A high-voltage substation panoramic state monitoring method based on remote sensing image technology

Abstract

This invention discloses a panoramic monitoring method for high-voltage substations based on remote sensing image technology, comprising the following steps: acquisition and interpretation of remote sensing images: acquiring multi-temporal remote sensing images of the substation and its surrounding area to identify areas suspected of actual changes; cross-scale refined inspection: collecting high-resolution visualization information containing controlled equipment or suspected changed features; multi-source data fusion and 3D reconstruction: receiving and evaluating the data quality of visualization information collected by the panoramic acquisition device, and inputting the multi-view visualization information into a preset 3D reconstruction model to generate a 3D real-world model of the monitored target and extracting complete target feature data; equipment status depth prediction: inputting the complete target feature data and synchronous data obtained from similar substation systems into a preset fault prediction model, and outputting a depth diagnosis result containing fault location and potential causes based on a dynamically adjusted similarity threshold.

Description

Technical Field

[0001] This invention relates to the field of high-voltage substation monitoring technology, and in particular to a panoramic monitoring method for high-voltage substations based on remote sensing image technology. Background Technology

[0002] A panoramic monitoring method for high-voltage substations based on remote sensing image technology integrates key technologies such as high-resolution remote sensing satellite / UAV image acquisition, intelligent image processing algorithms, multi-source data fusion analysis, and a dynamic risk early warning platform. This constructs an intelligent monitoring system that evolves from "local point-based monitoring" to "global area-based perception." Specifically, it encompasses functional modules such as automatic identification of equipment surface defects, precise location of thermal anomaly areas, dynamic assessment of vegetation intrusion risk, real-time monitoring of construction activities, and prediction of meteorological disaster impacts. This improves monitoring coverage and spatiotemporal resolution while enabling full-cycle tracking of equipment status and dynamic perception of environmental changes. Early warning of safety hazards has profound significance not only in enhancing the operational safety of high-voltage substations, reducing the labor intensity and maintenance costs of manual inspections, and improving fault response speed and handling efficiency—all direct operational benefits—but also in promoting the deep integration and innovation of remote sensing technology with the power industry, facilitating the large-scale application of artificial intelligence in the field of power equipment monitoring, and constructing a new paradigm of integrated "air-space-ground" smart grid monitoring. Ultimately, it achieves a value leap from "passive fault repair" to "proactive risk prevention and control," providing key technical support and scientific decision-making basis for smart grid construction, power equipment condition assessment, and the safety assurance of new energy grid connection.

[0003] While existing technologies utilize remote sensing imagery for its wide coverage and ability to detect macroscopic changes, their limited resolution only indicates a change in a specific location, failing to identify the exact object causing the change. Conversely, microscopic images acquired by ground robots or drones, though rich in detail, lack a macroscopic view and are typically used for blind inspections along pre-defined routes. This makes it difficult to specifically verify particular discrepancies identified by remote sensing, resulting in fragmented monitoring results and hindering the formation of a comprehensive understanding. Therefore, this paper proposes a panoramic monitoring method for high-voltage substations based on remote sensing imagery technology. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a panoramic monitoring method for high-voltage substations based on remote sensing image technology.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A panoramic monitoring method for high-voltage substations based on remote sensing image technology includes the following steps:

[0007] Remote sensing image acquisition and interpretation: Acquire multi-temporal remote sensing images of the substation and its surrounding area, extract elevation and spectral change information of ground objects based on the remote sensing images, and identify and remove false change patches caused by vegetation growth or seasonal changes through a preset elevation-spectral joint change threshold to obtain suspected real change areas.

[0008] Cross-scale refined inspection: Based on the spatial coordinates of the suspected real change area, it is calibrated as a spatial semantic anchor point in a pre-constructed unified spatial reference system, and at least one panoramic acquisition device is automatically dispatched to the precise location indicated by the spatial semantic anchor point to acquire high-resolution visualization information containing the controlled device or suspected changed ground features.

[0009] Knowledge verification and feedback optimization: The specific land cover types identified in the high-resolution visualization information collected by the panoramic acquisition device are compared and verified with the preliminary interpretation results of the remote sensing images. When the verification results are inconsistent, the verification information is used as feedback data to optimize the subsequent remote sensing image interpretation algorithm model.

[0010] Multi-source data fusion and 3D reconstruction: Receive and evaluate the data quality of the visualization information collected by the panoramic acquisition device, dynamically adjust the fusion weight of the visualization information under different perspectives based on the data quality evaluation results, and input the visualization information from multiple perspectives into a preset 3D reconstruction model to generate a 3D real-scene model of the monitored target and extract complete target feature data;

[0011] Equipment status deep prediction: The complete target feature data and synchronous data obtained from the same type of substation system are input into the preset fault prediction model. The fault prediction model is embedded with physical information constraints of the controlled equipment, which is used to correct the target feature data in combination with the operating conditions of the equipment in this substation, and outputs a deep diagnosis result containing fault location and potential causes based on a dynamically adjusted similarity threshold.

[0012] Furthermore, in the steps of acquiring and interpreting remote sensing images, extracting the elevation change information of ground features specifically involves:

[0013] By processing interferometric synthetic aperture radar images or stereo image pairs, the amount of change of ground features in the vertical direction can be obtained.

[0014] The elevation change information is combined with the spectral change information obtained through multispectral image analysis to distinguish between illegal buildings with significant vertical abrupt changes and spectral pseudo-changes with weak vertical changes.

[0015] Specifically, for the substation scenario, the elevation-spectral joint change threshold includes a dynamically adjusted substation environment correction factor. This correction factor is generated based on the statistical characteristics of the land cover type in the area surrounding the substation and is used to suppress the interference caused by abnormal surface thermal radiation due to heat dissipation from substation equipment in the spectral change information.

[0016] Furthermore, identifying and eliminating false change patches involves inputting remote sensing images from two consecutive time phases into a pre-trained Siamese neural network. The Siamese neural network includes a feature extraction layer with shared weights and directly outputs pixel-level change types. During the training process of the Siamese neural network, a negative sample library for substation scenarios is constructed. The negative sample library includes image sequences of shadow displacement of substation equipment under different lighting conditions, as well as image sequences of temporary ground disturbances around the substation caused by equipment maintenance. This enhances the network's ability to distinguish between false shadow detections and temporary disturbance false changes.

[0017] Furthermore, the construction of the unified spatial reference system specifically involves:

[0018] Real-time data collected during the execution of the task using the panoramic acquisition device;

[0019] Construct a 3D point cloud map;

[0020] The three-dimensional point cloud map is spatially registered with the orthorectified remote sensing image with high precision.

[0021] Before spatial registration, an electromagnetic interference suppression step for substation scenarios is included: acquiring electromagnetic field distribution data during substation operation, and compensating and correcting the pose estimation results of the panoramic acquisition device based on the electromagnetic field distribution data to eliminate drift errors caused by strong electromagnetic environment to sensor data.

[0022] Furthermore, the data quality assessment of the visualization information specifically involves: using a no-reference image quality assessment algorithm to calculate the ambiguity, signal-to-noise ratio, and information entropy of the visualization information, and generating dynamic weighting coefficients to guide data fusion based on the joint analysis of environmental data and indicators.

[0023] Specifically, for highly reflective metal surfaces in substation scenarios, the no-reference image quality assessment algorithm also includes a highlight region detection and evaluation submodule. This submodule analyzes the specular reflection components in the image to identify image saturation regions caused by reflections from equipment surfaces such as metal bushings and busbars, and reduces the weight of such regions in subsequent 3D reconstruction.

