Intelligent early warning method and system for dangerous source based on visual detection of power transmission line

Abstract

The application discloses a dangerous source intelligent early warning method and system based on power transmission line visual detection, and relates to the technical field of visual detection. The method comprises the following steps: collecting a power transmission line image set; identifying a dangerous source to determine dangerous source position information; combining the position information with wind deflection and dancing simulation to evaluate a risk distance; constructing an adversarial generation sample set and performing multi-scale fusion to determine spatial edge response and point cloud sparsity features, and complete risk grading; evaluating a safety margin and making a multi-level early warning decision in combination with the risk level. The technical problem of low early warning accuracy and untimely response of the dangerous source caused by the difficulty in accurately positioning the dangerous source in a complex dynamic environment and the dependence on static data in the existing power transmission line dangerous source detection is solved, the technical effect of accurately positioning the dangerous source based on three-dimensional reconstruction and back projection, and risk grading and multi-level early warning in combination with wind deflection dynamic simulation is achieved, and the operation safety and emergency response capability of the power transmission line are improved.

Description

Technical Field

[0001] This invention relates to the field of visual inspection technology, and more specifically to a method and system for intelligent early warning of hazards based on visual inspection of power transmission lines. Background Technology

[0002] As the power grid continues to expand, transmission lines, as the main arteries of power transmission, face increasingly complex operating environments. Field lines often traverse forest areas, mining areas, and densely populated construction zones, encountering multiple external risks such as tree growth, large construction machinery operations, and floating debris snagging the lines. These hazards are characterized by dynamic changes and highly random spatial distribution; if they get too close to the conductors, they can easily trigger discharge trips, posing a serious threat to the safe and stable operation of the power grid. Traditional transmission line inspections mainly rely on regular manual patrols or helicopter aerial inspections, assessing hazard risks through visual inspection or simple distance measuring tools. However, in complex field environments, factors such as terrain obstructions and limited field of vision lead to significant deviations in the spatial distance perception of inspection personnel, making it difficult to accurately quantify the actual distance between hazard sources and conductors. While existing image processing-based detection methods can identify hazard categories, most are limited to two-dimensional planar analysis and lack coupling consideration of dynamic conditions such as conductor wind deflection and galloping, and temperature sag changes. They are significantly deficient in terms of three-dimensional spatial positioning accuracy and dynamic risk assessment. The technical root cause lies in the lack of effective modeling capabilities for three-dimensional spatial structures and the failure to integrate the influence of environmental dynamic factors on conductor wind deflection and galloping, resulting in insufficient early warning capabilities for potential hazards and failing to meet the actual needs of intelligent operation and maintenance of transmission lines. Summary of the Invention

[0003] This application provides an intelligent early warning method and system for hazardous sources based on visual inspection of transmission lines. The key is to address the technical obstacles of insufficient spatial positioning of hazardous sources and dynamic risk assessment caused by the characteristics of image data in the complex operating environment of transmission lines, such as multi-view, weak depth constraints, and significant dynamic changes in the environment. By introducing data augmentation algorithms based on deep learning 3D reconstruction and adversarial generative networks, and combining them with a collaborative processing mechanism of back-projection positioning, multi-scale feature fusion, and dynamic simulation of wind deflection galloping, along with the construction of a hazardous source feature database and a risk distance quantification assessment process, we can achieve accurate assessment of the risk distance between hazardous sources and conductors, thereby improving the accuracy, real-time performance, and response capability of early warning.

[0004] The first aspect of this application provides a method for intelligent early warning of hazards based on visual inspection of transmission lines, the method comprising: Panoramic camera equipment is deployed along the transmission line to acquire multi-angle images along the towers while the line is in operation, obtaining a set of transmission line images including conductors and the surrounding environment. Deep learning algorithms are used to identify hazard sources, followed by image stereo vision algorithms to determine the location markers of the hazard sources. Based on the location markers, dynamic simulations are performed using wind drift analysis and environmental data to assess the risk distance characteristics between the hazard sources and the conductors. Risk distance characteristics corresponding to hazard sources of different risk levels are recorded and a hazard source feature database is configured. Based on the transmission line image set and the hazard source feature database, an adversarial generative sample set is constructed. Multi-scale feature fusion is performed on the adversarial generative sample set to determine the spatial edge response characteristics and point cloud sparsity characteristics of the hazard sources, and contact risk is classified to determine the contact risk level. The safety margin under the line's operating state is assessed, and multi-level early warning decisions are made based on the contact risk level.

[0005] A second aspect of this application provides an intelligent early warning system for hazardous sources based on visual inspection of transmission lines, the system comprising: Image Acquisition Module: Panoramic cameras are deployed along the transmission line to acquire multi-angle images along the towers while the line is in operation, obtaining a set of transmission line images including conductors and the surrounding environment; Hazard Source Identification Module: Deep learning algorithms are used to identify hazard sources, and then image stereo vision algorithms are combined to determine the location marker information of the hazard sources; Dynamic Simulation Module: Based on the location marker information, combined with wind deflection galloping analysis and environmental data, dynamic simulation is performed to evaluate the risk distance characteristics between the hazard source and the conductor; Feature Input Module: The risk distance characteristics corresponding to hazard sources of different risk levels are input and a hazard source feature database is configured; Risk Classification Module: Based on the transmission line image set and the hazard source feature database, an adversarial generation sample set is constructed, and multi-scale feature fusion is performed on the adversarial generation sample set to determine the spatial edge response characteristics and point cloud sparsity characteristics of the hazard source, and contact risk classification is performed to determine the contact risk level; Multi-level Early Warning Module: The safety margin under the line's operating state is evaluated, and multi-level early warning decisions are made in combination with the contact risk level.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: First, panoramic cameras are deployed along the transmission line to continuously capture images of the towers and surrounding environment from multiple angles, obtaining complete data on the line's operational scene. Then, deep learning algorithms are used to identify hazard sources, combined with image stereo vision algorithms to determine the location markers of these hazard sources. Next, the distance relationship between the hazard source and the conductor is assessed by combining spatial location information with wind drift and environmental factors. Then, the feature information corresponding to different risk levels is organized to establish a hazard source feature database. Then, adversarial examples are generated using image data and database information, and the spatial structural features and point cloud characteristics of the hazard sources are extracted through multi-scale feature fusion to achieve contact risk classification. Finally, based on the current safety margin and risk level of the line, graded early warning decisions are made, thereby improving the safety monitoring and response capabilities of the transmission line operation. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0008] Figure 1 This is a schematic diagram of the intelligent early warning method for hazardous sources based on visual inspection of power transmission lines provided in an embodiment of this application.

[0009] Figure 2 This is a schematic diagram of the structure of an intelligent early warning system for hazardous sources based on visual inspection of power transmission lines, provided in an embodiment of this application.

[0010] Figure labeling: Image acquisition module 11, Hazard source identification module 12, Dynamic simulation module 13, Feature input module 14, Risk classification module 15, Multi-level early warning module 16. Detailed Implementation

[0011] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0012] Example 1, as Figure 1 As shown, this application provides an intelligent early warning method for hazardous sources based on visual inspection of transmission lines, wherein the method includes: Panoramic camera equipment is deployed along power transmission lines to collect images from multiple angles along the towers while the line is in operation, acquiring a set of images of the power transmission line including the conductors and the surrounding environment.

[0013] In this embodiment, during the actual operation of the transmission line, a panoramic camera with a 360° field of view is fixedly installed on the crossarm of the tower, in the middle of the tower body, or at a key crossing point. This allows the camera to simultaneously cover the conductor's path and the surrounding environment within a certain range. By setting timed or continuous video acquisition modes, high-frequency, multi-angle imaging of the line is achieved without power interruption. During acquisition, pan-tilt-zoom (PTZ) control or multi-lens collaboration can be used to supplementary shooting at different heights and orientations, effectively acquiring image information of conductor sag, insulator posture changes, and potential hazards such as surrounding construction machinery, vegetation, and foreign objects. Simultaneously, the timestamp, equipment location information, and posture parameters of the images are recorded, providing foundational data for subsequent multi-view registration and 3D reconstruction. By summarizing this image information, a transmission line image set with spatial coverage integrity and temporal continuity is formed, providing a reliable data foundation for hazard identification and risk assessment.

[0014] Deep learning algorithms are used to identify hazards, and then image stereo vision algorithms are combined to determine the location marking information of the hazards.

