Power distribution network fault location method, system, device and medium

Abstract

The application relates to the technical field of power distribution network operation and maintenance, and discloses a power distribution network fault positioning method, system, device and medium. The method can stably extract abnormal visual information under the condition of uneven illumination, complex background and weak local features by implementing gray self-adaptive enhancement on the inspection image; subsequently, abnormal visual objects in the image coordinate system are converted into suspected fault sections with actual electrical implications by using collected associated information and feeder topology operation information, thereby overcoming the problem that ordinary visual detection results only stay at the target frame, mask or abnormal category level and cannot directly serve line maintenance; further, a fault evidence subgraph is constructed by taking the suspected fault section as the center to correlate visual features, device operation states and topological relations, and each suspected fault section is comprehensively inferred through a graph neural network, so that the visual evidence can participate in fault judgment along the actual device connection and power supply path, and the accuracy and stability of fault positioning are improved.

Description

Technical Field

[0001] This invention relates to the field of power distribution network operation and maintenance technology, and in particular to a method, system, equipment and medium for locating faults in power distribution networks. Background Technology

[0002] With the widespread application of drone inspections, tower cameras, and digital operation and maintenance systems in power distribution networks, fault identification and location using image data has become an important technological direction. However, for on-site operation and maintenance, simply identifying abnormal targets on the image plane is insufficient to support maintenance decisions. The key is whether abnormal visual evidence can be stably mapped to specific nodes, edges, or fault sections in the actual feeder topology, and whether the location accuracy can be maintained under topology changes, node loss, and complex environmental noise.

[0003] Existing similar solutions are generally divided into two categories: The first category uses a general visual model for image defect detection, which can identify cracks, burns, arcs or foreign objects, but most of them stop at the target box, mask or anomaly category, and it is difficult to directly map to the fault section in the actual distribution network topology; The second category uses graph neural networks to process the relationship between equipment status and topology, which can utilize graph structure information, but its input is mostly electrical quantity, state quantity or statistical feature, and it lacks a special mechanism to backmap image anomaly targets to topology nodes and edges.

[0004] However, the aforementioned existing technologies can usually only output the location or category of abnormal targets within the image coordinate system, and it is difficult to further map abnormal targets into specific nodes, edges or sections in the feeder diagram. Under the circumstances of changes in distribution network topology, changes in the status of sectional switches, visual obstruction and changes in shooting angle, the accuracy of fault location will be greatly reduced and the robustness is poor. Summary of the Invention

[0005] In view of this, in order to solve the above-mentioned technical problems, the present invention provides a method, system, device and medium for fault location in power distribution networks.

[0006] The first aspect of the present invention provides a method for locating faults in a distribution network, the method comprising: Acquire inspection images of the power distribution network, the acquisition association information corresponding to the inspection images, and the topology operation information of the corresponding feeders; The inspection image is subjected to grayscale adaptive enhancement, and abnormal visual objects and their visual features are extracted from the enhanced inspection image. Based on the collected association information and the topology operation information, the abnormal visual object is matched with the topology object in the corresponding feeder to determine at least one suspected faulty section. Centered on the suspected faulty section, the visual features, equipment operating status and topological relationships corresponding to the suspected faulty section are associated to construct a fault evidence subgraph; The fault evidence subgraph of at least one of the suspected fault sections is input into a pre-trained graph neural network model to obtain the fault probability corresponding to each of the suspected fault sections, and the target fault section is determined based on the fault probability.

[0007] In one example, the grayscale adaptive enhancement of the inspection image and the extraction of abnormal visual objects and their visual features from the enhanced inspection image include: The inspection image is normalized for brightness and suppressed for noise to obtain a basic grayscale image; Based on the brightness, texture and edge features of different local regions in the base grayscale image, grayscale enhancement parameters for the corresponding local regions are determined, and the base grayscale image is locally enhanced based on the grayscale enhancement parameters to obtain the enhanced inspection image. Based on the enhanced inspection image, a fault candidate region is extracted, and the abnormal visual object is determined by combining the correlation between the fault candidate region and the corresponding equipment target. The texture features, morphological features, brightness distribution features, and contextual features of the abnormal visual object are extracted to obtain the visual features.

[0008] In one example, the step of extracting fault candidate regions from the enhanced inspection image and determining the abnormal visual object by combining the correlation between the fault candidate regions and the corresponding equipment targets includes: The edge response information and hotspot saliency information corresponding to the enhanced inspection image are extracted, and the edge response information and hotspot saliency information are fused with the enhanced inspection image to obtain a fault sensitivity map; Identify equipment targets from the fault sensitivity map and extract fault candidate regions; Based on the relative relationship between the fault candidate region and the device target, the fault candidate region is filtered, and the retained fault candidate region is determined as the abnormal visual object.

[0009] In one example, the step of performing image-topology alignment between the abnormal visual object and the corresponding topology object in the feeder, based on the acquired correlation information and the topology operation information, to determine at least one suspected faulty section, includes: The field of view coverage area corresponding to the inspection image is determined based on the collected association information, and candidate topology objects within the field of view coverage area are determined from the corresponding feeder based on the topology operation information. Based on the position, size, orientation, and spatial relationship between the abnormal visual object and neighboring devices in the inspection image, the abnormal visual object is matched with the candidate topology object to obtain at least one candidate mapping object. Based on the power supply direction of the corresponding feeder, the equipment arrangement order, the upstream and downstream relationship of the equipment, and the segmented switch boundary, the candidate mapping object is checked for topology consistency. Based on the results of the topology consistency check, the feeder segments corresponding to the candidate mapping objects that meet the topology consistency conditions are identified as the suspected fault segments.

