Method, system and equipment for evaluating state of insulator and medium

Abstract

The invention relates to the technical field of power system inspection, and discloses an insulator state evaluation method, system and device and a medium, and the method comprises the steps: collecting the multi-dimensional observation data of a target insulator in real time to construct a real-time observation data set; performing point cloud registration on the real-time observation data set by adopting a semantic segmentation model, generating a target grid model based on a point cloud registration result, and optimizing the target grid model through a reinforcement learning algorithm to obtain an insulator three-dimensional model; obtaining an electric field abnormal point of the target insulator, combining the electric field abnormal point with the insulator three-dimensional model, performing electric field-defect space correlation analysis through an absolute coordinate mapping method based on ORC positioning and space geometric calibration, obtaining a correlation analysis result, performing multi-modal fusion analysis processing on the correlation analysis result, and obtaining a state evaluation result of the target insulator; while the high-efficiency and real-time requirements of electric power system inspection are met, comprehensive and accurate evaluation of the state of the insulator is also realized.

Description

Technical Field

[0001] This invention relates to the field of power system inspection technology, and in particular to a method, system, device and medium for assessing the condition of insulators. Background Technology

[0002] The inspection of insulators in power systems is a key link in ensuring the safe operation of the power grid. With the development of drone technology, drone inspection has become a new and efficient inspection method.

[0003] However, most existing drone inspection methods can only provide two-dimensional images, which cannot comprehensively and accurately assess the condition of insulators. Furthermore, they suffer from low detection accuracy and high false positive rate when facing complex environments and subtle defects. At the same time, they have high computational complexity and insufficient real-time performance when processing large amounts of data, failing to meet the high efficiency and real-time requirements of power system inspection.

[0004] Therefore, it is evident that the current technology for drone inspection of insulators suffers from problems such as limited data types, difficulty in comprehensively assessing the insulator's condition, high computational complexity when processing large amounts of data, insufficient real-time performance, and inability to meet the high efficiency and real-time requirements of power system inspection. This has become a pressing technical problem for those skilled in the art. Summary of the Invention

[0005] This invention provides a method, system, device, and medium for assessing the condition of insulators, solving the problems of existing technologies where unmanned aerial vehicle (UAV) inspections of insulators use only a single type of data, making it difficult to comprehensively assess the condition of insulators, and the computational complexity is high when processing large amounts of data, resulting in insufficient real-time performance and failing to meet the high efficiency and real-time requirements of power system inspections.

[0006] To address the aforementioned technical problems, the first aspect of this invention provides a method for assessing the condition of insulators, comprising: Real-time acquisition of multi-dimensional observation data of the target insulator, and construction of a real-time observation dataset based on the multi-dimensional observation data; A semantic segmentation model is used to perform point cloud registration on the real-time observation dataset, and a target mesh model is generated based on the point cloud registration results. The target mesh model is then optimized using a reinforcement learning algorithm to obtain a three-dimensional model of the insulator. The electric field anomaly points of the target insulator are obtained and combined with the three-dimensional model of the insulator. The electric field-defect spatial correlation analysis is performed by the absolute coordinate mapping method based on ORC positioning and spatial geometric calibration to obtain the correlation analysis results. Based on the correlation analysis results, multimodal fusion analysis is performed to obtain the state assessment results of the target insulator.

[0007] As one preferred embodiment, the real-time acquisition of multi-dimensional observation data of the target insulator, and the construction of a real-time observation dataset based on the multi-dimensional observation data, includes: The image data, point cloud data, electric field data, and position data of the target insulator are collected in real time to serve as the multidimensional observation data; The multidimensional observation data is cleaned and feature extracted to obtain HOG and color features, curvature and normal vector features, and electric field features. The HOG and color features, the curvature and normal vector features, and the electric field features are weighted and fused to obtain the real-time observation dataset.

[0008] As one preferred embodiment, the point cloud registration of the real-time observation dataset using a semantic segmentation model includes: Based on the location data, the corresponding point cloud data in the real-time observation dataset is transformed to the same world coordinate system to obtain transformed point cloud data; The converted point cloud data is downsampled using a voxel grid, and several key sampling points are extracted from the downsampling results using the FPS algorithm. A spherical neighborhood is constructed with each of the aforementioned key sampling points as the center, and local features of all key sampling points within each spherical neighborhood are extracted to obtain multiple sets of local features; A multilayer perceptron is used to encode the local feature set corresponding to each spherical neighborhood to obtain a multidimensional local feature descriptor. Calculate the cosine similarity between the multidimensional local feature descriptors, and select key sampling points with similarity results greater than a preset similarity threshold as valid corresponding point pairs; The effective corresponding point pairs are iteratively optimized using the ICP algorithm to obtain the target registration point cloud and the source registration point cloud as the point cloud registration result.

[0009] As one preferred embodiment, the generation of the target mesh model based on the point cloud registration results includes: Principal component analysis is used to calculate the normal vectors of the target registration point cloud and the source registration point cloud respectively for time-series consistency calibration, resulting in the target calibration normal vector and the source calibration normal vector. The historical three-dimensional mesh model corresponding to the target insulator is retrieved, and the micro-texture parameters and macro-topological features of the historical three-dimensional mesh model are extracted to construct a historical feature library; Based on the historical feature library and historical 3D mesh model of the source registration point cloud and its corresponding target registration point cloud, paired training data is constructed to train the generative adversarial network and obtain the trained generative adversarial network. The target calibration normal vector and the source calibration normal vector are processed by the Poisson surface reconstruction algorithm to obtain the initial mesh model; The initial mesh model is input into the trained generative adversarial network for texture optimization to obtain the target mesh model.

[0010] As one preferred embodiment, the step of optimizing the target mesh model using a reinforcement learning algorithm to obtain a three-dimensional insulator model includes: Defect detection is performed on the target mesh model to obtain defect regions; Using the Laplace smoothing factor as the agent, a target reward function is constructed based on the surface roughness of the target mesh model and the retention rate of the defect region; The Laplace smoothing factor is initialized so that the agent optimizes the target mesh model of the non-defect region based on the initialized Laplace smoothing factor, and calculates the current reward in real time based on the target reward function; The agent adaptively optimizes the initialized Laplace smoothing factor based on the current reward, so that the agent optimizes the target mesh model of the defect region based on the adaptively optimized Laplace smoothing factor. The optimization steps for the target mesh model in non-defective regions and the target mesh model in defective regions are executed iteratively until the preset termination condition is met. The final optimized target mesh models are then combined to obtain the three-dimensional model of the insulator.

