated, wildfire-dominated, or combined-dominated. The damage development trend after 1 hour is estimated based on the temporal change rate to form a complete damage feature dataset.

[0057] Step 606: Integrate the icing characteristic parameters, surface discharge intensity characteristic parameters, wildfire characteristic parameters, and structural damage characteristic parameters to form a disaster characteristic parameter set. Specifically, this includes: using a dual-index structure with spatial coordinates as the primary key and disaster type as the secondary key to integrate multiple types of parameters, solving the problem of scattered parameters across multiple disaster types; using WGS-84 latitude and longitude accurate to 0.00001 degrees as the core primary key to associate icing, discharge, wildfire, and damage parameters at the same spatial location, forming a spatially associated data group; within the data group, using disaster type as the secondary key to distinguish between single disasters and... The correlation between multiple disasters is investigated; complete data source information is supplemented during integration: multi-source data sources, feature extraction timestamps, and algorithm confidence levels. Among them, multi-source data sources include drone number, sensor type and number, feature extraction timestamps are accurate to milliseconds, and algorithm confidence levels are such as 95% confidence level for ice thickness extraction; a weighted voting method (with algorithm accuracy as the weight) is used to determine the final value for data conflicts; the integrated parameter set is stored in JSON format, containing six primary fields: basic information, ice, discharge, wildfire, damage, and source tracing, supporting fast retrieval.

[0058] Step 607: Normalize the disaster feature parameter set to obtain a standardized multi-hazard parameter vector. Specifically, this includes: performing normalization on the disaster feature parameter set to eliminate dimensional differences and provide standardized input for disaster pattern matching; classifying parameters by type: numerical parameters such as ice thickness and temperature are normalized to the 0-1 range using min-max normalization, with the value range determined based on historical extreme values ​​and engineering thresholds, such as ice thickness 0 to 50 mm and wildfire temperature 0 to 1000℃; grade-based parameters such as discharge intensity and damage level are encoded in an ordered manner, such as discharge levels 1 to 5 corresponding to 0.2, 0.4, 0.6, 0.8, and 1.0; and categorized parameters (such as ice type and vegetation type) are encoded using unique thermal encoding, such as rime encoding [1, 0, 0]. All standardized parameters are arranged in a fixed order as a 128-dimensional vector, with 32 dimensions for ice, 16 for discharge, 32 for wildfire, 32 for damage, and 16 for confidence level. Elements are retained to four decimal places, and a CRC-32 checksum is added to prevent data distortion.

[0059] Step 608: Perform disaster pattern matching based on the standardized multi-hazard parameter vector to obtain disaster identification results and their confidence levels, including disaster type and spatial coordinates. Specifically, this includes: performing disaster pattern matching based on the standardized multi-hazard parameter vector using a dual identification mechanism of template matching and machine learning classification. Both the matching template library and the classification model are constructed based on historical data from the power grid disaster database over the past 5 years, covering pattern features of 12 types of single disasters and 8 types of compound disasters. The template library consists of standardized vectors of 100,000 typical disaster samples and supports monthly online updates. Template matching: calculate the cosine similarity between the input vector and the template vector; if the similarity is greater than or equal to... Templates with a similarity of 0.85 are marked as candidates, and the type with the highest proportion among the candidate templates is taken as the preliminary identification result. At the same time, the LightGBM classification model trained with 50,000 sets of samples is called, with an accuracy of 94%, to output the prediction probability of various disasters. The highest value is taken as the auxiliary identification result. If the two are consistent, the disaster type is determined. If they are inconsistent, the one with the higher confidence (the template is the highest similarity, and the model is the highest prediction probability) is taken. The final result includes the disaster type, the coordinates of the core area, and the confidence of each sub-disaster (retaining two decimal places). Among them, disaster types such as conductor icing, insulator discharge, and minor tower damage with a confidence of less than 0.7 are marked as pending on-site verification.

