Method and device for evaluating spatiotemporal distribution of equipment failure loss based on lightweight deep learning and discrete grid technology

Abstract

This application discloses a method and apparatus for assessing the spatiotemporal distribution of equipment failure losses based on lightweight deep learning and discrete grid technology, relating to the field of power system technology, which can balance identification efficiency and accuracy. The method includes: acquiring multi-source remote sensing image data of the power distribution network in the disaster area after a typhoon; dividing the multi-source remote sensing image data into grids to obtain multiple grid cells; performing data quality screening and data cleaning within the multiple grid cells to obtain multiple high-quality grid cells; and performing multi-source data fusion and standardization processing to obtain multiple standardized grid images; using a lightweight deep learning convolutional neural network to extract and classify disaster features from the multiple standardized grid images to determine the disaster identification results; and performing inverse mapping and spatial stitching on the disaster identification results corresponding to each standardized grid image based on the geospatial encoding of each standardized grid image to obtain and output the evaluation results.

Description

Technical Field

[0001] This application relates to the field of power system technology, and in particular to a method and apparatus for assessing the spatiotemporal distribution of equipment failure losses based on lightweight deep learning and discrete grid technology. Background Technology

[0002] Frequent typhoon disasters often lead to serious faults such as power distribution lines collapsing, towers tilting, and insulator damage, directly affecting the safe and stable operation of the power system. After a typhoon passes, quickly and accurately identifying the type and spatial distribution of the disaster in the power distribution network is a key prerequisite for carrying out emergency repairs and restoring power supply.

[0003] In related technologies, when identifying power distribution network disasters after a typhoon, optical data of the disaster area is typically acquired through satellite remote sensing imagery. After preprocessing the images, such as radiometric calibration, atmospheric correction, and geometric registration, machine learning classifiers like edge detection, morphological filtering, or support vector machines are used to extract the geometric and spectral features of power distribution network towers, lines, insulators, and other facilities from the images, thereby identifying disaster types such as collapse, tilting, and damage. Simultaneously, high-resolution local images are acquired through drone aerial photography, and the power distribution network facilities in the drone images are labeled and disaster-determined on a target-by-target basis using manual visual interpretation or pixel-based supervised classification methods. A few technologies introduce the concept of discrete grids, dividing a large study area into regular rectangular grid units according to a fixed spatial scale. Satellite or drone images within each grid are independently cropped and stored. Then, conventional deep learning convolutional neural networks are used to perform forward inference on each grid unit, i.e., convolution, pooling, and fully connected operations are performed on each grid image to extract deep-level features, ultimately outputting the disaster type and confidence level of the power distribution network facilities within that grid.

[0004] However, the applicant recognizes that the relevant technology has at least the following technical problems in its implementation: Relying on single satellite remote sensing images or local UAV images is susceptible to interference from factors such as clouds and weather, which can lead to the loss of key damage features. While UAV images have high resolution, their coverage is limited, making it difficult for the two to complement each other. This results in the inability to fully capture the damage features of the power distribution network in the disaster area. At the same time, although some technologies have introduced discrete grids for simple block division, the model parameters are large and the computational cost is high. When processing large-scale disaster images block by block, the large amount of data can easily cause memory overflow or slow inference speed, making it difficult to meet the stringent timeliness requirements of emergency repairs. It is difficult to balance recognition efficiency and accuracy. Summary of the Invention

[0005] In view of this, this application provides a method and apparatus for assessing the spatiotemporal distribution of equipment failure losses based on lightweight deep learning and discrete grid technology. The main purpose is to solve the problem that it is currently difficult to meet the stringent timeliness requirements of emergency repairs and that it is difficult to balance identification efficiency and accuracy.

[0006] According to a first aspect of this application, a method for assessing the spatiotemporal distribution of equipment failure losses based on lightweight deep learning and discrete grid technology is provided, the method comprising: Satellite remote sensing image data and UAV image data of the power distribution network in the disaster area after the typhoon passed were acquired as multi-source remote sensing image data. Based on discrete grid technology, the multi-source remote sensing image data was divided into multiple grid units according to multi-level nested grids. Data quality screening and data cleaning are performed within the multiple grid cells to obtain multiple high-quality grid cells. Then, multi-source data fusion and standardization processing are performed on the multiple high-quality grid cells to obtain multiple standardized grid images corresponding to the multiple high-quality grid cells. A lightweight deep learning convolutional neural network is used to extract and classify disaster features from the multiple standardized grid images, and the disaster identification result corresponding to each of the multiple standardized grid images is determined. Based on the geospatial encoding of each standardized grid image, the disaster identification results corresponding to each standardized grid image are inversely mapped and spatially stitched to obtain and output the spatiotemporal distribution assessment results of equipment failure losses in a large-scale area.

[0007] According to a second aspect of this application, a device for assessing the spatiotemporal distribution of equipment failure losses based on lightweight deep learning and discrete grid technology is provided. The device includes: The partitioning module is used to acquire satellite remote sensing image data and UAV image data of the power distribution network in the disaster area after the typhoon passes as multi-source remote sensing image data. Based on discrete grid technology, the multi-source remote sensing image data is partitioned into multiple grid units according to multi-level nested grids. The fusion module is used to perform data quality screening and data cleaning within the multiple grid cells to obtain multiple high-quality grid cells, and to perform multi-source data fusion and standardization processing on the multiple high-quality grid cells to obtain multiple standardized grid images corresponding to the multiple high-quality grid cells. The identification module is used to extract and classify disaster features from the multiple standardized grid images using a lightweight deep learning convolutional neural network, and to determine the disaster identification result corresponding to each of the multiple standardized grid images; The generation module is used to perform inverse mapping and spatial stitching on the disaster identification results corresponding to each standardized grid image based on the geospatial encoding of each standardized grid image, so as to obtain and output the spatiotemporal distribution assessment results of equipment failure loss in a large-scale area.

[0008] According to a third aspect of this application, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in any of the first aspects above.

[0009] According to a fourth aspect of this application, a computer-readable storage medium is provided, on which a computer program is stored, wherein the computer program, when executed by a processor, implements the steps of the method described in any one of the first aspects above.

[0010] By employing the aforementioned technical solutions, this application provides a method and apparatus for assessing the spatiotemporal distribution of equipment failure losses based on lightweight deep learning and discrete grid technology. This application effectively addresses the problems of single data sources being susceptible to interference and incomplete coverage by fusing satellite remote sensing and UAV multi-source imagery and performing multi-level nested partitioning, quality screening, and standardized fusion based on discrete grid technology. It can comprehensively capture various damage characteristics of the distribution network. Simultaneously, it replaces conventional large-parameter models with lightweight deep learning convolutional neural networks, significantly reducing computational burden and inference latency under a grid-level parallel processing framework, avoiding memory overflow, and meeting the timeliness requirements of emergency repairs. Furthermore, by combining grid-level data augmentation and spatial inverse mapping stitching, it improves disaster identification accuracy while achieving efficient assessment of large-scale areas, ultimately balancing identification efficiency and accuracy, and providing reliable technical support for emergency repairs of distribution networks after typhoons.

[0011] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0012] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This illustration shows a flowchart of a method for assessing the spatiotemporal distribution of equipment failure losses based on lightweight deep learning and discrete grid technology, as provided in an embodiment of this application. Figure 2This illustration shows a schematic diagram of a lightweight deep learning convolutional neural network structure provided in an embodiment of this application; Figure 3 A schematic diagram of the module structure of a bottleneck residual module provided in an embodiment of this application is shown; Figure 4 This illustration shows a flowchart of another method for evaluating the spatiotemporal distribution of equipment failure losses based on lightweight deep learning and discrete grid technology, provided in an embodiment of this application. Figure 5 This illustration shows a schematic diagram of a device for assessing the spatiotemporal distribution of equipment failure losses based on lightweight deep learning and discrete grid technology, according to an embodiment of this application. Figure 6 A schematic diagram of the device structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0013] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application 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 application will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0014] This application provides a method for assessing the spatiotemporal distribution of equipment failure losses based on lightweight deep learning and discrete grid technology, such as... Figure 1 As shown, the method includes: S10: Acquire satellite remote sensing image data and UAV image data of the power distribution network in the disaster area after the typhoon passes as multi-source remote sensing image data. Based on discrete grid technology, the multi-source remote sensing image data is divided into multiple grid units according to multi-level nested grids.

[0015] The technical solution of this application embodiment can be applied to a loss assessment system. In this application embodiment, the loss assessment system first acquires satellite remote sensing image data and UAV image data of the power distribution network in the disaster area after the typhoon passes as multi-source remote sensing image data. Among them, satellite remote sensing images have the advantage of wide coverage but are easily interfered with by clouds, while UAV images have high resolution but limited coverage. The loss assessment system will divide these two types of data into grids according to multi-level nested grids based on discrete grid technology. Specifically, the loss assessment system will first convert the satellite and UAV images into raster data with geographic coordinate mapping relationships, and then adopt a three-level nested grid division strategy, using the 6° universal transverse Mercator projection zone as the first-level grid, 10km×10km as the second-level grid, and 1km×1km as the third-level grid, to establish a unique mapping relationship between grid codes and geographic space, thereby dividing the large-scale disaster area into multiple standardized grid units, each grid unit having a unique geographic space code.

[0016] In the above process, through multi-level nested grid partitioning, this embodiment of the application unifies massive heterogeneous multi-source image data into a standardized spatial reference framework, laying a structured foundation for subsequent data filtering, fusion, and parallel processing, while avoiding the data explosion problem caused by directly processing large-scale images. For example, assuming a typhoon landfall area covers approximately 1000 square kilometers, it can be divided into 1000 1km×1km grid units through the above process. Each grid unit is simultaneously associated with satellite image slices and UAV image slices within that area.

[0017] S20: Perform data quality screening and data cleaning within multiple grid cells to obtain multiple high-quality grid cells, and perform multi-source data fusion and standardization processing on multiple high-quality grid cells to obtain multiple standardized grid images corresponding to multiple high-quality grid cells.

[0018] In this embodiment, the loss assessment system first performs data quality screening and data cleaning on each obtained grid cell. Specifically, for each level-3 grid cell, its comprehensive quality evaluation index is calculated. This index is obtained by weighted summation of the proportion of cloud cover in the image within the grid, the proportion of effective coverage of the image, and the standard deviation of the raster pixel values. Grid cells with comprehensive quality evaluation indices below a preset threshold are marked as low-quality data and removed, thus obtaining multiple high-quality grid cells. Subsequently, the loss assessment system performs multi-source data fusion on the retained high-quality grid cells. Using a preset uniform processing resolution, satellite imagery and UAV imagery within the grid are resampled and scaled to eliminate resolution differences. Then, a pixel-level adaptive weighted fusion method is used to dynamically allocate fusion weights based on the signal-to-noise ratio and geometric resolution contribution of each data source, fusing the wide-area background information of the satellite imagery with the detailed damage features of the UAV imagery at the pixel level. The fused data then undergoes radiometric correction, geometric correction, pixel value normalization, and size normalization processing, finally outputting a standardized grid image that meets the input requirements of a lightweight deep learning model.

[0019] In this way, the above process eliminates low-quality grid data affected by clouds, noise, and other interference through quality screening, ensuring the reliability of subsequent identification. Furthermore, multi-source data fusion leverages the complementary advantages of satellite and UAV imagery, enabling the grid image to possess both wide coverage and high resolution. Standardization eliminates differences in data dimensions and sizes, providing a unified input format for efficient inference in deep learning models. For example, suppose that within a 1km×1km grid, satellite imagery is partially blurred due to thin clouds, while UAV imagery clearly covers the power grid lines within that grid. The loss assessment system first evaluates the quality index of this grid, retaining it as a high-quality grid because the cloud cover percentage is below a threshold. Then, it weightedly fuses the geographical background of the satellite imagery with the line details of the UAV imagery to generate a clear, standardized image with geographic coordinates.

[0020] S30: Use a lightweight deep learning convolutional neural network to extract and classify disaster features from multiple standardized grid images, and determine the disaster identification result corresponding to each standardized grid image in the multiple standardized grid images.