[0024] Furthermore, generating a 3D real-world model of the monitored target involves inputting visualization information obtained from different perspectives by ground-based and aerial acquisition devices, after data quality screening, into a pre-defined 3D reconstruction model. Using voxel rendering or splatting technology, a 3D model of the controlled equipment is reconstructed in virtual space. This 3D reconstruction model is a substation-specific reconstruction model built using 3D Gaussian sputtering. The 3D reconstruction model incorporates structural prior constraints in the objective function of the 3D Gaussian sputtering optimization. These structural prior constraints include: a geometric regularization term based on the standard CAD model of substation equipment, used to suppress free Gaussian points generated in highly reflective areas; and an occlusion inference term based on the physical location relationship of the equipment, used to compensate for the geometric structure of occluded areas in dense cable scenarios using a ray casting algorithm.

[0025] Furthermore, the embedded physical information constraints specifically include:

[0026] The physical operating equations of the equipment; when the fault prediction model analyzes the synchronous data of the same type of substation, it calls the corresponding physical equations for thermodynamic or kinematic simulation correction based on the real-time operating load and equipment model parameters of the controlled equipment in this substation.

[0027] Specifically, the thermodynamic or kinematic simulation correction includes:

[0028] The surface temperature distribution data of the controlled equipment extracted from the three-dimensional real scene model is assimilated with the theoretical temperature field obtained by solving thermodynamic equations based on the current load and environmental parameters of the equipment to generate the corrected equipment temperature state quantity.

[0029] The displacement and deformation of key components of the controlled equipment extracted from the three-dimensional real-scene model are compared with the theoretical deformation calculated by thermal expansion model or mechanical stress model based on the current operating state of the equipment to generate physical residual characteristics, so as to distinguish between normal thermal expansion and contraction and fault deformation of the equipment.

[0030] Furthermore, the dynamically adjusted similarity threshold is generated based on the health degradation curve of the controlled equipment. The health degradation curve is associated with the equipment's full lifecycle file. The health degradation curve is parameterized using a Weibull distribution-based equipment aging model. The model parameters are updated online according to the equipment model, operating history data, and the physical residual characteristics. When the similarity between the target feature data and the preset standard data is lower than the dynamic threshold of the current health state, an alarm message containing the fault location is generated, and the equipment's full lifecycle file and the corresponding physical residual characteristic historical curve are automatically retrieved for root cause analysis of the fault.

[0031] A panoramic monitoring system for high-voltage substations based on remote sensing image technology includes:

[0032] Remote sensing interpretation module: acquires and analyzes multi-temporal remote sensing images, and outputs suspected change areas and their spatial semantic anchors after being screened by joint elevation-spectral thresholding;

[0033] The collaborative scheduling module communicates with the remote sensing interpretation module. The collaborative scheduling module receives the spatial semantic anchor points and schedules at least one panoramic acquisition device to perform refined inspection tasks.

[0034] Multi-source fusion module: Receives and evaluates the data quality of the visualization information collected by the panoramic acquisition device, and generates a three-dimensional real-scene model of the monitored target and complete target feature data based on the preset three-dimensional reconstruction model;

[0035] The intelligent diagnostic module communicates with the multi-source fusion module. The intelligent diagnostic module inputs the target feature data into a fault prediction model embedded with physical information constraints and outputs a deep diagnostic result based on dynamic threshold adjustment.

[0036] The present invention has the following beneficial effects:

[0037] In this invention, the coordinates of problems discovered by remote sensing images are automatically associated with the inspection tasks of ground or aerial robots. When remote sensing images detect macroscopic anomalies, i.e., areas of suspected real change, the spatial coordinates of these areas are used to mark them as spatial semantic anchors in a pre-constructed unified spatial reference system. The system can automatically dispatch drones or robots to the precise location indicated by these spatial semantic anchors for refined data collection. This allows the monitoring results to jump from discovering changes in a certain area to confirming the specific nature of those changes, thus forming a global understanding of the internal and external environment of the substation. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the steps of a panoramic monitoring method for high-voltage substations based on remote sensing image technology proposed in this invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Please see Figure 1 As shown, this invention is a panoramic monitoring method for high-voltage substations based on remote sensing image technology, comprising the following steps:

[0041] Remote sensing image acquisition and interpretation: Acquire multi-temporal remote sensing images of the substation and its surrounding area, extract elevation and spectral change information of ground objects based on the remote sensing images, and identify and remove false change patches caused by vegetation growth or seasonal changes through a preset elevation-spectral joint change threshold to obtain suspected real change areas.

[0042] Cross-scale refined inspection: Based on the spatial coordinates of the suspected real change area, it is calibrated as a spatial semantic anchor point in a pre-constructed unified spatial reference system, and at least one panoramic acquisition device is automatically dispatched to the precise location indicated by the spatial semantic anchor point to acquire high-resolution visualization information containing the controlled device or suspected changed ground features.

[0043] Knowledge verification and feedback optimization: The specific land cover types identified in the high-resolution visualization information collected by the panoramic acquisition device are compared and verified with the preliminary interpretation results of the remote sensing images. When the verification results are inconsistent, the verification information is used as feedback data to optimize the subsequent remote sensing image interpretation algorithm model.

[0044] Multi-source data fusion and 3D reconstruction: Receive and evaluate the data quality of the visualization information collected by the panoramic acquisition device, dynamically adjust the fusion weight of the visualization information under different perspectives based on the data quality evaluation results, and input the visualization information from multiple perspectives into a preset 3D reconstruction model to generate a 3D real-scene model of the monitored target and extract complete target feature data;

[0045] Equipment status deep prediction: The complete target feature data and synchronous data obtained from the same type of substation system are input into the preset fault prediction model. The fault prediction model is embedded with physical information constraints of the controlled equipment, which is used to correct the target feature data in combination with the operating conditions of the equipment in this substation, and outputs a deep diagnosis result containing fault location and potential causes based on a dynamically adjusted similarity threshold.

[0046] In one embodiment, the extraction of elevation change information of ground features in the remote sensing image acquisition and interpretation step specifically involves:

[0047] By processing interferometric synthetic aperture radar images or stereo image pairs, the amount of change of ground features in the vertical direction can be obtained.

[0048] The elevation change information is combined with the spectral change information obtained through multispectral image analysis to distinguish between illegal buildings with significant vertical abrupt changes and spectral pseudo-changes with weak vertical changes.

[0049] Specifically, for the substation scenario, the elevation-spectral joint change threshold includes a dynamically adjusted substation environment correction factor. This correction factor is generated based on the statistical characteristics of the land cover type in the area surrounding the substation and is used to suppress the interference caused by abnormal surface thermal radiation due to heat dissipation from substation equipment in the spectral change information.