[0060] Step 609: Based on the disaster identification results, associate the power grid topology data in the digital twin model to assess the impact range and severity level of the disaster. Specifically, this includes: based on the disaster identification results, calling the power grid topology data (i.e., line connections, power supply radius, equipment capacity, etc.) through the digital twin model interface, and combining this with geographic information data to perform an impact range and severity level assessment. The geographic information data includes topography, vegetation, and meteorology. The impact range is analyzed using a two-dimensional approach: equipment association and geographic diffusion. The equipment association dimension traces the equipment chain affected by the disaster through topological traversal; the geographic diffusion dimension is based on the disaster... The predicted range of characteristics, such as wildfires, is predicted using a fan-shaped model to predict the 1-hour spread range, and icing combined with temperature and humidity to predict the expansion boundary. After superposition, a comprehensive impact range is formed and presented as a vector map. The severity level is based on the "Technical Standard for Emergency Response to Disasters in Power Systems" and adopts a multi-factor weighted score of 0 to 100 points: core disaster weight 0.3, derivative disaster weight 0.2, hub substation weight 0.2, main line weight 0.15, etc., corresponding to four levels: minor 0 to 25 points, general 26 to 50 points, severe 51 to 75 points, and extremely severe 76 to 100 points. For compound disasters, the level superposition principle is adopted.

[0061] Step 610: Integrate the disaster type, spatial coordinates, impact range, severity level, and confidence level to obtain a quantitative disaster assessment report. Specifically, this includes: integrating the identification results and assessment conclusions to generate a structured quantitative report that supports both machine analysis and human reading. The report contains seven core chapters: cover, abstract, basic information, feature parameter details, impact assessment, response recommendations, and data traceability. The cover includes a number such as Report-IceFire-20251207-003, the generation time accurate to the second, and the edge computing node number. The abstract includes the disaster type, location, severity level, range, and confidence level. Basic information includes the discovery time, equipment number, and geographical environment. Feature parameter details present various parameters and their confidence levels in a table. The impact assessment includes a list of affected equipment, load statistics, and a range map. Response recommendations include differentiated solutions, such as immediate power outage for a level 4 disaster. Data traceability includes the original data source and algorithm version. The report is output in both PDF (including satellite imagery and other visual charts) and JSON (system interaction) formats, automatically stored in the edge computing node database, and pushed to the power grid emergency command platform within 3 seconds, meeting minute-level response requirements.

[0062] In a preferred embodiment of the present invention, step 700 involves mapping the quantitative disaster assessment report to a digital twin model to obtain a disaster scenario; based on the disaster scenario, integrating power grid topology and traffic network information for path planning to obtain a comprehensive response plan; and based on the comprehensive response plan, automatically generating and issuing coordinated response instructions, including: Step 701 involves mapping the disaster type, spatial coordinates, impact range, severity level, and confidence level from the quantitative disaster assessment report to the digital twin model to obtain the disaster scene. Specifically, this includes: extracting structured data from the quantitative disaster assessment report, including disaster type identifiers, WGS-84 coordinates of the core and peripheral areas of the disaster, the impact range represented by vector polygons, severity levels 1 to 4, and confidence levels between 0 and 1; calling the scene mapping interface of the digital twin model, using the unique identifier of the power grid equipment as the association key, and binding the above data with the corresponding power facilities and geographical terrain entities in the model; after receiving the data, the digital twin model uses the Unity3D engine's material rendering module to perform differentiated rendering on the disaster-affected area, where severity levels 1 to 4 correspond to green, yellow, orange, and red semi-transparent overlays respectively, and areas with a confidence level greater than or equal to 0.9 are marked with a highlighted border; simultaneously, a disaster information floating window is created in the model, and clicking on the corresponding area displays complete parameters such as disaster type, occurrence time, and associated facility number, ultimately forming a visualized and interactive disaster scene.