[0021] In this embodiment, the loss assessment system inputs each standardized grid image into a lightweight deep learning convolutional neural network. This lightweight deep learning convolutional neural network uses the MobileNetV3 Small architecture, whose core feature is the use of depthwise separable convolutions instead of conventional convolutions, significantly reducing the number of model parameters and computational cost. The lightweight deep learning convolutional neural network first extracts low-dimensional features such as edges and textures from the image through an initial convolutional layer. Then, it performs deep feature extraction through multiple stacked bottleneck residual modules. Each bottleneck residual module combines depthwise separable convolutions and a compression-excitation attention mechanism, automatically focusing on damage features of the distribution network facilities, such as tower tilting, line collapse, and insulator damage. Finally, the lightweight deep learning convolutional neural network outputs the disaster category probability distribution corresponding to each standardized grid image through global average pooling and fully connected layers, thereby determining the disaster identification result of the distribution network within that grid.

[0022] Thus, in the above process, the lightweight network design enables the model to perform fast inference on ordinary computing devices, avoiding the memory overflow and computational lag issues that conventional large-parameter models encounter when processing large-scale images. Simultaneously, the attention mechanism can automatically uncover deep damage features in multi-source fused images, significantly improving recognition accuracy in complex disaster scenarios compared to traditional edge detection or machine learning methods. For example, when a standardized grid image is input into the lightweight deep learning convolutional neural network, it automatically extracts the tilt angle features of towers and the fracture texture features of lines within milliseconds, outputting a probability of 0.92 for the "tower tilt" category within the grid, with probabilities for all other categories below 0.1.

[0023] S40: Based on the geospatial coding of each standardized grid image, the disaster identification results corresponding to each standardized grid image are inversely mapped and spatially stitched to obtain and output the spatiotemporal distribution assessment results of equipment failure losses in a large-scale area.

[0024] In this embodiment, the loss assessment system performs inverse mapping and spatial stitching based on the disaster identification results corresponding to each standardized grid image and the established grid geospatial coding. Specifically, firstly, for the disaster category probability vector of each grid, the loss assessment system automatically retrieves its original three-level nested grid unique code, uses the inverse mapping function to calculate the actual spatial range of the grid cell in the universal transverse Mercator projection coordinate system based on the grid code, and then converts it to WGS-84 latitude and longitude coordinates, thereby realizing the binding of disaster perception attributes with physical geographical location; then, on a large-scale geographic background, the inverse calculation results of all grids are spatially stitched according to latitude and longitude boundaries. For overlapping coverage areas that may exist at the edges of adjacent grids, the loss assessment system uses a probability-weighted fusion strategy to determine the final disaster attribute of the area; finally, based on the geographic information system protocol, the stitched and fused results are output as a standard format file, forming a large-scale equipment failure loss spatiotemporal distribution assessment result covering the entire disaster area.

[0025] In this way, during the above process, the dimensionless classification probabilities output by the deep learning model are transformed into the distribution of disaster entities with precise geographic coordinates through inverse mapping, achieving a closed-loop mapping from the perceived results to the physical world. Simultaneously, spatial stitching and overlapping area fusion strategies ensure the continuity and consistency of the overall assessment results, and the final generated map can be directly used to guide emergency repair decisions. For example, assuming a grid identifies a "line collapse" disaster, with its geospatial code corresponding to the latitude and longitude range [118.5°E~118.6°E, 31.2°N~31.3°N], the loss assessment system will bind this disaster label to the geographic range and stitch it with the identification results of adjacent grids, ultimately outputting a GIS map file containing the locations and types of all collapsed lines.

[0026] Optionally, in this embodiment, satellite remote sensing image data and UAV image data of the power distribution network in the disaster area after a typhoon are acquired as multi-source remote sensing image data. Based on discrete grid technology, the multi-source remote sensing image data is divided into multiple grid cells according to a multi-level nested grid. This includes converting the satellite remote sensing image data and UAV image data into raster data with geographic coordinate mapping relationships using the following formula to obtain multi-source raster data.

[0027] in, For the first Data source in geographic coordinates The raster pixel value at that location, For the first Rasterization mapping function for data source class, For the first The original pixel values ​​of the data source in its own pixel coordinate system. For the first The set of rasterization transformation parameters for a data source. The value can be 1 or 2. A value of 1 corresponds to satellite remote sensing image data, and A value of 2 corresponds to UAV image data. A three-level nested grid partitioning strategy is adopted. Following the formula below, the geographic space of the power distribution network monitoring area is subdivided using a 6° universal transverse Mercator projection zone as the first-level grid, 10km×10km as the second-level grid, and 1km×1km as the third-level grid. Multi-source raster data is then mapped to the partitioned geographic space. This multi-level nested grid method divides the multi-source remote sensing image data into multiple grid cells.

[0028] in, A unique code for a three-level nested grid. For the mesh coding mapping function of the RDCRMG architecture, For the first Geographic parameters with a first-order projection and grid. For the first Spatial parameters of a second-level 10km×10km grid, For the first The subdivision parameters of a 1km×1km third-level grid, For the first Data source in geographic coordinates The raster pixel value at that location.

[0029] In this embodiment, the loss assessment system acquires satellite remote sensing image data and UAV image data of the disaster area's power distribution network after the typhoon as raw multi-source data input. Based on RDCRMG (Raster Data Cleaning and Reconstruction Multi-Grid Architecture) discrete grid technology, it first performs rasterization of the multi-source raw data, that is, it uniformly converts heterogeneous multi-source data such as raster format satellite remote sensing images and pixel format UAV images into georeferenced raster data adapted to the RDCRMG architecture, establishing a one-to-one mapping relationship between raster pixel values ​​and actual geographic coordinates. Specifically, this can be achieved using the following formula 1: Formula 1:

[0030] in, for Data source in geographic coordinates The raster pixel values ​​at that location enable data to be correlated with geographic coordinates. The binding, The value can be 1 or 2. A value of 1 corresponds to satellite remote sensing image data, and The value of 2 corresponds to the drone image data; For the first The original pixel values ​​of the data source in its own pixel coordinate system; For the first The rasterization transformation parameter set for the data source includes projection parameters, pixel resolution parameters, and coordinate offset parameters; For the first The rasterization mapping function for the data source class enables a non-linear transformation from raw data to georeferenced raster data, bringing the data into the geographic coordinate system. Through the above rasterization process, the original image, which originally lacked geographic location information, is converted into raster data with a clear location in the Earth's latitude and longitude coordinate system, laying the foundation for subsequent multi-source data fusion.

[0031] Subsequently, the loss assessment system, based on the hierarchical grid design principle of the RDCRMG architecture, employs a three-level nested grid partitioning strategy to divide the geographic space of the distribution network monitoring area into grids. Specifically, it uses the 6° Universal Transverse Mercator (UTM) projection zone as the first-level grid, 10km×10km as the second-level grid, and 1km×1km as the third-level grid, establishing a unique mapping relationship between grid codes and geographic space. This achieves grid-based partitioning and data block management of the monitoring area. The mapping process between grid codes and geographic coordinates is shown in Formula 2 below: Formula 2:

[0032] in, A unique code for a three-level nested grid, indexed by the first-level grid. Secondary grid index and three-level grid index Together constitute; This is a grid encoding mapping function for the RDCRMG architecture, implementing geographic coordinates. Conversion to a unique grid code; For the first Geographic parameters with a first-order projection and grid; For the first Spatial parameters of a second-level 10km×10km grid; For the first The subdivision parameters of a third-level 1km×1km grid; For the aforementioned Data source in geographic coordinates The raster pixel values ​​at each location. Through a three-level nested grid division, the loss assessment system forces all data into a standardized 1km×1km three-level grid, thus completing the spatial reference unification of multi-source data.

[0033] In this process, through the RDCRMG architecture's rasterization and multi-level nested grid partitioning, heterogeneous, spatially unreferenced satellite and UAV imagery in large-scale disaster areas is transformed into standardized, geocoded grid units. This enables unified spatial reference and block management of multi-source data, effectively solving the problems of massive data integration difficulties and resolution mismatches. It provides a structured, efficient, and parallelizable data foundation for subsequent quality screening, fusion processing, and deep learning inference. For example, suppose a typhoon passes through an area of ​​approximately 2000 square kilometers, containing 3 satellite images and 500 UAV aerial photographs. Through the above process, all images are converted into 1km×1km raster data blocks, each with a unique code, such as [UTM50-10-5] representing the grid in the 10th row and 5th column within the 50th UTM band. This organizes the originally scattered and disordered multi-source data into 2000 standardized grid units, with each grid unit simultaneously storing corresponding satellite image slices and UAV image slices.

[0034] Optionally, in this embodiment, data quality screening and data cleaning are performed within multiple grid cells to obtain multiple high-quality grid cells, including: unique encoding based on a three-level nested grid. Locate each grid cell and use the following formula, based on the rasterized pixel values ​​in each grid cell. Determine the first Pixel values ​​of three-level grid units ,

[0035] in, Mesh encoding mapping function for RDCRMG architecture The inverse mapping, These are the first-level, second-level, and third-level grid indices; based on the rasterized pixel values ​​in each grid cell, the cloud cover percentage within each grid cell is calculated. Percentage of effective coverage of images within the grid and the standard deviation of the grid pixel values. The comprehensive quality evaluation index for each grid cell is calculated using the following formula.

[0036] in, For the first The comprehensive quality evaluation index corresponding to each grid cell The weight coefficients for each item and Determine the preset grid data quality threshold. According to the following formula, combined with the grid data quality threshold Low-quality data is removed from multiple grid cells to obtain multiple high-quality grid cells.

[0037] in, For multiple high-quality mesh cells, the first In the first high-quality grid cell Raster data of the class, This indicates that the corresponding grid cell is low-quality data and should be cleared.

[0038] In this embodiment, the loss assessment system first uses a unique code based on a three-level nested grid. Position each grid cell, where These are level 1, level 2, and level 3 grid indices, respectively. The loss assessment system utilizes the grid encoding mapping function of the RDCRMG architecture. inverse mapping Based on the rasterized pixel values ​​in each grid cell Determine the first Pixel values ​​of three-level grid units Its calculation formula is as follows: Formula 3: Formula 3: In formula 3, This is the inverse mapping function, used to convert the unique grid code back to the corresponding geographic coordinate range; For the first Data source in geographic coordinates The raster pixel values ​​at each location. Through the above process, the raster data contained in each grid cell can be accurately extracted.

[0039] Next, the loss assessment system will combine the rasterized pixel values ​​in each grid cell to calculate the proportion of cloud cover in that grid cell. Effective coverage ratio of images and the standard deviation of raster pixel values Among these, cloud cover percentage reflects the proportion of pixels within the grid that are obscured by clouds, effective coverage percentage indicates the degree of coverage of effective information (non-background, non-noise) in the image, and standard deviation characterizes the dispersion of pixel values ​​and is used to identify pixel distortion. Based on these three indicators, the loss assessment system uses a weighted summation method to calculate the comprehensive quality evaluation index for each grid cell. The calculation formula is shown in Formula 4 below: Formula 4:

[0040] In formula 4, These are the weighting coefficients corresponding to the cloud cover percentage, effective coverage percentage, and standard deviation, respectively, satisfying... The specific value can be pre-calibrated based on the actual data characteristics. The smaller the value, the higher the quality of the grid data.

[0041] Subsequently, the loss assessment system will determine the pre-defined grid data quality threshold for the RDCRMG architecture. Low-quality data is removed from multiple grid cells according to the rules shown in Formula 5 below: Formula 5:

[0042] In formula 5, For the first time after cleaning In the first high-quality grid cell Raster data of the class; This indicates that the corresponding grid cell is low-quality data and has been removed. In other words, when the comprehensive quality evaluation index of a certain third-level grid cell is low... Greater than the threshold If a grid cell has issues such as excessive cloud cover, insufficient effective coverage, or pixel distortion, it is marked as low-quality data and removed; conversely, if... Less than or equal to If the desired result is obtained, the mesh cell will be retained as a high-quality mesh cell for subsequent fusion processing.

[0043] Thus, in the above process, by constructing a comprehensive quality evaluation index that includes cloud cover, effective coverage, and pixel dispersion, and setting a unified threshold for automated screening, low-quality grid data affected by factors such as weather and sensor noise can be effectively eliminated, avoiding the negative impact of this data on subsequent feature extraction and disaster identification, thereby improving the reliability and accuracy of the overall assessment. At the same time, the grid-level independent quality screening mechanism can ensure that the data quality of each grid unit is controllable, providing a clean input foundation for multi-source data fusion.