[0010] In one example, the step of constructing a fault evidence subgraph by associating the visual features, equipment operating status, and topological relationships corresponding to the suspected fault section with the suspected fault section as the center includes: Centered on the suspected fault section, determine the local topology range associated with the suspected fault section along the upstream and downstream directions of the corresponding feeder; Extract the equipment objects, equipment operating status and spatial location information associated with the suspected fault section from the local topology range, and bind the abnormal visual objects and their visual features to the corresponding equipment objects or line segments; Based on the binding results, nodes representing the device object, abnormal visual object, device operating status, and spatial location are constructed, and edges are established between the corresponding nodes according to the electrical connection relationship, spatial relationship, image adjacency relationship, power supply direction relationship, and switch segmentation relationship within the local topology range; By combining the nodes, edges, and corresponding visual features, device operating status, and topological relationships, a fault evidence subgraph corresponding to the suspected faulty section is obtained.

[0011] In one example, the step of inputting the fault evidence subgraph of at least one of the suspected fault segments into a pre-trained graph neural network model to obtain the fault probability corresponding to each of the suspected fault segments, and determining the target fault segment based on the fault probability, includes: For each suspected faulty section, node features, edge attributes, and topological relationships between nodes are extracted from the corresponding fault evidence subgraph. Based on the location of the suspected fault section in the corresponding fault evidence sub-graph, the local transmission range associated with the suspected fault section is determined, and based on the electrical path length, power supply direction and segment switch status between nodes within the local transmission range, the information transmission constraints between nodes are determined. The node features, edge attributes, topological relationships, and information transmission constraints are input into the pre-trained graph neural network model to obtain the fault probability corresponding to each suspected fault segment. Based on the fault probability corresponding to each of the suspected fault sections, the suspected fault sections are compared, and the suspected fault sections that meet the preset fault probability conditions are determined as the target fault sections.

[0012] In one example, the step of determining the local transmission range associated with the suspected fault segment based on its position in the corresponding fault evidence subgraph, and determining the information transmission constraints between nodes based on the electrical path length, power supply direction, and sectionalizing switch status within the local transmission range, includes: Centered on the node associated with the suspected fault segment in the corresponding fault evidence subgraph, expand outward according to a preset topology hop count to determine the local node set participating in the reasoning of the suspected fault segment, and determine the range corresponding to the local node set as the local propagation range; For the node pairs within the local transmission range, the electrical path length between the node pairs, the consistency of the power supply direction between the node pairs, and the segment switch status on the connection path are determined according to the current topology of the corresponding feeder. For a first node pair where the power supply path is reachable and the power supply direction is consistent, the information transmission weight of the first node pair is determined according to the corresponding electrical path length. For a second node pair where the connection path is interrupted by a sectionalizing switch or the power supply direction is inconsistent, the information transmission weight of the second node pair is reduced. The information transmission constraints are generated based on the information transmission weights of all the node pairs.

[0013] Secondly, the present invention also provides a power distribution network fault location system, the system comprising: The information acquisition module is used to acquire inspection images of the distribution network, the acquisition association information corresponding to the inspection images, and the topology operation information of the corresponding feeders; The grayscale enhancement module is used to perform adaptive grayscale enhancement on the inspection image and extract abnormal visual objects and their visual features from the enhanced inspection image. The suspected fault location module is used to perform image and topology alignment between the abnormal visual object and the corresponding topology object in the feeder based on the collected association information and the topology operation information, and to determine at least one suspected fault section. The evidence subgraph construction module is used to construct a fault evidence subgraph by associating the visual features, equipment operating status and topological relationships corresponding to the suspected fault section with the suspected fault section as the center. The target fault location module is used to input the fault evidence subgraph of at least one of the suspected fault sections into a pre-trained graph neural network model to obtain the fault probability corresponding to each of the suspected fault sections, and to determine the target fault section based on the fault probability.

[0014] Thirdly, the present invention also provides an electronic device, the electronic device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the power distribution network fault location method as described in the first aspect.

[0015] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed, implements the steps of the power distribution network fault location method as described in the first aspect.

[0016] As can be seen from the above technical solutions, this invention, by implementing grayscale adaptive enhancement on inspection images, can stably extract abnormal visual information under conditions of uneven lighting, complex backgrounds, and weak local features. Subsequently, by utilizing the acquired correlation information and feeder topology operation information, abnormal visual objects in the image coordinate system are converted into suspected fault sections with actual electrical meaning, overcoming the problem that ordinary visual detection results only remain at the level of target boxes, masks, or abnormal categories and cannot directly serve line maintenance. Furthermore, a fault evidence subgraph is constructed by associating visual features, equipment operating status, and topological relationships with the suspected fault sections as the center, and a graph neural network is used to perform comprehensive reasoning on each suspected fault section, enabling visual evidence to participate in fault judgment along the actual equipment connection and power supply path. Therefore, this solution can reduce the positioning deviation caused by the disconnect between image recognition and electrical topology, and improve the accuracy and stability of section-level fault location under conditions of distribution network topology changes, changes in sectional switch status, visual obstruction, and changes in shooting angle. Attached Figure Description

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

[0018] Figure 1 This is an application environment diagram of a power distribution network fault location method provided in an embodiment of the present invention; Figure 2 A flowchart of a power distribution network fault location method provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a power distribution network fault location system provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The power distribution network fault location method provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, terminal 101 communicates with server 102 via a network. A data storage system can store the data that server 102 needs to process. The data storage system can be integrated onto server 102 or placed on a cloud or other network server. Terminal 101 or server 102 executes a power distribution network fault location method, which includes: acquiring inspection images of the power distribution network, corresponding acquisition association information, and topology operation information of the corresponding feeders; performing grayscale adaptive enhancement on the inspection images and extracting abnormal visual objects and their visual features from the enhanced images; based on the acquisition association information and topology operation information, performing image-topology alignment between the abnormal visual objects and the corresponding topology objects in the feeders to determine at least one suspected fault section; using the suspected fault section as the center, associating the visual features, equipment operating status, and topology relationships corresponding to the suspected fault section to construct a fault evidence subgraph; inputting the fault evidence subgraph of at least one suspected fault section into a pre-trained graph neural network model to obtain the fault probability corresponding to each suspected fault section, and determining the target fault section based on the fault probability.