[0063] Step 702: Based on the disaster scenario, and integrating power grid topology and transportation network constraints, construct a multi-objective optimization model for emergency response. Specifically, this includes: building a hierarchical multi-objective optimization model for emergency response based on the disaster scenario, with a target layer, criterion layer, and solution layer. The target layer is defined as a comprehensive optimization that minimizes emergency response time, resource scheduling costs, and power grid power loss. The criterion layer provides quantitative evaluation indicators for each objective, including quantitative values ​​for response time, scheduling costs, and power loss. The solution layer covers all possible resource scheduling paths, operation sequences, and power grid adjustment strategies. Model training uses historical data from the past five years of power grid disaster emergency response as a sample set. The sample includes 2000 sets of optimization objective parameter constraints and optimal response results under different disaster types and levels. During training, the model is first... The model parameters are initialized using sample data. The weight coefficients of each criterion layer are adjusted using the objective function error as feedback signal. Training stops when the prediction error of 50 consecutive samples is less than or equal to 5%, ensuring the model adapts to actual emergency scenarios. Three types of constraints are input during model construction: first, power grid topology constraints, taken from the power grid topology layer of the digital twin model, including the power supply priority of hub substations, the load limit of transmission lines, and fault isolation boundaries; second, traffic network constraints, derived from the road spatial distribution vector dataset, covering the road level, real-time congestion coefficient, and bridge load-bearing limits for emergency vehicle passage; and third, resource constraints, including the location, equipment configuration, and skill level of emergency repair teams, the remaining power and operating radius of drone inspection groups, and the storage locations and reserves of emergency supplies. The weights of each optimization objective are determined using the analytic hierarchy process (AHP), with response time weighted at 0.4, power supply loss weighted at 0.35, and scheduling cost weighted at 0.25. Combining the trained parameters and constraints, the mathematical expression and boundary conditions of the multi-objective optimization model are constructed, clarifying the value range of each variable and the comprehensive calculation logic of the objective function.

[0064] Step 703: Based on the multi-objective optimization model, perform optimization calculations to generate a virtual integrated solution that includes resource scheduling paths, operation sequences, and power grid adjustment strategies. Specifically, this includes: using an improved genetic algorithm to solve the multi-objective optimization model, with an initial population size of 100, a crossover probability of 0.7, and a mutation probability of 0.05; calculating the comprehensive performance of each candidate solution using a fitness function; and stopping the calculation when the fitness value fluctuation is less than or equal to 0.01 for 10 consecutive generations. The generated virtual integrated solution includes three core components: the resource scheduling path uses Dijkstra's algorithm. Based on real-time traffic data planning, the departure time, route, and estimated arrival time of each repair team and material transport vehicle are determined; the operation sequence is logically ordered according to fault isolation, hidden danger investigation, and emergency repair and restoration, prioritizing the handling tasks of hub substations and power supply lines of important users, and marking the start and end times and connection relationships of each task; the power grid adjustment strategy is specified down to the switch operation sequence, load transfer path, and backup line switching scheme, clarifying the execution object and parameter requirements of each step of the operation; the scheme is stored in XML format, containing a four-level data structure of resource ID, task content, time node, and constraints.

[0065] Step 704: Perform data matching and task decoupling between the virtual integrated response plan and the virtual resource library in the digital twin model to obtain a task-resource mapping relationship; based on the task-resource mapping relationship, obtain an independent executable task instruction sequence, specifically including: the virtual resource library of the digital twin model pre-stores attribute data of various physical resources, including the repair team number and skill matching table, UAV equipment ID and operational capability parameters, emergency vehicle license plate and transport specifications, etc.; associate and match the virtual integrated response plan with the resource library data, the matching rules being task type and resource capability adaptation, task location and resource location proximity, and task time and resource idle time period overlap; through the task decoupling module, the complex tasks in the plan are broken down into independent sub-tasks, such as line repair being broken down into on-site safety warning, fault point location, wire replacement, insulation testing, etc., each sub-task having a clearly defined unique execution subject; the generated task instruction sequence is in JSON array format, each instruction containing execution terminal ID, task code, operation parameters, execution time limit, and safety threshold, wherein the operation parameters are specific to directly executable details such as the UAV's flight altitude and the usage specifications of repair tools.