[0044] For example, suppose that within a 1km×1km grid, satellite imagery shows a cloud cover ratio as high as 60%, an effective coverage ratio of only 30%, and a large pixel standard deviation. The overall quality evaluation index is then calculated. The value is 0.85, which is higher than the grid data quality threshold of 0.6. Then the grid cell is marked as They were then discarded and not included in the subsequent fusion and recognition process; while the other grid had a cloud coverage rate of 5%, effective coverage of 95%, and normal pixel standard deviation. If the value is 0.15, it is reserved as a high-quality grid cell for subsequent multi-source data fusion.

[0045] Optionally, in this embodiment, multiple high-quality grid cells are subjected to multi-source data fusion and standardization processing to obtain multiple standardized grid images corresponding to the multiple high-quality grid cells, including: using a preset unified processing resolution. The following formula is used to reconstruct the resolution function. Resampling and scaling are performed on the original heterogeneous data within each high-quality grid cell to obtain the first... Within the first high-quality grid cell Normalized pixel matrix after resolution reconstruction of the data source ,

[0046] in, For multiple high-quality mesh cells, the first In the first high-quality grid cell Raster data of the class; using the following formula, according to the fusion weights The first Pixel-level adaptive weighted fusion of multi-source data within a high-quality grid cell yields the first... The fused raster pixel values ​​corresponding to each high-quality grid cell

[0047] in, For the first The raster pixel values ​​obtained by fusing multi-source data within a high-quality grid cell. The fusion weights are adaptively determined by the signal-to-noise ratio and geometric resolution contribution of the data source; according to the following formula, for the ... Radiometric fine correction is performed on the fused raster pixel values ​​corresponding to each high-quality grid cell.

[0048] in, For the first The radiometrically corrected raster pixel values ​​within a high-quality grid cell. They are respectively from the first The radiation correction gain coefficient and offset coefficient are determined by the measured radiation values ​​of ground control points within each high-quality grid cell. Simultaneously, a quadratic polynomial correction model is used to correct the geographic coordinates of each high-quality grid cell according to the following formula.

[0049] in, These are the geometrically corrected geographic coordinates. The geographic coordinates before correction. to The quadratic polynomial geometric correction parameters are based on ground control point calibration. The raster data of each high-quality grid cell in the corrected high-quality grid cells are standardized to obtain multiple standardized grid images corresponding to multiple high-quality grid cells.

[0050] In this embodiment, the loss assessment system first performs resolution matching and spatial alignment of multi-source images for each high-quality mesh cell retained after cleaning. Specifically, it needs to utilize the unified processing resolution preset by the RDCRMG architecture. Through resolution reconstruction function The original heterogeneous data within each level-3 grid are resampled and scaled to reconstruct the scale, and the satellite imagery and UAV imagery are uniformly mapped to the RDCRMG standard raster coordinate system to eliminate pixel misalignment caused by differences in sensor sampling intervals. For the specific calculation formula, please refer to Formula 6 below: Formula 6:

[0051] in, For the first Within the first grid cell The normalized pixel matrix of the data source after resolution reconstruction; For the first time after cleaning In the first high-quality grid cell The raster data of the class, where the process described above is similar. When the value is 1, it corresponds to the satellite image. When the value is 2, it corresponds to the drone image; This is a resolution reconstruction function based on grid-center geographic coordinate mapping, responsible for resampling heterogeneous images to a uniform resolution. .

[0052] Next, the loss assessment system employs a pixel-level adaptive weighted fusion method to integrate multi-source features within the grid, fully extracting the wide-area ground cover background from satellite remote sensing imagery and the equipment detail damage features from UAV imagery, achieving information complementarity through dynamic weight allocation. The fusion calculation formula is shown in Formula 7 below: Formula 7:

[0053] In formula 7, For the first The final raster pixel value after fusing multi-source data within a grid; For the first The fusion weights for different data sources are adaptively determined by the signal-to-noise ratio and geometric resolution contribution of each data source. Specifically, high-resolution UAV imagery that provides core damage features is assigned a higher gain weight; while for satellite imagery that provides geographic base map information, the weights are used to ensure the integrity of the background topology.

[0054] After fusion, the loss assessment system performs radiometric and geometric fine corrections on the fused raster data to eliminate systematic errors caused by atmospheric scattering, sensor response bias, terrain undulation, and imaging attitude bias, ensuring the radiometric accuracy and geospatial consistency of the raster data. The radiometric fine correction uses a linear correction model; the specific formula is shown in Formula 8 below. Formula 8:

[0055] in, For the first The raster pixel values ​​within each grid cell after radiometric fine correction; These are the radiation correction gain coefficient and offset coefficient of the grid cell, respectively, which are calibrated by the measured radiation values ​​of the ground control points within the grid.

[0056] Simultaneously, the loss assessment system performs geographic coordinate correction on each high-quality grid cell, i.e., geometric fine correction, specifically using a quadratic polynomial correction model, as shown in Formula 9 below: Formula 9:

[0057] In formula 9, These are the geometrically corrected geographic coordinates. Geographic coordinates before correction; to The quadratic polynomial geometric correction parameters are based on ground control point calibration and are determined by the geographic coordinates of feature points such as distribution network towers and insulators within the grid.

[0058] Finally, the loss assessment system standardizes the corrected raster data, including pixel value normalization and size normalization, and finally outputs a standardized grid image with uniform specifications to adapt to the input requirements of subsequent lightweight deep learning models.

[0059] Thus, in the above process, resolution reconstruction and adaptive weighted fusion can complement the advantages of satellite wide-area background and UAV detailed features, and the generated fused image has both high recognition and low noise characteristics. At the same time, radiometric and geometric fine correction can eliminate systematic errors between multi-source data, ensuring data consistency and geographic accuracy. In addition, standardization processing provides a uniform input format for deep learning models, which significantly improves the accuracy and efficiency of subsequent disaster feature extraction and identification.

[0060] For example, assuming a high-quality 1km × 1km grid, satellite imagery provides the overall terrain and tower distribution background for the area, while UAV imagery clearly shows the tilt details of a particular tower, the loss assessment system first resamples both to a uniform resolution, such as 0.1m / pixel. Then, it assigns weights to the UAV imagery and satellite imagery for weighted fusion, resulting in a fused image that preserves the geographical background while highlighting the tower tilt characteristics. Subsequently, using the known GPS coordinates of the towers within the grid as ground control points, it calibrates radiometric correction coefficients and geometric correction parameters to eliminate biases caused by atmospheric conditions and imaging angles. Finally, it outputs a standardized 224×224 pixel image, which can be directly input into a lightweight deep learning convolutional neural network for disaster identification.

[0061] Optionally, in this embodiment, the raster data of each high-quality grid cell in the corrected plurality of high-quality grid cells is standardized, including: for each high-quality grid cell, the pixel value of the raster data of the corrected high-quality grid cell is normalized using the following linear normalization formula.

[0062] in, For normalized raster data, These represent the maximum and minimum pixel values ​​of the raster data within the corresponding high-quality grid cell, respectively; the following formula is used based on the target image width required by the training model. and target image height Calculate the size scaling factor for high-quality mesh cells. ,

[0063] in, The original width and height of the merged raster data in the high-quality grid cell are given; the scaled image width is calculated for the high-quality grid cell using the following formula. and the height of the scaled image ,

[0064] Referring to the formula below, zero-fill is applied to regions where the high-quality mesh cells exceed the scaling size to obtain a normalized mesh image corresponding to the high-quality mesh cells.

[0065] in, To standardize the pixel values ​​of a grid image. To standardize the pixel coordinates of the image, This is a floor operation.

[0066] In this embodiment, the loss assessment system first performs pixel value normalization on the raster data within each high-quality grid cell after radiometric and geometric correction. This eliminates dimensional differences between different data sources due to imaging conditions and sensor variations, thus avoiding adverse effects on subsequent deep learning model training. Specifically, in this embodiment, normalization uses a linear mapping formula to linearly transform the original pixel values ​​to the [0, 1] interval. The calculation formula is shown in Formula 10 below: Formula 10:

[0067] in, The data is normalized raster data; For the first Radiometrically corrected raster pixel values ​​within each grid cell; These represent the maximum and minimum pixel values ​​within the grid cell, respectively. Through transformation using Equation 10, all pixel values ​​are unified to the same numerical range to eliminate the influence of dimensions.

[0068] After pixel value normalization, the raster data is further normalized in size according to the input requirements of lightweight deep learning models such as MobileNetV3 Small. Specifically, the target image width required by the model needs to be determined first. and target image height ,For example The value can be set to 224. The value can be set to 224, which simultaneously retrieves the original width of the fused raster data within the current high-quality grid cell. and original height To ensure that the image remains undistorted after scaling and retains as much information as possible, the loss assessment system employs a proportional scaling strategy and calculates the size scaling factor. For the specific calculation formula, please refer to Formula 11 below: Formula 11:

[0069] In formula 11, and These represent the scaling ratios in the width and height directions, respectively, and the smaller of the two values ​​is taken as the uniform scaling factor. This ensures that the scaled image can be fully contained within the target size while maintaining the original aspect ratio, avoiding stretching distortion.

[0070] Based on scaling factor The loss assessment system will then continue to calculate the scaled image width. and height The rounding down operation ensures that the result is an integer pixel value. For the specific calculation logic, please refer to Formula 12 below: Formula 12:

[0071] in, This represents the floor function. and This refers to the actual width and height of the scaled image, whose values ​​do not exceed the dimensions. and .

[0072] Finally, the loss assessment system generates a standardized raster image according to the following rules: for the scaled image region, the normalized raster data is mapped to the corresponding pixel coordinates through bilinear interpolation or nearest neighbor sampling; for regions exceeding the scaled size, i.e., the parts of the target size not filled by the scaled image, zero-padding is performed, and the calculation logic is shown in Formula 13 below: Formula 13:

[0073] in, These are the standardized image pixel values; To standardize the pixel coordinates of the image, the value range is... , .when Located in the scaled effective area, that is , Within the time frame, take the corresponding position in the normalized image. Pixel values; when When a pixel is located outside the valid region, its value is set to 0. After the above processing, the final output is a standardized grid image that meets the input requirements of lightweight deep learning convolutional neural networks such as the MobileNetV3 Small model. It is usually a uniform image with 3 channels (RGB) and 224×224 pixels.

[0074] In this way, pixel value normalization eliminates the dimensional differences between different image sources, making model training more stable and convergent faster. The proportional scaling and zero-padding strategy not only preserves the key geometric features of the original image, but also strictly meets the fixed input size requirements of the deep learning model, avoiding information loss due to image cropping or deformation errors caused by stretching, and providing input data with uniform format and controllable quality for subsequent high-precision disaster feature extraction and classification recognition.

[0075] For example, suppose the original size of the fused image of a certain high-quality grid cell is... Equal to 1200 pixels Equal to 800 pixels, target size is 224×224×224, calculate the scaling factor. Equals 0.1867, scaled size It equals 224. The height of the scaled image is 149. Since the height of the scaled image is less than 224, (224-149) / 2, or approximately 37 rows of zero pixels, are added to both the top and bottom sides in the vertical direction. This results in a 224×224 normalized image, where the top 37 rows and the bottom 38 rows are black backgrounds, and the middle 149 rows retain the scaled content of the original fused image.

[0076] In this embodiment, optionally, before using a lightweight deep learning convolutional neural network to extract and classify disaster features from multiple standardized grid images and determine the disaster identification result corresponding to each standardized grid image, the method further includes: acquiring historical satellite remote sensing image data and historical UAV image data of the power distribution network in the disaster area after the typhoon as sample multi-source image data; based on discrete grid technology, dividing the sample multi-source image data into grids according to multi-level nested grids, filtering data quality, and cleaning data to obtain multiple high-quality sample grid units, and performing multi-source data fusion and standardization processing on the multiple high-quality sample grid units to obtain multiple sample standardized grid images corresponding to the multiple high-quality sample grid units; performing sample category statistics on the multiple sample standardized grid images to obtain the number of sample disaster categories, and when the number of sample disaster categories is lower than a preset sample number threshold, performing cross-grid data augmentation processing on the multiple sample standardized grid images to obtain virtual disaster grid samples; constructing a standardized dataset using the multiple sample standardized grid images and virtual disaster grid samples, and training a lightweight deep learning convolutional neural network using the standardized dataset.