[0050] It should be noted that the specific analysis process for extracting elevation change information of ground features is as follows:

[0051] Substation thermal radiation anomaly area identification: First, based on thermal infrared band data from multi-temporal remote sensing images, the area affected by heat dissipation from substation equipment is identified. Let the thermal infrared images at time phases T1 and T2 be respectively... and Calculate the thermal radiation difference diagram between the two time phases. : ,in, These are the pixel coordinates. The preset thermal radiation change threshold is used. Extracting areas of abnormal thermal radiation : Considering that substation equipment (such as transformers and reactors) continuously generates heat during operation, the surface temperature of the surrounding area will be significantly higher than that of the natural surface. Therefore, The areas are mainly distributed within a certain buffer zone around the main equipment of the substation;

[0052] Statistical feature extraction of land cover types in the substation surrounding area: To construct a dynamic correction factor, it is necessary to obtain the distribution of land cover types in the substation surrounding area. Based on multispectral imagery of the same temporal phase, supervised classification methods (such as support vector machine or random forest) are used to classify the substation surrounding area into the following land cover types: equipment area (including concrete foundations), hardened ground, bare soil, grassland, and water bodies. The area proportion of each land cover type in the substation surrounding area is then statistically analyzed. Where $k$ represents the land cover type index, and the statistical characteristics of spectral changes for each land cover type during periods of no historical change are extracted, including the mean spectral change. and standard deviation This serves as a baseline for background changes of this land cover type under normal conditions: ,in, This represents the intensity of the spectral change of the k-th land cover on the nth historical sample (e.g., the Euclidean distance obtained from change vector analysis). This represents the number of historical samples of this type of land cover;

[0053] Generation of dynamic substation environmental correction factors: Based on the above-mentioned anomaly regions of thermal radiation and statistical characteristics of land cover, dynamic substation environmental correction factors are constructed. This correction factor is a spatially variable coefficient that affects the original spectral variation information. Above, generate corrected spectral change information. : ,in, Defined as: ,in, As an indicator function, when a cell It belongs to the region of abnormal thermal radiation. The value is 1 if the condition is met, and 0 otherwise. This function ensures that the correction factor only applies to the area affected by device heat dissipation. The thermal radiation anomaly intensity factor represents the change in thermal radiation of the current pixel relative to a preset threshold. The intensity of this factor means that the more intense the thermal radiation anomaly, the greater the correction magnitude. For pixels Type of land cover The mean spectral variation represents the level of spectral background variation of this type of land cover under normal conditions caused by natural factors (such as vegetation growth and soil moisture changes). For pixels Type of land cover The standard deviation of spectral variation represents the range of fluctuations in the spectral background variation of this type of land cover. This is a global correction factor, with a value ranging from 0.5 to 1.5, calibrated according to the size of the substation. This is the thermal radiation index factor, ranging from 0.5 to 2.0, used to control the nonlinear effect of thermal radiation intensity on the correction amplitude. At that time, areas with strong thermal radiation receive a greater correction weight. This is the confidence coefficient, ranging from 1 to 3, used to control the degree of conservatism in the correction. The larger the value, the more the correction tends to cover the upper limit of the normal variation range of this type of land cover;

[0054] Elevation-Spectral Joint Variation Threshold Reconstruction: Based on Corrected Spectral Variation Information The decision space for joint change detection is reconstructed, assuming elevation change information is... The elevation-spectral joint change threshold is then reconstructed as follows:

[0055] Criteria for determining genuine changes (e.g., newly added illegal constructions): Among them, the spectral dynamic threshold Adaptively adjusts based on the pixel's location: ,in, The threshold for basic spectral variation is determined by statistical methods. This is a dynamic adjustment term, reflecting the level of background change for different land cover types under natural conditions. To dynamically adjust the strength coefficient, the value ranges from 0.2 to 1.0. This is the confidence interval coefficient, with a value range of 1 to 2;

[0056] False change removal and output: For those that meet the requirements and The region, if it is located in a region of abnormal thermal radiation. If the spectral change is within a certain range and the thermal radiation anomaly intensity factor is large, then the change may be caused by an abnormal surface temperature due to equipment heat dissipation. The system will mark it as a pseudo-change caused by thermal radiation interference and remove it from the suspected real change area.

[0057] In one embodiment, identifying and eliminating false change patches specifically involves: inputting remote sensing images from two consecutive time phases into a pre-trained Siamese neural network. The Siamese neural network includes a feature extraction layer with shared weights and directly outputs pixel-level change types. During the training process of the Siamese neural network, a negative sample library for substation scenarios is constructed. The negative sample library includes image sequences of shadow displacement of substation equipment under different lighting conditions, as well as image sequences of temporary ground disturbances around the substation caused by equipment maintenance. This is used to enhance the network's ability to distinguish between false shadow detections and temporary disturbance false changes.

[0058] It should be noted that the specific analytical process for identifying and removing pseudo-change patches is as follows:

[0059] Construction of a negative sample library for substation scenarios: False changes are subdivided into multiple categories, and the network's discriminative ability is enhanced through targeted samples.

[0060] Acquisition and annotation of shadow displacement image sequences: High-voltage substations contain a large number of tall equipment (such as transformer bushings, surge arresters, and frames), whose projections shift significantly with changes in solar altitude and azimuth angles. This shadow displacement is easily misinterpreted as ground feature changes in spectral change detection.

[0061] Data acquisition: Select typical equipment areas of the substation and collect multiple sets of high-resolution remote sensing images at different time points (such as morning, noon, and afternoon) to form a time series image set;

[0062] Labeling strategy: Pixel-level labels are applied to shadow areas in each image, categorized as "shadow area." Simultaneously, areas where shadow positions change across different temporal images of the same device region are labeled as pseudo-shadow displacement changes. Let the 1st... The shadowed area in the temporal image is , No. The shadowed area in the temporal image is Then the shadow displacement pseudo-change region Defined as: That is, the shaded area from arrive In the area where displacement occurs, the physical features (such as the ground surface and equipment foundations) within that area do not actually change.

[0063] Acquisition and annotation of temporary ground disturbance image sequences: During substation equipment maintenance and repair work, temporary disturbances may occur on the ground around the equipment, such as the parking of construction machinery, the stacking of materials, and the erection of temporary fences. These disturbances will return to their original state after the maintenance is completed, which is a typical example of short-term pseudo-changes.

[0064] Data acquisition: During equipment maintenance, multi-temporal images are acquired before and after maintenance, and images are acquired after maintenance is completed and the site is restored as a baseline.

[0065] Labeling strategy: The temporary changes caused by maintenance activities in the images during the maintenance period are labeled at the pixel level. The labeling category is temporary disturbance pseudo-change. At the same time, the labeled area has been restored to its original state in the images after the maintenance is completed, so as to ensure that the network learns the change-recovery time pattern.

[0066] Composition of the negative sample library: The constructed negative sample library contains three types of image pairs with pixel-level annotations, as shown in the table below:

[0067] Shadow displacement samples Equipment shadow area A Equipment shadow area B Shadow displacement pseudo-change Learn the spectral characteristics of shadow displacement to avoid misjudging it as an illegal structure. Temporary perturbation samples ground before maintenance Ground disturbance during maintenance Temporary disturbance pseudo-change Learn the characteristics of temporary disturbances to avoid misjudging them as newly added features. Real change samples No illegal construction areas Newly added illegal construction areas Real Changes Learning requires focusing on the real characteristics of change. Unchanged samples Equipment area Equipment area No change Learning background stable region characteristics ;

[0069] Siamese neural network architecture design: A Transformer-based feature extraction backbone network is adopted, with the following structure:

[0070] Shared weighted feature extraction layer: Let the two input temporal images be respectively and ,in , For image size, The number of channels (e.g., RGB three-channel or multispectral channels) is used as the input for two images, which are structurally identical and share weights. Feature extraction network : ,in, For the extracted depth feature map, The downsampling factor (preferred in this invention) ), Number of feature channels (preferred) );

[0071] Differential Feature Fusion Layer: Performs differential calculations on the dual-temporal feature maps to generate differential feature maps. A learnable difference metric module is employed, which performs a non-linear mapping of feature differences through 1×1 convolution: ,in, It is a 1×1 convolutional layer used to project differential features onto a feature space suitable for classification;

[0072] Pixel-level classification decoder: converts differential feature maps The input image is fed into a decoder network, which performs upsampling and pixel-by-pixel classification to output a probability map of variation types of the same size as the input image. The decoder employs a progressive upsampling structure. ,in, , Set the number of change type categories (set) These correspond to: Category 0 - No change, Category 1 - Pseudo-change in shadow displacement, Category 2 - Pseudo-change in temporary perturbation, and Category 3 - True change, respectively. For each pixel position ( The network outputs the probability distributions for each category: The final change type is determined by taking the category with the highest probability: ;

[0073] Loss function design for substation scenarios: A weighted cross-entropy loss function is adopted, assigning higher weights to negative sample classes.