[0015] In one embodiment, after acquiring a set of images of the transmission line, a deep learning algorithm is used for hazard identification and classification. Specifically, the original images are preprocessed with distortion correction, illumination normalization, and image enhancement. Then, an improved YOLOv8 or Faster R-CNN is selected as the target detection framework. Deformable convolutions are introduced into the feature extraction network to enhance the geometric adaptability to targets such as slender conductors and irregular tree obstacles. During the training phase, a dataset labeled with hazards such as vehicles, tree obstacles, and foreign objects is used. Data augmentation is used to improve the model's generalization ability, and a feature pyramid network is used to fuse deep semantic features and shallow detail features to ensure sensitivity to small foreign objects and distant vehicles. The network finally outputs the bounding box, class label, and confidence score for each detected target, retaining detection results with confidence scores higher than a set threshold as input for subsequent processing. Subsequently, a stereo vision algorithm is used to calculate the three-dimensional spatial location of the identified hazards, and then the absolute spatial location of the hazards is determined through back projection. For continuously acquired image sequences, Kalman filtering is used to temporally smooth and predict the spatial location of hazard sources, eliminating location jumps caused by single-frame mismatches. For hazard sources with complex shapes, such as tree obstacles, a point cloud set of their spatial contours is constructed by accumulating point clouds from multiple frames. Finally, hazard source location marking information with category labels, 3D spatial coordinates, spatial dimensions, motion state, and confidence level is output, laying the data foundation for subsequent wind drift simulation and risk distance assessment.

[0016] Based on location marking information, dynamic simulations are performed using wind deflection galloping analysis and environmental data to assess the risk distance characteristics between hazard sources and conductors.

[0017] In one embodiment, based on the results of 3D reconstruction and back-projection positioning, the 3D location marker information of the hazard source in the transmission line scenario is obtained. Subsequently, based on the obtained hazard source location marker information, spatial modeling of the hazard source is performed, extracting its geometric contour parameters and dynamic motion trajectory. These geometric parameters include the shape, size, and boundary features of the hazard source, while the motion trajectory is obtained based on historical data or real-time tracking information, describing the positional changes and movement trends of the hazard source in space. Then, combined with real-time environmental data, the wind deflection angle and galloping amplitude of the conductor under wind load are calculated. The wind deflection angle represents the angle of deviation of the conductor relative to its original position under wind force; the galloping amplitude reflects the amplitude and change of the conductor's vibration under wind force. Afterwards, based on the spatial contour parameters and dynamic motion trajectory of the hazard source, combined with the wind deflection angle and galloping amplitude, the dynamic displacement process of the conductor under wind load is simulated. This process considers the conductor's elasticity, wind intensity, and changes in wind speed and direction, simulating the real-time displacement of the conductor. Simultaneously, using the motion state of the hazard source and the dynamic displacement of the conductor, the minimum spatial distance between the hazard source and the conductor is calculated. Finally, based on the calculated minimum spatial distance, the risk distance characteristic is determined. This risk distance characteristic quantifies the proximity between the hazard source and the conductor and serves as the basis for subsequent risk assessment to determine whether there is a potential contact risk or the possibility of the hazard source approaching the conductor.

[0018] Furthermore, based on location marker information, combined with wind deflection galloping analysis and environmental data, the method also includes dynamic simulation: Based on location marker information, spatial modeling is performed on the hazard source, and the spatial contour parameters and dynamic motion trajectory of the hazard source are extracted. By using real-time wind speed, wind direction, and conductor mechanical parameters, the conductor deflection angle and galloping amplitude under wind load are determined. Based on the spatial contour parameters and dynamic motion trajectory of the hazard source, the conductor deflection angle and galloping amplitude under wind load are used to simulate the dynamic offset process of the conductor under wind load and the real-time motion state of the hazard source, calculate the minimum spatial distance between the hazard source and the conductor, and determine the risk distance characteristics.

[0019] Preferably, when spatially modeling a hazard source based on location marker information, a three-dimensional geometric model of the hazard source is first constructed using its three-dimensional spatial coordinate sequence and spatial contour point cloud. Specifically, the minimum bounding box of the hazard source is used as the basic expression of its spatial contour. The spatial dimensions of the hazard source are accurately described by the length, width, and height parameters of the bounding box. At the same time, the spatial attitude of the hazard source is fitted using multi-frame temporal point cloud data. For hazard sources with regular shapes, such as vehicles, their overall contour and orientation angle are extracted. For hazard sources with irregular shapes, such as tree barriers, point cloud clustering and convex hull algorithms are used to construct their refined spatial contour model. On this basis, based on continuously acquired multi-frame location marker information, the dynamic motion trajectory of the hazard source is extracted through temporal analysis, including the direction of motion, velocity vector, and acceleration characteristics. For mobile hazard sources, such as construction machinery, their motion equations are established to describe the continuous change process of spatial position. For static hazard sources, such as tree barriers, the slow deformation trend of their contour that may occur with growth or external influences is recorded. Subsequently, in calculating the dynamic behavior of the conductor, environmental data was collected using real-time wind speed sensors and wind vanes deployed on the towers. Combined with the conductor's own mechanical parameters, including conductor type, cross-sectional area, mass per unit length, elastic modulus, span, and suspension point height difference, a static spatial model of the conductor was established using the catenary equation. Based on this, wind load calculation formulas from aerodynamics were introduced using real-time wind speed and direction data to determine the wind load per unit length on the conductor, thereby solving for the horizontal offset and wind deflection angle of the conductor under wind load. For galloping phenomena, considering the conductor's torsional stiffness and icing condition, an elliptical trajectory model or the Den Hartog galloping mechanism was used to calculate the conductor's galloping amplitude and frequency, thus obtaining the dynamic offset process of the conductor under wind load and outputting the spatial coordinates of each point on the conductor at any given time.

[0020] Subsequently, based on the spatial contour parameters and dynamic trajectory of the aforementioned hazard source, as well as the wind deflection angle and galloping amplitude of the conductor under wind load, a safe distance calculation model considering the dynamic behavior of both parties is established. In the time dimension, the entire simulation period is sampled at discrete time steps. Within each time step, the spatial curve equation after the conductor's dynamic offset and the spatial position and contour boundary of the hazard source after its movement are calculated. A spatial nearest distance search algorithm is used to traverse the discrete point set of the conductor and the contour point set of the hazard source, calculating the minimum Euclidean distance between each point of the conductor and the contour surface of the hazard source. For hazard sources with complex contours, such as tree barriers, the distance between the hazard source contour point cloud and the discrete points of the conductor is calculated point-by-point, and KD-trees are used to accelerate the nearest neighbor search, improving computational efficiency. After completing the distance calculation for all time steps, the minimum spatial distance within the entire simulation period is extracted as the risk distance feature between the hazard source and the conductor. Simultaneously, the time and location of the occurrence of this minimum distance, as well as the corresponding conductor offset state and hazard source movement state, are recorded, providing accurate quantitative basis for subsequent risk level classification and early warning decisions.

[0021] Furthermore, based on location marker information, combined with wind deflection galloping analysis and environmental data, the method also includes dynamic simulation: If the hazard identification result determines that a hazard exists, a multi-angle synchronized image is determined based on the timestamp corresponding to the hazard; the multi-angle synchronized image is used to perform three-dimensional reconstruction through an image stereo vision algorithm to generate a three-dimensional spatial point cloud model; and a back projection algorithm is used to determine the location marker information of the hazard in the three-dimensional spatial point cloud model.

[0022] Preferably, after the deep learning algorithm identifies hazards in the transmission line images, if a hazard is identified, the system will extract multi-angle synchronized images from the transmission line image set acquired by the panoramic camera equipment based on the timestamp corresponding to the hazard, ensuring sufficient visual information for subsequent 3D reconstruction. Subsequently, the extracted multi-angle synchronized images are preprocessed, including image distortion correction, brightness equalization, noise suppression, time synchronization, and multi-view registration, to eliminate the impact of differences in imaging from different lenses and fluctuations in ambient lighting on the subsequent reconstruction results. Image distortion correction can be achieved through radial and tangential distortion correction models based on camera calibration parameters; brightness equalization can be achieved through histogram equalization or adaptive histogram equalization; noise suppression can be achieved through Gaussian filtering, median filtering, or nonlocal mean denoising; time synchronization can be achieved through time alignment based on a unified timestamp or a hardware-triggered synchronization mechanism; and multi-view registration can be achieved through multi-view geometric correction and extrinsic parameter optimization methods based on feature point matching, such as SIFT / SURF feature matching combined with RANSAC. Simultaneously, the camera's intrinsic and extrinsic parameters corresponding to each image are read. The camera's intrinsic parameters include at least the focal length, principal point coordinates, and distortion coefficients, while the camera's extrinsic parameters include at least the rotation matrix and translation vector of the camera relative to the tower coordinate system or the line coordinate system, thereby establishing a mapping relationship between the image plane and the actual space.