[0021] Terminal 101 can be, but is not limited to, various personal computers, laptops, smartphones, and tablets.

[0022] Server 102 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides cloud computing services.

[0023] like Figure 2 As shown in the embodiment of this application, a method for locating faults in a distribution network is provided, which is applied to... Figure 1 Taking terminal 101 or server 102 as an example, the explanation includes the following steps S1 to S5. Wherein: Step S1: Obtain inspection images of the distribution network, the acquisition association information corresponding to the inspection images, and the topology operation information of the corresponding feeders.

[0024] Inspection images refer to images captured by photographing power distribution network lines, towers, and their auxiliary equipment. These images can be collected by drones, tower cameras, or mobile inspection terminals. Inspection images can be single frames or sequences of images captured continuously by the same inspection equipment. The equipment covered by the inspection images can include conductors, insulators, switches, crossarms, surge arresters, and other power distribution equipment requiring fault inspection.

[0025] Acquisition association information describes the acquisition location, viewing angle, and corresponding equipment object of the inspection images. It may include at least one or more of the following: shooting time, trackpoint, shooting orientation, gimbal attitude, focal length, tower number, and equipment identification. The trackpoint determines the spatial location of the inspection equipment during image acquisition; the shooting orientation and gimbal attitude determine the camera's observation direction; the focal length determines the field of view and scale relationships within the image; and the tower number and equipment identification are used to initially associate the inspection images with the equipment register and feeder topology.

[0026] Topology operation information reflects the connection relationships and operational status of the corresponding feeder at the time of image acquisition during inspection. This information can include the feeder topology, connection relationships between equipment nodes and line segments, tower geographic coordinates, equipment operating status, sectionalizing switch status, equipment ledger information, and historical fault records. This topology operation information does not refer solely to fixed line design diagrams, but rather preferably uses current topology data that reflects current interconnection and power transfer, sectionalizing switch switching, and line modification status.

[0027] When obtaining the above information, the corresponding feeder can be determined first based on the feeder identification, tower number, or shooting location carried in the inspection image. Then, the topology operation information of the feeder can be retrieved from the distribution automation system, geographic information system, or equipment ledger. If the inspection image does not directly carry the feeder identification, the candidate feeder can be determined based on the positional relationship between the shooting location and the geographical coordinates of each feeder tower. The corresponding feeder can then be determined by combining the shooting direction and the tower arrangement.

[0028] Step S2: Perform grayscale adaptive enhancement on the inspection image, and extract abnormal visual objects and their visual features from the enhanced inspection image.

[0029] Among them, grayscale adaptive enhancement refers to the enhancement processing that automatically adjusts the enhancement parameters according to the grayscale distribution characteristics of the inspection image itself to highlight the grayscale differences in the fault area. It can adapt to inspection images obtained under different lighting and shooting conditions, and avoid the problems of under-enhancement or over-enhancement caused by fixed enhancement parameters.

[0030] Anomaly visual objects can be either an anomalous region associated with a single device target or an anomalous region spanning adjacent devices or line segments. The output can include the location, outline, main direction, corresponding device target, and visual features of the anomalous region.

[0031] Visual features are characteristics that describe the visual appearance of abnormal areas, including one or more of color features, texture features, shape features, and gradient features. They can be used to distinguish different types of faults and anomalies and help determine the severity of the fault.

[0032] Step S3: Based on the collected correlation information and topology operation information, perform image and topology alignment between the abnormal visual object and the corresponding topology object in the feeder to identify at least one suspected faulty section.

[0033] Image-to-topology alignment refers to mapping abnormal visual objects in the image coordinate system to equipment nodes, line segments, or feeder sections in the distribution network topology. This process does not simply search for similar equipment based on the category of the target in the image, but comprehensively uses factors such as shooting location, shooting angle, tower number, feeder geographical coordinates, equipment arrangement order, and current switch status to verify spatial and electrical relationships.

[0034] Topology objects can be equipment nodes, tower nodes, sectionalizing switch nodes, line segments, or feeder sections formed by adjacent nodes and line segments in the feeder topology.

[0035] Suspected fault segments are candidate segments obtained by mapping abnormal visual objects and need to enter the subsequent fault evidence subgraph construction and graph neural network inference. They are not the same as the already confirmed target fault segments.

[0036] Step S4: Centered on the suspected faulty section, associate the visual features, equipment operating status and topological relationships corresponding to the suspected faulty section to construct a fault evidence subgraph.

[0037] Among them, the fault evidence subgraph is a local graph structure built around a suspected faulty section, which is used to organize the evidence scattered in image data, equipment operation data and topology data into nodes, edges and their attributes that can be processed by graph neural networks.

[0038] The fault evidence subgraph differs from the complete feeder topology. The complete feeder topology may contain a large number of nodes that are not directly related to the current suspected faulty section, while the fault evidence subgraph only extracts local nodes and edges that are related to the fault judgment of the current suspected faulty section, in order to reduce the interference of irrelevant nodes on model inference.

[0039] Step S5: Input the fault evidence subgraph of at least one suspected faulty section into the pre-trained graph neural network model to obtain the fault probability corresponding to each suspected faulty section, and determine the target faulty section based on the fault probability.