[0074] Category weight design: Based on the number of samples in each category and the importance of the scene in the negative sample library, design a category weight vector. Considering that shadow displacement and temporary disturbances are common in substation scenarios and are prone to causing false detections, the following settings are made: Preferably, (No change) (Shadow displacement) (Temporary disturbance) (Real changes);

[0075] Loss function definition: For a batch of training samples, the overall loss function is defined as follows: ,in, Total number of pixels Number of categories ( ), The weights of category k, The true label (one-hot encoded) for pixel i. Predict the probability that pixel i belongs to category k for the network;

[0076] Network training and optimization:

[0077] Training data organization: Mix the constructed negative sample library and positive sample library (real change samples) in a certain proportion to form a training set. The proportion of each type of sample in each batch should be as balanced as possible to avoid the network being biased to predict false changes due to the scarcity of real change samples.

[0078] Training strategy: The Adam optimizer is used for network training, with an initial learning rate set to... The learning rate is dynamically adjusted using a cosine annealing strategy. During training, data augmentation operations such as random cropping, rotation, and flipping are performed on the images.

[0079] Convergence criterion: When the loss value on the validation set no longer decreases for 10 consecutive epochs, or the classification accuracy reaches a preset threshold (e.g., overall accuracy > 95%, shadow displacement class recall > 90%), training is stopped and the optimal model parameters are saved.

[0080] Change detection and spurious change removal in practical applications: This involves analyzing two consecutive temporal remote sensing images to be detected. and Input the trained Siamese neural network, perform forward propagation, and obtain the change classification map. : False change removal based on classification graph:

[0081] Shadow displacement pseudo-change removal: when At that time, the pixel is marked as a pseudo-change in shadow displacement and is not included in the subsequent statistics of suspected change areas;

[0082] Temporary perturbation spurious change removal: when At that time, the pixel is marked as a temporary perturbation pseudo-change and is not included in the subsequent statistics of suspected change areas;

[0083] Extracting real changes: only when Only then was the pixel confirmed as a suspected real change.

[0084] Finally, the extracted real change patches are subjected to morphological post-processing to generate vector result files containing attributes such as change patch boundaries, spatial coordinates, and change types, which serve as input for subsequent cross-scale refined inspections.

[0085] In one embodiment, constructing the unified spatial reference system specifically involves:

[0086] Real-time data collected during the execution of the task using the panoramic acquisition device;

[0087] Construct a 3D point cloud map;

[0088] The three-dimensional point cloud map is spatially registered with the orthorectified remote sensing image with high precision.

[0089] Before spatial registration, an electromagnetic interference suppression step for substation scenarios is included: acquiring electromagnetic field distribution data during substation operation, and compensating and correcting the pose estimation results of the panoramic acquisition device based on the electromagnetic field distribution data to eliminate drift errors caused by strong electromagnetic environment to sensor data.

[0090] It should be noted that the specific analytical process for constructing a unified spatial reference system is as follows:

[0091] Constructing a 3D point cloud map: Utilizing real-time data collected by panoramic acquisition devices (such as drones / robots equipped with LiDAR or binocular cameras) during task execution, a 3D point cloud map is constructed using Simultaneous Localization and Mapping (SLAM) technology. This specifically includes:

[0092] Feature extraction and matching: for consecutive frames of images and Extract local feature points such as Scale Invariant Feature Transform (SIFT) or Speed-Up Robust Feature Transform (SURF) to establish the correspondence between feature points in adjacent frames. Let... for Feature point set, for The feature point set is used to obtain the set of matching point pairs through brute-force matching or fast nearest neighbor search algorithms. ;

[0093] Pose estimation and 3D point triangulation: based on matching point pair sets By using epipolar geometry constraints to calculate the essential matrix E between adjacent frames, the camera motion of the current frame relative to the previous frame can be solved. (Rotation matrix and translation vector) Based on the principle of motion reconstruction structure, triangulation measurement is performed on successfully matched feature point pairs to calculate their three-dimensional coordinates in the camera coordinate system. The initial global pose values ​​are provided by the inertial measurement unit and the global navigation satellite system. By minimizing the reprojection error through nonlinear optimization, the 3D point cloud in the local coordinate system is transformed to the global coordinate system. The reprojection error equation is expressed as: ,in, The coordinates of the feature points observed on the image. These are the coordinates of the three-dimensional point in space corresponding to this point. For camera projection function, For robust kernel functions;

[0094] Point cloud map generation: As the device moves, the above process is repeated continuously, gradually adding the newly calculated 3D point coordinates to the map to form a sparse or dense 3D point cloud map. Each point Associate it with its corresponding visual feature descriptor and global coordinates;

[0095] Electromagnetic interference suppression for substation scenarios:

[0096] Electromagnetic field distribution data acquisition: Electromagnetic field distribution data of the substation during operation is acquired through one or a combination of the following methods: A pre-established electromagnetic field simulation model of the substation is used; based on information such as the substation's equipment layout, voltage level, and current load, Maxwell's equations are solved using the finite element method to generate a three-dimensional electromagnetic field distribution map of the entire substation. , representing the magnetic induction intensity vector at each point in space; fixed electromagnetic sensors are deployed at key locations in the substation to collect electromagnetic field intensity data in real time, and a dynamic electromagnetic field distribution field is generated through interpolation algorithms;

[0097] Modeling the impact of electromagnetic interference on pose estimation: Assume that the raw angular velocity measurement value output by the inertial measurement unit of the panoramic acquisition device at time t is... The true angular velocity is The measurement error of angular velocity caused by electromagnetic field is This error is related to the electromagnetic field strength at the location of the equipment. And it is related to the current attitude of the device. This invention establishes the following error model: ,in, The electromagnetic sensitivity coefficient matrix of the sensor is obtained through pre-calibration. Gaussian white noise; similarly, the geomagnetic vector measured by the magnetometer... With the true geomagnetic vector The relationship is: ,in, The electromagnetic coupling coefficient matrix of the magnetometer. For measuring noise;

[0098] Pose estimation compensation and correction: Based on the above error model, the pose estimation results are compensated and corrected. Specifically, electromagnetic field distribution data is used as prior information and introduced into the cost function of the SLAM back-end optimization, forming a joint optimization problem for electromagnetic compensation. ,in, and The weighting coefficients are used to balance the reprojection error and the electromagnetic compensation term. By solving the above optimization problem, the device pose after electromagnetic interference compensation is obtained. ;

[0099] Dynamic Compensation and Online Update: Since changes in substation load lead to dynamic changes in the electromagnetic field distribution, this invention also supports an online update mechanism. When a sudden change in electromagnetic sensor readings or a sudden increase in the covariance of pose estimation is detected, the system automatically triggers the reacquisition or simulation calculation of electromagnetic field distribution data and updates the error model parameters. and This enables adaptive adjustment of compensation correction;

[0100] High-precision spatial registration with remote sensing imagery, followed by electromagnetic interference suppression and acquisition of compensated pose estimation results. Then, the constructed 3D point cloud map Map{SLAM} is spatially registered with the orthorectified remote sensing image with high precision.