[0023] After preprocessing, a stereo vision algorithm is used to reconstruct the 3D image set of the transmission line. Specifically, firstly, scale-invariant feature transformation or accelerated robust feature extraction algorithms are used to extract a large number of feature points from each image and generate descriptor vectors. Then, an approximate nearest neighbor search algorithm is used to match feature points between images to establish cross-view feature point correspondences. Based on the matching results, epipolar geometry and the five-point method are used to estimate the initial camera pose, and then triangulation is used to calculate the 3D spatial coordinates corresponding to the feature points, forming an initial sparse point cloud. To eliminate accumulated errors, bundle adjustment is used to globally optimize the camera intrinsic and extrinsic parameters and 3D point coordinates, minimizing reprojection errors and obtaining accurate camera parameters and a sparse point cloud structure. On this basis, a multi-view stereo matching algorithm is used for dense reconstruction. Based on the optimized camera parameters, adjacent view images are selected as references for each view image, and then the disparity value of each pixel is calculated based on a semi-global matching algorithm to generate a high-precision depth map. During depth map calculation, erroneous matching points are eliminated through consistency checks and left / right disparity consistency constraints, and median filtering and Gaussian filtering are used to smooth the depth map. Next, the depth maps from all views are back-projected into 3D space using the camera projection matrix, generating local point clouds for each view. An iterative nearest-point algorithm is then used to register and fuse these local point clouds, forming a dense 3D point cloud model covering the entire scene. This 3D point cloud model contains spatial geometric information about conductors, towers, and the surrounding environment. The point cloud density can be adaptively adjusted according to the hazardous area, with denser sampling in key areas where hazardous sources may occur, ensuring that the geometric details of critical areas are fully preserved.

[0024] To facilitate subsequent hazard location, the 3D point cloud model undergoes coordinate unification and semantic segmentation. Specifically, the point cloud model is transformed into a unified line coordinate system, with the X-axis extending along the line direction, the Y-axis representing the transverse direction, and the Z-axis representing the elevation direction. Then, a point cloud segmentation network or cluster-based segmentation method is used to classify and label the conductor regions, tower regions, ground regions, vegetation regions, construction machinery regions, and foreign object regions in the point cloud, obtaining point cloud subsets corresponding to different categories of targets. For point cloud subsets identified as hazard sources, their bounding boxes, centroid coordinates, top elevation, edge contours, and volume range can be further extracted as the basis data for subsequent location marking.

[0025] After obtaining the 3D spatial point cloud model, a target detection network or instance segmentation network is used to identify the hazard source region in the 2D image, obtaining the pixel-level region mask, bounding box, or key point position of the hazard source in the images from various viewpoints. Then, based on the image coordinates of each pixel within the hazard source region and combined with the camera's intrinsic parameters, the corresponding imaging ray direction is obtained. Next, combined with the camera's extrinsic parameters, the imaging ray is transformed into the line coordinate system, forming a spatial ray originating from the camera's optical center and passing through the target pixel. This spatial ray is then intersected with the aforementioned 3D spatial point cloud model through nearest neighbor matching or ray casting search to determine the 3D point cloud set corresponding to that pixel. For the same hazard source observed simultaneously from multiple viewpoints, the 3D point sets obtained by backprojection from different viewpoints can be cross-validated and fused to remove outliers and retain point cloud regions with high spatial consistency, thereby improving positioning accuracy. Then, location marker information is extracted from the fused hazard source point cloud region. This location marker information includes at least the hazard source centroid coordinates, the coordinates of the highest point, the coordinates of the bottom grounding point, the hazard source edge contour point set, its length, width, and height dimensions in three-dimensional space, and its orientation relative to the conductor and tower. The centroid coordinates can be obtained by averaging the three-dimensional coordinates of the hazard source point cloud point set; the coordinates of the highest point can be obtained by searching for the maximum value point along the Z-axis; and the edge contour point set can be extracted using the convex hull algorithm, α-shape algorithm, or boundary tracking algorithm. Finally, the location marker information of the hazard source in the three-dimensional point cloud model is output and unified with the three-dimensional trajectory of the conductor to the same spatial coordinate system. This provides basic data support for subsequent risk distance calculations, dynamic simulations, and contact risk classification between the hazard source and the conductor, thereby achieving accurate reconstruction and highly reliable calibration of the hazard source's spatial location.

[0026] Enter the risk distance characteristics corresponding to hazard sources of different risk levels and configure the hazard source characteristic database.

[0027] In one embodiment, during historical inspections, the identified hazards are first classified by risk level, and their actual distances to conductors under different operating conditions and their changes are statistically analyzed. For each risk level, corresponding risk distance features are extracted, including key parameters such as minimum safe clearance, dynamic approach distance, and distance change trends. These risk distance features are then structured and combined with hazard type and size range to construct a unified hazard feature vector. A mapping relationship is then established between different risk levels and their corresponding risk distance features, forming standardized hazard feature entries. These feature entries are then categorized and stored according to hazard type and risk level, and an indexing mechanism is established to support rapid retrieval and matching, thereby constructing a hazard feature database. This database is used for comparative analysis and risk assessment of newly emerging hazards during subsequent identification and evaluation processes, improving the accuracy and consistency of risk assessment.

[0028] Based on the transmission line image set and the hazard source feature database, an adversarial generation sample set is constructed. Multi-scale feature fusion is performed on the adversarial generation sample set to determine the spatial edge response features and point cloud sparsity features of the hazard source, and contact risk classification is performed to determine the contact risk level.

[0029] In one embodiment, after acquiring a set of transmission line images, potential hazard areas in the images are first preliminarily identified and their features extracted. Geometric contours, texture distribution, and depth-related features are extracted from these images. These features are then matched against a pre-built hazard feature database to determine the categories of potential target hazards in the current scene and their corresponding features. Based on this, a conditional adversarial generative network (GAN) is used to augment the identified hazard samples. By introducing different meteorological conditions, viewpoint changes, and scale perturbations, a set of adversarial generative samples containing diverse scene features is generated to improve the model's adaptability to complex environments and edge cases. Subsequently, the adversarial generative sample is input into the feature extraction backbone network. Shallow convolutional units extract edge, corner, and texture details of the target; mid-layer networks extract local shape and contour continuity features; and deep networks extract semantic features of hazard categories, spatial occupancy, and scene association features with the conductor. Then, using feature pyramid structures, skip connection mechanisms, or attention fusion modules, feature maps at different scales are aligned, weighted, and fused. This allows the model to identify local edge information of small, distant hazard sources while preserving the overall spatial structure information of large construction machinery or tall vegetation, thus forming a unified multimodal representation of hazard sources. Next, for identified potential hazard sources, a contact risk assessment index system is constructed by combining multidimensional spatial parameters, spatial edge response features, and point cloud sparsity features. By weighting different features, the proximity and potential contact probability between the hazard source and the conductor are assessed. Based on a preset risk grading standard, it is classified into different contact risk levels, providing a reliable basis for subsequent multi-level early warning and improving the accuracy and robustness of transmission line safety monitoring.

[0030] Furthermore, based on the transmission line image set and combined with the hazard source feature database, an adversarial generative sample set is constructed, the method comprising: Risk distance features corresponding to different risk levels of hazard sources from historical inspections are collected, hazard source feature vectors are configured, and a three-level classification and storage is performed against the size, height, and dynamic behavior of construction machinery and vegetation to obtain the hazard source feature database. The three-level classification and storage adopts a distributed storage architecture. Geometric contour features, texture distribution features, and depth response features are extracted from the transmission line image set. Based on the geometric contour features, texture distribution features, and depth response features, comparative analysis is performed in the hazard source feature database to determine the target hazard source features. The target hazard source features are data augmented through a conditional adversarial generative network to construct an adversarial generative sample set including those without hazard markers or those with hazard source markers.

[0031] Optionally, when collecting risk distance features corresponding to hazard sources of different risk levels from historical inspections, the system first extracts a large number of inspection records from the historical inspection database, including manual inspection reports, drone inspection images, and tripping accident analysis reports. For each historical record, hazard source cases with clearly marked risk levels are selected. Risk levels are typically divided into Level 1, Level 2, and Level 3 risks, corresponding to the historical measured minimum distance between the hazard source and the conductor, accident occurrence, and emergency response records, respectively. For each case, its risk distance features are extracted, including the static minimum spatial distance between the hazard source and the conductor, the dynamic minimum spatial distance considering wind deflection, the spatial size parameters of the hazard source, the type of hazard source, and the working environment data at that time. Based on this, each hazard source case is represented as a multi-dimensional feature vector, with feature dimensions including the geometric size features, spatial location features, dynamic behavior features, environmental association features, and historical risk level of the hazard source. The feature vector is normalized to eliminate the influence of dimensions, ensuring the accuracy of subsequent similarity calculations. Subsequently, it is stored in a three-level classification according to different hazard source categories such as construction machinery and vegetation. The primary classification is based on the type of hazard source, divided into major categories such as construction machinery, vegetation, and foreign objects. Construction machinery is further subdivided into cranes, tower cranes, excavators, etc.; vegetation into trees, shrubs, bamboo, etc.; and foreign objects into kites, balloons, plastic film, etc. The secondary classification is based on the spatial dimensions and height of the hazard source. For example, construction machinery is classified into large, medium, and small machinery according to boom length and operating height; vegetation is classified into tall trees, medium trees, and low shrubs according to tree height and crown width. The tertiary classification is based on the dynamic behavior characteristics of the hazard source. Construction machinery is classified into stationary operation type, moving approach type, and rotating boom type according to movement mode; vegetation is classified into fast-growing type, stable type, and wind-prone type according to growth and swaying characteristics. Each primary classification corresponds to different risk distance feature templates and warning threshold configurations. The storage architecture adopts a distributed storage architecture, deploying the hazard source feature database in a distributed file system or distributed database cluster. The system shards data based on the geographical region attributes of hazard sources, distributing hazard source feature data from different transmission line sections across multiple storage nodes to achieve load balancing and parallel access. For frequently accessed "hot" data, such as feature vectors of high-risk hazard sources recently discovered during inspections, it is stored in a memory cache layer to improve retrieval efficiency; for "cold" data, such as previously processed low-risk hazard source features, it is stored on low-cost, high-capacity storage nodes. At the indexing level, a distributed vector index is established for hazard source feature vectors, employing approximate nearest neighbor retrieval algorithms based on locality-sensitive hashing or hierarchical navigable small-world graphs. This ensures that when faced with newly collected hazard source features, similar cases can be quickly retrieved within the distributed storage architecture, achieving efficient feature matching and risk level reference.