[0040] The graph neural network model can perform information transmission and feature aggregation based on the topological structure of the fault evidence subgraph. It fully utilizes multi-source fault evidence contained in node features, edge attributes, and topological relationships, and directly outputs the fault probability of the corresponding suspected fault segment through end-to-end inference, without the need for manually designed rules to fuse multi-source evidence. The pre-trained graph neural network model can be trained on a labeled fault dataset. During training, the fault evidence subgraph is used as input, and the label of whether the corresponding suspected fault segment has actually experienced a fault is used as the optimization objective to adjust the network parameters of the model, so that the model learns the mapping relationship between multi-source fault evidence and fault probability.

[0041] It should be noted that this application, by implementing grayscale adaptive enhancement on inspection images, can stably extract abnormal visual information even under conditions of uneven lighting, complex backgrounds, and weak local features. Subsequently, by utilizing acquired correlation information and feeder topology operation information, abnormal visual objects in the image coordinate system are transformed into suspected fault sections with actual electrical meaning. This overcomes the problem that ordinary visual detection results only remain at the level of target boxes, masks, or anomaly categories and cannot directly serve line maintenance. Furthermore, a fault evidence subgraph is constructed by associating visual features, equipment operating status, and topological relationships with the suspected fault sections as the center. A graph neural network is then used to perform comprehensive reasoning on each suspected fault section, enabling visual evidence to participate in fault judgment along the actual equipment connections and power supply paths. Therefore, this scheme can reduce the positioning deviation caused by the disconnect between image recognition and electrical topology, and improve the accuracy and stability of section-level fault location under conditions of distribution network topology changes, changes in sectional switch status, visual obstruction, and changes in shooting angle.

[0042] In some embodiments, grayscale adaptive enhancement is performed on the inspection image, and abnormal visual objects and their visual features are extracted from the enhanced inspection image. This includes: performing brightness normalization and noise suppression on the inspection image to obtain a base grayscale image; determining grayscale enhancement parameters for corresponding local regions based on the brightness features, texture features, and edge features of different local regions in the base grayscale image, and performing local enhancement on the base grayscale image based on the grayscale enhancement parameters to obtain an enhanced inspection image; extracting fault candidate regions from the enhanced inspection image, and determining abnormal visual objects by combining the correlation between the fault candidate regions and the corresponding equipment targets; and extracting the texture features, morphological features, brightness distribution features, and contextual features of the abnormal visual objects to obtain visual features.

[0043] Brightness normalization is used to reduce the overall brightness differences caused by different shooting times, weather conditions, backlighting, and exposure parameters, so that similar fault features have a relatively consistent grayscale representation in different inspection images. Brightness normalization can adjust overly bright and overly dark areas according to the overall brightness distribution or regional brightness distribution of the inspection image, rather than uniformly adjusting all pixels to the same brightness.

[0044] Noise suppression is used to remove random noise caused by image sensors, transmission compression, flight vibrations, or environmental particles, while preserving fine-grained structures such as wire edges, crack edges, and discharge trajectories as much as possible. Noise suppression should not over-smooth fine-line wires or small cracks, otherwise it may lead to missed detection of subsequent fault candidate areas.

[0045] Local regions can be divided using fixed-size image windows, device target areas, or salient partitions. For each local region, its corresponding brightness features, texture features, and edge features are determined. Brightness features characterize the lightness and darkness of the local region and the brightness variations; texture features characterize the density of grayscale variations and the complexity of the structure within the local region; edge features characterize conductor outlines, crack boundaries, ablation boundaries, and other locations of structural abrupt changes.

[0046] Gray-level enhancement parameters can include gray-level mapping intensity, local contrast enhancement level, and edge enhancement level. When determining gray-level enhancement parameters, different processing strategies can be adopted based on the image state of different local areas. For example: for local areas that are generally dark and have weak texture, local contrast can be increased to reveal cracks or ablation boundaries obscured by shadows; for areas with fine linear structures and weak edge gradients, edge enhancement can be increased to highlight broken wire strands, fine cracks, or flashover trajectories; for locally bright areas, it is advisable to avoid further significant increases in brightness, instead focusing on preserving the boundary between the bright area and its surrounding areas to identify bright traces formed by arcing or local overheating; for leafy or complex background areas with high texture density, the enhancement amplitude can be limited to avoid mistakenly enhancing background textures as fault features.

[0047] Based on the determined grayscale enhancement parameters, each local region in the base grayscale image is enhanced separately, and the enhanced local regions are then recombined according to their original spatial positions to obtain the enhanced inspection image, i.e., the grayscale enhanced image. The "adaptive" aspect in this process refers to the enhancement parameters changing with local brightness, texture, and edge conditions, rather than applying fixed parameters uniformly to the entire image.

[0048] Local adaptive enhancement can take into account different areas such as highlights, shadows, low contrast and complex backgrounds in the same image, highlighting fault-related visual features such as arcing, cracks, ablation, foreign object overlap and wire damage, while reducing the over-enhancement of irrelevant backgrounds.

[0049] In one implementation, fault candidate regions are extracted from the enhanced inspection image, and abnormal visual objects are determined by combining the correlation between the fault candidate regions and the corresponding equipment targets. This includes: extracting edge response information and hotspot saliency information corresponding to the enhanced inspection image, and fusing the edge response information and hotspot saliency information with the enhanced inspection image to obtain a fault sensitivity map; identifying equipment targets from the fault sensitivity map and extracting fault candidate regions; filtering the fault candidate regions according to the relative relationship between the fault candidate regions and the equipment targets, and determining the retained fault candidate regions as abnormal visual objects.

[0050] Specifically, firstly, edge response information is extracted from the grayscale enhancement image. This edge response information is used to characterize the locations of grayscale abrupt changes, highlighting conductor edges, insulator outlines, crack boundaries, burn-off boundaries, and flashover trajectories. Secondly, hotspot saliency information is extracted to characterize areas with significant brightness differences or degrees of saliency relative to their surroundings. Then, the edge response information and hotspot saliency information are fused with the grayscale enhancement image to form a fault sensitivity map. This fault sensitivity map is not simply a grayscale conversion result, but rather an enhanced image that highlights thin-line conductors, small cracks, bright arcs, and localized burn-off traces. For example, arc areas can be highlighted using hotspot saliency information; broken conductor strands and small cracks can be highlighted using edge response information; and localized ablation can be expressed using combined information from grayscale distribution, edge contours, and texture changes.