[0011] As one preferred embodiment, the electric field anomaly points of the target insulator are obtained and combined with the three-dimensional model of the insulator. An electric field-defect spatial correlation analysis is then performed using an absolute coordinate mapping method based on ORC positioning and spatial geometric calibration to obtain the correlation analysis results, including: Based on the relationship between the electric field data and the preset electric field data, the electric field anomaly point of the target insulator is determined, and the position data corresponding to the electric field anomaly point is converted to the world coordinate system to be mapped to the three-dimensional model of the insulator. The K-nearest neighbor algorithm is used to find several vertices within a preset distance of the electric field anomaly point in the mapped 3D model of the insulator, and the average coordinate of each vertex is used as the mapping point. The image data and the point cloud data are associated using the SIFT feature matching method to obtain associated coordinates. Based on the associated coordinates and the three-dimensional model of the insulator, the three-dimensional spatial coordinates of the defect area are extracted using the Canny edge detection method to obtain the defect area coordinate set. Calculate the spatial distance between each mapping point and the coordinate set of the defect region, and analyze the spatial correlation between each electric field anomaly point and the coordinate set of the defect region based on the spatial distance calculation results to obtain the correlation analysis results.

[0012] As one preferred embodiment, the multimodal fusion analysis based on the correlation analysis results to obtain the state assessment result of the target insulator includes: The coordinates of electric field anomaly points and defect regions that are associated with the analysis results are extracted from the correlation analysis results, and used as target anomaly points and target defect regions. The average deviation rate of electric field anomalies, defect area, and longest defect length are determined based on the target anomaly points and the target defect regions to quantify the associated anomaly scores. The random forest algorithm is used to perform multimodal fusion analysis on the target anomaly points, the target defect regions, and the associated anomaly scores to obtain a comprehensive score, and the state assessment result of the target insulator is determined based on the comprehensive score.

[0013] A second aspect of the present invention provides a system for assessing the condition of an insulator, comprising: The dataset construction module is used to collect multidimensional observation data of the target insulator in real time and construct a real-time observation dataset based on the multidimensional observation data; The model building module is used to perform point cloud registration on the real-time observation dataset using a semantic segmentation model, generate a target mesh model based on the point cloud registration results, and optimize the target mesh model through a reinforcement learning algorithm to obtain a three-dimensional model of the insulator. The correlation analysis module is used to obtain the electric field anomaly points of the target insulator, and combine them with the three-dimensional model of the insulator. The electric field-defect spatial correlation analysis is performed by the absolute coordinate mapping method based on ORC positioning and spatial geometric calibration to obtain the correlation analysis results. The state assessment module is used to perform multimodal fusion analysis based on the correlation analysis results to obtain the state assessment results of the target insulator.

[0014] A third aspect of the present invention provides an electronic device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the insulator state assessment method as described above.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium comprising a stored computer program, wherein when the device containing the computer-readable storage medium executes the computer program, it implements the insulator condition assessment method as described above.

[0016] Compared with the prior art, the beneficial effects of the embodiments of the present invention are as follows: The accuracy and reliability of defect detection were improved by collecting and processing multidimensional data. Through point cloud registration and reinforcement learning optimization, the generated 3D model of the insulator is not only highly accurate, but also retains key defect geometric features through optimization algorithms, avoiding excessive smoothing. The mapping method based on absolute coordinates ensures that electric field anomalies can be associated with specific locations on the 3D model with millimeter-level accuracy. The causal relationship between electric field anomalies and surface defects is established through spatial correlation analysis, revealing the actual impact of defects on electrical performance. Machine learning algorithms such as random forests are used to perform weighted fusion and comprehensive evaluation of multidimensional features, making the evaluation results accurate and accompanied by quantitative scores and causal inferences. This achieves a comprehensive and accurate assessment of the insulator condition, providing a direct and scientific basis for formulating differentiated maintenance strategies. By integrating multidimensional data acquisition, high-precision 3D reconstruction, multimodal data correlation, and intelligent decision-making, a complete closed loop from the physical world to the digital model and then to the diagnostic conclusion is constructed, which greatly improves detection efficiency and greatly meets the requirements of power system inspection for high efficiency and real-time performance. Attached Figure Description

[0017] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method for assessing the condition of an insulator according to a certain embodiment of the present invention; Figure 2 This is a structural diagram of a condition assessment system for insulators provided in a certain embodiment of the present invention; Figure 3 This is a structural diagram of an electronic device provided in a certain embodiment of the present invention; Figure label: Among them, 10 is the dataset construction module; 20 is the model construction module; 30 is the correlation analysis module; 40 is the state assessment module; 5000 is the electronic device; 5001 is the processor; 5002 is the bus; 5003 is the memory; and 5004 is the transceiver. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings and examples. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] In this invention description, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In this invention description, unless otherwise stated, "a plurality of" means two or more. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. Those skilled in the art will understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0021] In the description of this invention, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is merely for describing specific embodiments and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0022] To address the shortcomings of existing technologies, such as the limited data types of insulators inspected by drones, the inability to provide three-dimensional structural information, the difficulty in comprehensively assessing the insulator's condition, the high computational complexity and insufficient real-time performance when processing large amounts of data, and the inability to meet the high efficiency and real-time requirements of power system inspections, as well as the insufficient data security during data transmission and storage, making them susceptible to interference and attacks that could lead to data loss or leakage, this invention utilizes a control system combining a drone and a ground control center connected to the drone's data to achieve multi-dimensional data acquisition, processing, and condition assessment of insulators. This improves the real-time performance, accuracy, and reliability of insulator defect detection.

[0023] In one embodiment, such as Figure 1 As shown, the first aspect of the present invention provides a method for assessing the condition of insulators, which is applied in a ground control center, comprising: S1. Real-time acquisition of multi-dimensional observation data of the target insulator, and construction of a real-time observation dataset based on the multi-dimensional observation data; In one embodiment, step S1 includes: The image data, point cloud data, electric field data, and position data of the target insulator are collected in real time to serve as the multidimensional observation data; The multidimensional observation data is cleaned and feature extracted to obtain HOG and color features, curvature and normal vector features, and electric field features. The HOG and color features, the curvature and normal vector features, and the electric field features are weighted and fused to obtain the real-time observation dataset.

[0024] Specifically, this invention utilizes a drone equipped with a high-sensitivity fiber optic electric field sensor, lidar, binocular vision camera, and ORC positioning device to collect real-time electric field data, point cloud data, image data, and position data of a target insulator from different perspectives, and uses this data as multi-dimensional observation data. All sensors are fixed to the same mounting bracket on the drone, recording relative coordinate offsets to ensure spatial consistency of the data. During multi-source data acquisition, the drone performs multi-angle flight photography of the target insulator, the lidar acquires three-dimensional point cloud data of 100,000-200,000 discrete points per group, the binocular vision camera acquires image data, the fiber optic electric field sensor measures the surrounding electric field distribution at a 10Hz frequency, and the ORC positioning device provides absolute coordinates with an error of ±0.05m, ensuring spatial accuracy of the data acquisition.