[0032] Next, candidate target regions are screened from the acquired transmission line image set. Regions near conductors, tower perimeters, line corridors, and protruding ground features are separated from the images as the initial analysis scope for hazard source identification. Multidimensional feature extraction is then performed on each candidate region to obtain geometric contour features, texture distribution features, and depth response features. Geometric contour features characterize the target's shape and structure, and can be extracted through edge detection, region segmentation, or convolutional neural network feature encoding. Taking the Canny edge detection and contour analysis method as an example, the candidate regions are first grayscaled and smoothed using Gaussian filtering for noise reduction. Then, the Canny edge detection algorithm is used to extract the target edges, resulting in a binary edge map. Based on the edge map, connected component analysis and contour tracking algorithms are used to extract the complete contour curve. The contour is then approximated as a polygon, and its minimum bounding rectangle and convex hull region are calculated. The aspect ratio, orientation angle, area, and perimeter are calculated based on the bounding rectangle. The degree of contour expansion is calculated based on the convex hull. Simultaneously, the top contour undulation features are obtained by analyzing the height changes of points on the contour. By judging the elongation ratio of the contour, the extensibility features of the slender structure are extracted, thereby distinguishing the robotic arm, the wire-like foreign object, and the blocky target.

[0033] Texture distribution features are used to characterize the texture density, grayscale variation patterns, and local repetitive structures of a target surface. These features can be obtained through local binary mode, Gabor filtering, and gray-level co-occurrence matrix (GLCM). Taking GLCM as an example, candidate regions are first converted to grayscale images, and grayscale compression is performed to reduce computational complexity. A GLCM is constructed in multiple directions with a fixed pixel spacing. Then, multiple statistical features, including contrast, energy, entropy, and correlation, are calculated based on the GLCM. By averaging or weighting the statistical results from different directions, a unified texture feature is obtained to distinguish between highly regular textures (such as mechanical metal surfaces), highly random textures (such as foliage and vegetation), and low-texture regions (such as the sky background).

[0034] Depth response features are used to characterize the spatial hierarchy and relative distance changes of a target. These features can be obtained through multi-view depth estimation networks, disparity calculation, or 3D reconstruction results. Taking a multi-view depth estimation network as an example, the multi-view images corresponding to candidate regions are first input into a pre-trained depth estimation network to obtain pixel-level depth maps. Then, depth values ​​within the candidate regions are extracted to form depth sub-regions. Within these sub-regions, the mean depth, variance depth, and gradient depth are calculated. Simultaneously, depth boundary features are obtained by detecting abrupt depth changes to identify the spatial boundary between the target and the background or guide wires. Furthermore, depth continuity is statistically analyzed to distinguish between continuous entities and discrete structures, ultimately forming depth features that reflect the spatial proximity of the target relative to guide wires, towers, and the ground.

[0035] After completing the above feature extraction, geometric contour features, texture distribution features, and depth response features are encoded according to a unified dimension to form the matching feature representation of the candidate target. Specifically, geometric contour features are encoded as structural description vectors, texture distribution features as texture response vectors, and depth response features as spatial distance vectors. These are then combined through concatenation, weighted splicing, or feature fusion networks to form a comprehensive feature representation. Next, the comprehensive feature representation is input into a hazard source feature database for comparative analysis. Specifically, feature vectors, historical scene feature templates, and risk level association features of different categories of hazard sources are retrieved from the database, and the matching degree between the current candidate target and various hazard source samples is calculated using Euclidean distance or cosine similarity. During the matching process, a coarse matching can be performed first according to the major hazard source categories, classifying candidate targets into construction machinery, vegetation, foreign objects, or non-hazardous background categories. Then, a fine-grained comparison is performed within each corresponding category to determine the set of hazard source features closest to the current candidate target. Finally, the set of features with the highest matching degree to the current candidate target is selected from the hazard source feature database as the target hazard source features for subsequent training sample construction and risk identification.

[0036] After determining the characteristics of the target hazard source, to improve the model's adaptability to complex, low-sample, and ambiguous boundary scenarios, the target hazard source characteristics are augmented using a conditional adversarial generative network (GAN). Specifically, a GAN consisting of a generator and a discriminator is constructed. The generator uses the target hazard source characteristics as the core conditional input and further combines geometric contour features, texture distribution features, depth response features, scene background information, meteorological condition labels, and risk category labels to generate augmented samples. Augmented samples can be synthetic images containing hazard sources, or corresponding depth maps, contour maps, or pseudo-point cloud projections synchronized with images. The discriminator is used to determine the authenticity and category consistency of the input samples, that is, to determine whether the sample is a real sample or a generated sample, and to determine whether the category, contour shape, and spatial depth distribution of the hazard source in the sample are consistent with the conditional input. Through adversarial training between the generator and the discriminator, the generated samples gradually approximate the hazard source distribution characteristics in real transmission line scenarios. For vegetation-related hazards, samples showing canopy expansion, branch and leaf swaying, and changes in shading should be emphasized; for construction machinery-related hazards, samples showing boom extension, angle changes, and small targets at long distances should be emphasized; for foreign object-related hazards, samples showing slender hanging objects, irregular shapes, and mixed backgrounds should be emphasized.

[0037] After completing the adversarial generation (ADG) process, an ADG sample set is constructed. This set consists of two parts: one part comprises unmarked samples (images containing only normal scenes such as conductors, towers, ground, sky, and vegetation, excluding hazardous targets that could threaten the power line), used as base samples for model learning of normal power line scenarios; the other part comprises samples with hazard markers (images clearly labeled with hazard categories, bounding boxes, pixel masks, risk locations, or spatially proximate areas), used as positive samples for model learning hazard identification and localization. Both unmarked and hazard-marked samples include real-world samples and synthetic samples enhanced by the conditional adversarial generative network. Ultimately, this results in an ADG sample set that combines scene diversity, category balance, and risk representativeness, providing ample training and validation data for subsequent multi-scale feature fusion, hazard identification modeling, and contact risk classification.

[0038] Furthermore, the target hazard source characteristics are augmented using a conditional adversarial generative network, the method comprising: Based on the target hazard source characteristics, a conditional adversarial generative network is constructed using geometric contour features, texture distribution features, and depth response features as conditional inputs, based on hazard source category and meteorological condition label; during the training process of the adversarial generative network, Wasserstein distance and gradient penalty terms are introduced.

[0039] Optionally, the characteristics of the identified target hazards are first uniformly encoded, transforming the geometric contour features, texture distribution features, and depth response features in the target hazard features into fixed-dimensional conditional feature vectors. At the same time, the hazard category labels and meteorological condition labels are discretely encoded or embedded. The hazard category labels include at least construction machinery, vegetation, floating foreign objects, and other external intrusions; the meteorological condition labels include at least sunny, cloudy, windy, rainy, foggy, daytime, and nighttime scene labels. Subsequently, the aforementioned feature vectors are concatenated with the label embedding results to form a conditional input vector, which drives the conditional adversarial generative network to generate enhanced samples that match the specified hazard category and meteorological conditions. This conditional adversarial generative network includes a generator, a discriminator, and a conditional embedding module. The conditional embedding module maps geometric contour features, texture distribution features, depth response features, hazard category labels, and meteorological condition labels to a unified feature space and outputs a conditional control vector. The generator receives a random noise vector and a conditional control vector and outputs hazard enhancement samples that meet the conditional constraints. The discriminator receives real samples or generated samples and the same conditional control vector, scores the authenticity of the input samples, and determines whether the sample is consistent with the input conditions.