[0051] The fault sensitivity map is used for subsequent identification of equipment targets and extraction of fault candidate regions. Equipment targets can include conductors, insulators, switches, crossarms, and surge arresters.

[0052] Fault candidate regions refer to image regions in the enhanced image that have abnormal brightness, abnormal edges, abnormal shapes, or abnormal salience.

[0053] Fault candidate regions can include locally highlighted areas, edge breakage areas, abnormally suspended areas, discharge trajectories, smoke trails, and tree contact areas. Generally, fault candidate regions can be extracted using threshold-based anomaly detection, pre-trained anomaly segmentation models, or object detection models.

[0054] For fault signs that are difficult to accurately represent with a single rectangular detection box, such as arcing, smoke, flashover trajectories, and tree contact areas, pixel-level salient regions can be used as fault candidate regions, and the center position, coverage outline, main direction, and area distribution of these regions can be determined. This avoids the rectangular detection box containing too much background, which would reduce the accuracy of subsequent alignment.

[0055] The correlation between the equipment target and the fault candidate area can be determined based on their relative position, overlap, distance, area ratio, and direction. For example, a discharge trajectory located on the surface of an insulator or adjacent to the edge of an insulator can be associated with the corresponding insulator; an abnormal edge distributed along the direction of conductor extension and overlapping with the conductor outline can be associated with the corresponding conductor; and an abnormal object suspended below and in contact with a conductor can be associated with that conductor.

[0056] When identifying anomalous visual objects, non-fault disturbances can be ruled out by combining the number of consecutive frames and environmental context information. Highlighted areas that appear for a short time and change rapidly with the viewing angle may be sunlight reflections; areas that sway synchronously with the leaves in the background and do not form stable contact with the line equipment may be background interference; shadowed areas that appear only in a single frame and have no stable association with any equipment target can be excluded as anomalous visual objects.

[0057] For example, if a bright area is always located at the same end of an insulator in three consecutive frames and its spatial position remains consistent with the insulator as the camera moves, then the bright area can be identified as an abnormal visual object; if the bright area moves rapidly on different equipment surfaces as the shooting angle changes, it can be screened out as a reflective disturbance.

[0058] Anomalies can be either an anomalous region associated with a single device target or an anomalous region that spans adjacent devices or line segments. The output can include the location, outline, main direction, corresponding device target, and visual features of the anomalous region.

[0059] Texture features are used to characterize the grayscale variation patterns within and near the boundaries of abnormal visual objects, and can be used to distinguish between smooth reflective areas, rough ablation areas, and fine linear crack areas. Morphological features are used to characterize the area, outline, length-width relationship, main direction, and extension shape of abnormal visual objects, and can be used to distinguish between point-like arcs, linear cracks, sheet-like burns, and strip-like smoke.

[0060] Brightness distribution features are used to characterize the internal brightness of abnormal visual objects and their brightness difference with surrounding areas. They can be used to identify visual manifestations such as high-brightness discharge, localized overheating, and dark ablation. Contextual features are used to characterize the relationship between abnormal visual objects and equipment targets, adjacent equipment, the sky, trees, and the direction of lines. This allows the model to not only identify the abnormal area itself but also determine whether the abnormal area is located in a location with fault significance.

[0061] The extracted visual features, together with the image location and device association of the abnormal visual object, constitute the input for subsequent image and topology alignment, and are written into the visual target node or corresponding device node of the fault evidence subgraph in step S4.

[0062] In some embodiments, based on the collected association information and topology operation information, image-topology alignment is performed between abnormal visual objects and corresponding topology objects in the feeder to determine at least one suspected fault segment. This includes: determining the field of view coverage area corresponding to the inspection image based on the collected association information, and determining candidate topology objects within the field of view coverage area from the corresponding feeder based on the topology operation information; matching the abnormal visual object with the candidate topology objects based on the position, size, orientation, and spatial relationship between the abnormal visual object and adjacent equipment in the inspection image to obtain at least one candidate mapping object; performing topology consistency verification on the candidate mapping objects based on the power supply direction, equipment arrangement order, upstream and downstream relationships of the equipment, and segmented switch boundaries of the corresponding feeder; and determining the feeder segment corresponding to the candidate mapping object that meets the topology consistency conditions as a suspected fault segment based on the topology consistency verification result.

[0063] Specifically, the location of the camera equipment can be determined based on the flight path or shooting position of the inspection equipment, and the observation direction and coverage area of ​​the camera equipment can be determined based on the shooting orientation, gimbal attitude, and focal length. Then, the field of view coverage area is projected onto the geographic space of the corresponding feeder, and towers, equipment nodes, and line segments that may appear in the current inspection image are screened from the feeder topology to form a candidate topology object set.

[0064] During the screening process, topological objects that are clearly outside the field of view can be excluded first by using spatial location. Then, the candidate range can be narrowed down by combining tower number, equipment identification, and feeder direction. For example, if the image is taken upstream of tower T12 and the shooting direction is towards T12, tower T12 and its upstream and downstream adjacent segments can be retained first, while towers located behind the shooting equipment or with a significantly different shooting direction can be excluded.

[0065] For a set of candidate topological objects, geometric matching can be performed based on the relative position, size ratio, main direction, occlusion relationship, and combination of neighboring devices of the anomalous visual objects in the image.