[0077] In this embodiment, before using a lightweight deep learning convolutional neural network to extract disaster features and classify standardized grid images, the network needs to be pre-trained. Therefore, the loss assessment system first acquires historical satellite remote sensing image data and historical UAV image data of the power distribution network in the disaster area after a typhoon, as sample multi-source image data. This historical data covers the damage to the power distribution network under different typhoon events and terrain conditions, including various disaster types such as tower tilting, line collapse, and insulator damage. Then, based on discrete grid technology, the loss assessment system uses the same method as in steps S10 and S20 to perform gridding, data quality screening, and data cleaning on these sample multi-source image data according to a multi-level nested grid, obtaining multiple high-quality sample grid units. Each high-quality sample grid unit then undergoes multi-source data fusion and standardization processing to generate a corresponding sample standardized grid image, ensuring that the training samples and subsequent actual application data maintain a high degree of consistency in spatial reference, resolution, and format.

[0078] After generating standardized grid images, the loss assessment system performs sample category statistics on all these images, counting the number of samples corresponding to each type of disaster, such as tower tilting, line collapse, and insulator damage. When the number of samples for a certain disaster category falls below a preset threshold, that category is identified as a minority disaster. To avoid overfitting or identification bias caused by class imbalance during model training, the loss assessment system automatically triggers cross-grid data augmentation. Cross-grid data augmentation utilizes the spatial topological adjacency relationships between discrete grids to extract features from the central grid containing minority disasters and its adjacent grids. Through spatial concatenation and weighted fusion, virtual disaster grid samples carrying real disaster characteristics and diverse backgrounds are generated, effectively expanding the number of minority samples without increasing the cost of on-site data collection.

[0079] Subsequently, the loss assessment system combines all existing standardized grid images with newly generated virtual disaster grid samples to construct a standardized dataset. This dataset covers balanced samples of various disasters, and each sample is a standardized image of uniform specifications, such as 3 channels and 224×224 pixels. Finally, this standardized dataset is used to train a lightweight deep learning convolutional neural network, such as MobileNetV3 Small. Through forward propagation, calculation of cross-entropy loss, and backpropagation, the network parameters are iteratively updated, enabling the model to fully learn the deep features of power distribution network disasters in multi-source fused images.

[0080] Thus, by introducing historical multi-source image data and adopting the same discrete grid processing procedure, the consistency of the feature space between the training data and the actual application data can be guaranteed in the above process. The cross-grid data augmentation mechanism can effectively alleviate the difficulties in typhoon disaster sample collection and the problem of sparse minority class samples, so that the trained model can still maintain a high recognition accuracy and generalization ability when facing disaster types with scarce samples. At the same time, training a lightweight network using a balanced standardized dataset can improve the robustness of disaster recognition while ensuring the model is lightweight and efficient, laying a solid foundation for rapid and accurate assessment in subsequent real-world scenarios.

[0081] For example, assuming only 50 "collapsed line" samples are collected from historical data, while there are 2000 normal samples, and the preset threshold is set to 100, the loss assessment system will generate 60 virtual collapsed line samples from the center grid of the existing collapsed line samples and its adjacent background grid through cross-grid data augmentation, expanding the collapsed line sample to 110, which meets the threshold requirement. The expanded dataset is then used to train the MobileNetV3 Small model, which significantly improves the accuracy of the model in identifying collapsed lines and avoids missed detections due to insufficient samples.

[0082] In this embodiment, optionally, sample category statistics are performed on multiple standardized grid images to obtain the number of sample disaster categories. If the number of sample disaster categories is lower than a preset sample number threshold, cross-grid data augmentation processing is performed on the multiple standardized grid images to obtain virtual disaster grid samples. This includes: counting the number of sample disaster categories for each disaster category in the multiple standardized grid images; when the number of sample disaster categories for a target disaster category is lower than the sample number threshold, marking the grid cell containing the target disaster category as the central disaster grid; based on the encoding mapping rules of discrete grids, and according to the row and column space index of the central disaster grid under the three-level grid division, retrieving the spatial topological adjacency of the central disaster grid as shown in the following expression. Neighborhood grid set ,

[0083] in, Central disaster grid Row and column space indexing under the RDCRMG three-level grid partitioning This is the offset of the spatial index. For the retrieved neighborhood grid cells; from the retrieved Neighborhood grid set A neighborhood grid is randomly selected as the background grid, and the normalized sample image matrix of the background grid is extracted as the background grid image matrix. Simultaneously, extract the standardized sample image matrix of the central disaster grid. Standardized sample image matrix for the central disaster grid Apply a geometric transformation function based on the center symmetry of the mesh. The transformed disaster feature matrix is ​​obtained. Obtain the binary spatial mask matrix of the disaster target. In this context, the pixel value of the disaster area is 1 and the pixel value of the background area is 0, represented by a binarized spatial mask matrix. From the central disaster grid, morphological edge detection or threshold segmentation are used. Extract from; extract the transformed disaster feature matrix. Background grid image matrix with background grid Spatial cascading and weighted fusion are performed, and the virtual grid sample image matrix is ​​calculated according to the virtual sample generation formula shown below. Cross-grid data augmentation processing was performed on multiple standardized grid images to obtain virtual disaster grid samples.

[0084] in, This represents the Hadamard product of a matrix.

[0085] In this embodiment, the loss assessment system first counts the number of samples for each disaster category in all standardized grid images. When the number of samples for a target disaster category is lower than a preset sample count threshold, the grid cell containing that category is marked as the central disaster grid. The loss assessment system then uses the RDCRMG discrete grid encoding mapping rules and the row and column space indexes of the central disaster grid under the three-level grid partitioning. Automatically retrieve its spatial topological adjacency Neighborhood grid set For the specific search formula, please refer to Formula 14 below: Formula 14:

[0086] In formula 14, Central disaster grid The row and column space index; The offset of the spatial index, which can be -1, 0, or 1 but not all of them can be 0 at the same time; This refers to the retrieved neighborhood grid cells. In this embodiment of the application, the number of neighborhoods is... The value is set to 8, meaning that the system searches for the eight neighboring grids surrounding the central grid, specifically the top, bottom, left, right, and four diagonal directions. The loss assessment system randomly selects a neighboring grid from the retrieved set as the background grid, and extracts the standardized sample image matrix of this background grid, denoted as . Simultaneously, the standardized sample image matrix of the central disaster grid is extracted and denoted as... .

[0087] Next, a geometric transformation function based on the grid's central symmetry is applied to the normalized sample image matrix of the central disaster grid. For example, random rotation, flipping, or affine transformation can be used to enrich the disaster perspective and obtain the transformed disaster feature matrix. Simultaneously, the binarized spatial mask matrix of the disaster target is obtained from the central disaster grid. The mask matrix is ​​extracted using morphological edge detection or thresholding methods, where pixel values ​​in disaster areas are set to 1 and pixel values ​​in background areas are set to 0. Then, the transformed disaster feature matrix and the background grid image matrix are spatially concatenated and weighted, and the virtual grid sample image matrix is ​​calculated according to the virtual sample generation formula shown in Formula 15 below. : Formula 15:

[0088] In formula 15, This represents the Hadamard product of a matrix, which is the element-wise multiplication of corresponding positions. Formula 15 means that in the disaster area, i.e., the mask... The position marked as 1 represents the transformed disaster feature; in the background area, i.e., the mask... The image pixels of the background grid are taken at position 1. In this way, the generated virtual sample can retain the morphological characteristics of the real disaster target, while also obtaining the natural terrain background of the adjacent grid, making the synthesized sample more consistent with the actual geographical environment.

[0089] Finally, the loss assessment system will generate virtual mesh samples. The data are uniformly incorporated into the original standardized dataset, resulting in the final, complete standardized dataset after data augmentation and cross-grid sample information sharing, as shown in the expression in Formula 16 below: Formula 16:

[0090] The final complete standardized dataset in Formula 16 It will be used to train a lightweight deep learning convolutional neural network.

[0091] In this way, by utilizing the topological adjacency relationship of discrete grids for cross-grid data augmentation, the limitations of traditional single-image random cropping and flipping augmentation methods can be overcome. The generated virtual samples not only expand the number of samples for sparse disaster categories, but also introduce the geographical background of real adjacent grids, making the synthetic samples more environmentally realistic and diverse. This can effectively alleviate the problems of difficult typhoon disaster sample collection and uneven distribution, prevent the model from overfitting during training, and significantly improve the generalization ability and recognition robustness of lightweight deep learning models in complex and unseen areas.

[0092] For example, suppose the "insulator damage" category in the training set has only 30 samples, below the threshold of 100. We select a central grid cell whose 8-neighborhood includes a pure background grid. We apply a 90° rotation to this central grid cell to obtain the transformed disaster features and extract the binarized mask of the damaged insulator. After fusing according to the formula, a virtual sample is generated: the damaged insulator is located on the new background. We repeat this process to generate 70 virtual samples, bringing the total number of samples in this category to 100. The finally trained model can maintain high recognition accuracy when facing insulator damage under different backgrounds.

[0093] Optionally, in this embodiment, a standardized dataset is constructed using multiple sample standardized grid images and virtual disaster grid samples, and a lightweight deep learning convolutional neural network is trained using the standardized dataset, including: constructing a standardized dataset from multiple sample standardized grid images and virtual disaster grid samples. It initializes an untrained lightweight deep learning convolutional neural network and loads the pre-trained weight parameters of the lightweight deep learning convolutional neural network; it then standardizes the dataset. The input is fed into an untrained lightweight deep learning convolutional neural network. Based on this untrained lightweight deep learning convolutional neural network, the input standardized dataset is processed through standard convolutional layers with a stride of 2, combined with batch normalization and the h-swish activation function, according to the following formula. Preliminary feature extraction is performed to obtain the initial convolutional layer output feature map. ,

[0094] in, A standard convolutional layer with a stride of 2 uses a 3×3 convolutional kernel. Based on the 11-layer bottleneck residual module in a lightweight deep learning convolutional neural network, the feature map output of the initial convolutional layer is processed. Deep feature extraction is performed, where each bottleneck residual module employs a lightweight structure combining depthwise separable convolutions and 1×1 pointwise convolutions, and embeds a compression-excitation attention mechanism. For the th... Each bottleneck residual module inputs the feature map using the following formula. After dimensionality upscaling via 1×1 pointwise convolution, spatial feature extraction via 3×3 depthwise separable convolution, and weighting via a compression-excitation attention mechanism, the feature is reduced by another 1×1 pointwise convolution and finally combined with the input feature map. Perform residual connections and output feature maps. ,

[0095] in, It is a 3×3 depthwise separable convolution with a stride of [missing value]. The value is either 1 or 2 and is used to control feature map downsampling. The feature map is constructed using a 1×1 pointwise convolution. The SE (Search-Enabled Attention) mechanism employs global average pooling and two fully connected layers to adaptively weight channel features. After the output of the final bottleneck residual module, the feature map dimension is expanded using a 1×1 pointwise convolution layer according to the following formula, and the h-swish activation function is used to map the feature dimensions, resulting in the feature map. ,

[0096] in, Output feature maps for the bottleneck residual module of layer 11; perform global average pooling on the feature maps according to the following formula. Spatial dimensionality compression is performed to obtain a global one-dimensional feature vector. ,

[0097] in, For feature map All channels in the space coordinates , To calculate the mean after doubly summing the pixel values ​​across all 7 spatial dimensions; the global one-dimensional feature vector... The input is fed into two fully connected layers in a lightweight deep learning convolutional neural network. The first fully connected layer combines batch normalization and the h-swish activation function to process the global one-dimensional feature vector. A nonlinear transformation of the features is performed, and the output of the second fully connected layer in the two-layer fully connected layer is a logits vector that matches the number of disaster categories. Then, the logits vector is converted into a predicted probability distribution for each category using the Softmax activation function according to the following formula.

[0098] in, This is the classification layer weight matrix. For bias vectors, For global one-dimensional feature vectors, The Softmax activation function is used. For the output of the first The predicted probability of disasters is calculated using the cross-entropy loss function shown in the following formula. The model parameters of the lightweight deep learning convolutional neural network during training are optimized by iteratively updating the model parameters through backpropagation, resulting in a fully trained lightweight deep learning convolutional neural network.

[0099] in, For the number of disaster categories, For real labels coding, For the lightweight deep learning convolutional neural network prediction during the training process, the first... Probability of disasters.

[0100] In this embodiment, the loss assessment system first combines all the standardized grid images of the samples with the generated virtual disaster grid samples to form a standardized dataset. It initializes an untrained lightweight deep learning convolutional neural network, which can adopt the MobileNetV3 Small standard architecture and loads the weight parameters of the network pre-trained on the ImageNet dataset to achieve transfer learning of feature extraction capabilities, avoiding the problems of slow convergence speed and poor generalization caused by training from scratch.