[0025] Targeted filtering algorithms are employed for different data types. For example, statistical filtering is used to remove outliers (excluding the mean plus three standard deviations) from point cloud data; Gaussian filtering is used to suppress noise in image data; and sliding window filtering is used to eliminate transient pulses in electric field data, removing invalid information from all data. Key features are then extracted from the cleaned data: a quadratic surface is fitted to the neighborhood of the point cloud, and the surface curvature is calculated using the radius of curvature to distinguish between the insulator's "skirt" and "cap"; principal component analysis is used to reduce the dimensionality of the neighborhood point cloud, and the eigenvector corresponding to the minimum eigenvalue is taken as the normal vector for subsequent surface fitting in 3D reconstruction. The surface curvature and normal vector are combined to obtain curvature and normal vector features characterizing geometric anomalies (used to detect deformation, cracks, etc.); the image is divided into 2×2 cell blocks, each cell being 8×8 pixels, and the gradient direction histogram of each cell is calculated to generate a 128-dimensional / block feature vector for identifying surface contamination and cracks. Texture was analyzed, and the mean hue (H) and variance of saturation (S) were extracted in the HSV space. The S value of the contaminated area is usually 15%-30% higher than that of the normal area. Based on this, the HSV space features were extracted and combined with the 128-dimensional / block feature vector to obtain HOG and color features that characterize surface texture and color changes (used to detect dirt, damage, etc.). The peak electric field and electric field distortion rate in the electric field data were extracted, and the maximum electric field value at the edge and middle of the insulator skirt was statistically analyzed. The peak electric field value at the edge of a normal 110kV insulator skirt is about 15kV / m. Based on this, the deviation rate between the measured electric field curve and the normal curve was calculated, and the electric field peak and electric field distortion rate constituted the electric field features characterizing the electrical insulation state.

[0026] The extracted features are normalized to ensure they are of the same magnitude. Then, based on prior knowledge or experimental analysis, weights are assigned to different features (in this embodiment, the preferred weights for HOG and color features, curvature and normal vector features, and electric field features are 0.3, 0.4, and 0.3, respectively). For example, if surface defects are of greater concern, HOG and color features are given higher weights; if deformation is of greater concern, curvature and normal vector features are given higher weights; if electrical performance is of greater concern, electric field features are given higher weights. These can be adjusted according to actual needs. Finally, the weighted features are concatenated with their corresponding location data to form a comprehensive feature vector, constituting the real-time observation dataset.

[0027] This invention, by simultaneously acquiring image, point cloud, electric field, and location data, can comprehensively perceive the insulator's state from different dimensions (visual, geometric, electrical, and spatial), avoiding the limitations of a single data source. Through key feature extraction, it can reflect the insulator's health status from various perspectives. Weighted fusion of different features allows for the assignment of different weights based on their importance, resulting in a fused feature set that better highlights key information and improves the accuracy of subsequent state assessments. After data cleaning and feature extraction, a multi-feature fusion real-time observation dataset is obtained, laying a solid foundation for subsequent semantic segmentation, point cloud registration, 3D model construction, and state assessment.

[0028] S2. A semantic segmentation model is used to perform point cloud registration on the real-time observation dataset, and a target mesh model is generated based on the point cloud registration results. The target mesh model is then optimized using a reinforcement learning algorithm to obtain a three-dimensional model of the insulator. In one embodiment, the step of performing point cloud registration on the real-time observation dataset using a semantic segmentation model includes: Based on the location data, the corresponding point cloud data in the real-time observation dataset is transformed to the same world coordinate system to obtain transformed point cloud data; The converted point cloud data is downsampled using a voxel grid, and several key sampling points are extracted from the downsampling results using the FPS algorithm. A spherical neighborhood is constructed with each of the aforementioned key sampling points as the center, and local features of all key sampling points within each spherical neighborhood are extracted to obtain multiple sets of local features; A multilayer perceptron is used to encode the local feature set corresponding to each spherical neighborhood to obtain a multidimensional local feature descriptor. Calculate the cosine similarity between the multidimensional local feature descriptors, and select key sampling points with similarity results greater than a preset similarity threshold as valid corresponding point pairs; The effective corresponding point pairs are iteratively optimized using the ICP algorithm to obtain the target registration point cloud and the source registration point cloud as the point cloud registration result.

[0029] Specifically, this invention first initializes the point cloud data from multiple perspectives. That is, based on the absolute coordinate data with an error of ±0.05m provided by the ORC positioning device, each group of point cloud data containing 100,000 to 200,000 discrete points collected by the UAV from different flight perspectives is transformed from the local coordinate system to the same world coordinate system to obtain transformed point cloud data, achieving preliminary alignment. At this time, the error of the overlapping area of ​​the point cloud is about 5mm.

[0030] Next, voxel mesh downsampling was employed, with the voxel size set to 0.5 mm to downsample the converted point cloud data. This reduced the number of point clouds to 60%-70% of the original size while preserving the key geometric features of the insulator, effectively reducing subsequent computational complexity. The PointNet++ semantic segmentation model was then used to continue multi-scale feature learning on the point cloud data. This model has a three-level structure: sampling layer, grouping layer, and feature extraction layer. In the sampling layer, 5000 key sampling points were selected from the downsampled point cloud using the FPS (Farthest Point Sampling) algorithm as the centers for local feature extraction. In the grouping layer, a spherical neighborhood with a radius of 2 mm was constructed centered on each key sampling point, and the local features of all key sampling points within each spherical neighborhood were obtained, resulting in multiple local feature sets. In the feature extraction layer, the local feature sets of the point cloud corresponding to each spherical neighborhood were feature-encoded using an MLP (Multilayer Perceptron) to generate a 128-dimensional local feature descriptor, which accurately represents the local geometric features of the insulator.

[0031] Subsequently, feature matching is performed, and the cosine similarity between local feature descriptors of point clouds under different viewpoints is calculated (when calculating, any viewpoint can be selected as the target viewpoint point cloud, and the cosine similarity between local feature descriptors of other non-target viewpoints—that is, the local feature descriptors of the source viewpoint point cloud and the local feature descriptors of the target viewpoint point cloud—is calculated). Key sampling points with similarity ≥ 0.92 are selected as valid corresponding point pairs, thereby reducing the mismatch rate to below 2%.

[0032] In the fine registration stage, the ICP algorithm is used for iterative optimization. First, based on effective corresponding point pairs, the error function is minimized using the least squares method to obtain the 3×3 rotation matrix and the 3×1 translation vector. This process is expressed by the following formula: In the formula, It is the error function; Point cloud for target viewpoint; Point cloud from the source viewpoint; It is a 3×3 rotation matrix; It is a 3×1 translation vector.

[0033] Then, the coordinate transformation of the source viewpoint point cloud is performed according to the solved (R,t), and the average distance between the transformed point cloud and the target point cloud is calculated. This process is repeated iteratively until the average distance is ≤0.008mm, and finally the registration error is <3mm and the point cloud overlap rate is ≥98%. The target viewpoint point cloud and the source viewpoint point cloud that meet these conditions are used as the target registration point cloud and the source registration point cloud, respectively, thus forming the point cloud registration result.