[0066] For example, if an abnormal visual object is located at the lower end of an identified insulator in an image and maintains a preset spatial relationship with the conductor below the insulator, it can be preferentially matched to the insulator or its connection position; if an abnormal visual object extends along the conductor direction between two adjacent towers, it can be preferentially matched to the line segment between the two towers; when an abnormal visual object crosses a switchgear and an adjacent conductor, candidate mapping objects for the switch node, the adjacent line segment, and the segment between them can be generated respectively.

[0067] When an abnormal visual object covers only the area near a single device or a single line segment, and its image location and device combination relationship can uniquely correspond to a topology object, a unique candidate mapping object can be directly formed. When an abnormal visual object spans adjacent devices, crosses line segment boundaries, or cannot be clearly identified as belonging to an upstream device, downstream device, or intermediate segment due to occlusion, multiple candidate mapping objects can be formed for subsequent topology consistency verification.

[0068] Thus, abnormal regions in the image are transformed into one or more candidate nodes, candidate line segments, or candidate sections with clear electrical semantics, providing a basis for subsequent topology verification.

[0069] Topology consistency checks are used to exclude candidate mapping objects that are geometrically close but do not conform to the physical relationship of the feeder. During the check, the following can be considered separately: Whether the candidate mapping object is located on the actual power supply path of the current feeder; Whether the upstream and downstream relationships of the candidate mapping objects are consistent with the device arrangement direction in the image; Does the order of the poles corresponding to the candidate mapping objects match the movement order of the inspection equipment and the order in which the equipment appears in the image? The segmentation switch for whether the candidate mapping object is disconnected is isolated from the corresponding region of the current image; Does the candidate mapping object cross a feeder segment boundary that it should not cross?

[0070] For example, geometric matching yields two candidate results, segment A and segment B, but the current segment switch status indicates that there is no reachable power supply path between segment B and the upstream device in the image, so segment B can be excluded. As another example, if the device appearance order in a continuous inspection image is towers T12, T13, and T14, and a certain candidate result corresponds to the reverse order of T14, T13, and T12, then the consistency of the device arrangement order in this candidate result is low.

[0071] When multiple candidate mapping objects pass the basic topology verification, further consistency disambiguation can be performed based on power supply direction consistency, tower sequence consistency, path reachability, and cross-frame observation stability. These factors form the basis for evaluating the consistency of candidate mapping objects, and those with higher consistency are retained for subsequent processing. Specifically, multiple candidate mapping objects can be sorted according to their consistency scores, and the segment corresponding to the candidate mapping object with the highest score can be selected as the suspected fault segment. Alternatively, multiple candidate segments with scores higher than a preset threshold can be retained and collectively entered into the subsequent fault inference step.

[0072] When the same abnormal visual object appears repeatedly in adjacent inspection frames, different shooting angles, or adjacent tower images, cross-frame consensus fusion can be performed based on the spatial overlap relationship, equipment association relationship, tower order, and continuous occurrence status of the abnormal visual object in each frame.

[0073] Cross-frame consensus fusion is not simply averaging the candidate results of each frame, but rather determining whether candidate mappings in multiple frames point to the same device, adjacent segments, or the same feeder segment. For candidate segments that remain stable in consecutive frames, their retention priority can be increased; for candidate segments that appear only in a single frame and are inconsistent with the preceding and following frames, their priority can be decreased or they can be treated as candidates to be reviewed.

[0074] Based on the cross-frame consensus fusion results, segment anchor point representations corresponding to suspected faulty segments can be generated. These segment anchor point representations can include upstream anchor points, downstream anchor points, segment center indexes, and segment sequence codes. The upstream and downstream anchor points define the segment's boundaries within the feeder; the segment center index identifies the center position of the currently suspected faulty segment; and the segment sequence code reflects the segment's arrangement order along the feeder's power supply direction. The segment anchor point representations are not independent outputs but can be further written into the fault evidence subgraph to determine subsequent local propagation ranges, node position representations, and information propagation paths, thereby enabling the image and topology alignment results to continue constraining graph neural network inference.

[0075] In some embodiments, a fault evidence subgraph is constructed by associating the visual features, equipment operating status, and topological relationships corresponding to the suspected fault section with the suspected fault section as the center. This includes: determining the local topological range associated with the suspected fault section along the upstream and downstream directions of the corresponding feeder with the suspected fault section as the center; extracting the equipment objects, equipment operating status, and spatial location information associated with the suspected fault section from the local topological range, and binding the abnormal visual objects and their visual features with the corresponding equipment objects or line segments; constructing nodes representing the equipment objects, abnormal visual objects, equipment operating status, and spatial location based on the binding results, and establishing edges between the corresponding nodes based on the electrical connection relationships, spatial relationships, image adjacency relationships, power supply direction relationships, and switch segmentation relationships within the local topological range; and combining the nodes, edges, and corresponding visual features, equipment operating status, and topological relationships to obtain the fault evidence subgraph corresponding to the suspected fault section.

[0076] This involves using the equipment node, line segment, or segment anchor point corresponding to the suspected faulty section as the center, and expanding upstream and downstream according to a preset topology hop count to obtain the local topology range. The preset topology hop count can be pre-set based on the feeder structure, segment length, and the range of adjacent equipment that may be affected by the fault.

[0077] During the expansion process, the current status of the sectionalizing switches should be taken into account. For topology paths blocked by disconnected switches, expansion along those paths can be stopped; branches with opposite power supply directions or not belonging to the current feeder's effective power supply path can be excluded from the local topology scope. The resulting local topology scope retains evidence of upstream and downstream connections around the suspected fault section while avoiding the introduction of irrelevant nodes isolated by switches or located too far away.

[0078] For example, if the line segment between towers T13 and T14 is considered a suspected fault section, T12, T13 and their corresponding equipment can be extracted upstream, and T14, T15 and their corresponding equipment can be extracted downstream. If the sectionalizing switch downstream of T15 is in an open state, then that switch can be used as a local boundary, and the fault cannot be extended further downstream.