[0040] The generator can employ an encoder-decoder backbone structure or a generation structure with progressive upsampling. Taking the encoder-decoder structure as an example, random noise vectors are concatenated with conditional control vectors, mapped to a low-resolution high-dimensional feature tensor via a fully connected layer, and then spatial resolution is gradually restored through multiple cascaded deconvolutional layers, upsampling convolutional layers, or residual blocks. A conditional modulation mechanism is introduced during the feature restoration process at each level, allowing hazard category and meteorological condition information to continuously participate in control at different scale feature layers. This conditional modulation mechanism can be implemented using conditional batch normalization, feature linear modulation, or channel attention gating, so that the generator retains the structural features of the target hazard while adapting to the corresponding environmental representation when outputting samples. For cases where scene background needs to be preserved, a U-Net-type skip connection structure can be introduced into the generator to directly transmit shallow edge texture information to the decoder, thereby enhancing the reconstruction capability of wires, towers, sky background, and hazard boundary details. The generator's input conditions can be functionally categorized into structural conditions, appearance conditions, and spatial conditions. Structural conditions, comprised of geometric contour features and hazard category labels, constrain the shape of hazard sources in the generated samples. Appearance conditions, comprised of texture distribution features and meteorological condition labels, constrain brightness variations, surface texture density, shadow states, and visibility variations in the generated samples. Spatial conditions, comprised of depth response features, constrain the spatial hierarchy between hazard sources and conductors, towers, and the background. To enhance the coupled modeling capability between multiple conditions, a multi-branch input structure can be set in the generator. The first branch extracts structural condition features, the second branch extracts texture and meteorological coupling features, and the third branch extracts depth spatial features. The outputs of each branch are concatenated or attention-weighted by an intermediate fusion unit before being fed into the main generation path to complete sample generation. This intermediate fusion unit can employ channel attention units, spatial attention units, or cross-attention modules to improve the information coordination capability between different modal conditions.

[0041] The discriminator can employ a multi-layer convolutional discriminant structure or a dual-branch structure combining local block discrimination and global discrimination. Specifically, the discriminator performs two checks: firstly, it determines the authenticity of the input sample, outputting a score indicating that the sample belongs to the real sample distribution; secondly, it determines the consistency between the input sample and the conditional control vector to prevent generated samples from appearing realistic but not matching the specified hazard category, geometric contour, or meteorological conditions. The discriminator typically includes several layers of stride convolutional layers, normalization layers, and non-linear activation layers, with the discrimination result output at the network's end through a global pooling layer and a fully connected layer. For conditional consistency discrimination, the conditional control vector can be concatenated with the discriminator's intermediate layer features after embedding mapping, or the conditional information can be projected onto the discrimination result space using projection discrimination, thereby strengthening the conditional constraint effect.

[0042] During network training, the Wasserstein distance is introduced as the optimization objective between the generated and real distributions. In this process, the discriminator no longer outputs traditional binary classification probabilities, but instead outputs a continuous score indicating the realism of the input samples. This score approximates the Wasserstein distance between the real and generated sample distributions. The generator's training objective is to make the generated samples as close as possible to the real samples under the discriminator's score, while the discriminator's training objective is to maximize the score difference between the real and generated samples. Compared to the Jensen-Shannon divergence commonly used in traditional GANs, the Wasserstein distance can provide stable gradients even when the real and generated distributions have little overlap, thus reducing mode collapse and gradient vanishing problems during training. It is more suitable for complex scenarios in power transmission lines where the distribution of hazard source samples is unbalanced, the class differences are large, and the background changes are complex. To satisfy the Wasserstein distance constraint on the discriminator's 1-Lipschitz continuity, a gradient penalty term is introduced during training. This allows for linear interpolation between real and generated samples in each iteration, constructing interpolated samples. These interpolated samples are then input into the discriminator, and the gradient norm of the discriminator output relative to the interpolated sample input is calculated. The degree to which this gradient norm deviates from 1 is added as a penalty value to the discriminator's loss function. This gradient penalty term is used to suppress the discriminator function from changing too rapidly, avoiding the insufficient expressive power problem caused by traditional weight pruning methods. Thus, the discriminator maintains stronger discriminative performance while ensuring the Lipschitz constraint.

[0043] In the specific training process, image samples with hazard category labels and meteorological condition labels, along with their corresponding geometric contour features, texture distribution features, and depth response features, are first read from the real sample set and used as real conditional samples to input into the discriminator. Simultaneously, the generator receives random noise and inputs with the same conditions to generate corresponding hazard-enhanced samples, which are then input into the discriminator for scoring. Next, the discriminator parameters are updated based on the real sample scores, generated sample scores, and gradient penalty terms. After several discriminator updates, the discriminator parameters are fixed, and the generator parameters are updated in reverse using the discriminator's scoring results for the generated samples. Through multiple rounds of iterative training, the generator is able to gradually generate enhanced samples that closely resemble the real sample distribution in terms of hazard morphology, texture representation, depth relationships, and meteorological adaptability. Ultimately, the trained conditional adversarial generative network can generate enhanced hazard samples that are highly consistent with real transmission line scenarios, given the target hazard features, hazard category labels, and meteorological condition labels. These enhanced samples are used to expand the training dataset of subsequent hazard identification models, thereby improving the model's recognition accuracy, generalization ability, and risk classification stability in complex weather, complex backgrounds, low sample counts, and ambiguous boundary scenarios.

[0044] Furthermore, the method for determining the spatial edge response characteristics and point cloud sparsity characteristics of a hazard source, and performing contact risk classification to determine the contact risk level, includes: For identified potential hazards, a risk distance quantification assessment unit is set up by combining multi-dimensional spatial parameters, spatial edge response characteristics, and point cloud sparsity characteristics. The multi-dimensional spatial parameters include the top elevation of the hazard, horizontal offset, and dynamic approach rate. In the risk distance quantification assessment unit, a weighting coefficient is introduced for weighted correction. The weighting coefficient is determined based on the degree of influence of each parameter on the conductor discharge risk in historical tripping accidents. Using the risk distance quantification assessment unit, combined with the minimum allowable safe distance standard for conductors, the contact risk level is output.

[0045] Optionally, firstly, three-dimensional spatial parameters are extracted from the identified potential hazards. These multi-dimensional spatial parameters include the hazard's top elevation, horizontal offset, and dynamic approach rate. The hazard's top elevation characterizes the vertical height of its highest point relative to the ground reference plane or the origin of the tower coordinate system; the horizontal offset characterizes the degree of offset of the hazard relative to the centerline of the conductor projection or the tower centerline in the transverse direction; and the dynamic approach rate characterizes the speed at which the hazard moves towards the conductor in continuous time. Simultaneously, spatial edge response features and point cloud sparsity features of the potential hazards are extracted. Spatial edge response features characterize the salience and abrupt changes of the hazard's boundary contour in space; and point cloud sparsity features characterize the density and structural integrity of the hazard's point cloud distribution in three-dimensional space. Subsequently, the multi-dimensional spatial parameters, spatial edge response features, and point cloud sparsity features are input into a risk distance quantification assessment unit, which employs a weighted calculation model. Specifically, the basic risk distance value is first calculated based on the spatial relationship between the hazard source and the conductor. This basic risk distance value includes the vertical clearance between the top of the hazard source and the lowest point of the conductor, the minimum horizontal clearance between the edge of the hazard source and the outer edge of the conductor, and the minimum three-dimensional Euclidean distance between the hazard source and the conductor. Then, the top elevation of the hazard source, the horizontal offset, the dynamic approach rate, the spatial edge response characteristics, and the point cloud sparsity characteristics are normalized and converted into comparable dimensionless parameters. Among them, the closer the top elevation is to the conductor elevation, the greater its contribution to the risk; the smaller the horizontal offset, the closer the hazard source is to the conductor projection area, and the greater its contribution to the risk; the greater the dynamic approach rate, the more obvious the trend of the hazard source continuously approaching the conductor, and the greater its contribution to the risk; the stronger the spatial edge response, the clearer the boundary of the hazard source and the more obvious the protruding parts, and the stronger the characterization ability of partial discharge approach risk; the higher the point cloud sparsity or the more obvious the local void ratio, the more likely the hazard source has slender protruding parts or incomplete observation areas, which need to be adversely corrected during risk assessment. Based on this, each feature can be mapped to a corresponding risk sub-value according to a preset function, and together with the basic risk distance, they constitute the comprehensive risk calculation input.