[0034] This invention transforms point clouds to the same world coordinate system based on location data, providing a foundation for subsequent multi-view point cloud fusion and ensuring spatial consistency. Voxel mesh downsampling reduces data volume while preserving geometric features, improving processing speed. The FPS algorithm ensures uniform distribution of key points, covering the entire point cloud. Constructing spherical neighborhoods to extract local features captures the subtle geometric structure of the point cloud, providing excellent representation of complex-shaped insulators. Multilayer perceptrons encode local features, generating discriminative local feature descriptors, which facilitates subsequent matching. Cosine similarity is used to measure the similarity of feature descriptors, and a threshold is set to filter reliable corresponding point pairs, reducing the false matching rate. The ICP algorithm iteratively optimizes based on effective corresponding point pairs, further refining the registration results and ensuring high accuracy in the final registration.

[0035] In one embodiment, generating the target mesh model based on the point cloud registration result includes: Principal component analysis is used to calculate the normal vectors of the target registration point cloud and the source registration point cloud respectively for time-series consistency calibration, resulting in the target calibration normal vector and the source calibration normal vector. The historical three-dimensional mesh model corresponding to the target insulator is retrieved, and the micro-texture parameters and macro-topological features of the historical three-dimensional mesh model are extracted to construct a historical feature library; Based on the historical feature library and historical 3D mesh model of the source registration point cloud and its corresponding target registration point cloud, paired training data is constructed to train the generative adversarial network and obtain the trained generative adversarial network. The target calibration normal vector and the source calibration normal vector are processed by the Poisson surface reconstruction algorithm to obtain the initial mesh model; The initial mesh model is input into the trained generative adversarial network for texture optimization to obtain the target mesh model.

[0036] Specifically, this invention uses PCA (Principal Component Analysis) to calculate the normal vector of each point in the registered complete point cloud. Specifically, it constructs a covariance matrix for each point's 10 neighboring points, solves for the eigenvector corresponding to the smallest eigenvalue, which is the normal vector of that point, and retrieves the normal vector data from the last three historical inspections of the point cloud to calibrate the normal vector of the current point cloud. If the angle between the normal vector of a current point and the normal vector of a historical point at the same location is greater than 15° (determined as an anomaly, possibly caused by point cloud jitter), then the mean of the historical normal vectors is used to correct the current point cloud's normal vector. The front normal vector ensures that the normal vectors all point outwards from the insulator, achieving temporal consistency calibration of the point cloud and avoiding the "face flipping" problem in subsequent mesh generation. The normal vectors of the calibrated point cloud are used as the corresponding target calibration normal vectors or source calibration normal vectors. Alternatively, the overall centroid of each point cloud is calculated, and it is determined whether the angle between the normal vector and the "point-centroid line" satisfies <90°. If it does not satisfy the condition, the reverse normal vector is corrected to ensure that the normal vectors of all points point outwards from the insulator, avoiding the "face flipping" problem in subsequent mesh generation.

[0037] The training phase of the Generative Adversarial Network (GAN) then begins: From the historical database of the ground control center, three-dimensional mesh models from the last three inspections of the same model and installation location as the current insulator are retrieved. Key texture features of each mesh model are analyzed: micro-texture parameters: curvature distribution of triangular facets (e.g., mean curvature of the steel cap surface, standard deviation of curvature of the umbrella skirt glaze), and mean texture grayscale (texture information based on binocular camera image mapping); macro-topological features: fixed structural parameters such as the number of umbrella skirts, steel cap diameter, and porcelain insulator column height (serving as topological constraints for texture generation, preventing GAN from generating textures that do not conform to physical structures). The extracted historical texture features are then normalized (e.g., curvature values ​​are normalized to the [0,1] interval). A historical feature library is constructed according to "insulator regions": for example, the steel cap region library stores the curvature distribution and texture grayscale variance of the historical steel cap surface (typical value: mean curvature 0.8). (Grayscale variance < 10); Umbrella skirt region library: stores the concavity and convexity features of the historical umbrella skirt glaze (such as the natural concavity and convexity curvature range of 0.3-0.5mm) and edge texture regularity; Porcelain bottle column library: stores the longitudinal texture spacing of the column surface (typical value: 5-8mm / channel); This feature library will be used as the "temporal texture constraint" for training the GAN network to ensure that the generated mesh texture conforms to the long-term stable surface regularity of the insulator.

[0038] Based on the historical feature database and historical 3D mesh model of the source registration point cloud and its corresponding target registration point cloud, paired training data was constructed, collecting a total of 1000 sets (covering historical inspection data of different types of insulators and different years of operation). The input consisted of the registered point cloud (including normal vectors and temporal texture feature database index); the labels were the corresponding historical real mesh models (obtained by a 0.05mm precision ground 3D scanner to ensure texture authenticity). This data was then used to train a generative adversarial network (GAN). The generator in this network uses a 3-layer deconvolutional network. The input is the feature matrix of the registered point cloud (containing geometric features and temporal texture features), and the output is an initial mesh candidate model (approximately 100,000 triangular facets). The core function is to learn the surface features from the temporal texture feature database. Patterns (such as the metallic luster texture of a steel cap, the uneven glaze of an umbrella skirt); the discriminator uses a 3-layer convolutional network, with the input being the candidate mesh output by the generator and the historical real mesh. By comparing the texture parameters (curvature distribution, grayscale mean) and topology (patch connection method) of the two, it outputs a realism score (0-1, 1 indicating complete consistency with the real mesh); by alternately training the generator and discriminator, the sum of the texture difference loss (using the L2 loss function) and the topology difference loss (using the IoU loss function) between the generated mesh and the real mesh is minimized until the discriminator's recognition accuracy for the generated mesh and the real mesh approaches 50% (i.e., the texture of the mesh generated by the generator is not significantly different from the historical real texture), thus obtaining the completed generative adversarial network.

[0039] The preprocessed point cloud (including normal vectors) is input into the Poisson surface reconstruction algorithm, the core of which is to construct a 3D implicit function. , so that its gradient The implicit function surface is aligned with the point cloud normal vector direction (ensuring it fits the point cloud geometry), and then extracted using the Marching Cubes algorithm. The isosurfaces are used to generate an initial mesh model—this initial mesh contains approximately 100,000 triangular facets, which can reproduce surface details down to 1mm.

[0040] The initial mesh is input into the pre-trained GAN generator. Based on the constraints of the temporal texture feature library, the generator optimizes the initial mesh as follows: Texture supplementation: Add 0.2-0.3mm micro-bumps to the umbrella skirt glaze area in accordance with historical patterns (referencing the curvature distribution of the umbrella skirt area in the temporal texture library); Topology correction: Adjust the surface structure of the connection between the steel cap and the porcelain bottle column to match the topology of the historical mesh (such as the transition curvature between the bottom of the steel cap and the column, avoiding surface breakage); Finally, the optimized mesh model containing 180,000 triangular facets is output, which can clearly present 0.2mm-level surface bump features (such as the subtle depressions of early micro-cracks), thus obtaining the target mesh model.