[0101] Specifically, control points of ground features (such as corner points of substation buildings and feature points of fixed equipment) corresponding to those in the 3D point cloud map are extracted from the remote sensing image, and the following spatial transformation relationship is established: ,in, The coordinates are in the remote sensing image coordinate system. The coordinates are in the coordinate system of the 3D point cloud map. and These are the rotation matrix and translation vector, respectively. By solving the above equations, high-precision spatial registration between the 3D point cloud map and the remote sensing image is achieved, ultimately establishing a unified spatial reference system. In this reference system, each spatial point has unique, semantically aligned 3D coordinates, providing a precise navigation and positioning basis for subsequently identifying suspected change areas discovered by remote sensing as spatial semantic anchor points and scheduling inspection equipment.

[0102] In one embodiment, the data quality assessment of the visualization information specifically involves: using a no-reference image quality assessment algorithm to calculate the ambiguity, signal-to-noise ratio, and information entropy of the visualization information, and generating dynamic weighting coefficients to guide data fusion based on the joint analysis of environmental data and indicators.

[0103] Specifically, for highly reflective metal surfaces in substation scenarios, the no-reference image quality assessment algorithm also includes a highlight region detection and evaluation submodule. This submodule analyzes the specular reflection components in the image to identify image saturation regions caused by reflections from equipment surfaces such as metal bushings and busbars, and reduces the weight of such regions in subsequent 3D reconstruction.

[0104] In one embodiment, generating a 3D real-world model of the monitored target involves inputting visualization information obtained from different perspectives by ground-based and aerial acquisition devices, after data quality screening, into a preset 3D reconstruction model. Using voxel rendering or splatting technology, a 3D model of the controlled equipment is reconstructed in virtual space. This 3D reconstruction model is a substation-specific reconstruction model built using 3D Gaussian sputtering. The 3D reconstruction model incorporates structural prior constraints in the 3D Gaussian sputtering optimization objective function. These structural prior constraints include: a geometric regularization term based on the standard CAD model of substation equipment, used to suppress free Gaussian points generated in highly reflective areas; and an occlusion inference term based on the physical location relationship of the equipment, used to compensate for the geometric structure of occluded areas in dense cable scenarios using a ray casting algorithm.

[0105] It should be noted that the specific analysis process for the optimized 3D Gaussian sputtering reconstruction for substation scenarios is as follows:

[0106] Multi-view image acquisition and preprocessing: Image sequences of the monitored target (such as transformers, circuit breakers, etc.) are simultaneously acquired from multiple perspectives using ground-based acquisition devices (such as inspection robots) and aerial acquisition devices (such as drones). During the acquisition process, the system records the camera pose parameters of each frame. Including rotation matrix Translation vector The collected images, after undergoing the aforementioned data quality assessment and screening, form a multi-view image set. , as input data for 3D reconstruction;

[0107] Scene representation initialization based on 3D Gaussian sputtering: The substation scene is explicitly represented using 3D Gaussian sputtering technology. Each 3D Gaussian primitive in the scene is defined by the following parameters:

[0108] Central location: ;

[0109] Covariance matrix: (This indicates the shape and orientation of the Gaussian ellipsoid);

[0110] Opacity: ;

[0111] Spherical harmonic coefficients: (Used for color representation related to viewing angle);

[0112] Initially, a sparse point cloud is generated by restoring the structure through motion, with each point initialized with a Gaussian element and the covariance matrix initialized as an isotropic sphere.

[0113] Constructing a structural prior constraint term for substation scenarios: The core innovation of this invention lies in introducing a structural prior constraint term into the optimization objective function of three-dimensional Gaussian sputtering to overcome the special technical challenges of substation scenarios, and the total loss function. Image reconstruction loss and structural prior constraints constitute: ,in, This is a geometric regularization term used to suppress free Gaussian points generated in highly reflective regions. This is an occlusion inference term used to compensate for the geometry of areas obstructed by dense cables. and These are the corresponding balance weight coefficients;

[0114] Geometric regularization term Suppressing Free Gaussian Points in Highly Reflective Areas: Metal surfaces prevalent in substations (such as transformer bushings and busbars) exhibit high specular reflection characteristics, resulting in highlight areas in multi-view images. In general 3D Gaussian sputtering, these highlight areas are often incorrectly reconstructed as free Gaussian points distributed around the real surface, causing geometric distortion.

[0115] The specific steps for applying geometric regularization based on the standard CAD model of substation equipment are as follows:

[0116] Equipment CAD model import and registration: For controlled equipment of known models in the substation (such as OSFPSZ-240000 / 220 transformer), its standard CAD model is imported in advance, and the CAD model is initially aligned with the initialized Gaussian point cloud through a coarse registration algorithm (such as the iterative nearest point algorithm based on feature points).

[0117] Distance constraint regularization: Geometric regularization term Force the position of each Gaussian element It should be close to the nearest neighbor CAD model surface, while allowing a certain offset to accommodate minor actual deformations of the equipment surface (such as changes in coating thickness or slight deformation). This constraint is expressed as: Where K is the total number of Gaussian elements. For points on the surface of the CAD model, The shortest distance from the center of the Gaussian element to the surface of the CAD model;

[0118] Highlight region adaptive weights: weight coefficients Dynamically adjust based on the image reflectivity characteristics of the region corresponding to each Gaussian pixel. Calculate the specular confidence of the projection region corresponding to each Gaussian pixel by analyzing the specular reflection component of each pixel in the input image (e.g., using a specular separation algorithm). Highlight area ( Gaussian elements close to 1 are subject to stronger geometrical regularity constraints, thus suppressing their drift away from the real surface, while diffuse reflection regions ( The constraints (close to 0) are relatively weak, allowing the preservation of actual, small geometric features: ,in, This is a preset hyperparameter for regularization strength;

[0119] Obscuring reasoning terms : Compensation for the geometry of densely cabled areas: Substations contain a large number of densely cabled structures (such as busbars, down conductors, and jumpers). These slender structures are easily obscured by the equipment itself or other cables from a single viewpoint, resulting in geometric defects or breaks in the 3D reconstruction.

[0120] The occlusion inference term based on the physical location relationship of devices is introduced, and the specific steps are as follows:

[0121] Physical location relationship modeling: Based on the electrical wiring diagram and equipment layout diagram of the substation, construct a physical location relationship diagram between the controlled equipment. , among which, nodes The supporting structure representing the equipment or cable, edge Represents the connection relationship of cables, with each edge associated with its spatial geometric parameters (such as the catenary equation parameters of the cable).