[0046] In the risk distance quantification assessment unit, a weighting coefficient is introduced for weighted correction. This weighting coefficient is determined based on the influence of each parameter on the conductor discharge risk in historical trip accident samples. Specifically, data on hazard source type, minimum distance between the hazard source and the conductor at the time of the accident, top elevation difference, horizontal offset, approach velocity, environmental conditions, discharge location, and accident consequence level from historical trip accidents are first collected. Correlation analysis, regression analysis, or entropy weighting methods are used to statistically analyze the above data to determine the contribution of each parameter to the discharge risk, and corresponding weighting coefficients are generated accordingly. Then, each normalized risk sub-value is multiplied by its corresponding weight, and weighted summation or weighted nonlinear fusion is performed to obtain a comprehensive risk score. This comprehensive risk score can be represented as the final quantified result of the basic risk distance assessment after correction by various risk factors. This ensures that the risk assessment considers not only the current geometric distance but also the comprehensive impact of hazard source height characteristics, lateral offset state, movement trend, and structural boundary integrity on conductor discharge risk. After weighted correction, the contact risk level is output using the risk distance quantification assessment unit in conjunction with the minimum permissible safe distance standard for conductors. Specifically, the minimum permissible safe distance standard for conductors under the corresponding voltage level, line type, and operating conditions is read. This minimum permissible safe distance standard serves as a risk assessment threshold. The equivalent risk distance corresponding to the comprehensive risk score is then compared with the minimum permissible safe distance standard. When the equivalent risk distance is greater than the safe distance threshold and maintains a significant safety margin, it is classified as a low contact risk level. When the equivalent risk distance is close to the safe distance threshold, or although it has not reached a dangerous contact state but shows a clear dynamic approach trend, it is classified as a medium contact risk level. When the equivalent risk distance is less than or equal to the minimum permissible safe distance standard, it is classified as a high contact risk level. Finally, this contact risk level is output and used as the basis for subsequent multi-level early warning decisions. This enables quantitative calculation and classification of the risk of potential hazards approaching conductors, improving the accuracy, objectivity, and timeliness of risk identification and early warning.

[0047] Furthermore, the method also includes: The feature descriptors of the hazards are compared with the feature vectors of the hazards in the hazard feature database using cosine similarity analysis to obtain the feature matching degree. A hazard confidence matrix is ​​constructed based on the feature matching degree, and the matrix elements of the hazard confidence matrix represent the matching confidence of hazards of different risk levels in the current transmission line scenario. Threshold segmentation is performed on the hazard confidence matrix to determine the potential hazards, which are hazards with a matching confidence degree higher than the confidence threshold.

[0048] Optionally, based on the multi-scale feature fusion results, hazard source feature descriptors are first extracted from candidate hazard source regions in the current transmission line image set. These hazard source feature descriptors are multi-dimensional feature vectors that uniformly represent the current candidate targets, comprehensively reflecting information such as the spatial edge response, point cloud sparsity, geometric structure, texture distribution, and depth level of the candidate target. Subsequently, hazard source feature vectors corresponding to different hazard source categories and risk levels are retrieved from the hazard source feature database. These hazard source feature vectors are standardized vectors constructed based on historical inspection samples, corresponding to different types of hazard sources such as construction machinery, vegetation, and foreign objects, and further distinguishing between low-risk, medium-risk, and high-risk levels. Next, cosine similarity analysis is performed between the hazard source feature descriptors corresponding to the current candidate targets and each set of hazard source feature vectors in the database. The cosine value of the angle between the hazard source feature descriptors and each hazard source feature vector is calculated as the feature matching degree, measuring the degree of directional consistency between the two in the feature space. When the cosine value is closer to 1, it indicates that the current candidate target is more similar to the corresponding database sample in terms of feature composition, and the matching degree is higher. When the cosine value is closer to 0 or a lower value, it indicates that the current candidate target is significantly different from the corresponding database sample, and the matching degree is lower. After obtaining the feature matching degrees of each group, a hazard source confidence matrix is ​​constructed based on these feature matching degrees. This hazard source confidence matrix can be constructed in a two-dimensional manner according to the hazard source category-risk level. In this matrix, the rows are used to represent different hazard source categories, the columns are used to represent different risk levels, and each element in the matrix is ​​used to represent the matching confidence between a certain category and a certain risk level of hazard source in the current transmission line scenario and the current candidate target. Then, this hazard source confidence matrix is ​​analyzed, and the elements with higher confidence in the matrix are selected as key discrimination objects. Then, the hazard source confidence matrix is ​​subjected to threshold segmentation processing, that is, each element in the matrix is ​​compared with a preset confidence threshold. When the matching confidence level corresponding to a matrix element is higher than the confidence level threshold, a valid matching relationship is determined between the current candidate target and the hazard source of that category and risk level. When the matrix element is lower than the confidence level threshold, the matching relationship is considered invalid or lacks sufficient confidence. After threshold segmentation, potential hazards are identified. These potential hazards are candidate targets where at least one matrix element in the hazard source confidence matrix is ​​higher than the confidence level threshold. For the same candidate target, if multiple matrix elements are higher than the threshold, the hazard source category and risk level corresponding to the matrix element with the highest confidence level can be further selected as the final matching result for that candidate target. Thus, candidate targets in the current transmission line scenario can be screened as potential hazards with clear hazard attributes and risk level tendencies.Finally, the system outputs the potential hazards and their corresponding hazard categories and risk levels, providing basic input for subsequent risk distance quantitative assessment, contact risk classification, and multi-level early warning decision-making, thereby achieving highly reliable screening and accurate identification of hazards in power transmission line scenarios.

[0049] Furthermore, the method further includes performing cosine similarity analysis between the hazard source feature descriptors and the hazard source feature vectors in the hazard source feature database. Based on the multi-scale feature fusion results, a three-dimensional graph structure corresponding to various hazard source regions is constructed. The nodes of the three-dimensional graph structure represent the local point cloud sparsity and local spatial edge response, and the edge weights of the three-dimensional graph structure are determined by spatial proximity and semantic similarity. A global pooling operation is performed on the three-dimensional graph structure to generate hazard source feature descriptors.

[0050] Optionally, the multi-scale feature fusion results obtained in the previous steps are first used to divide the hazard source regions, separating the 3D regions corresponding to construction machinery, vegetation, floating foreign objects, and other potential hazards from the transmission line scene. For each type of hazard source region, its corresponding 3D point cloud is locally divided according to a preset neighborhood rule, forming multiple local point set units. Each local point set unit corresponds to a graph node. This division can be performed using voxel partitioning, K-nearest neighbor clustering, radius neighborhood clustering, or superpoint segmentation, so that each node corresponds to a spatial sub-region in the local structure of the hazard source. For each node, the local point cloud sparsity and local spatial edge response are extracted as node attributes. Texture features, depth response features, or category semantics corresponding to the local region can also be embedded and written into the node feature vector to enhance the node's ability to represent the local structure and semantics of the hazard source. After completing node construction, if two nodes satisfy a preset spatial adjacency condition in 3D space, a graph edge is established between them. This spatial adjacency condition can be achieved using Euclidean distance threshold determination, K-nearest neighbor connection, or radius neighborhood connection. For each graph edge, an edge weight is assigned, determined jointly by spatial proximity and semantic similarity. Spatial proximity characterizes the closeness of two nodes in 3D space, and can be determined based on the Euclidean distance between the centroids of the nodes, the angle between local normal vectors, or the overlap ratio of their neighborhoods. The closer the distance and the more consistent the directions, the higher the spatial proximity. Semantic similarity characterizes the consistency of the category or structure of two nodes in the fused feature space, and can be obtained by comparing the similarity between the texture features, depth features, edge features, or point cloud density features of the two nodes. Spatial proximity and semantic similarity are then weighted and fused according to a preset ratio to obtain the final edge weight. This ensures that the graph structure not only reflects the spatial connectivity between local regions of a hazard source but also reflects its semantic consistency. Thus, the constructed 3D graph structure can characterize the local composition, edge variation patterns, and overall structural continuity of a hazard source in 3D space.

[0051] After constructing the 3D graph structure, graph node features are propagated and updated based on graph neural networks. Specifically, the local point cloud sparsity, local spatial edge response, and additional semantic features of each node are used as initial node features and input into graph convolutional layers, graph attention layers, or message passing networks. Through information interaction between nodes and their neighboring nodes, the feature representation of each node is iteratively updated. During message passing, the influence of neighboring nodes on the current node is controlled by edge weights. The larger the edge weight, the closer the corresponding neighboring nodes are spatially and the more related they are semantically, and the higher the proportion of information transmitted. After several layers of graph feature updates, each node not only retains the structural information of its own local region but also integrates the contextual information of its surrounding neighborhood. This allows the node features to more completely represent the local structure and overall topological relationship of the hazard source. For example, for slender continuous targets such as construction machinery booms, the graph structure can connect multiple local nodes along the spatial extension direction; for vegetation targets, the graph structure can reflect the spatial characteristics of scattered leaf edges and uneven canopy density; for floating foreign objects, the graph structure can highlight their irregular edges, sparse point clouds, and local isolation.