[0041] This invention improves the accuracy of normal vector calculation by using principal component analysis (PCA) and performing temporal consistency calibration, thus providing more accurate geometric information for subsequent surface reconstruction. It utilizes historical 3D mesh models to construct a historical feature library, drawing upon microscopic textures and macroscopic topological features from historical data to make the reconstructed model more consistent with the characteristics of real objects. By using generative adversarial networks (GANs) to optimize the texture of the initial mesh model, more realistic and detailed surface textures are generated, improving the model's appearance quality. It combines Poisson surface reconstruction and deep learning, balancing geometric accuracy and texture realism. By constraining GANs with temporal texture features, it solves the problems of poor texture realism and detail loss in traditional mesh generation, providing a high-precision geometric foundation for subsequent mesh-based defect detection.

[0042] In one embodiment, optimizing the target mesh model using a reinforcement learning algorithm to obtain a three-dimensional insulator model includes: Defect detection is performed on the target mesh model to obtain defect regions; Using the Laplace smoothing factor as the agent, a target reward function is constructed based on the surface roughness of the target mesh model and the retention rate of the defect region; The Laplace smoothing factor is initialized so that the agent optimizes the target mesh model of the non-defect region based on the initialized Laplace smoothing factor, and calculates the current reward in real time based on the target reward function; The agent adaptively optimizes the initialized Laplace smoothing factor based on the current reward, so that the agent optimizes the target mesh model of the defect region based on the adaptively optimized Laplace smoothing factor. The optimization steps for the target mesh model in non-defective regions and the target mesh model in defective regions are executed iteratively until the preset termination condition is met. The final optimized target mesh models are then combined to obtain the three-dimensional model of the insulator.

[0043] Specifically, this invention optimizes the model parameters of the target mesh network through "multi-objective loss function design + dynamic weight adjustment + regularization constraint", enabling the network to accurately learn temporal texture features and geometric structure rules, and ultimately enabling the generated 3D model of insulator to have both 0.2mm level detail reproduction and temporal consistency.

[0044] To avoid convergence difficulties caused by random parameter fluctuations in the early stages of training, this invention requires targeted initialization of the core parameters (weights, biases) of the generator and discriminator: The generator adopts a 3-layer deconvolutional network (input layer → hidden layer 1 → hidden layer 2 → output layer), and the parameter initialization rules for each layer are as follows: Weight parameters: Kaiming normal initialization is adopted (designed for ReLU activation function), mean μ=0, standard deviation σ is the square root of the quotient of 2 and the number of input channels, to ensure that the variance of the output of each layer is consistent and to avoid gradient vanishing; Bias parameters: initialized to 0.01 (a small positive value) to avoid activation failure caused by the initial output of neurons being 0; Deconvolution kernel size: 3×3×3 (three-dimensional convolution kernel, matching the spatial dimension of the point cloud), stride set to 1 (to ensure that texture details are not lost). The discriminator employs a 3-layer convolutional network (input layer → hidden layer 1 → hidden layer 2 → output layer). Parameter initialization rules are as follows: Weight parameters: uniform initialization using Xavier (designed for the Sigmoid activation function), with values ​​ranging from the square root of the quotient of -6 and the sum of the input and output channels, to the square root of the quotient of 6 and the sum of the input and output channels, ensuring that the mean and variance of the input and output are consistent; Bias parameters: initialized to 0 (to avoid introducing bias into the discriminator's initial judgment); Convolutional kernel size: 5×5×5 (slightly larger than the generator, enhancing the ability to recognize the global topology of the grid), with a stride of 2 (gradually compressing the feature map to extract global features).

[0045] After initialization, 10 sets of historical grid data are randomly sampled for pre-verification to ensure that the deviation between the grid topology (such as the number of umbrella skirts and the diameter of the steel cap) initially output by the generator and the real grid is ≤5% (if the deviation is too large, the initialization parameters are readjusted).

[0046] Subsequently, mesh defect detection was conducted, and the target mesh model was analyzed: by fitting the quadratic surface of the mesh patch neighborhood, the radius of curvature R was solved, and thus the surface curvature was obtained. Set the curvature change rate threshold to >0.5 When the rate of curvature change of a certain area of ​​the mesh exceeds the threshold, it is determined to be a potential defect area, and the mesh vertex coordinates of the area are marked with a marking accuracy of ≤0.1mm, while other areas are non-defect areas.

[0047] Set up a reinforcement learning environment: Using the mesh model to be optimized as the environment, output the surface roughness of the current mesh model from the environment. ) and defect area retention rate (Calculated by comparing the changes in the geometric features of the defect region before and after optimization) is used as the state, and the Laplacian smoothing factor λ (range [0.1, 0.6]) is used as the agent to control the degree of mesh smoothing; the objective reward function is... ,(in The initial value is approximately 1.5 μm (≥95%), ensuring that defect details are preserved while reducing surface roughness.

[0048] In the initial iteration, the Laplace smoothing factor is initialized, i.e., λ=0.6 is set in the non-defect region, so that the agent performs Laplace smoothing on the vertices of the target mesh model in the non-defect region using the Deep Deterministic Policy Gradient (DDPG) algorithm based on 0.6. The formula is as follows: (in For the vertex coordinates to be optimized, Given the coordinates of its k neighboring vertices (k is usually 8-12), calculate the current reward based on the target reward function.

[0049] When optimizing to the marked defect region, the reinforcement learning model adaptively optimizes the initialized Laplace smoothing factor based on the current reward fed back by the reward function, that is, automatically reduces λ to 0.1 to reduce the smoothness and avoid blurring the defect details; The optimization steps for the target mesh model in the non-defect region and the target mesh model in the defect region are executed iteratively until the preset termination condition is met. That is, when the reward value fluctuation of three consecutive iterations is ≤0.02 and the current surface roughness reaches the predetermined target (e.g., ≤1.0μm) and the defect region retention rate is ≥95%, the optimization stops. The optimized non-defect region and defect region are combined to obtain the final three-dimensional model of the insulator.

[0050] This invention identifies defect regions through defect detection, enabling subsequent optimization to differentiate between defective and non-defective regions, avoiding over-smoothing of defective areas and preserving important defect details. Using a Laplacian smoothing factor as the agent, reinforcement learning adaptively adjusts the smoothing factor, allowing for bold smoothing in non-defective regions to reduce surface roughness, while cautious smoothing in defective regions to preserve details. The reward function considers both surface roughness and defect retention rate, ensuring that smoothing improves model quality without losing defect features, satisfying the dual requirements of surface smoothness and defect preservation in insulator detection. Iterative optimization with a set termination condition allows the optimization process to proceed automatically until a stable state is reached, ensuring the reliability of the optimization results. This invention executes the algorithm sequentially from "point cloud registration → mesh generation → model optimization," combining deep learning and reinforcement learning techniques to generate a high-precision 3D model that accurately reflects the geometry and surface details of the insulator.