[0101] Then, the loss assessment system will standardize the dataset. The input image is fed into an untrained lightweight deep learning convolutional neural network. The model first performs preliminary feature extraction on the input image through a standard convolutional layer with a stride of 2. This convolutional layer uses a 3×3 kernel and 16 output channels, combining batch normalization (BN) and the h-swish activation function to complete low-dimensional feature mapping and non-linear transformation of the input image, extracting basic visual features such as edges and textures. The output feature map size is 16×112×112, achieving spatial dimension downsampling and feature dimension expansion. Specifically, the calculation process is shown in Equation 17 below: Formula 17:

[0102] in, This indicates a standard convolution operation using a 3×3 kernel and a stride of 2; BN stands for batch normalization; h-swish is the MobileNetV3-specific activation function, defined as... , where ReLU6 is a modified linear unit with an upper limit of 6.

[0103] Subsequently, based on the 11-layer bottleneck residual module (bneck) stacked structure of MobileNetV3 Small, the feature maps output from the initial convolutional layers are processed. Deep feature extraction is performed. Each bneck module employs a lightweight structure combining depthwise convolution and 1×1 pointwise convolution, and embeds a compression-excitation (SE) attention mechanism. Specifically, depthwise separable convolution decomposes standard convolution into channel-wise depthwise convolution and pointwise convolution, significantly reducing the number of parameters; the SE attention mechanism uses global average pooling and two fully connected layers to adaptively weight channel features, enabling the model to focus on key damage features of distribution network facilities such as tower tilting and line collapse. For the first... A bottleneck residual module, input feature map After dimensionality upscaling via 1×1 pointwise convolution, spatial feature extraction via 3×3 depthwise separable convolution, and weighting via SE attention mechanism, followed by dimensionality reduction via 1×1 pointwise convolution, the result is finally compared with the input feature map. Perform residual connections and output feature maps. The calculation process is shown in Formula 18 below: Formula 18:

[0104] In formula 18, It is a 3×3 depthwise separable convolution with a stride of [missing value]. Choose 1 or 2 to control the feature map downsampling; The algorithm employs a 1×1 pointwise convolution; SE (Search Engine) is a compression-enhanced attention mechanism that generates channel weights through global average pooling and two fully connected layers, scaling the feature map output by the depthwise separable convolution. The first layer uses ReLU activation, and the second layer uses h-sigmoid activation. After stacking 11 bottleneck residual modules, the output feature map size is 96×7×7, completing the extraction and fusion of multi-scale deep features and fully capturing the morphology, texture, and damage characteristics of power distribution network facilities.

[0105] After the output of the final bottleneck residual module, the feature map dimension is expanded to 576×7×7 through a 1×1 pointwise convolutional layer according to the following formula 19, and the h-swish activation function is used to further map the feature dimensions to obtain the feature map. : Formula 19:

[0106] in, Output feature map for the bottleneck residual module of layer 11.

[0107] Subsequently, global average pooling is applied according to the following formula 20. Spatial dimensionality compression is performed, transforming the 576×7×7 feature map into a 576×1×1 global one-dimensional feature vector. To achieve global aggregation and feature compression of spatial information: Formula 20:

[0108] In formula 20, Representation of feature map All channels in the space coordinates Pixel value at; This is to take the mean after doubly summing all pixel values ​​in the 7×7 spatial dimension.

[0109] The loss assessment system will use the global one-dimensional feature vector The input is fed into two fully connected layers. The first fully connected layer performs a nonlinear transformation of features using batch normalization and the h-swish activation function. The second fully connected layer outputs a logits vector that matches the number of disaster categories, which can be one of four: normal, tower tilt, line collapse, or insulator damage. Then, according to Equation 21, the logits vector is converted into a predicted probability distribution for each category using the Softmax activation function. Formula 21:

[0110] in, This is the classification layer weight matrix. For bias vectors, For global one-dimensional feature vectors, The softmax activation function outputs the first... Predicted probability of disasters .

[0111] Meanwhile, during the model training phase, the cross-entropy loss function shown in Equation 22 is used to optimize the model parameters: Formula 22:

[0112] In formula 22, Number of disaster categories; For real labels Encoding, when the sample belongs to the first Class Time The value is 1 if it is set to 1, and 0 otherwise. The model predicts the first Disaster probability classification. By iteratively updating the network parameters through backpropagation, a lightweight deep learning convolutional neural network is obtained, which can achieve fast inference of 3×224×224 normalized images and accurate output of power distribution network disaster categories.

[0113] Thus, by utilizing the pre-trained MobileNetV3 Small model for transfer learning through the above process, the training convergence speed can be significantly accelerated and the generalization ability improved. At the same time, the combination of depthwise separable convolution and SE attention mechanism can automatically focus on the deep damage characteristics of distribution network facilities while ensuring lightweight efficiency. Moreover, the design of global average pooling and fully connected layers makes the model output directly correspond to the disaster category. The training process uses cross-entropy loss for end-to-end optimization. The final trained model can quickly and accurately identify various distribution network faults in large-scale disaster areas, providing reliable technical support for emergency repairs.

[0114] For example, suppose the constructed standardized dataset contains 2000 samples, including 1800 normal samples, 100 tilted tower samples, 60 collapsed line samples, 40 insulator-damaged samples, and 100 virtual samples generated through data augmentation, ensuring that each category has more than 100 samples. The loss assessment system inputs the dataset into the MobileNetV3 Small model. After the aforementioned forward propagation calculation, it outputs a 4-dimensional probability vector for each sample. The difference between the predicted and true labels is calculated using cross-entropy loss, and the network parameters are updated through backpropagation using the Adam optimizer. After training for 20 epochs, the model achieves a 94% accuracy rate in recognizing tilted towers and a 91% accuracy rate in receding line heights on the validation set, with a single image inference time of only 15 milliseconds, meeting real-time requirements.

[0115] In practical applications, the network structure of lightweight deep learning convolutional neural networks is as follows: Figure 2 As shown, the MobileNetV3 Small architecture is used. The network input is a 224×224 pixel 3-channel image, i.e., 224×224×3. The input first passes through a standard convolutional layer with a stride of 2, namely Conv2_s2, followed by a batch normalization (BN) layer and an h-swish activation function (hswish). Subsequently, the network stacks 11 bottleneck residual modules, i.e., Bneck Small:11, which use a structure combining depthwise separable convolutions and attention mechanisms for deep feature extraction. After the 11 bottleneck residual modules, a global average pooling layer, i.e., Avg_pool, is applied to compress the spatial dimension of the feature map. The pooled feature vector passes through a fully connected layer (FC), a batch normalization layer (BN), and an h-swish activation function (hswish) in sequence, and finally passes through another fully connected layer (FC) to output a classification result with a dimension of 1×1000, corresponding to the 1000 categories of the ImageNet dataset. Figure 2 The network structure shown achieves efficient feature extraction and classification through a lightweight convolutional design and attention mechanism.

[0116] Furthermore, in practical applications, the module structure of the bottleneck residual module is as follows: Figure 3 As shown, the input of the module is a size of The feature tensor is first input through a 1×1 expansion convolution, followed by a batch normalization (BN) layer and an activation function (h-swish or ReLU). The expanded feature map then enters a depthwise convolution layer with a kernel size of 3×3 or 5×5. This layer is also followed by a batch normalization layer and an h-swish or ReLU activation function for spatial feature extraction, with a stride denoted as s. After the depthwise convolution, a squeeze-and-excitation (SE) module is connected. This module first compresses the feature map into channel description vectors through global average pooling, then passes it through a 1×1 fully connected layer (FC) with a ReLU activation function, and another 1×1 fully connected layer with an h-sigmoid activation function to generate scaling weights for each channel, and performs channel scaling on the output of the depthwise convolution. After weighting by the SE module, the feature map is then passed through a 1×1 dimensionality reduction convolution (projection convolution) and a batch normalization layer to transform the number of channels into the target dimension. Finally, the module output size is... The feature tensor.

[0117] Optionally, in this embodiment of the application, based on the geospatial coding of each standardized grid image, the disaster identification results corresponding to each standardized grid image are inversely mapped and spatially stitched to obtain and output the spatiotemporal distribution assessment results of equipment failure losses within a large-scale area, including: The disaster identification result corresponding to each standardized grid image is uniquely encoded with the original three-level nested grid corresponding to the standardized grid image. Binding is performed to achieve the binding of disaster perception attributes and grid spatial attributes; According to the following formula, a spatial inverse mapping based on discrete grid coding is performed on each normalized grid image to determine the actual spatial extent of each normalized grid image in the universal transverse Mercator projection coordinate system.

[0118] in, Mesh encoding mapping function for RDCRMG architecture The inverse mapping, A unique code for a three-level nested grid. For the first Geographic parameters with a first-order projection and grid. For the first Spatial parameters of a second-level 10km×10km grid, For the first The subdivision parameters of a 1km×1km three-level grid; The coordinate transformation parameters are used to convert the actual spatial extent of each standardized grid image in the universal Mercator projection coordinate system into the geographic boundary in the preset latitude and longitude coordinate system. ; Against a large-scale geographical background, based on the geographical boundaries corresponding to each standardized grid image, the disaster identification results corresponding to each standardized grid image are spatially stitched together to obtain and output the spatiotemporal distribution assessment results of equipment failure losses. For overlapping coverage areas at the edges of adjacent grids, a probability-weighted fusion strategy is used according to the following formula to determine the final disaster attributes of the overlapping coverage areas. ,

[0119] in, An overlapping set of grids covering a specific area. For the first Confidence weights for a standardized grid image In order to target the The first standardized grid image inferred from the first The original probability of disasters.

[0120] In this embodiment, firstly, a lightweight deep learning convolutional neural network infers and outputs a disaster category probability vector for any standardized input image. Automatically retrieve the unique code of the original three-level nested grid corresponding to the standardized grid image. ,in , , These are first-level, second-level, and third-level grid indexes, which enable one-to-one linking of disaster perception attributes with grid spatial attributes. This associates the dimensionless classification probabilities output by the model with grid cells that have clear spatial locations, laying the foundation for subsequent geographic coordinate transformation.

[0121] Next, a spatial inverse mapping based on discrete grid coding is performed, specifically utilizing the RDCRMG architecture grid coding mapping function. inverse mapping According to grid coding and its included first-level projection zone index Secondary management grid index and three-level data processing grid index The actual spatial extent of the mesh element in the Universal Transverse Mercator (UTM) projection coordinate system is calculated inversely. The inverse mapping formula is given in Formula 23 below: Formula 23:

[0122] In formula 23, A unique code for a three-level nested grid; For the first Geographic parameters of a first-level projection zone grid, such as the central meridian of the projection zone and the latitude of the origin; For the first Spatial parameters of a second-level 10km×10km grid, such as the coordinates of the grid's starting point; For the first The meshing parameters of a 1km×1km level 3 grid, such as the local offset within the grid, can be obtained through this inverse mapping. The minimum and maximum X and Y coordinate values ​​of this grid cell in the UTM coordinate system can then be obtained. and This determines the rectangular geographical extent of the grid.

[0123] Subsequently, the loss assessment system uses preset coordinate transformation parameters, such as a seven-parameter Bursa model or a three-parameter translation, to convert the UTM projected coordinates into the true geographic boundary in the WGS-84 latitude and longitude coordinate system, thus obtaining... This conversion enables disaster identification results to be compatible with general geographic information systems (GIS), facilitating subsequent visualization and decision analysis.

[0124] Against a large-scale geographical backdrop, the loss assessment system spatially stitches together the disaster identification results of all grids based on the latitude and longitude boundaries corresponding to each standardized grid image. For areas where adjacent grid edges may overlap, such as due to duplicate sampling of boundary pixels during grid division, the loss assessment system employs a probability-weighted fusion strategy to determine the final disaster attribute of that area. Specifically, the fusion formula is shown in Formula 24 below: Formula 24: in, A collection of all grid cells covering a specific overlapping region; For the first The confidence weight of each grid cell can be set based on the quality evaluation index of that grid or the confidence score of the model output. For the first The first grid inference derived from the first grid The original probability of a disaster type is the logits or direct probability value before the Softmax output. Equation 24 means that for each overlapping location, the weighted probability sum of each disaster type across all grids covering that location is calculated, and the disaster type with the largest sum is taken as the final disaster attribute for that location. This method effectively eliminates inconsistencies or noise at the splicing boundaries, ensuring the consistency and continuity of the global evaluation results.