[0066] Step 705: Through the 5G private network and satellite relay communication link, the sequence of independently executable task instructions is sent in real time to the corresponding physical execution terminal, driving it to execute coordinated response instructions. Specifically, this includes: constructing a link adaptive selection sending mechanism; collecting signal strength, bandwidth, and latency parameters of the 5G private network and satellite relay link in real time through a link status monitoring module; when the execution terminal is located in a 5G coverage area along the power grid, the 5G private network is prioritized for instruction sending, utilizing its low latency characteristics to ensure timely instruction transmission, with latency controlled within 500ms; when the terminal is located in mountainous areas or other areas with weak 5G signals (i.e., signal strength less than -90dBm), it automatically switches to the satellite relay link, ensuring instruction integrity through anti-interference coding technology. Before instruction sending, a CRC-32 algorithm is used for verification, and the receiving terminal provides feedback on the verification result. If a transmission error occurs, a retransmission mechanism is triggered. The execution terminals include UAV flight control systems, emergency repair terminals, and power grid dispatch terminals. After receiving the instructions, the terminals automatically parse and execute them, while simultaneously uploading the execution status to the digital twin model in real time, forming a closed-loop control of instruction sending, execution, and feedback.

[0067] like Figure 2 As shown, embodiments of the present invention also provide a high-altitude unmanned aerial vehicle (UAV) remote sensing system for integrated identification of multiple disasters in power grids, comprising: The module is used to build a digital twin model for the power grid area to be inspected by integrating historical disaster data, geographic information and satellite imagery. The risk assessment module is used to perform risk clustering and quantitative assessment of transmission line corridors based on the risk distribution output by the digital twin model, and obtain a risk level quantification matrix; and obtain a differentiated data acquisition strategy parameter table based on the risk level quantification matrix. The collaborative acquisition module is used to perform parameterized optimization of UAV swarm missions using a differentiated acquisition strategy parameter table to obtain a collaborative observation scheme; and to schedule the UAV swarm to acquire multi-source remote sensing data streams based on the collaborative observation scheme. The transmission module is used to transmit multi-source remote sensing data streams to edge computing nodes in real time via 5G private network and satellite relay communication links; The fusion module is used by edge computing nodes to receive multi-source remote sensing data streams and perform spatiotemporal registration to obtain spatiotemporal registered data; based on the spatiotemporal registered data, combined with differentiated acquisition strategy parameters, adaptive correction and feature-level fusion are performed to obtain multi-dimensional fused data; The analysis module is used to perform multi-scale feature analysis on multi-dimensional fused data to obtain a multi-scale feature set; based on the multi-scale feature set, parallel disaster feature parsing is performed to obtain a disaster feature parameter set; and a comprehensive analysis of the disaster feature parameter set is performed to obtain a quantitative disaster assessment report. The response module maps the quantitative disaster assessment report to a digital twin model to obtain the disaster scenario; based on the disaster scenario, it integrates power grid topology and traffic network information to perform route planning and obtain a comprehensive response plan; based on the comprehensive response plan, it automatically generates and issues collaborative response instructions.

[0068] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for comprehensive identification of multiple hazards in power grids using high-altitude unmanned aerial vehicle (UAV) remote sensing, characterized in that: The method includes: Step 100: For the power grid area to be inspected, a digital twin model is constructed by integrating historical disaster data, geographic information and satellite imagery; Step 200: Based on the risk distribution output by the digital twin model, perform risk clustering and quantitative assessment on the transmission line corridor to obtain a risk level quantification matrix; and obtain a differentiated acquisition strategy parameter table based on the risk level quantification matrix. Step 300: Optimize the parameters of the UAV swarm mission using the differentiated acquisition strategy parameter table to obtain a collaborative observation scheme; and schedule the UAV swarm to acquire multi-source remote sensing data streams based on the collaborative observation scheme. Step 400: The multi-source remote sensing data stream is transmitted to the edge computing node in real time through the 5G private network and satellite relay communication link; Step 500: The edge computing node receives multi-source remote sensing data streams and performs spatiotemporal registration to obtain spatiotemporal registration data; based on the spatiotemporal registration data, combined with differentiated acquisition strategy parameters, adaptive correction and feature-level fusion are performed to obtain multidimensional fused data; Step 600: Perform multi-scale feature analysis on the multi-dimensional fused data to obtain a multi-scale feature set; perform parallel disaster feature parsing based on the multi-scale feature set to obtain a disaster feature parameter set; and conduct comprehensive analysis on the disaster feature parameter set to obtain a quantitative disaster assessment report. Step 700: Map the quantitative disaster assessment report to the digital twin model to obtain the disaster scenario; based on the disaster scenario, integrate power grid topology and traffic network information to perform path planning and obtain a comprehensive disposal plan.