[0051] S3. Obtain the electric field anomaly points of the target insulator and combine them with the three-dimensional model of the insulator. Perform electric field-defect spatial correlation analysis by absolute coordinate mapping method based on ORC positioning and spatial geometric calibration to obtain the correlation analysis results. In one embodiment, step S3 includes: Based on the relationship between the electric field data and the preset electric field data, the electric field anomaly point of the target insulator is determined, and the position data corresponding to the electric field anomaly point is converted to the world coordinate system to be mapped to the three-dimensional model of the insulator. The K-nearest neighbor algorithm is used to find several vertices within a preset distance of the electric field anomaly point in the mapped 3D model of the insulator, and the average coordinate of each vertex is used as the mapping point. The image data and the point cloud data are associated using the SIFT feature matching method to obtain associated coordinates. Based on the associated coordinates and the three-dimensional model of the insulator, the three-dimensional spatial coordinates of the defect area are extracted using the Canny edge detection method to obtain the defect area coordinate set. Calculate the spatial distance between each mapping point and the coordinate set of the defect region, and analyze the spatial correlation between each electric field anomaly point and the coordinate set of the defect region based on the spatial distance calculation results to obtain the correlation analysis results.

[0052] This invention is based on the normal electric field distribution curve calibrated in the laboratory. Calculate its relationship with electric field data Deviation rate between and will When the error exceeds 15%, the point is identified as an electric field anomaly. Based on the error between sensors, the location data corresponding to the electric field anomaly is converted to the world coordinate system. Since the fiber optic electric field sensor and the lidar are fixed on the same mounting bracket, and the relative coordinate offset is recorded (e.g., the center of the electric field sensor relative to the lidar: X=20mm, Y=5mm, Z=0mm), the relative coordinate deviation is <0.5mm. ORC outputs the absolute coordinates of the mounting bracket in real time. Calculate lidar coordinates Coordinates of electric field sensor And the electric field anomalies are mapped to the three-dimensional model of the insulator.

[0053] Since the coordinates of the 3D model are completely consistent with the absolute coordinates of the lidar, this invention uses the K-nearest neighbor algorithm (K=3) to find several vertices in the mapped 3D model of the insulator that are within a preset distance from the electric field anomaly point, that is, the 3 vertices in the model that are closest to the electric field anomaly point, and calculates the average coordinates of these vertices as mapping points. The mapping error is ≤1mm. The mapping points are connected to form a closed region, which is marked with red highlight on the model.

[0054] The pixel coordinates of the binocular vision camera image are associated with the LiDAR point cloud coordinates using the SIFT feature matching method to obtain the 3D associated coordinates of the image feature points in the point cloud. The Canny edge detection algorithm and gradient operator are then employed. , : The gradient magnitude is calculated based on the gradient operator. The dual thresholds are a high threshold of 80 and a low threshold of 40 for edge extraction, as well as morphological expansion of the 3×3 structuring element. It can detect cracks as small as 2mm and damage of 5mm² with a defect positioning accuracy of ≤1mm. Combined with the associated three-dimensional coordinates, the three-dimensional spatial coordinate set of the defect area is obtained, i.e., the defect area coordinate set, with a positioning accuracy of ≤1mm.

[0055] Calculate the minimum spatial distance d between the mapping point corresponding to each electric field anomaly and all defect regions in the defect region coordinate set. When d ≤ 5 mm, it is determined that the electric field anomaly and the defect region are spatially correlated. When d > 5 mm, it is determined to be an independent anomaly. Statistically analyze the correlation of all electric field anomalies to form the correlation analysis results, including: the number of correlated anomalies, the correlated defect regions, the average spatial distance, and the correlation strength index.

[0056] This invention integrates electric field data, image data, and point cloud data, leveraging their respective advantages to improve the comprehensiveness and accuracy of detection. It utilizes ORC positioning and world coordinate system transformation to ensure that electric field anomalies and defect areas are correlated within a unified coordinate system, resulting in high spatial accuracy. Through the K-nearest neighbor algorithm and SIFT feature matching, electric field anomalies are correlated with visual defects in three-dimensional space, accurately determining whether they originate from the same problem. Combining fiber optic electric field data and image processing technology, and employing absolute coordinate mapping and spatial distance correlation analysis, it comprehensively evaluates the electrical performance status and surface defects of insulators, providing more comprehensive detection results.

[0057] S4. Based on the correlation analysis results, perform multimodal fusion analysis to obtain the state assessment results of the target insulator; In one embodiment, step S4 includes: The coordinates of electric field anomaly points and defect regions that are associated with the analysis results are extracted from the correlation analysis results, and used as target anomaly points and target defect regions. The average deviation rate of electric field anomalies, defect area, and longest defect length are determined based on the target anomaly points and the target defect regions to quantify the associated anomaly scores. The random forest algorithm is used to perform multimodal fusion analysis on the target anomaly points, the target defect regions, and the associated anomaly scores to obtain a comprehensive score, and the state assessment result of the target insulator is determined based on the comprehensive score.

[0058] Specifically, this invention reads data from the correlation analysis results database, filters out electric field anomaly points and defect areas with a spatial distance of less than or equal to 5 mm as target anomaly points and target defect areas; and extracts the average deviation rate of electric field anomalies based on the target anomaly points and target defect areas (calculating the average deviation rate of all sampling points in the electric field anomaly area where the target anomaly point is located, reflecting the severity of electric field distortion). The defects are calculated as follows: 1) Defect area (S, calculated from the three-dimensional coordinates of the target defect region, in square millimeters, representing the scale of the defect); 2) Maximum defect length (L, calculated from the maximum Euclidean distance between any two points in the defect region, in millimeters, reflecting the extent of defect expansion); and 3) Correlation anomaly score is calculated based on these extracted data. (50mm² is the threshold for the area of ​​a minor defect, and 20mm is the threshold for the length of a minor defect), the score range is [0,100]; for example, "abnormal electric field in the middle of the umbrella skirt ( The model is associated with a 3mm×15mm crack and scores 68. When the association is successful, the cause is inferred. When it is an independent anomaly, it is determined to be an internal insulation problem. The model is presented with "red highlight (electric field anomaly) + yellow dashed box (associated defect)" and includes quantitative features and diagnostic conclusions.

[0059] Subsequently, a random forest algorithm was used to train a weight model based on historical fault data. The weights for electric field anomalies were determined to be 0.4, surface defects 0.3, and associated anomaly scores 0.3. These weights were then used to perform multimodal fusion analysis on the target anomaly points, target defect areas, and associated anomaly scores, also known as formaldehyde fusion, to obtain a comprehensive score. An SVM algorithm was then used for comprehensive classification, with the feature vector being "area of ​​electric field anomaly region (weight 0.4) + area of ​​defect (weight 0.3) + grayscale difference of defect (weight 0.3)". The training set contained 500 sets of insulator data, including 200 normal sets, 200 sets of slightly defective sets, and 100 sets of severely defective sets. The accuracy of the test set reached 96%.