[0052] After obtaining the updated 3D graph structure features, a global pooling operation is performed on the 3D graph structure. The global pooling operation is used to compress and aggregate node-level features into a graph-level feature representation, so that the entire hazard source area corresponds to a unified high-dimensional feature description vector. This pooling operation can be implemented by global average pooling, global max pooling, weighted attention pooling, or a combination of multiple pooling methods. Among them, global average pooling is used to reflect the average structural features of the overall hazard source; global max pooling is used to highlight the most significant edge response, the highest sparsity region, or the strongest local structural features; attention pooling can adaptively allocate aggregation weights according to the importance of nodes, so that node regions that are more sensitive to contact risks occupy a larger proportion in the final descriptor. For example, for nodes at the end of the robotic arm near the conductor, nodes at the upper edge of the vegetation canopy, or nodes at the edge of foreign objects, attention weights can be used to increase their contribution to the global description. Through the aforementioned global pooling operation, node information, topological relationships, and key local features in the entire 3D graph structure are compressed and integrated to generate a hazard source feature descriptor. This hazard source feature descriptor is a global abstract expression of the current hazard source region, including the shape distribution features, local edge variation features, point cloud sparsity distribution features, and topological connectivity features of the hazard source in 3D space. The generated hazard source feature descriptor can serve as the basic input for subsequent hazard source feature matching, similarity analysis, confidence matrix construction, and potential hazard source screening. It can effectively improve the ability to express the complex 3D structure and local boundary differences of hazard sources, thereby improving the accuracy of hazard source identification and the stability of feature matching.

[0053] Assess the safety margin under the operating conditions of the line and make multi-level early warning decisions based on the contact risk level.

[0054] In one embodiment, after determining the contact risk level of a hazardous source, the actual minimum distance and predicted minimum distance between the hazardous source and the conductor are compared with the minimum allowable safe distance at the corresponding voltage level based on the obtained risk distance characteristics to calculate the current safety margin value. After obtaining the safety margin value, it is fused with the contact risk level for determination. When the safety margin is in the sufficient range and the contact risk level is low, normal operation status or prompt information is output; when the safety margin is close to the critical range or the contact risk level increases to medium risk, a level one warning is triggered to alert relevant personnel; when the safety margin enters the hazardous range or the contact risk level reaches high risk, a level two warning is triggered, and on-site intervention measures are recommended; when the safety margin is extremely low and the contact risk level reaches extremely high risk, the highest level warning is triggered, and if necessary, the dispatch system is linked to take load limiting, power outage, or emergency response measures. Simultaneously, the early warning results can be correlated with the time dimension to smooth out the warning status across multiple consecutive moments, avoiding frequent false alarms caused by instantaneous fluctuations. Furthermore, targeted handling suggestions can be output based on the type of hazard source. For example, low-level warnings are recorded and tracked, medium-level warnings are manually reviewed and inspected, and high-level or emergency warnings trigger on-site handling, remote commands, or coordinated protection measures. Ultimately, through joint analysis of safety margin and contact risk level, graded and dynamic early warning decisions can be achieved, thereby improving the accuracy of transmission line operation risk identification and the timeliness of early warning response.

[0055] Furthermore, the method for assessing the safety margin under the operating conditions of the line includes: The safety margin is determined based on the relative magnitude of the risk distance feature and the safety distance threshold, and is used to quantify the current safety margin.

[0056] Preferably, the risk distance characteristics between the hazard source and the conductor are first obtained based on the risk assessment results. Then, the safety distance threshold corresponding to the transmission line is obtained. This safety distance threshold is pre-set based on the line voltage level, operating specifications, and environmental conditions, and is used to limit the minimum permissible safe distance between the conductor and external targets. Subsequently, the risk distance characteristics and the safety distance threshold are compared and calculated. That is, by calculating the difference between the actual distance or predicted minimum distance and the safety distance threshold, or by proportionalizing the two, a quantitative index representing the size of the safety margin is obtained. When the risk distance is greater than the safety distance threshold, it indicates that there is still a certain safety margin, and the safety margin is positive. When the risk distance is close to or less than the safety distance threshold, it indicates that the safety margin has decreased or even entered a dangerous state, and the safety margin tends to zero or is negative. Finally, the safety margin is output as a quantitative index to uniformly represent the safety margin between the conductor and the hazard source under the current operating state of the transmission line, providing a basis for subsequent contact risk classification and multi-level early warning decisions, thereby realizing an intuitive depiction and dynamic assessment of the line safety status.

[0057] In summary, the embodiments of this application have at least the following technical effects: First, panoramic camera equipment is deployed along the transmission line to acquire multi-angle images along the towers while the line is in operation, obtaining an image set of the transmission line including the conductors and the surrounding environment. Then, based on location marker information, combined with wind deflection and galloping analysis and environmental data, dynamic simulation is performed to assess the risk distance characteristics between hazard sources and conductors. Next, the risk distance characteristics corresponding to hazard sources of different risk levels are recorded and a hazard source feature database is configured. Then, based on the transmission line image set and the hazard source feature database, an adversarial generative sample set is constructed. Multi-scale feature fusion is performed on the adversarial generative sample set to determine the spatial edge response characteristics and point cloud sparsity characteristics of the hazard sources, and contact risk is classified to determine the contact risk level. Finally, the safety margin under the line's operating state is assessed, and multi-level early warning decisions are made based on the contact risk level. This invention solves the technical problem that existing hazard source detection methods for transmission lines rely on static data, making it difficult to accurately locate hazards in complex dynamic environments, resulting in low early warning accuracy and untimely response. It achieves the technical effect of accurately locating hazards based on 3D reconstruction and back projection, combined with wind deflection dynamic simulation for risk classification and multi-level early warning, thereby improving the operational safety and emergency response capabilities of transmission lines.

[0058] Example 2 is based on the same inventive concept as the intelligent early warning method for hazardous sources based on visual detection of transmission lines in the previous examples, such as... Figure 2 As shown, this application provides an intelligent early warning system for hazardous sources based on visual inspection of transmission lines, wherein the system includes: Image Acquisition Module 11: A panoramic camera is deployed along the transmission line to acquire multi-angle images along the towers while the line is in operation, obtaining a set of transmission line images including conductors and the surrounding environment; Hazard Source Identification Module 12: A deep learning algorithm is used to identify hazard sources, and then a stereo vision algorithm is used to determine the location marker information of the hazard sources; Dynamic Simulation Module 13: Based on the location marker information, dynamic simulation is performed in conjunction with wind deflection and galloping analysis and environmental data to evaluate the risk distance characteristics between the hazard sources and the conductors; Feature Input Module 14: The risk distance characteristics corresponding to hazard sources of different risk levels are input, and a hazard source feature database is configured; Risk Classification Module 15: Based on the transmission line image set and the hazard source feature database, an adversarial generation sample set is constructed, and multi-scale feature fusion is performed on the adversarial generation sample set to determine the spatial edge response characteristics and point cloud sparsity characteristics of the hazard sources, and contact risk classification is performed to determine the contact risk level; Multi-level Early Warning Module 16: The safety margin under the line's operating state is evaluated, and multi-level early warning decisions are made in conjunction with the contact risk level.

[0059] Furthermore, the dynamic simulation module 13 is used to perform the following methods: Based on location marker information, spatial modeling is performed on the hazard source, and the spatial contour parameters and dynamic motion trajectory of the hazard source are extracted. By using real-time wind speed, wind direction, and conductor mechanical parameters, the conductor deflection angle and galloping amplitude under wind load are determined. Based on the spatial contour parameters and dynamic motion trajectory of the hazard source, the conductor deflection angle and galloping amplitude under wind load are used to simulate the dynamic offset process of the conductor under wind load and the real-time motion state of the hazard source, calculate the minimum spatial distance between the hazard source and the conductor, and determine the risk distance characteristics.

[0060] Furthermore, the feature input module 14 is used to perform the following method: If the hazard identification result determines that a hazard exists, a multi-angle synchronized image is determined based on the timestamp corresponding to the hazard; the multi-angle synchronized image is used to perform three-dimensional reconstruction through an image stereo vision algorithm to generate a three-dimensional spatial point cloud model; and a back projection algorithm is used to determine the location marker information of the hazard in the three-dimensional spatial point cloud model.

[0061] Furthermore, the risk classification module 15 is used to perform the following method: Risk distance features corresponding to different risk levels of hazard sources from historical inspections are collected, hazard source feature vectors are configured, and a three-level classification and storage is performed against the size, height, and dynamic behavior of construction machinery and vegetation to obtain the hazard source feature database. The three-level classification and storage adopts a distributed storage architecture. Geometric contour features, texture distribution features, and depth response features are extracted from the transmission line image set. Based on the geometric contour features, texture distribution features, and depth response features, comparative analysis is performed in the hazard source feature database to determine the target hazard source features. The target hazard source features are data augmented through a conditional adversarial generative network to construct an adversarial generative sample set including those without hazard markers or those with hazard source markers.

[0062] Furthermore, the risk classification module 15 is used to perform the following method: For identified potential hazards, a risk distance quantification assessment unit is set up by combining multi-dimensional spatial parameters, spatial edge response characteristics, and point cloud sparsity characteristics. The multi-dimensional spatial parameters include the top elevation of the hazard, horizontal offset, and dynamic approach rate. In the risk distance quantification assessment unit, a weighting coefficient is introduced for weighted correction. The weighting coefficient is determined based on the degree of influence of each parameter on the conductor discharge risk in historical tripping accidents. Using the risk distance quantification assessment unit, combined with the minimum allowable safe distance standard for conductors, the contact risk level is output.