2. The method for comprehensive identification of multiple disasters in a power grid using high-altitude unmanned aerial vehicle (UAV) remote sensing according to claim 1, characterized in that, Step 100 includes: Collect multi-source basic geographic data covering the power grid area to be inspected. The multi-source basic geographic data includes: high-resolution visible light and synthetic aperture radar remote sensing images, high-precision digital elevation model data, historical archived disaster point distribution data, and regional meteorological background field data including temperature, humidity and wind speed. The high-resolution visible light and synthetic aperture radar remote sensing images are subjected to radiometric calibration, atmospheric correction and geometric fine correction to obtain a standardized remote sensing image base; the high-precision digital elevation model data and regional meteorological background field data are subjected to format standardization and spatial interpolation to obtain a topographic and meteorological environment dataset that is spatially aligned with the standardized remote sensing image base. Based on the standardized remote sensing image base, a multi-scale semantic feature tensor is obtained through forward propagation calculation using a pre-trained deep convolutional neural network. Based on the multi-scale semantic feature tensor, region proposal and target classification are performed to complete the initial localization of key power facilities. Pixel-level instance segmentation is performed on the initial localization to obtain a contour mask. The contour mask is vectorized and spatial coordinate transformed to obtain a vector dataset of the spatial distribution of power facilities.

3. The high-altitude UAV remote sensing power grid multi-hazard comprehensive identification method according to claim 2, characterized in that, Step 100 also includes: Based on the standardized remote sensing image base, multi-scale contextual features are extracted and fused, and pixel-by-pixel semantic segmentation is performed to obtain the road pixel-level classification confidence distribution; the road pixel-level classification confidence distribution is subjected to topology optimization and denoising processing to obtain a binarized road mask; the binarized road mask is subjected to skeleton extraction and vectorization processing to obtain a road spatial distribution vector dataset; By integrating the aforementioned power facility spatial distribution vector dataset, road spatial distribution vector dataset, terrain and meteorological environment dataset, and historically archived disaster point distribution data, a multi-dimensional spatiotemporal background database is constructed, which integrates terrain elevation, facility location, road network connectivity, meteorological conditions, and historical disaster information. Based on the multi-dimensional spatiotemporal background database, data is organized and spatially correlated to obtain integrated scene data; based on the integrated scene data, the electrical topology relationship of the power grid is fused to construct a data-topology correlation structure; based on the data-topology correlation structure, three-dimensional geometric reconstruction and dynamic rendering are performed to obtain a visual simulation environment; simulation and risk inference rules are configured in the visual simulation environment to complete the construction of an integrated air-space-ground digital twin model.

4. The method for comprehensive identification of multiple disasters in a high-altitude unmanned aerial vehicle (UAV) remote sensing power grid according to claim 3, characterized in that, Step 200 includes: Based on the risk distribution output by the digital twin model, the spatial proximity of each risk point within the transmission line corridor is calculated, and a risk spatial density distribution field is constructed. Based on the risk spatial density distribution field, the core risk point set is identified and marked by setting the neighborhood radius and the minimum number of points threshold; Based on the core risk point set, clusters are expanded and merged through connectivity analysis and density reachability judgment to generate an initial risk cluster partition, which simultaneously includes clusters and noise points. For each cluster in the initial risk clustering partition, a statistical consistency test of the internal risk feature vector is performed, and the noise points at the boundary are re-discriminated to obtain the risk clustering partition set. For each risk clustering unit in the risk clustering partition set, the terrain, vegetation and historical disaster features within its spatial range are extracted based on the digital twin model to form a multi-dimensional risk feature vector; The multi-dimensional risk feature vector is normalized to generate a standardized risk feature vector; A weighted assessment is performed based on the standardized risk feature vector to generate a quantitative risk level value for each unit. The quantified values ​​of all units are organized into a risk level quantification matrix according to spatial topological relationships; Based on the risk level quantification matrix, combined with the preset UAV platform performance parameters and sensor physical constraints, the quantification values ​​of each unit are transformed and boundary verified through the preset risk-policy mapping function to obtain the optimized policy parameter set of each unit. The optimized policy parameter set includes at least spatial sampling density, sensor operating frequency and flight altitude. The optimization strategy parameter sets of all units are organized into a differentiated acquisition strategy parameter table according to their spatial index.