[0122] Ray projection occlusion detection: for each Gaussian element Starting from its position, a ray is emitted towards the optical center of each camera. ,in Given the optical center coordinates of the i-th camera, use a ray casting algorithm to detect whether the ray is occluded by other known structures (including reconstructed primitives or structures predicted by other physical models);

[0123] Occlusion Compensation Regularization: Occlusion Inference Term To penalize Gaussian cells that are continuously occluded from multiple viewpoints and produce erroneous geometry in informationless regions, this regularization term, based on structural priors provided by the physical location graph, guides Gaussian cells to align in an ordered manner along the expected cable path in the occluded region. ,in, Let Gaussian elements be the indicator functions, and when viewed from the i-th perspective... The value is 1 when the object is occluded, and 0 otherwise. For physical location relationship diagram The location of Gaussian elements is predicted by the associated cable geometry model (such as the catenary equation);

[0124] Specifically, for Gaussian elements belonging to cable-like structures, their predicted positions... Determined by the equation of the catenary between two adjacent support nodes: Where 'a' is a parameter related to cable tension and weight per unit length. The coordinates are the lowest point of the catenary. Using this equation, even in areas where the cable is visually completely obscured, the system can infer the continuous direction of the cable based on its physical location, guiding Gaussian elements to align along the correct trajectory and filling in geometric gaps caused by obstruction.

[0125] Joint optimization and model update: Image reconstruction loss With structural prior constraints and Combined, they form the total loss function. An adaptive moment estimation optimizer is used for iterative optimization, and the parameters of the Gaussian elements are updated in each iteration. ;

[0126] During the optimization process, geometric regularization term Continuously suppress the free Gaussian points generated in the highlight areas, causing them to gradually converge towards the surface of the CAD model, thus obscuring the inference terms. In the obstructed area, the Gaussian elements are guided to align along the cable path predicted by the physical model to ensure the geometric integrity of the slender structure;

[0127] 3D Real-Scene Model Generation and Target Feature Extraction: After optimization and convergence, the final 3D Gaussian sputtering scene representation is obtained. Using Gaussian sputtering rendering technology, high-fidelity images of the controlled device can be generated in real-time from any virtual perspective. Based on this, complete target feature data is extracted from the reconstructed 3D model, including but not limited to:

[0128] Geometric features: three-dimensional dimensions, spatial location, and surface contour point cloud of key components of the equipment;

[0129] Deformation characteristics: By comparing with the CAD model, the displacement and deformation of key parts are calculated;

[0130] Temperature characteristics: Map temperature information onto the surface of a 3D model to generate a temperature distribution field on the surface of the device.

[0131] In one embodiment, the embedded physical information constraints specifically include:

[0132] The physical operating equations of the equipment; when the fault prediction model analyzes the synchronous data of the same type of substation, it calls the corresponding physical equations for thermodynamic or kinematic simulation correction based on the real-time operating load and equipment model parameters of the controlled equipment in this substation.

[0133] Specifically, the thermodynamic or kinematic simulation correction includes:

[0134] The surface temperature distribution data of the controlled equipment extracted from the three-dimensional real scene model is assimilated with the theoretical temperature field obtained by solving thermodynamic equations based on the current load and environmental parameters of the equipment to generate the corrected equipment temperature state quantity.

[0135] The displacement and deformation of key components of the controlled equipment extracted from the three-dimensional real-scene model are compared with the theoretical deformation calculated by thermal expansion model or mechanical stress model based on the current operating state of the equipment to generate physical residual characteristics, so as to distinguish between normal thermal expansion and contraction and fault deformation of the equipment.

[0136] It should be noted that the specific analysis process for the correction is as follows:

[0137] Visual Feature Extraction Based on 3D Reality Model: Two types of key visual features are extracted from the 3D reality model of the monitored target generated by multi-source data fusion and 3D reconstruction.

[0138] Surface temperature distribution data: Using the color information associated with each Gaussian point or voxel in the 3D real-world model and the fusion results of infrared thermal imaging data, the surface temperature field distribution of key components such as transformer tanks, bushings, and circuit breaker moving contact shells is extracted. Let the extracted temperature observation value be... ,in, Represents position coordinates in three-dimensional space. Represents a timestamp;

[0139] Displacement and Deformation of Key Components: By comparing 3D reality models of the same equipment at different time stamps, an iterative nearest-point algorithm is used for point cloud registration to calculate the displacement vectors of key components (such as transformer winding leads and circuit breaker insulation rods) in 3D space. With deformation ;

[0140] Theoretical baseline value generation based on physical operating equations: For different types of controlled equipment, the corresponding physical operating equations are retrieved from the physical information database. Combined with the real-time operating load and model parameters of the equipment in this substation, the theoretical baseline value under the current operating conditions is calculated. Specifically:

[0141] Thermodynamic simulation correction of power transformers: The internal heating and cooling processes of power transformers can be described by a thermal circuit model based on thermoelectric analogy. This involves establishing the temperature rise of the transformer's top oil. Differential equation with respect to the load factor K(t): ,in, The oil time constant (determined by the equipment model parameters). This refers to the steady-state oil temperature rise under rated load. The ratio of load loss to no-load loss under rated load, where n is an index reflecting the cooling method (e.g., under ONAN cooling method). ), For real-time load factor, To obtain the real-time current from the SCADA system, the current ambient temperature is included. The oil temperature rise obtained from the solution Adding these together, we obtain the theoretical reference value for the top layer oil of the transformer: Furthermore, based on the hot spot temperature calculation in the transformer thermal circuit model, the theoretical temperature field distribution at the hot spot location of the transformer winding can be solved. As a result of visual observations The theoretical benchmark for comparison;

[0142] Kinematic simulation correction of circuit breakers: For circuit breakers, the opening and closing process of their operating mechanism can be described by a second-order mechanical vibration equation. Establish the displacement of the moving contact. With operating power Differential equations between: ,in, The quality of the moving contact (determined by the equipment model parameters). The damping coefficient is (fitted from historical data). The spring stiffness is determined by the equipment model parameters. The theoretical displacement-time curve of the moving contact displacement is obtained by solving the equation, which represents either electromagnetic force or spring force (calculated based on real-time control signals) during the normal opening and closing process of the equipment. To address the thermal expansion and contraction effect of equipment during operation, a temperature-based thermal expansion model is introduced: ,in, For components at reference temperature The original length below, The coefficient of linear expansion of the material (determined by the equipment model). The surface temperature of the component is extracted from the 3D reality model, from which the theoretical deformation of the component due to normal thermal expansion and contraction is obtained: ;

[0143] Data assimilation and physical residual feature generation: Visual observation data is fused with theoretical values ​​from physical models to generate core feature quantities for fault identification. The specific steps are as follows:

[0144] Temperature field data assimilation: A Kalman filter framework is used to assimilate the theoretical and observed temperature fields. The system state variables are assumed to be the actual temperature distribution of the equipment. Its state transition equation is obtained by discretizing the thermodynamic differential equation: ,in Here is the state transition matrix. The heat source input matrix, For the heat generated by the equipment, For process noise, the observation equation is: ,in For the observation matrix, To observe the noise, the fused corrected temperature state quantity is obtained through the prediction and update steps of Kalman filtering. : ,in The Kalman gain matrix is ​​given, and the temperature physical residual is calculated simultaneously. This residual characterizes the deviation between the actual temperature and the physical model's prediction. When a local fault occurs inside the equipment (such as a winding short circuit or oil circuit blockage), this residual will exhibit a systematic shift.