[0079] Equipment objects, equipment operating status, and spatial location information can be extracted from a local topology. Equipment objects can include conductors, insulators, switches, crossarms, surge arresters, and towers; equipment operating status can include switch open / closed status, equipment status indicators, and other operating statuses related to the equipment; spatial location information can include tower coordinates, relative positions of equipment, line routes, and actual distances between nodes.

[0080] Binding refers to writing abnormal visual objects and their visual features into the corresponding device objects or line segments based on the image and topology alignment results obtained in step S3. For example, when an abnormal visual object is mapped to an insulator device node, its texture features, morphological features, and brightness distribution features can be written into the insulator node or a visual target association edge can be established with the node; when an abnormal visual object is mapped to a line segment, it can be bound to the corresponding line segment node.

[0081] When an anomalous visual object corresponds to multiple suspected sections that have not yet been fully disambiguated, a corresponding fault evidence subgraph can be constructed for each suspected faulty section. The binding relationship between the anomalous visual object and the corresponding candidate device or line segment can be preserved in each subgraph. Subsequently, a graph neural network determines the fault probability based on the comprehensive evidence in each subgraph.

[0082] The nodes in the fault evidence subgraph can include at least: Equipment nodes are used to characterize equipment objects such as conductors, insulators, switches, crossarms, surge arresters, or towers; Visual target nodes are used to represent abnormal visual objects extracted from inspection images; The operating status node is used to characterize the current operating status of the device or the on / off status of the switch; Spatial location nodes are used to represent tower coordinates, section locations, or other spatial positioning information.

[0083] In some implementations, visual characteristics, operating status, and spatial location can be directly used as attributes of equipment nodes or line segment nodes, without establishing separate nodes of the corresponding type. Regardless of the form used, it should be ensured that visual anomaly evidence can be clearly associated with its corresponding electrical equipment or line segment.

[0084] The following edges can be established between the corresponding nodes: Electrical connection edge is used to characterize the actual electrical connection relationship between equipment nodes or line segments; Spatial relationship edge, used to characterize the relative distance and spatial orientation relationship between equipment or towers; Image adjacency edges are used to characterize the proximity relationship of abnormal visual objects in the same or adjacent images; Power supply direction edge is used to characterize the upstream and downstream relationships between nodes; The switch segment edge is used to characterize the segmented switch and its impact on the connectivity of adjacent segments.

[0085] Therefore, visual features are no longer ordinary inputs that are independent of topological data, but are entered into the fault evidence subgraph through node binding and edge relationships, enabling subsequent models to aggregate visual evidence along actual equipment relationships and power supply paths.

[0086] Each node's features can include visual features, electrical features, and spatial features. Visual features can include texture, edges, shape, and saliency features in the grayscale enhanced image; electrical features can include equipment type, rated parameters, and switch status; spatial features can include tower coordinates, relative distance, and line direction.

[0087] Each edge can be configured with one or more edge attributes, including electrical path length, sectionalizing switch blocking marker, inspection view correlation, and historical fault propagation correlation.

[0088] The electrical path length can be determined by the number of line segments contained in the shortest power supply path between two nodes in the current topology, or by the corresponding actual line length. The blocking marker of a segment switch can be determined based on the switch's open / closed state and whether it is located at the current segment boundary. When the switch is open, it indicates that the corresponding path is blocked. The correlation of inspection viewpoints can be determined based on the viewpoint deviation between adjacent frames, the spatial overlap of images, and the consistency of tower order. It is used to reflect different viewpoint observations of whether two visual targets may be the same fault signs. The correlation of historical fault propagation can be determined based on the co-occurrence of the same equipment combination, similar topology conditions, and similar fault types in historical fault records.

[0089] The aforementioned edge attributes, after being processed with a unified dimension, can be input into a graph neural network model to determine the degree of information transmission between nodes.

[0090] When combining nodes, edges, and their corresponding visual features, device operating status, and topological relationships, a graph data structure can be used to store node sets, edge sets, node attributes, edge attributes, and suspected fault segment identifiers. The suspected fault segment identifier indicates the candidate location object corresponding to the current fault evidence subgraph, allowing the model to output the fault probability of that suspected fault segment.

[0091] Fault evidence Figure 1 On the one hand, it serves as a direct input to the graph neural network model, enabling the model to jointly analyze visual anomalies, equipment operating status, and feeder topology; on the other hand, it is used to limit the subsequent local transmission range and information transmission constraints, so that the model prioritizes aggregating evidence around suspected fault sections that conforms to the actual power supply path.

[0092] In some embodiments, inputting a fault evidence subgraph of at least one suspected fault segment into a pre-trained graph neural network model to obtain the fault probability corresponding to each suspected fault segment, and determining the target fault segment based on the fault probability, includes: for each suspected fault segment, extracting node features, edge attributes, and topological relationships between nodes from the corresponding fault evidence subgraph; determining the local transmission range associated with the suspected fault segment based on the position of the suspected fault segment in the corresponding fault evidence subgraph, and determining the information transmission constraints between nodes based on the electrical path length, power supply direction, and segmented switch status between nodes within the local transmission range; inputting the node features, edge attributes, topological relationships, and information transmission constraints into the pre-trained graph neural network model to obtain the fault probability corresponding to each suspected fault segment; comparing each suspected fault segment based on the fault probability corresponding to each suspected fault segment, and determining the suspected fault segment that meets the preset fault probability condition as the target fault segment.

[0093] Among them, node features can be extracted according to node type. For device nodes, device type, device parameters, device operating status, spatial location, and visual features associated with them can be extracted; for abnormal visual object nodes, texture features, morphological features, brightness distribution features, and contextual features can be extracted; for switch nodes, switch open / closed status and the segment boundary they are located in can be extracted.