[0125] Finally, the loss assessment system will output the above stitched and merged results into a standard format file, such as GeoJSON, Shapefile, or GeoTIFF, based on the Geographic Information System (GIS) protocol. This will present the precise categories, quantities, and geographical distribution of damaged facilities in the entire power distribution network after the typhoon on a large-scale map, forming a visualized map for power emergency repair decision-making.

[0126] In this way, during the above process, the classification results of the deep learning model can be accurately restored to real geographic coordinates through grid coding inverse mapping, realizing a closed-loop mapping from the perception results to the physical world. At the same time, the probability weighted fusion strategy can effectively solve the problem of repeated coverage at the grid splicing edges and avoid the ambiguity of disaster attributes. The final generated GIS standard file can be directly imported into various emergency command platforms, providing intuitive and quantitative spatial information support for the dispatch of repair teams, material allocation and power restoration decisions, significantly improving the practicality and operability of the power distribution network disaster assessment after the typhoon.

[0127] For example, suppose the disaster identification result of a standardized grid image is "tower tilt", and its corresponding grid code is: This represents the 3rd UTM zone, the 5th secondary grid, and the 2nd tertiary grid. Through inverse mapping calculations, the range of this grid in the UTM coordinate system is (500000m, 3100000m) to (501000m, 3101000m), which translates to [118.5°E~118.51°E, 31.2°N~31.21°N] in WGS-84 latitude and longitude. This grid overlaps with the adjacent grid to its right by 10 pixels. The right grid is identified as "normal" with a confidence weight of 0.8, while this grid is identified as "tower tilt" with a weight of 0.9. After weighted fusion, the overlapping area is ultimately determined to be "tower tilt." All grids are stitched together to generate a GeoJSON file, and all tower tilt locations are marked with red icons on the electronic map to guide repair personnel to the site as quickly as possible.

[0128] In summary, the technical solution framework in the embodiments of this application is as follows: Figure 4As shown, the loss assessment system consists of an input layer, a data processing layer, and an output layer. The input layer acquires remote sensing satellite data and UAV data of the disaster-stricken power distribution network after the typhoon as multi-source inputs. The data processing layer first integrates the multi-source data, fusing satellite and UAV imagery. Then, based on discrete grid technology, it performs block processing on the large-scale data to achieve gridded division and management; sample augmentation is performed at the grid level to expand the diversity of disaster samples; subsequently, the MobileNetV3 Small lightweight model is used for feature extraction and disaster identification and classification. The output layer generates a complete distribution map of the power distribution network disaster, presenting the assessment results of the spatiotemporal distribution of equipment failure losses within a large-scale area. Thus, the technical solution in this embodiment combines lightweight deep learning convolutional neural networks with discrete grid technology, fusing remote sensing optical satellite and UAV multi-source imagery for the post-typhoon power distribution network disaster identification scenario, breaking through the existing technical bottlenecks in multiple dimensions such as identification efficiency, accuracy, data adaptability, and scenario applicability. Discrete grid technology divides large-scale disaster areas into standardized grid units, enabling block-based parallel processing of multi-source images. This solves the memory overflow and computational lag issues that arise when lightweight deep learning directly processes large-area images. Simultaneously, the lightweight network simplifies model parameters through lightweight designs such as depthwise separable convolutions, resulting in grid-level inference speeds far exceeding those of traditional deep learning. The combination of these two technologies improves disaster image processing efficiency several times over, meeting the rapid reconnaissance needs of emergency repairs. Discrete grid technology completes the standardized integration and quality optimization of multi-source images, removing interference information such as clouds and fog, providing high-quality feature input for lightweight deep learning. Lightweight convolutional neural networks can automatically mine the texture, morphology, and deep-seated features of disaster damage in power distribution network facilities, replacing traditional manual feature extraction. This effectively solves the problems of missed or misjudged minor disasters such as tower tilting and line collapse, significantly improving recognition accuracy and achieving refined fault location. Discrete grid technology enables grid-level data augmentation, facilitating cross-grid sample information sharing and virtual sample generation. This overcomes the difficulties and uneven distribution of typhoon disaster samples. Lightweight deep learning is used to specifically train the augmented grid-level samples, allowing the model to adapt to power distribution network scenarios with different terrains and disaster types. This avoids overfitting and maintains stable recognition performance in areas where no samples have been collected, effectively addressing the sample sparsity problem and enhancing generalization ability. Discrete grid technology establishes a standardized multi-source data processing framework, unifying the spatial reference and resolution of satellite and UAV imagery. This solves the problems of data heterogeneity and the difficulty of fusion. Lightweight deep learning jointly learns the multi-source features after fusion within the grid, fully utilizing the wide coverage of satellite imagery and the high resolution of UAV imagery to comprehensively capture the disaster characteristics of the power distribution network. Compared to single data sources, the recognition results are more comprehensive and accurate, perfectly adapting to multi-source data fusion and fully leveraging the complementary value of imagery.

[0129] The method provided in this application, by fusing satellite remote sensing and UAV multi-source imagery and performing multi-level nested partitioning, quality screening, and standardized fusion based on discrete grid technology, can effectively solve the problems of single data source being susceptible to interference and incomplete coverage. It can comprehensively capture various damage characteristics of the distribution network. At the same time, it uses a lightweight deep learning convolutional neural network to replace the conventional large parameter model. Under the grid-level parallel processing framework, it can significantly reduce the computational burden and inference latency, avoid memory overflow, and meet the timeliness requirements of emergency repair. Furthermore, by combining grid-level data augmentation and spatial inverse mapping stitching, it can improve the accuracy of disaster identification while achieving efficient assessment of large-scale areas. Ultimately, it balances identification efficiency and accuracy, providing reliable technical support for emergency repair of the distribution network after a typhoon.

[0130] Furthermore, as Figure 1 In a specific implementation of the method, this application provides a device for assessing the spatiotemporal distribution of equipment failure losses based on lightweight deep learning and discrete grid technology, such as... Figure 5 As shown, the device includes: a division module 501, a fusion module 502, an identification module 503, and a generation module 504.

[0131] The partitioning module 501 is used to acquire satellite remote sensing image data and UAV image data of the power distribution network in the disaster area after the typhoon passes as multi-source remote sensing image data. Based on discrete grid technology, the multi-source remote sensing image data is divided into grids according to multi-level nested grids to obtain multiple grid units. The fusion module 502 is used to perform data quality screening and data cleaning within the multiple grid cells to obtain multiple high-quality grid cells, and to perform multi-source data fusion and standardization processing on the multiple high-quality grid cells to obtain multiple standardized grid images corresponding to the multiple high-quality grid cells. The identification module 503 is used to extract and classify disaster features from the plurality of standardized grid images using a lightweight deep learning convolutional neural network, and to determine the disaster identification result corresponding to each of the plurality of standardized grid images; The generation module 504 is used to perform inverse mapping and spatial stitching on the disaster identification results corresponding to each standardized grid image based on the geospatial coding of each standardized grid image, so as to obtain and output the spatiotemporal distribution assessment results of equipment failure loss in a large-scale area.

[0132] In specific application scenarios, the segmentation module 501 is used to convert the satellite remote sensing image data and the UAV image data into raster data with geographic coordinate mapping relationships using the following formula, so as to obtain multi-source raster data.

[0133] in, For the first Data source in geographic coordinates The raster pixel value at that location, For the first Rasterization mapping function for data source class, For the first The original pixel values ​​of the data source in its own pixel coordinate system. For the first The set of rasterization transformation parameters for a data source. The value can be 1 or 2. A value of 1 corresponds to the satellite remote sensing image data, and A value of 2 corresponds to the UAV image data. A three-level nested grid partitioning strategy is adopted, using the following formula: a 6° universal transverse Mercator projection zone as the first-level grid, 10km×10km as the second-level grid, and 1km×1km as the third-level grid. The geographic space of the power distribution network monitoring area is then subdivided, and the multi-source raster data is mapped to the partitioned geographic space. This multi-level nested grid partitioning of the multi-source remote sensing image data yields multiple grid units.

[0134] in, A unique code for a three-level nested grid. For the mesh coding mapping function of the RDCRMG architecture, For the first Geographic parameters with a first-order projection and grid. For the first Spatial parameters of a second-level 10km×10km grid, For the first The subdivision parameters of a 1km×1km third-level grid, For the first Data source in geographic coordinates The raster pixel value at that location.

[0135] In specific application scenarios, the fusion module 502 is used for unique encoding based on a three-level nested grid. Each of the grid cells is located, and the following formula is used to determine the rasterized pixel values ​​in each grid cell. Determine the first Pixel values ​​of three-level grid units ,

[0136] in, Mesh encoding mapping function for RDCRMG architecture The inverse mapping, These are classified as first-level, second-level, and third-level grid indices; based on the rasterized pixel values ​​in each grid cell, the proportion of cloud cover within each grid cell is calculated. Percentage of effective coverage of images within the grid and the standard deviation of the grid pixel values. The comprehensive quality evaluation index for each grid cell is calculated using the following formula.

[0137] in, For the first The comprehensive quality evaluation index corresponding to each grid cell The weight coefficients for each item and Determine the preset grid data quality threshold. According to the following formula, combined with the grid data quality threshold Low-quality data is removed from the plurality of grid cells to obtain the plurality of high-quality grid cells.

[0138] in, For the plurality of high-quality mesh cells, the first In the first high-quality grid cell Raster data of the class, This indicates that the corresponding grid cell is low-quality data and should be cleared.

[0139] In specific application scenarios, the fusion module 502 is used to utilize a preset unified processing resolution. The following formula is used to reconstruct the resolution function. The original heterogeneous data within each of the high-quality grid cells is resampled and scaled to obtain the first... Within the first high-quality grid cell Normalized pixel matrix after resolution reconstruction of the data source ,

[0140] in, For the plurality of high-quality mesh cells, the first In the first high-quality grid cell Raster data of the class; using the following formula, according to the fusion weights The first Pixel-level adaptive weighted fusion of multi-source data within a high-quality grid cell yields the first... The fused raster pixel values ​​corresponding to each high-quality grid cell

[0141] in, For the first The raster pixel values ​​obtained by fusing multi-source data within a high-quality grid cell. The fusion weights are adaptively determined by the signal-to-noise ratio and geometric resolution contribution of the data source; according to the following formula, for the ... Radiometric fine correction is performed on the fused raster pixel values ​​corresponding to each high-quality grid cell.

[0142] in, For the first The radiometrically corrected raster pixel values ​​within a high-quality grid cell. They are respectively from the first The radiation correction gain coefficient and offset coefficient are determined by the measured radiation values ​​of ground control points within each high-quality grid cell. Simultaneously, a quadratic polynomial correction model is used to correct the geographic coordinates of each high-quality grid cell according to the following formula.

[0143] in, These are the geometrically corrected geographic coordinates. The geographic coordinates before correction. to The quadratic polynomial geometric correction parameters are based on ground control point calibration. The raster data of each high-quality grid cell in the corrected high-quality grid cells are standardized to obtain multiple standardized grid images corresponding to the multiple high-quality grid cells.

[0144] In specific application scenarios, the fusion module 502 is used to perform pixel value normalization processing on the raster data of the corrected high-quality grid cell for each high-quality grid cell using the following linear normalization formula.

[0145] in, For normalized raster data, These represent the maximum and minimum pixel values ​​of the raster data within the corresponding high-quality grid cell, respectively; the following formula is used based on the target image width required by the training model. and target image height Calculate the size scaling factor for the high-quality mesh cells. ,

[0146] in, The original width and height of the merged raster data in the high-quality grid cell are given; the scaled image width of the high-quality grid cell is calculated using the following formula. and the height of the scaled image ,

[0147] Referring to the following formula, zero-fill is applied to the regions of the high-quality mesh cells that exceed the scaling size to obtain the normalized mesh image corresponding to the high-quality mesh cells.

[0148] in, To standardize the pixel values ​​of a grid image. To standardize the pixel coordinates of the image, This is a floor operation.