[0063] Furthermore, the risk classification module 15 is used to perform the following method: The feature descriptors of the hazards are compared with the feature vectors of the hazards in the hazard feature database using cosine similarity analysis to obtain the feature matching degree. A hazard confidence matrix is ​​constructed based on the feature matching degree, and the matrix elements of the hazard confidence matrix represent the matching confidence of hazards of different risk levels in the current transmission line scenario. Threshold segmentation is performed on the hazard confidence matrix to determine the potential hazards, which are hazards with a matching confidence degree higher than the confidence threshold.

[0064] Furthermore, the risk classification module 15 is used to perform the following method: Based on the multi-scale feature fusion results, a three-dimensional graph structure corresponding to various hazard source regions is constructed. The nodes of the three-dimensional graph structure represent the local point cloud sparsity and local spatial edge response, and the edge weights of the three-dimensional graph structure are determined by spatial proximity and semantic similarity. A global pooling operation is performed on the three-dimensional graph structure to generate hazard source feature descriptors.

[0065] Furthermore, the risk classification module 15 is used to perform the following method: Based on the target hazard source characteristics, a conditional adversarial generative network is constructed using geometric contour features, texture distribution features, and depth response features as conditional inputs, based on hazard source category and meteorological condition label; during the training process of the adversarial generative network, Wasserstein distance and gradient penalty terms are introduced.

[0066] Furthermore, the multi-level early warning module 16 is used to perform the following method: The safety margin is determined based on the relative magnitude of the risk distance feature and the safety distance threshold, and is used to quantify the current safety margin.

[0067] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for intelligent early warning of hazards based on visual inspection of transmission lines, characterized in that, The method includes: Panoramic camera equipment is deployed along power transmission lines to collect images from multiple angles along the towers while the line is in operation, acquiring a set of images of the power transmission line including the conductors and the surrounding environment; Deep learning algorithms are used to identify hazards, and then image stereo vision algorithms are combined to determine the location marking information of the hazards. Based on location marking information, dynamic simulations are performed using wind deflection galloping analysis and environmental data to assess the risk distance characteristics between hazard sources and conductors. Enter the risk distance characteristics corresponding to hazard sources of different risk levels and configure the hazard source characteristic database; Based on the transmission line image set and the hazard source feature database, an adversarial generation sample set is constructed. Multi-scale feature fusion is performed on the adversarial generation sample set to determine the spatial edge response features and point cloud sparsity features of the hazard source, and contact risk classification is performed to determine the contact risk level. Assess the safety margin under the operating conditions of the line and make multi-level early warning decisions based on the contact risk level.

2. The intelligent early warning method for hazardous sources based on visual inspection of transmission lines as described in claim 1, characterized in that, The method for assessing the safety margin under the operating conditions of the line includes: The safety margin is determined based on the relative magnitude of the risk distance feature and the safety distance threshold, and is used to quantify the current safety margin.

3. The intelligent early warning method for hazardous sources based on visual inspection of transmission lines as described in claim 1, characterized in that, Based on location marker information, combined with wind deflection and galloping analysis and environmental data, dynamic simulation is performed. The method further includes: Based on location marker information, spatial modeling is performed on the hazard source, and the spatial contour parameters and dynamic motion trajectory of the hazard source are extracted; The wind deflection angle and galloping amplitude of the conductor under wind load are determined by real-time wind speed, wind direction and conductor mechanical parameters; Based on the spatial profile parameters and dynamic motion trajectory of the hazard source, the wind deflection angle and galloping amplitude of the conductor under wind load, the dynamic offset process of the conductor under wind load and the real-time motion state of the hazard source are simulated, the minimum spatial distance between the hazard source and the conductor is calculated, and the risk distance characteristics are determined.

4. The intelligent early warning method for hazardous sources based on visual inspection of transmission lines as described in claim 1, characterized in that, The method involves using a deep learning algorithm to identify hazards, followed by combining this with an image stereo vision algorithm to determine the location marking information of the hazards. If the hazard identification result determines that a hazard exists, multi-angle synchronized images are determined based on the timestamp corresponding to the hazard. The multi-angle synchronized images are used to perform three-dimensional reconstruction through an image stereo vision algorithm to generate a three-dimensional spatial point cloud model. The back-projection algorithm is used to determine the location marking information of the hazard source in the three-dimensional spatial point cloud model.

5. The intelligent early warning method for hazardous sources based on visual inspection of transmission lines as described in claim 4, characterized in that, Based on the transmission line image set and combined with the hazard source feature database, an adversarial generative sample set is constructed, the method comprising: Collect risk distance features corresponding to hazard sources of different risk levels from historical inspections, configure hazard source feature vectors, and perform three-level classification and storage by referring to the size, height and dynamic behavior of construction machinery and vegetation to obtain the hazard source feature database. The three-level classification and storage adopts a distributed storage architecture. Geometric contour features, texture distribution features, and depth response features are extracted from the transmission line image set. Based on the geometric contour features, texture distribution features, and depth response features, a comparative analysis is performed in the hazard source feature database to determine the target hazard source features; The target hazard source features are augmented using a conditional adversarial generative network to construct an adversarial generative sample set that includes either no hazard markers or hazard source markers.

6. The intelligent early warning method for hazardous sources based on visual inspection of transmission lines as described in claim 4, characterized in that, The method involves determining the spatial edge response characteristics and point cloud sparsity characteristics of a hazard source, and then classifying the contact risk to determine the contact risk level. For the identified potential hazards, a risk distance quantification assessment unit is set up by combining multi-dimensional spatial parameters, spatial edge response characteristics and point cloud sparsity characteristics. The multi-dimensional spatial parameters include the top elevation of the hazard, horizontal offset and dynamic approach rate. In the risk distance quantification assessment unit, a weighting coefficient is introduced for weighted correction. The weighting coefficient is determined based on the degree of influence of each parameter on the conductor discharge risk in historical tripping accidents. Using the aforementioned risk distance quantification assessment unit, combined with the minimum permissible safe distance standard for conductors, the contact risk level is output.

7. The intelligent early warning method for hazardous sources based on visual inspection of transmission lines as described in claim 6, characterized in that, The method further includes: The feature matching degree is obtained by performing cosine similarity analysis between the hazard source feature descriptor and the hazard source feature vector in the hazard source feature database. A hazard source confidence matrix is ​​constructed based on the feature matching degree, and the matrix elements of the hazard source confidence matrix represent the matching confidence degree of hazard sources of different risk levels in the current transmission line scenario; The hazard source confidence matrix is ​​segmented by a threshold to determine the potential hazard sources, which are hazard sources with a matching confidence level higher than the confidence threshold.

8. The intelligent early warning method for hazardous sources based on visual inspection of transmission lines as described in claim 7, characterized in that, The method further includes performing cosine similarity analysis between the hazard source feature descriptor and the hazard source feature vector in the hazard source feature database. Based on the multi-scale feature fusion results, a three-dimensional graph structure corresponding to various hazard source areas is constructed. The nodes of the three-dimensional graph structure represent the local point cloud sparsity and local spatial edge response, and the edge weights of the three-dimensional graph structure are determined by spatial proximity and semantic similarity. A global pooling operation is performed on the three-dimensional graph structure to generate a hazard source feature descriptor.

9. The intelligent early warning method for hazardous sources based on visual inspection of transmission lines as described in claim 5, characterized in that, The target hazard source characteristics are augmented using a conditional adversarial generative network, and the method includes: Based on the target hazard source characteristics, a conditional adversarial generative network is constructed using geometric contour features, texture distribution features, and depth response features as conditional inputs, based on hazard source category and meteorological condition label. During the training of the adversarial generative network, Wasserstein distance and gradient penalty terms are introduced.

10. A hazardous source intelligent early warning system based on visual inspection of transmission lines, characterized in that, The system is used to implement the intelligent early warning method for hazardous sources based on visual inspection of transmission lines as described in any one of claims 1-9, the system comprising: Image acquisition module: The panoramic camera equipment is deployed along the transmission line to acquire multi-angle images along the towers while the line is in operation, obtaining a set of transmission line images including conductors and the surrounding environment; Hazard identification module: Uses deep learning algorithms to identify hazards, and then combines image stereo vision algorithms to determine the location marking information of the hazards; Dynamic simulation module: Based on location marker information, combined with wind deflection galloping analysis and environmental data, dynamic simulation is performed to assess the risk distance characteristics between the hazard source and the conductor; Feature Entry Module: This module allows users to enter the risk distance features corresponding to hazard sources of different risk levels and configures the hazard source feature database. Risk classification module: Based on the transmission line image set and combined with the hazard source feature database, construct an adversarial generation sample set, perform multi-scale feature fusion on the adversarial generation sample set, determine the spatial edge response features and point cloud sparsity features of the hazard source, and perform contact risk classification to determine the contact risk level. Multi-level early warning module: Assess the safety margin under the operating status of the line and make multi-level early warning decisions based on the contact risk level.

Images