[0145] Deformation data assimilation and physical residual generation: For deformation variables, the actual deformation obtained from visual observation is used to generate physical residuals. Theoretical deformation calculated by thermal expansion model For comparison, considering that the equipment is also subjected to electrodynamic forces and mechanical stresses, the total theoretical deformation is defined as follows: ,in For the deformation components solved by the mechanical stress model, generate the deformation physical residuals: ,when When the fluctuations are small and random, it indicates that the observed deformation can be explained by normal thermal expansion and contraction and mechanical stress, and is judged as normal fluctuation. When the deformation continues to increase and exceeds the preset threshold, it indicates that there is abnormal deformation that is beyond the expectations of the physical model (such as insulation paper shrinkage or permanent deformation caused by metal fatigue), which is judged as a precursor to faulty deformation.

[0146] Dynamic threshold setting and fault diagnosis: The dynamically adjusted similarity threshold is dynamically generated based on the statistical properties of the physical residual characteristics. Let the physical residual sequence... If it follows a normal distribution, then the dynamic threshold... Defined as: ,in, and These are the mean and standard deviation of the recent physical residuals, respectively. This is the confidence coefficient (e.g., 3 corresponds to a 99.7% confidence interval).

[0147] In one embodiment, the dynamically adjusted similarity threshold is generated based on the health degradation curve of the controlled device, which is associated with the device's full lifecycle profile. The health degradation curve is parameterized using a device aging model based on a Weibull distribution. The model parameters are updated online according to the device model, historical operating data, and the physical residual characteristics. When the similarity between the target feature data and the preset standard data is lower than the dynamic threshold of the current health state, an alarm message containing the fault location is generated, and the device's full lifecycle profile and the corresponding historical physical residual characteristic curve are automatically retrieved for root cause analysis of the fault.

[0148] It should be noted that the specific analysis process of dynamic similarity threshold and root cause analysis is as follows:

[0149] Health decline curve modeling based on Weibull distribution: Establishing a health index for each type of controlled equipment (such as transformers and circuit breakers) in the substation. The mathematical model of equipment health degradation over time t uses the Weibull distribution to parameterize the process, as the Weibull distribution can flexibly describe the failure characteristics of different stages, such as early failure, random failure, and wear-out failure. The cumulative distribution function F(t) and reliability function R(t) of the Weibull distribution are expressed as follows: Where t is the time the device has been running. Shape parameters reflect equipment failure modes ( <1 indicates early failure. =1 indicates an accidental failure. >1 represents consumption loss efficiency). As a scale parameter reflecting the characteristic lifespan of the equipment, a health degradation curve is defined based on the aforementioned reliability function. for: ,in, Set the initial health index of the device (usually set to 1). The value range is [0, 1], and its value decreases as time increases, representing the decline trend of the equipment's health status. This health decline curve is associated with the equipment's full life cycle file, including information such as equipment model, manufacturing date, rated parameters, and historical maintenance records, providing a basis for subsequent dynamic threshold calculation.

[0150] Online updating of model parameters based on physical residual characteristics: Traditional Weibull model parameters β and η are usually fitted based on statistical lifetime data of similar equipment, but cannot reflect the actual operating conditions differences of individual equipment. This invention introduces physical residual characteristics. The model parameters are updated online so that the health degradation curve can adaptively reflect the current true state of the equipment. The theoretical baseline value is calculated based on the thermodynamic or kinematic equations of the equipment, and the residual generated after comparing it with the actual observed value is represented as the physical residual characteristic. For example, for a transformer, its physical residual can be defined as: ,in, This refers to the actual observed top-layer oil temperature. The theoretical oil temperature, calculated based on thermodynamic equations under current load and environmental conditions, is obtained over a specific time period. , Within this model, a health correction factor λ(t) is calculated by monitoring the cumulative characteristics of the physical residuals. This factor is used to adjust the parameters of the Weibull model. Specifically, the cumulative residuals are defined... : ,in, The time-weighted function assigns higher weight to recent data, using the accumulated residuals within a sliding window. Compared with historical average residuals Comparison, dynamic adjustment of shape parameters : ,in, The initial shape parameters are based on the equipment model, and k is an adjustment coefficient. As the physical residual continues to increase (i.e., the actual state deviates more severely from the theoretical model),... > , A corresponding increase indicates an accelerated aging trend in the equipment; conversely, if the equipment is operating well, By maintaining or reducing the health status of an individual device through the online update mechanism described above, the health decline curve HI(t) can reflect the actual health status of the device in real time, rather than relying on the statistical average of similar devices.

[0151] Generation of dynamic similarity threshold: Based on the updated health decline curve HI(t), a dynamic similarity threshold δ(t) is generated for the current time t. This threshold is used to determine whether the similarity between the target feature data and the preset standard data is within the normal range. The formula for calculating the dynamic threshold δ(t) is as follows: ,in, The baseline similarity threshold is set at the time of equipment leaving the factory, usually 0.9 or preset according to the equipment type. HI(t) is the health index at the current moment, whose value decreases as the equipment ages, causing the threshold to decrease accordingly, allowing parameter drift of older equipment within a certain range. The operating noise tolerance characterizes the range of similarity changes caused by fluctuations in normal operating conditions. Based on recent physical residuals Dynamic calculation of standard deviation: ,in, The standard deviation of the physical residuals over the past N sampling points. With a preset coefficient (such as 2 or 3), this design allows for a moderate relaxation of the threshold when operating conditions fluctuate significantly, avoiding false alarms caused by normal fluctuations. The target feature data X(t) and the preset standard data... The similarity S(t) between them is calculated using cosine similarity or Mahalanobis distance: Alternatively, similarity normalized by weighted Euclidean distance can be used: ;

[0152] Fault diagnosis and root cause analysis: When the calculated similarity Below the current dynamic threshold At that time, that is < The system determines that the equipment is in an abnormal state and generates an alarm message containing the location of the fault. At the same time, the system automatically triggers the root cause analysis process, which includes:

[0153] Retrieve full lifecycle records: Obtain the equipment's model parameters, years of operation, historical maintenance records, past fault information, etc., to form the context of the equipment health record;

[0154] Historical curves of physical residual characteristics: Extract the historical change curves of physical residual ε(τ) on the time axis and analyze the trend characteristics of the residuals (such as sudden increases, gradual changes, periodic fluctuations, etc.).

[0155] Comprehensive analysis outputs diagnostic results: Based on the type of similarity anomaly (such as temperature feature decrease, deformation feature anomaly, etc.), the change pattern of residual curve, and historical events in the archive, an in-depth diagnostic report is generated.

[0156] A panoramic monitoring system for high-voltage substations based on remote sensing image technology includes:

[0157] Remote sensing interpretation module: acquires and analyzes multi-temporal remote sensing images, and outputs suspected change areas and their spatial semantic anchors after being screened by joint elevation-spectral thresholding;

[0158] The collaborative scheduling module communicates with the remote sensing interpretation module. The collaborative scheduling module receives the spatial semantic anchor points and schedules at least one panoramic acquisition device to perform refined inspection tasks.

[0159] Multi-source fusion module: Receives and evaluates the data quality of the visualization information collected by the panoramic acquisition device, and generates a three-dimensional real-scene model of the monitored target and complete target feature data based on the preset three-dimensional reconstruction model;

[0160] The intelligent diagnostic module communicates with the multi-source fusion module. The intelligent diagnostic module inputs the target feature data into a fault prediction model embedded with physical information constraints and outputs a deep diagnostic result based on dynamic threshold adjustment.