5. The high-altitude UAV remote sensing power grid multi-hazard comprehensive identification method according to claim 4, characterized in that, Step 300 includes: Based on the differentiated acquisition strategy parameter table, the flight range is estimated to obtain the initial flight range estimation matrix; Based on the initial range estimation matrix, tasks are bundled and assigned to obtain a conflict-free task assignment scheme. Based on the conflict-free task allocation scheme, the final track point sequence and sensor working sequence of each UAV are obtained through track smoothing and cooperative intersection planning. By integrating the aforementioned conflict-free task allocation scheme, final trackpoint sequence, and sensor operating timing, a collaborative observation scheme is constructed. Based on the aforementioned collaborative observation scheme, collaborative control commands for the UAV are parsed and generated; Based on the aforementioned collaborative control commands, the drone swarm is controlled to perform collaborative flight and synchronized sensor detection to obtain raw data; The original data is subjected to spatiotemporal synchronization and format standardization to obtain a multi-source remote sensing data stream.

6. The high-altitude UAV remote sensing power grid multi-hazard comprehensive identification method according to claim 5, characterized in that, Step 400 includes: The multi-source remote sensing data stream is parsed, classified, and prioritized to obtain a priority data sequence; the priority data sequence is then encapsulated, protocol-converted, and time-series reassembled to obtain a standardized data packet queue. Based on the preset link status and transmission strategy, the standardized data packet queue is allocated to the 5G private network link and the satellite relay communication link to obtain data packets allocated to the 5G private network link and the satellite relay communication link. For the data packets allocated to the 5G private network link, forward error correction coding and flow control based on the transmission control protocol are performed to obtain the 5G link transport stream; The data packets allocated to the satellite relay communication link are subjected to efficient compression, anti-interference channel coding, and adaptive modulation based on link quality to obtain the satellite link transmission stream; The 5G link transport stream and the satellite link transport stream are parsed, decoded, and their integrity is verified to obtain a set of valid data packets. Based on the timestamps and sequence numbers of the data packets in the set of valid data packets, cross-link alignment and deduplication fusion are performed to obtain a fused data packet stream. The fused data packet stream is restored to generate a reconstructed data stream consistent with the multi-source remote sensing data stream, and the reconstructed data stream is pushed to the edge computing node.

7. The method for comprehensive identification of multiple disasters in a high-altitude unmanned aerial vehicle (UAV) remote sensing power grid according to claim 6, characterized in that, Step 500 includes: The edge computing node receives the reconstructed data stream as a multi-source remote sensing data stream; The multi-source remote sensing data stream is parsed and its types are separated to obtain visible light image, infrared thermal image and lidar point cloud data stream; the spatiotemporal reference information of the visible light image, infrared thermal image and lidar point cloud data stream is extracted to obtain reference data with original spatiotemporal stamps. Based on the reference data with the original spatiotemporal stamp and the unified benchmark of the digital twin model, spatiotemporal registration data is obtained through coordinate transformation and time synchronization; Based on the spatiotemporal registration data, and combined with the spatial sampling density, sensor operating frequency and flight altitude parameters of each partition in the differentiated acquisition strategy parameter table, the geometric deformation and radiation deviation caused by terrain undulation and atmospheric conditions in the data are adaptively corrected to obtain the corrected data. Feature extraction is performed on the corrected data to obtain the corresponding spectral feature set, thermal infrared feature set, and three-dimensional structural feature set; Spatial alignment and scale normalization are performed on the spectral feature set, thermal infrared feature set, and three-dimensional structural feature set to obtain an aligned multi-source feature set; information from the aligned multi-source feature set is integrated to obtain multi-dimensional fused data.