[0060] Alternatively, a pre-trained random forest model can be used for fusion analysis. This model is trained using 500 sets of historical data, including 200 normal sets, 200 sets with minor defects, and 100 sets with severe defects. The model input consists of the three quantitative features mentioned above (average deviation rate of electric field anomalies, defect area, and longest defect length) and the associated anomaly score, and the output is a comprehensive score from 0 to 100. The random forest model determines the comprehensive score through voting among multiple decision trees, which can effectively avoid overfitting and handle complex relationships between features. The model achieves an accuracy of 96% on the test set.

[0061] Based on the comprehensive score, the system classifies the insulator condition into four levels: Excellent (90-100 points): The insulator is in good condition and requires no maintenance; Good (70-89 points): The insulator has minor defects and regular observation is recommended; Caution (50-69 points): The insulator has moderate defects and planned maintenance is recommended; Warning (0-49 points): The insulator has serious defects and immediate maintenance is recommended. For example, if an insulator has an abnormal electric field area of ​​300 mm² (score 50), a defect area of ​​60 mm² (score 40), and a defect grayscale difference of 60 (score 80), then the comprehensive score is 50×0.4+40×0.3+80×0.3=58, which is classified as "moderate defect" and requires attention. Finally, the comprehensive score, surface defects, and anomaly and defect association annotations are combined to obtain the condition assessment result of the target insulator. This result uses 3D rendering technology to intuitively display the 3D model of the insulator, supporting 1-10 times model scaling, arbitrary angle rotation, and cross-section along the XYZ axes. The internal structure of the insulator can be viewed, and electric field anomalies and associated defect areas are highlighted. The condition assessment result of the insulator can also be displayed through text, charts, and color coding.

[0062] This invention extracts truly correlated anomaly points and defect regions from correlation analysis results, avoiding interference from irrelevant data and improving the targeting of the assessment. It comprehensively quantifies the severity of correlated anomalies using three key indicators: average deviation rate of electric field anomalies, defect area, and longest defect length, providing an objective basis for condition assessment. The random forest algorithm is used to fuse multimodal features, capturing complex nonlinear relationships and improving the accuracy and reliability of condition assessment. Combining fiber optic electric field data and image processing technology, and employing the "absolute coordinate mapping method" with spatial distance correlation analysis, it comprehensively assesses the electrical performance status and surface defects of insulators, providing more comprehensive detection results. Simultaneously, it utilizes efficient data transmission from UAVs and ground control centers, combined with preprocessing and machine learning algorithms, to improve the real-time performance and efficiency of inspections.

[0063] It should be noted that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order requirement for the execution of these steps, and they can be executed in other orders.

[0064] In another embodiment, such as Figure 2 As shown, a second aspect of the present invention provides a system for assessing the condition of an insulator, comprising: The dataset construction module 10 is used to collect multidimensional observation data of the target insulator in real time and construct a real-time observation dataset based on the multidimensional observation data. The model building module 20 is used to perform point cloud registration on the real-time observation dataset using a semantic segmentation model, generate a target mesh model based on the point cloud registration results, and optimize the target mesh model through a reinforcement learning algorithm to obtain a three-dimensional model of the insulator. The correlation analysis module 30 is used to obtain the electric field anomaly points of the target insulator, and combine them with the three-dimensional model of the insulator. The electric field-defect spatial correlation analysis is performed by the absolute coordinate mapping method based on ORC positioning and spatial geometric calibration to obtain the correlation analysis results. The state assessment module 40 is used to perform multimodal fusion analysis based on the correlation analysis results to obtain the state assessment results of the target insulator.

[0065] It should be noted that the various modules in the aforementioned insulator condition assessment system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module. For specific limitations regarding the insulator condition assessment system, please refer to the limitations regarding the insulator condition assessment method described above; both have the same function and role, and will not be repeated here.

[0066] A third aspect of the present invention provides an electronic device comprising: Processor, memory, and bus; The bus is used to connect the processor and the memory; The memory is used to store operation instructions; The processor is configured to execute instructions by invoking the operation instructions, causing the processor to perform operations corresponding to a method for assessing the state of an insulator as shown in the first aspect of the present invention.

[0067] In one alternative embodiment, an electronic device is provided, such as Figure 3 As shown, Figure 3 The illustrated electronic device 5000 includes a processor 5001 and a memory 5003. The processor 5001 and the memory 5003 are connected, for example, via a bus 5002. Optionally, the electronic device 5000 may also include a transceiver 5004. It should be noted that in practical applications, the transceiver 5004 is not limited to one type, and the structure of this electronic device 5000 does not constitute a limitation on the embodiments of the present invention.

[0068] Processor 5001 may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in connection with this disclosure. Processor 5001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0069] Bus 5002 may include a path for transmitting information between the aforementioned components. Bus 5002 may be a PCI bus or an EISA bus, etc. Bus 5002 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 3 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0070] The memory 5003 may be a ROM or other type of static storage device capable of storing static information and instructions, RAM or other type of dynamic storage device capable of storing information and instructions, or it may be an EEPROM, CD-ROM or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto.

[0071] The memory 5003 is used to store application code that executes the present invention, and its execution is controlled by the processor 5001. The processor 5001 is used to execute the application code stored in the memory 5003 to implement the content shown in any of the foregoing method embodiments.

[0072] Among them, electronic devices include, but are not limited to: mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), and in-vehicle terminals (such as in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers.

[0073] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a method for assessing the state of an insulator as described in the first aspect of the present invention.

[0074] Another embodiment of the present invention provides a computer-readable storage medium storing a computer program that, when run on a computer, enables the computer to execute the corresponding content in the foregoing method embodiments.

[0075] Furthermore, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0076] In summary, this invention relates to the field of power system inspection technology, and discloses a method, system, equipment, and medium for assessing the condition of insulators. It constructs a real-time observation dataset by acquiring multi-dimensional observation data of the target insulator in real time; performs point cloud registration on the real-time observation dataset using a semantic segmentation model, and generates a target mesh model based on the point cloud registration results. The target mesh model is then optimized using a reinforcement learning algorithm to obtain a three-dimensional model of the insulator. Electric field anomalies of the target insulator are acquired and combined with the three-dimensional model. An electric field-defect spatial correlation analysis is performed using an absolute coordinate mapping method based on ORC positioning and spatial geometric calibration. The correlation analysis results are then subjected to multi-modal fusion analysis to obtain the condition assessment results of the target insulator. This approach not only meets the requirements of high efficiency and real-time operation for power system inspection, but also achieves a comprehensive and accurate assessment of the insulator's condition.