[0161] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A panoramic monitoring method for high-voltage substations based on remote sensing image technology, characterized in that, Includes the following steps: Remote sensing image acquisition and interpretation: Acquire multi-temporal remote sensing images of the substation and its surrounding area, extract elevation and spectral change information of ground objects based on the remote sensing images, and identify and remove false change patches caused by vegetation growth or seasonal changes through a preset elevation-spectral joint change threshold to obtain suspected real change areas. Cross-scale refined inspection: Based on the spatial coordinates of the suspected real change area, it is calibrated as a spatial semantic anchor point in a pre-constructed unified spatial reference system, and the panoramic acquisition device is automatically dispatched to the precise location indicated by the spatial semantic anchor point to collect high-resolution visualization information containing the controlled equipment or suspected changed features. Knowledge verification and feedback optimization: The specific land cover types identified in the high-resolution visualization information collected by the panoramic acquisition device are compared and verified with the preliminary interpretation results of the remote sensing images. When the verification results are inconsistent, the verification information is used as feedback data to optimize the subsequent remote sensing image interpretation algorithm model. Multi-source data fusion and 3D reconstruction: Receive and evaluate the data quality of the visualization information collected by the panoramic acquisition device, dynamically adjust the fusion weight of the visualization information under different perspectives based on the data quality evaluation results, and input the visualization information from multiple perspectives into a preset 3D reconstruction model to generate a 3D real-scene model of the monitored target and extract complete target feature data; Equipment status deep prediction: The complete target feature data and synchronous data obtained from the same type of substation system are input into the preset fault prediction model. The fault prediction model is embedded with physical information constraints of the controlled equipment, which is used to correct the target feature data in combination with the operating conditions of the equipment in this substation, and outputs a deep diagnosis result containing fault location and potential causes based on a dynamically adjusted similarity threshold.

2. The panoramic monitoring method for high-voltage substations based on remote sensing image technology according to claim 1, characterized in that, In the steps of acquiring and interpreting remote sensing images, extracting the elevation change information of ground features specifically involves: By processing interferometric synthetic aperture radar images or stereo image pairs, the amount of change of ground features in the vertical direction can be obtained. The elevation change information is combined with the spectral change information obtained through multispectral image analysis to distinguish between illegal buildings with significant vertical abrupt changes and spectral pseudo-changes with weak vertical changes. Specifically, for the substation scenario, the elevation-spectral joint change threshold includes a dynamically adjusted substation environment correction factor. This correction factor is generated based on the statistical characteristics of the land cover type in the area surrounding the substation and is used to suppress the interference caused by abnormal surface thermal radiation due to heat dissipation from substation equipment in the spectral change information.

3. The panoramic monitoring method for high-voltage substations based on remote sensing image technology according to claim 1, characterized in that, The specific steps for identifying and removing false change patches are as follows: remote sensing images from two consecutive time phases are input into a pre-trained Siamese neural network. The Siamese neural network includes a feature extraction layer with shared weights and directly outputs the change type at the pixel level. During the training process of the Siamese neural network, a negative sample library for substation scenarios is constructed. The negative sample library includes image sequences of shadow displacement of substation equipment under different lighting conditions, as well as image sequences of temporary ground disturbances around the substation caused by equipment maintenance. This is used to enhance the network's ability to distinguish between false shadow detections and false changes caused by temporary disturbances.

4. The panoramic monitoring method for high-voltage substations based on remote sensing image technology according to claim 1, characterized in that, The construction of the unified spatial reference system specifically involves: Real-time data collected during the execution of the task using the panoramic acquisition device; Construct a 3D point cloud map; The three-dimensional point cloud map is spatially registered with the orthorectified remote sensing image with high precision.

5. The panoramic monitoring method for high-voltage substations based on remote sensing image technology according to claim 1, characterized in that, The specific steps for evaluating the data quality of the visualized information are as follows: using a no-reference image quality assessment algorithm, the ambiguity, signal-to-noise ratio, and information entropy of the visualized information are calculated, and dynamic weighting coefficients are generated based on the joint analysis of environmental data and indicators to guide data fusion.

6. The panoramic monitoring method for high-voltage substations based on remote sensing image technology according to claim 5, characterized in that, The specific process of generating a 3D real-scene model of the monitored target is as follows: Visualized information obtained from different perspectives by ground acquisition equipment and aerial acquisition equipment, after data quality screening, is input into a preset 3D reconstruction model to reconstruct a 3D model of the controlled equipment in virtual space. The 3D reconstruction model is a substation-specific reconstruction model constructed based on 3D Gaussian sputtering. The 3D reconstruction model introduces a structural prior constraint term in the optimization objective function of 3D Gaussian sputtering. The structural prior constraint term includes: a geometric regularization term based on the standard CAD model of substation equipment, used to suppress free Gaussian points generated in highly reflective areas, and an occlusion inference term based on the physical position relationship of the equipment, used to compensate for the geometric structure of the occluded area in dense cable scenarios through ray casting algorithm.

7. The panoramic monitoring method for high-voltage substations based on remote sensing image technology according to claim 1, characterized in that, The embedded physical information constraints specifically include: The physical operating equations of the equipment; when the fault prediction model analyzes the synchronous data of the same type of substation, it calls the corresponding physical equations for thermodynamic or kinematic simulation correction based on the real-time operating load and equipment model parameters of the controlled equipment in this substation. Specifically, the thermodynamic or kinematic simulation correction includes: The surface temperature distribution data of the controlled equipment extracted from the three-dimensional real scene model is assimilated with the theoretical temperature field obtained by solving thermodynamic equations based on the current load and environmental parameters of the equipment to generate the corrected equipment temperature state quantity. The displacement and deformation of key components of the controlled equipment extracted from the three-dimensional real-scene model are compared with the theoretical deformation calculated by thermal expansion model or mechanical stress model based on the current operating state of the equipment to generate physical residual characteristics, so as to distinguish between normal thermal expansion and contraction and fault deformation of the equipment.

8. The panoramic monitoring method for high-voltage substations based on remote sensing image technology according to claim 1, characterized in that, The dynamically adjusted similarity threshold is generated based on the health degradation curve of the controlled equipment. The health degradation curve is associated with the equipment's full life cycle file. The health degradation curve is parameterized using a Weibull distribution-based equipment aging model. The model parameters are updated online according to the equipment model, operating history data, and the physical residual characteristics. When the similarity between the target feature data and the preset standard data is lower than the dynamic threshold of the current health state, an alarm message containing the fault location is generated, and the equipment's full life cycle file and the corresponding physical residual characteristic historical curve are automatically retrieved for root cause analysis of the fault.

9. A panoramic monitoring system for high-voltage substations based on remote sensing image technology, characterized in that, The system is used to perform the method as described in claim 1, including: Remote sensing interpretation module: acquires and analyzes multi-temporal remote sensing images, and outputs suspected change areas and their spatial semantic anchors after being screened by joint elevation-spectral thresholding; Collaborative scheduling module: Receives the spatial semantic anchor points and schedules the panoramic acquisition equipment to perform refined inspection tasks; Multi-source fusion module: Receives and evaluates the data quality of the visualization information collected by the panoramic acquisition device, and generates a three-dimensional real-scene model of the monitored target and complete target feature data based on the preset three-dimensional reconstruction model; Intelligent diagnostic module: Inputs the target feature data into a fault prediction model embedded with physical information constraints, and outputs a deep diagnostic result based on dynamic threshold adjustment.

Images