[0149] In specific application scenarios, the device further includes: The training module is used to acquire historical satellite remote sensing image data and historical UAV image data of the power distribution network in the disaster area after a typhoon as sample multi-source image data. Based on discrete grid technology, the sample multi-source image data is divided into grids according to multi-level nested grids, and data quality is filtered and data is cleaned to obtain multiple high-quality sample grid units. Multi-source data fusion and standardization processing are performed on the multiple high-quality sample grid units to obtain multiple sample standardized grid images corresponding to the multiple high-quality sample grid units. Sample category statistics are performed on the multiple sample standardized grid images to obtain the number of sample disaster categories. When the number of sample disaster categories is lower than a preset sample number threshold, cross-grid data augmentation processing is performed on the multiple sample standardized grid images to obtain virtual disaster grid samples. A standardized dataset is constructed using the multiple sample standardized grid images and the virtual disaster grid samples, and the lightweight deep learning convolutional neural network is trained using the standardized dataset.

[0150] In specific application scenarios, the training module is used to count the number of sample disaster categories for each disaster category in the multiple sample standardized grid images. When the number of sample disaster categories for a target disaster category is lower than the sample number threshold, the grid cell containing the target disaster category is marked as the central disaster grid. Based on the encoding mapping rules of discrete grids, and according to the row and column space index of the central disaster grid under the three-level grid division, the spatial topological adjacency of the central disaster grid as shown in the following expression is retrieved. Neighborhood grid set ,

[0151] in, Central disaster grid Row and column space indexing under the RDCRMG three-level grid partitioning This is the offset of the spatial index. For the retrieved neighborhood grid cells; from the retrieved Neighborhood grid set A neighborhood grid is randomly selected as the background grid, and the normalized sample image matrix of the background grid is extracted as the background grid image matrix. Simultaneously, the standardized sample image matrix of the central disaster grid is extracted. ; the standardized sample image matrix of the central disaster grid Apply a geometric transformation function based on the center symmetry of the mesh. The transformed disaster feature matrix is ​​obtained. Obtain the binary spatial mask matrix of the disaster target. In this context, the pixel value of the disaster area is 1 and the pixel value of the background area is 0. The binarized spatial mask matrix... From the central disaster grid, morphological edge detection or threshold segmentation are used. Extract from; extract the transformed disaster feature matrix. Background grid image matrix of the background grid Spatial cascading and weighted fusion are performed, and the virtual grid sample image matrix is ​​calculated according to the virtual sample generation formula shown below. The process involves performing cross-grid data augmentation on the multiple standardized grid images to obtain the virtual disaster grid samples.

[0152] in, This represents the Hadamard product of a matrix.

[0153] In specific application scenarios, the training module is used to construct the standardized dataset by combining the multiple sample standardized grid images and the virtual disaster grid samples. The untrained lightweight deep learning convolutional neural network is initialized, and the pre-trained weight parameters of the lightweight deep learning convolutional neural network are loaded; the standardized dataset is then used. The input is fed into the untrained lightweight deep learning convolutional neural network, which, based on the untrained lightweight deep learning convolutional neural network, passes through standard convolutional layers with a stride of 2, combined with batch normalization and the h-swish activation function, according to the following formula, to normalize the input dataset. Preliminary feature extraction is performed to obtain the initial convolutional layer output feature map. ,

[0154] in, The instruction states that the standard convolutional layer with a stride of 2 uses a 3×3 convolutional kernel. Based on the 11-layer bottleneck residual module in the lightweight deep learning convolutional neural network, the initial convolutional layer outputs a feature map. Deep feature extraction is performed, wherein each bottleneck residual module adopts a lightweight structure combining depthwise separable convolution and 1×1 pointwise convolution, and embeds a compression-excitation attention mechanism. For the th Each bottleneck residual module inputs the feature map using the following formula. After dimensionality upscaling via 1×1 pointwise convolution, spatial feature extraction via 3×3 depthwise separable convolution, and weighting via a compression-excitation attention mechanism, the feature is reduced by another 1×1 pointwise convolution and finally combined with the input feature map. Perform residual connections and output feature maps. ,

[0155] in, It is a 3×3 depthwise separable convolution with a stride of [missing value]. The value is either 1 or 2 and is used to control feature map downsampling. The feature map is constructed using a 1×1 pointwise convolution. The SE (Search-Enabled Attention) mechanism employs global average pooling and two fully connected layers to adaptively weight channel features. After the output of the final bottleneck residual module, the feature map dimension is expanded using a 1×1 pointwise convolution layer according to the following formula, and the h-swish activation function is used to map the feature dimensions, resulting in the feature map. ,

[0156] in, Output a feature map for the bottleneck residual module of layer 11; perform global average pooling on the feature map according to the following formula. Spatial dimensionality compression is performed to obtain a global one-dimensional feature vector. ,

[0157] in, For feature map All channels in the space coordinates , To calculate the mean value after double summing of pixel values ​​across all spatial dimensions of the 7×7 area; the global one-dimensional feature vector... The input is fed into two fully connected layers in the lightweight deep learning convolutional neural network. The first fully connected layer combines batch normalization and the h-swish activation function to process the global one-dimensional feature vector. A feature nonlinear transformation is performed, and the second fully connected layer outputs a logits vector matching the number of disaster categories. Then, the logits vector is converted into a predicted probability distribution for each category using the Softmax activation function according to the following formula.

[0158] in, This is the classification layer weight matrix. For bias vectors, For the global one-dimensional feature vector, The Softmax activation function is... For the output of the first The predicted probability of disasters is calculated using the cross-entropy loss function shown in the following formula. The model parameters of the lightweight deep learning convolutional neural network during training are optimized by iteratively updating the model parameters through backpropagation, resulting in the completed lightweight deep learning convolutional neural network.

[0159] in, For the number of disaster categories, For real labels coding, The lightweight deep learning convolutional neural network predicts the first [number] [number] [unit] during the training process. Probability of disasters.

[0160] In specific application scenarios, the generation module 504 is used to uniquely encode the disaster identification result corresponding to each standardized grid image with the original three-level nested grid corresponding to the standardized grid image. Binding is performed to achieve the binding of disaster perception attributes and grid spatial attributes; according to the following formula, a spatial inverse mapping based on discrete grid coding is performed on each of the standardized grid images to determine the actual spatial range of each of the standardized grid images in the universal transverse Mercator projection coordinate system.

[0161] in, Mesh encoding mapping function for RDCRMG architecture The inverse mapping, A unique code for a three-level nested grid. For the first Geographic parameters with a first-order projection and grid. For the first Spatial parameters of a second-level 10km×10km grid, For the first The subdivision parameters of a 1km×1km grid at level 3; and the conversion parameters to transform the actual spatial extent of each standardized grid image in the universal transverse Mercator projection coordinate system into the geographic boundary in the preset latitude and longitude coordinate system. Against the geographical background of the large-scale region, based on the geographical boundary corresponding to each standardized grid image, the disaster identification results corresponding to each standardized grid image are spatially stitched together to obtain and output the spatiotemporal distribution assessment result of the equipment failure loss. For overlapping coverage areas existing at the edges of adjacent grids, the final disaster attribute of the overlapping coverage area is determined using a probability-weighted fusion strategy according to the following formula. ,

[0162] in, An overlapping set of grids covering a specific area. For the first Confidence weights for a standardized grid image In order to target the The first standardized grid image inferred from the first The original probability of disasters.

[0163] The device provided in this application, by fusing satellite remote sensing and UAV multi-source imagery and performing multi-level nested partitioning, quality screening, and standardized fusion based on discrete grid technology, can effectively solve the problems of single data source being susceptible to interference and incomplete coverage. It can comprehensively capture various damage characteristics of the distribution network. At the same time, it uses a lightweight deep learning convolutional neural network to replace the conventional large parameter model. Under the grid-level parallel processing framework, it can significantly reduce the computational burden and inference latency, avoid memory overflow, and meet the timeliness requirements of emergency repair. Furthermore, by combining grid-level data augmentation and spatial inverse mapping stitching, it can improve the accuracy of disaster identification while achieving efficient assessment of large-scale areas. Ultimately, it balances identification efficiency and accuracy, providing reliable technical support for emergency repair of the distribution network after a typhoon.

[0164] It should be noted that other corresponding descriptions of the functional units involved in the device for assessing the spatiotemporal distribution of equipment failure losses based on lightweight deep learning and discrete grid technology provided in this application embodiment can be found in the following references. Figures 1 to 4 The corresponding descriptions in [the document] will not be repeated here.

[0165] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0166] The above embodiments and the technical features in the embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0167] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

[0168] In an exemplary embodiment, see Figure 6 The invention also provides a computer device including a bus, a processor, a memory, and a communication interface. It may also include an input / output interface and a display device, wherein the various functional units can communicate with each other via the bus. The memory stores a computer program, and the processor executes the program stored in the memory to perform the device fault loss spatiotemporal distribution assessment method based on lightweight deep learning and discrete grid technology described in the above embodiments.

[0169] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method for assessing the spatiotemporal distribution of device failure losses based on lightweight deep learning and discrete grid technology.

[0170] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented in hardware or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0171] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application.

[0172] Those skilled in the art will understand that the modules in the apparatus of the implementation scenario can be distributed within the apparatus of the implementation scenario as described, or they can be located in one or more apparatuses different from this implementation scenario, with corresponding changes. The modules of the above-described implementation scenario can be combined into one module, or they can be further divided into multiple sub-modules.

[0173] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of the implementation scenario.

[0174] The above disclosures are only a few specific implementation scenarios of this application. However, this application is not limited to these. Any variations that can be conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. A method for assessing the spatiotemporal distribution of equipment failure losses based on lightweight deep learning and discrete grid technology, characterized in that, include: Satellite remote sensing image data and UAV image data of the power distribution network in the disaster area after the typhoon passed were acquired as multi-source remote sensing image data. Based on discrete grid technology, the multi-source remote sensing image data was divided into multiple grid units according to multi-level nested grids. Data quality screening and data cleaning are performed within the multiple grid cells to obtain multiple high-quality grid cells. Then, multi-source data fusion and standardization processing are performed on the multiple high-quality grid cells to obtain multiple standardized grid images corresponding to the multiple high-quality grid cells. A lightweight deep learning convolutional neural network is used to extract and classify disaster features from the multiple standardized grid images, and the disaster identification result corresponding to each of the multiple standardized grid images is determined. Based on the geospatial coding of each standardized grid image, the disaster identification results corresponding to each standardized grid image are inversely mapped and spatially stitched to obtain and output the spatiotemporal distribution assessment results of equipment failure losses in a large-scale area.

2. The method according to claim 1, characterized in that, The satellite remote sensing image data and UAV image data of the power distribution network in the disaster area after the typhoon were acquired as multi-source remote sensing image data. Based on discrete grid technology, the multi-source remote sensing image data was divided into multiple grid cells according to a multi-level nested grid, including: The satellite remote sensing image data and the UAV image data are converted into raster data with geographic coordinate mapping using the following formulas to obtain multi-source raster data. in, For the first Data source in geographic coordinates The raster pixel value at that location, For the first Rasterization mapping function for data source class, For the first The original pixel values ​​of the data source in its own pixel coordinate system. For the first The set of rasterization transformation parameters for a data source. The value can be 1 or 2. A value of 1 corresponds to the satellite remote sensing image data, and A value of 2 corresponds to the image data of the UAV; A three-level nested grid partitioning strategy was adopted. Following the formula below, the geographic space of the power distribution network monitoring area was subdivided using a 6° universal transverse Mercator projection zone as the first-level grid, 10km×10km as the second-level grid, and 1km×1km as the third-level grid. The multi-source raster data was then mapped to the partitioned geographic space. This multi-level nested grid method was used to grid the multi-source remote sensing image data, resulting in multiple grid cells. in, A unique code for a three-level nested grid. For the RDCRMG architecture's mesh coding mapping function, For the first Geographic parameters with a first-order projection and grid. For the first Spatial parameters of a second-level 10km×10km grid, For the first The subdivision parameters of a 1km×1km third-level grid, For the first Data source in geographic coordinates The raster pixel value at that location.