8. The method for comprehensive identification of multiple disasters in a high-altitude unmanned aerial vehicle (UAV) remote sensing power grid according to claim 7, characterized in that, Step 600 includes: Multi-scale convolutional neural network features are extracted from multi-dimensional fused data to obtain a multi-scale feature set. Based on the multi-scale feature set, the three-dimensional point cloud and visible light image are analyzed to extract the icing feature parameters of the conductor ground wire. Based on the spatial location marked by the icing characteristic parameters, the hyperspectral and ultraviolet data of the corresponding region are analyzed in conjunction to obtain the surface discharge intensity characteristic parameters. Simultaneously process the multispectral and thermal infrared data in the multi-scale feature set to obtain wildfire characteristic parameters; By fusing the structural point cloud time series data associated with icing characteristic parameters and wildfire characteristic parameters, structural damage characteristic parameters are obtained through analysis. By integrating the aforementioned icing characteristic parameters, surface discharge intensity characteristic parameters, wildfire characteristic parameters, and structural damage characteristic parameters, a set of disaster characteristic parameters is formed; The disaster characteristic parameter set is normalized to obtain a standardized multi-hazard parameter vector; Based on the standardized multi-hazard parameter vector, disaster pattern matching is performed to obtain disaster identification results and their confidence levels, which include disaster type and spatial coordinates. Based on the disaster identification results, the power grid topology data in the digital twin model is correlated to assess the scope and severity of the disaster. By integrating the disaster type, spatial coordinates, impact range, severity level, and confidence level, a quantitative disaster assessment report is obtained.

9. The method for comprehensive identification of multiple disasters in a high-altitude unmanned aerial vehicle (UAV) remote sensing power grid according to claim 8, characterized in that, Step 700 includes: The disaster type, spatial coordinates, impact range, severity level and confidence level in the quantitative disaster assessment report are correlated and mapped to the digital twin model to obtain the disaster scenario; Based on disaster scenarios, a multi-objective optimization model for emergency response is constructed by integrating power grid topology and transportation network constraints. Based on a multi-objective optimization model, optimization calculations are performed to generate a virtual integrated solution that includes resource scheduling paths, operation sequences, and power grid adjustment strategies. The virtual integrated processing scheme is matched with the virtual resource library in the digital twin model to obtain a task-resource mapping relationship; based on the task-resource mapping relationship, an independent executable task instruction sequence is obtained. Through 5G private network and satellite relay communication link, the sequence of independently executable task instructions is sent to the corresponding physical execution terminal in real time, driving it to execute coordinated response instructions.

10. A high-altitude unmanned aerial vehicle (UAV) remote sensing system for integrated identification of multiple disasters in power grids, wherein the system implements the method as described in any one of claims 1 to 9, characterized in that, include: The module is used to build a digital twin model for the power grid area to be inspected by integrating historical disaster data, geographic information and satellite imagery. The risk assessment module is used to perform risk clustering and quantitative assessment of transmission line corridors based on the risk distribution output by the digital twin model, and obtain a risk level quantification matrix; and obtain a differentiated data acquisition strategy parameter table based on the risk level quantification matrix. The collaborative acquisition module is used to perform parameterized optimization of UAV swarm missions using a differentiated acquisition strategy parameter table to obtain a collaborative observation scheme. The UAV swarm is scheduled to collect multi-source remote sensing data streams based on the collaborative observation scheme; The transmission module is used to transmit multi-source remote sensing data streams to edge computing nodes in real time via 5G private network and satellite relay communication links; The fusion module is used by edge computing nodes to receive multi-source remote sensing data streams and perform spatiotemporal registration to obtain spatiotemporal registration data. Based on spatiotemporal registration data, combined with differentiated acquisition strategy parameters, adaptive correction and feature-level fusion are performed to obtain multidimensional fused data; The analysis module is used to perform multi-scale feature analysis on multi-dimensional fused data to obtain a multi-scale feature set; based on the multi-scale feature set, parallel disaster feature parsing is performed to obtain a disaster feature parameter set; and a comprehensive analysis of the disaster feature parameter set is performed to obtain a quantitative disaster assessment report. The response module is used to map the quantitative disaster assessment report to the digital twin model to obtain the disaster scenario; Based on disaster scenarios, route planning is performed by integrating power grid topology and traffic network information to obtain a comprehensive response plan.

Images