[0077] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on its differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above 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 the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0078] The above-described embodiments are merely preferred embodiments of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A method for assessing the condition of insulators, characterized in that, include: Real-time acquisition of multi-dimensional observation data of the target insulator, and construction of a real-time observation dataset based on the multi-dimensional observation data; A semantic segmentation model is used to perform point cloud registration on the real-time observation dataset, and a target mesh model is generated based on the point cloud registration results. The target mesh model is then optimized using a reinforcement learning algorithm to obtain a three-dimensional model of the insulator. The electric field anomaly points of the target insulator are obtained and combined with the three-dimensional model of the insulator. The electric field-defect spatial correlation analysis is performed by the absolute coordinate mapping method based on ORC positioning and spatial geometric calibration to obtain the correlation analysis results. Based on the correlation analysis results, multimodal fusion analysis is performed to obtain the state assessment results of the target insulator.

2. The method for assessing the condition of an insulator according to claim 1, characterized in that, The real-time acquisition of multi-dimensional observation data of the target insulator, and the construction of a real-time observation dataset based on the multi-dimensional observation data, includes: The image data, point cloud data, electric field data, and position data of the target insulator are collected in real time to serve as the multidimensional observation data; The multidimensional observation data is cleaned and feature extracted to obtain HOG and color features, curvature and normal vector features, and electric field features. The HOG and color features, the curvature and normal vector features, and the electric field features are weighted and fused to obtain the real-time observation dataset.

3. The method for assessing the condition of an insulator according to claim 2, characterized in that, The step of using a semantic segmentation model to perform point cloud registration on the real-time observation dataset includes: Based on the location data, the corresponding point cloud data in the real-time observation dataset is transformed to the same world coordinate system to obtain transformed point cloud data; The converted point cloud data is downsampled using a voxel grid, and several key sampling points are extracted from the downsampling results using the FPS algorithm. A spherical neighborhood is constructed with each of the aforementioned key sampling points as the center, and local features of all key sampling points within each spherical neighborhood are extracted to obtain multiple sets of local features; A multilayer perceptron is used to encode the local feature set corresponding to each spherical neighborhood to obtain a multidimensional local feature descriptor. Calculate the cosine similarity between the multidimensional local feature descriptors, and select key sampling points with similarity results greater than a preset similarity threshold as valid corresponding point pairs; The effective corresponding point pairs are iteratively optimized using the ICP algorithm to obtain the target registration point cloud and the source registration point cloud as the point cloud registration result.

4. The method for assessing the condition of an insulator according to claim 3, characterized in that, The generation of the target mesh model based on the point cloud registration results includes: Principal component analysis is used to calculate the normal vectors of the target registration point cloud and the source registration point cloud respectively, in order to perform temporal consistency calibration, and obtain the target calibration normal vector and the source calibration normal vector; The historical three-dimensional mesh model corresponding to the target insulator is retrieved, and the micro-texture parameters and macro-topological features of the historical three-dimensional mesh model are extracted to construct a historical feature library; Based on the historical feature library and historical 3D mesh model of the source registration point cloud and its corresponding target registration point cloud, paired training data is constructed to train the generative adversarial network and obtain the trained generative adversarial network. The target calibration normal vector and the source calibration normal vector are processed by the Poisson surface reconstruction algorithm to obtain the initial mesh model; The initial mesh model is input into the trained generative adversarial network for texture optimization to obtain the target mesh model.

5. The method for assessing the condition of an insulator according to claim 2, characterized in that, The optimization of the target mesh model using a reinforcement learning algorithm to obtain a three-dimensional insulator model includes: Defect detection is performed on the target mesh model to obtain defect regions; Using the Laplace smoothing factor as the agent, a target reward function is constructed based on the surface roughness of the target mesh model and the retention rate of the defect region; The Laplace smoothing factor is initialized so that the agent optimizes the target mesh model of the non-defect region based on the initialized Laplace smoothing factor, and calculates the current reward in real time based on the target reward function; The agent adaptively optimizes the initialized Laplace smoothing factor based on the current reward, so that the agent optimizes the target mesh model of the defect region based on the adaptively optimized Laplace smoothing factor. The optimization steps for the target mesh model in the non-defective region and the target mesh model in the defective region are executed iteratively until the preset termination condition is reached. The final optimized target mesh models are then combined to obtain the three-dimensional model of the insulator.

6. The method for assessing the condition of an insulator according to claim 5, characterized in that, The process involves acquiring the electric field anomaly points of the target insulator and combining them with the insulator's three-dimensional model. Then, an electric field-defect spatial correlation analysis is performed using an absolute coordinate mapping method based on ORC positioning and spatial geometric calibration. The correlation analysis results include: Based on the relationship between the electric field data and the preset electric field data, the electric field anomaly point of the target insulator is determined, and the position data corresponding to the electric field anomaly point is converted to the world coordinate system to be mapped to the three-dimensional model of the insulator. The K-nearest neighbor algorithm is used to find several vertices within a preset distance of the electric field anomaly point in the mapped 3D model of the insulator, and the average coordinate of each vertex is used as the mapping point. The image data and the point cloud data are associated using the SIFT feature matching method to obtain associated coordinates. Based on the associated coordinates and the three-dimensional model of the insulator, the three-dimensional spatial coordinates of the defect area are extracted using the Canny edge detection method to obtain the defect area coordinate set. Calculate the spatial distance between each mapping point and the coordinate set of the defect region, and analyze the spatial correlation between each electric field anomaly point and the coordinate set of the defect region based on the spatial distance calculation results to obtain the correlation analysis results.

7. The method for assessing the condition of an insulator according to claim 1, characterized in that, The multimodal fusion analysis based on the correlation analysis results yields a state assessment result for the target insulator, including: The coordinates of electric field anomaly points and defect regions that are associated with the analysis results are extracted from the correlation analysis results, and used as target anomaly points and target defect regions. The average deviation rate of electric field anomalies, defect area, and longest defect length are determined based on the target anomaly points and the target defect regions to quantify the associated anomaly scores. The random forest algorithm is used to perform multimodal fusion analysis on the target anomaly points, the target defect regions, and the associated anomaly scores to obtain a comprehensive score, and the state assessment result of the target insulator is determined based on the comprehensive score.

8. A condition assessment system for insulators, characterized in that, include: The dataset construction module is used to collect multidimensional observation data of the target insulator in real time and construct a real-time observation dataset based on the multidimensional observation data; The model building module is used to perform point cloud registration on the real-time observation dataset using a semantic segmentation model, generate a target mesh model based on the point cloud registration results, and optimize the target mesh model through a reinforcement learning algorithm to obtain a three-dimensional model of the insulator. The correlation analysis module is used to obtain the electric field anomaly points of the target insulator, and combine them with the three-dimensional model of the insulator. The electric field-defect spatial correlation analysis is performed by the absolute coordinate mapping method based on ORC positioning and spatial geometric calibration to obtain the correlation analysis results. The state assessment module is used to perform multimodal fusion analysis based on the correlation analysis results to obtain the state assessment results of the target insulator.

9. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the insulator condition assessment method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein when the device containing the computer-readable storage medium executes the computer program, it implements the insulator condition assessment method as described in any one of claims 1 to 7.

Images