3. The method according to claim 1, characterized in that, The process of filtering and cleaning data within the multiple grid cells to obtain multiple high-quality grid cells includes: Unique encoding based on a three-level nested grid Each of the grid cells is located, and the following formula is used to determine the rasterized pixel values ​​in each grid cell. Determine the first Pixel values ​​of three-level grid units , in, Mesh encoding mapping function for RDCRMG architecture The inverse mapping, These are first-level, second-level, and third-level grid indexes, respectively. By combining the rasterized pixel values ​​in each grid cell, the proportion of cloud cover within each grid cell is calculated. Percentage of effective coverage of images within the grid and the standard deviation of the grid pixel values. ; The comprehensive quality evaluation index for each grid cell is calculated using the following formula. in, For the first The comprehensive quality evaluation index corresponding to each grid cell The weight coefficients for each item and ; Determine the preset grid data quality threshold According to the following formula, combined with the grid data quality threshold Low-quality data is removed from the plurality of grid cells to obtain the plurality of high-quality grid cells. in, For the plurality of high-quality mesh cells, the first In the first high-quality grid cell Raster data of the class, This indicates that the corresponding grid cell is low-quality data and should be cleared.

4. The method according to claim 1, characterized in that, The process of fusing and standardizing the multiple high-quality grid cells to obtain multiple standardized grid images corresponding to the multiple high-quality grid cells includes: Using the preset uniform processing resolution The following formula is used to reconstruct the resolution function. The original heterogeneous data within each of the high-quality grid cells is resampled and scaled to obtain the first... Within the first high-quality grid cell Normalized pixel matrix after resolution reconstruction of the data source , in, For the plurality of high-quality mesh cells, the first In the first high-quality grid cell Raster data of the class; Using the following formula, according to the fusion weight The first Pixel-level adaptive weighted fusion of multi-source data within a high-quality grid cell yields the first... The fused raster pixel values ​​corresponding to each high-quality grid cell in, For the first The raster pixel values ​​obtained by fusing multi-source data within a high-quality grid cell. The fusion weights are adaptively determined by the signal-to-noise ratio and geometric resolution contribution of the data source; According to the following formula, for the first Radiometric fine correction is performed on the fused raster pixel values ​​corresponding to each high-quality grid cell. in, For the first The radiometrically corrected raster pixel values ​​within a high-quality grid cell. They are respectively from the first Radiation correction gain coefficient and offset coefficient calibrated by measured radiation values ​​of ground control points within a high-quality grid cell. Simultaneously, according to the following formula, a quadratic polynomial correction model is used to correct the geographic coordinates of each high-quality grid cell. in, These are the geometrically corrected geographic coordinates. The geographic coordinates before correction. to These are the quadratic polynomial geometric correction parameters based on ground control point calibration. The raster data of each of the corrected high-quality grid cells is standardized to obtain multiple standardized grid images corresponding to the multiple high-quality grid cells.

5. The method according to claim 4, characterized in that, The standardization process for the raster data of each of the multiple high-quality grid cells in the corrected set of high-quality grid cells includes: For each of the high-quality grid cells, the pixel values ​​of the raster data of the corrected high-quality grid cells are normalized using the following linear normalization formula. in, For normalized raster data, These represent the maximum and minimum pixel values ​​of the raster data within the corresponding high-quality grid cell, respectively. The following formula is used to determine the target image width required by the training model. and target image height Calculate the size scaling factor for the high-quality mesh cells. , in, The original width and height of the merged raster data in the high-quality grid cell; The scaled image width is calculated for the high-quality mesh cell using the following formula. and the height of the scaled image , Referring to the following formula, zero-fill is applied to the regions of the high-quality mesh cells that exceed the scaling size to obtain the normalized mesh image corresponding to the high-quality mesh cells. in, To standardize the pixel values ​​of a grid image. To standardize the pixel coordinates of the image, This is a floor operation.

6. The method according to claim 1, characterized in that, Before using a lightweight deep learning convolutional neural network to extract and classify disaster features from the multiple standardized grid images, and determining the disaster identification result corresponding to each of the multiple standardized grid images, the method further includes: Historical satellite remote sensing image data and historical UAV image data of the power distribution network in the disaster area after the typhoon passed were used as sample multi-source image data. Based on discrete grid technology, the sample multi-source image data is divided into grids according to multi-level nested grids, and data quality is screened and data is cleaned to obtain multiple high-quality sample grid units. Then, the multiple high-quality sample grid units are subjected to multi-source data fusion and standardization processing to obtain multiple sample standardized grid images corresponding to the multiple high-quality sample grid units. The sample categories of the multiple standardized grid images are statistically analyzed to obtain the number of sample disaster categories. If the number of sample disaster categories is lower than a preset sample number threshold, cross-grid data augmentation processing is performed on the multiple standardized grid images to obtain virtual disaster grid samples. A standardized dataset is constructed using the multiple standardized grid images and the virtual disaster grid samples, and the lightweight deep learning convolutional neural network is trained using the standardized dataset.

7. The method according to claim 6, characterized in that, The process of performing sample category statistics on the multiple standardized grid images to obtain the number of sample disaster categories, and then performing cross-grid data augmentation processing on the multiple standardized grid images to obtain virtual disaster grid samples when the number of sample disaster categories is lower than a preset sample number threshold, includes: The number of sample disaster categories for each disaster category in the multiple sample standardized grid images is counted. When the number of sample disaster categories for a target disaster category is lower than the sample number threshold, the grid cell containing the target disaster category is marked as the central disaster grid. Based on the coding mapping rules of discrete grids, and according to the row and column spatial indexes of the central disaster grid under the three-level grid division, the spatial topological adjacency of the central disaster grid, as shown in the following expression, is retrieved. Neighborhood grid set , in, Central disaster grid Row and column space indexing under the RDCRMG three-level grid partitioning This is the offset of the spatial index. For the retrieved neighborhood grid cells; From the retrieved Neighborhood grid set A neighborhood grid is randomly selected as the background grid, and the normalized sample image matrix of the background grid is extracted as the background grid image matrix. Simultaneously, the standardized sample image matrix of the central disaster grid is extracted. ; The standardized sample image matrix of the central disaster grid Apply a geometric transformation function based on the center symmetry of the mesh. The transformed disaster feature matrix is ​​obtained. ; Obtain the binary spatial mask matrix of the disaster target In this context, the pixel value of the disaster area is 1 and the pixel value of the background area is 0. The binarized spatial mask matrix... From the central disaster grid, morphological edge detection or threshold segmentation are used. Extract from; Transformed disaster feature matrix Background grid image matrix of the background grid Spatial cascading and weighted fusion are performed, and the virtual grid sample image matrix is ​​calculated according to the virtual sample generation formula shown below. The process involves performing cross-grid data augmentation on the multiple standardized grid images to obtain the virtual disaster grid samples. in, This represents the Hadamard product of a matrix.

8. The method according to claim 6, characterized in that, The process of constructing a standardized dataset using the multiple standardized grid images and the virtual disaster grid samples, and training the lightweight deep learning convolutional neural network using the standardized dataset, includes: The standardized dataset is composed of the multiple sample standardized grid images and the virtual disaster grid samples. , and initialize the untrained lightweight deep learning convolutional neural network, and load the pre-trained weight parameters of the lightweight deep learning convolutional neural network; The standardized dataset The input is fed into the untrained lightweight deep learning convolutional neural network, which, based on the untrained lightweight deep learning convolutional neural network, passes through standard convolutional layers with a stride of 2, combined with batch normalization and the h-swish activation function, according to the following formula, to normalize the input dataset. Preliminary feature extraction is performed to obtain the initial convolutional layer output feature map. , in, The instruction states that the standard convolutional layer with a stride of 2 uses a 3×3 convolutional kernel. ; Based on the 11-layer bottleneck residual module in the lightweight deep learning convolutional neural network, the initial convolutional layer outputs a feature map. Deep feature extraction is performed, wherein each bottleneck residual module adopts a lightweight structure combining depthwise separable convolution and 1×1 pointwise convolution, and embeds a compression-excitation attention mechanism. For the th Each bottleneck residual module inputs the feature map using the following formula. After dimensionality upscaling via 1×1 pointwise convolution, spatial feature extraction via 3×3 depthwise separable convolution, and weighting via a compression-excitation attention mechanism, the feature is reduced by another 1×1 pointwise convolution and finally combined with the input feature map. Perform residual connections and output feature maps. , in, It is a 3×3 depthwise separable convolution with a stride of [missing value]. The value is either 1 or 2 and is used to control feature map downsampling. It is a 1×1 pointwise convolution, and SE is a compression-excitation attention mechanism that achieves adaptive weighting of channel features through global average pooling and two fully connected layers; After the output of the final bottleneck residual module, the feature map dimension is expanded to the specified value using a 1×1 pointwise convolutional layer according to the following formula, and the h-swish activation function is used to map the feature dimensions to obtain the feature map. , in, Output feature map for the bottleneck residual module of layer 11; The feature map is processed using global average pooling according to the following formula. Spatial dimensionality compression is performed to obtain a global one-dimensional feature vector. , in, For feature map All channels in the space coordinates , The average value is obtained by doubly summing the pixel values ​​across all spatial dimensions of the 7×7 area. The global one-dimensional feature vector The input is fed into two fully connected layers in the lightweight deep learning convolutional neural network. The first fully connected layer combines batch normalization and the h-swish activation function to process the global one-dimensional feature vector. A feature nonlinear transformation is performed, and the second fully connected layer of the two fully connected layers outputs a logits vector that matches the number of disaster categories; The logits vector is transformed into a predicted probability distribution for each category using the Softmax activation function according to the following formula. in, This is the classification layer weight matrix. For bias vectors, For the global one-dimensional feature vector, The Softmax activation function is... For the output of the first The predicted probability of disasters; The cross-entropy loss function is shown in the following formula. The model parameters of the lightweight deep learning convolutional neural network during training are optimized by iteratively updating the model parameters through backpropagation, resulting in the completed lightweight deep learning convolutional neural network. in, For the number of disaster categories, For real labels coding, The lightweight deep learning convolutional neural network predicts the first [number] [number] [unit] during the training process. Probability of disasters.

9. The method according to claim 1, characterized in that, The step of inversely mapping and spatially stitching the disaster identification results corresponding to each standardized grid image based on the geospatial encoding of each standardized grid image to obtain and output the spatiotemporal distribution assessment results of equipment failure losses over a large-scale area includes: The disaster identification result corresponding to each standardized grid image is uniquely encoded with the original three-level nested grid corresponding to the standardized grid image. Binding is performed to achieve the binding of disaster perception attributes and grid spatial attributes; According to the following formula, a spatial inverse mapping based on discrete grid coding is performed on each of the standardized grid images to determine the actual spatial extent of each standardized grid image in the universal transverse Mercator projection coordinate system. in, Mesh encoding mapping function for RDCRMG architecture The inverse mapping, A unique code for a three-level nested grid. For the first Geographic parameters with a first-order projection and grid. For the first Spatial parameters of a second-level 10km×10km grid, For the first The subdivision parameters of a 1km×1km three-level grid; The actual spatial extent of each standardized grid image in the universal Mercator projection coordinate system is converted into the geographic boundary in the preset latitude and longitude coordinate system by using coordinate transformation parameters. ; Against the geographical background of the large-scale region, based on the geographical boundary corresponding to each standardized grid image, the disaster identification results corresponding to each standardized grid image are spatially stitched together to obtain and output the spatiotemporal distribution assessment result of the equipment failure loss. For overlapping coverage areas existing at the edges of adjacent grids, the final disaster attribute of the overlapping coverage area is determined using a probability-weighted fusion strategy according to the following formula. , in, An overlapping set of grids covering a specific area. For the first Confidence weights for a standardized grid image In order to target the The first standardized grid image inferred from the first The original probability of disasters.

10. A device for assessing the spatiotemporal distribution of equipment failure losses based on lightweight deep learning and discrete grid technology, characterized in that, include: The partitioning module is used to acquire satellite remote sensing image data and UAV image data of the power distribution network in the disaster area after the typhoon passes as multi-source remote sensing image data. Based on discrete grid technology, the multi-source remote sensing image data is partitioned into multiple grid units according to multi-level nested grids. The fusion module is used to perform data quality screening and data cleaning within the multiple grid cells to obtain multiple high-quality grid cells, and to perform multi-source data fusion and standardization processing on the multiple high-quality grid cells to obtain multiple standardized grid images corresponding to the multiple high-quality grid cells. The identification module is used to extract and classify disaster features from the multiple standardized grid images using a lightweight deep learning convolutional neural network, and to determine the disaster identification result corresponding to each of the multiple standardized grid images; The generation module is used to perform inverse mapping and spatial stitching on the disaster identification results corresponding to each standardized grid image based on the geospatial encoding of each standardized grid image, so as to obtain and output the spatiotemporal distribution assessment results of equipment failure loss in a large-scale area.

Images