[0094] Edge attributes can include electrical path length, sectionalizing switch blocking status, inspection view relevance, and historical fault propagation correlation. Topological relationships are used to indicate which nodes have edges between them and the type of relationship between the edges.

[0095] The node features, edge attributes, and topological relationships mentioned above are all derived from the fault evidence subgraph of the corresponding suspected faulty section.

[0096] The device node, line edge node, or segment center node corresponding to the suspected faulty segment in the fault evidence subgraph can be used as the central node. This central node is then expanded outwards according to a preset topology hop count to obtain a set of local nodes participating in the reasoning for the current suspected faulty segment. The range corresponding to this set of local nodes is the local propagation range.

[0097] "Topology hop count" refers to the number of edges traversed from the central node along the edges in the fault evidence subgraph to reach other nodes. For example, a one-hop node is a node directly connected to the central node, and a two-hop node is a node reached via an intermediate node. The preset topology hop count can be set according to the range of upstream and downstream devices to be included as needed.

[0098] The local propagation range is used to limit the evidence that the model needs to focus on, so that the suspected faulty section itself, adjacent equipment, adjacent line segments and directly related switch nodes have high inference relevance, while nodes that are far away or isolated by switches will not have an unreasonable impact on the current candidate section.

[0099] When using a graph transformer network in a graph neural network model, local region attention can be set based on the local propagation range. Specifically, a higher initial attention bias can be set for node pairs within a local node set during attention calculation, causing the model to prioritize visual evidence, operational status, and topological relationships around suspected faulty sections.

[0100] If a node and a suspected faulty section are located on the same line segment, adjacent line segment, or within a set of directly related switch nodes, the information aggregation priority between the node and the central node can be increased. If the node is far away or has been isolated by segmented switches, its initial impact can be reduced. The role of local area attention is to limit which evidence is analyzed in a focused manner, and the results formed continue to participate in the fault probability calculation, rather than being independent intermediate results.

[0101] For any pair of nodes within the local transmission range, the power supply path between them can be searched in the current topology of the corresponding feeder.

[0102] Electrical path length can be expressed as the number of line segments the power supply path traverses or the actual line length. The more line segments the path traverses or the longer the actual line, the weaker the electrical connection between the two nodes is generally.

[0103] Power supply direction consistency is used to determine whether the direction of information transmission between nodes matches the current upstream-to-downstream power supply direction of the feeder. For example, the status of upstream equipment can be associated with downstream equipment along the effective power supply path; for node relationships that clearly conflict with the current power supply direction, the degree of information transmission should be reduced.

[0104] The segment switch status is used to determine whether the connection path between two nodes is blocked. If the connection path passes through a segment switch that is in the open state, the two nodes do not form a continuous power supply path under the current operating mode; if the switch is in the closed state, the degree of information transmission can be further determined by combining the path length and power supply direction.

[0105] It is important to emphasize that the electrical path length, power supply direction, and sectional switch status should be determined based on the current topology at the time of inspection image acquisition or fault location, rather than solely based on a fixed design topology, in order to adapt to the operating mode after interconnection transfer or sectional switching.

[0106] In one implementation example, based on the location of the suspected faulty section in the corresponding fault evidence subgraph, the local transmission range associated with the suspected faulty section is determined. Then, based on the electrical path length, power supply direction, and sectionalizing switch status between nodes within the local transmission range, information transmission constraints between nodes are determined. This includes: using the node associated with the suspected faulty section in the corresponding fault evidence subgraph as the center, expanding outwards according to a preset topology hop count to determine the set of local nodes participating in the reasoning of the suspected faulty section, and defining the range corresponding to the set of local nodes as the local transmission range; for node pairs within the local transmission range, based on the current topology of the corresponding feeder, determining the electrical path length between node pairs, the consistency of power supply direction between node pairs, and the sectionalizing switch status on the connecting path; for a first node pair with a reachable power supply path and consistent power supply direction, determining the information transmission weight of the first node pair based on the corresponding electrical path length; for a second node pair with a disconnected sectionalizing switch or inconsistent power supply direction, reducing the information transmission weight of the second node pair; and generating information transmission constraints based on the information transmission weights of all node pairs.

[0107] For the first node pair where the power supply path is reachable and the power supply direction is consistent, the information transmission weight can be determined based on the corresponding electrical path length. When the electrical path is short, a relatively high information transmission weight can be retained; as the electrical path length increases, the information transmission weight can be gradually reduced, thus forming an information transmission relationship that attenuates according to path distance.

[0108] For second node pairs whose connection paths are interrupted by segmented switches or whose power supply directions are inconsistent, their information transmission weight can be reduced. In one implementation, it can be reduced to a smaller value; when it is necessary to strictly exclude paths without physical significance, information transmission between the node pairs can also be prohibited by masking. The above processing corresponds to a path attenuation masking mechanism, which attenuates or restricts the visibility between nodes according to their electrical and physical significance, rather than allowing all nodes in the fault evidence subgraph to arbitrarily engage in fully connected information exchange.

[0109] The information transmission weights of all node pairs within the local transmission range are organized into information transmission constraints. These information transmission constraints can serve as constraint inputs for node feature aggregation or attention calculation in graph neural networks, controlling the propagation of visual information, device operating status, and topological information along the physically effective path.

[0110] In this way, the local transmission range is used to determine the set of nodes participating in the reasoning, and the information transmission weight is used to determine the degree of transmission between nodes within the set. The two together serve the calculation of the failure probability of the suspected faulty section and will not form isolated features.

[0111] To enable the graph neural network model to simultaneously perceive graph structure, feeder direction, and device arrangement order, composite position codes can also be generated for nodes in the fault evidence subgraph.

[0112] First, a graph structure position code can be generated based on the Laplace spectrum characteristics of the fault evidence subgraph to characterize the relative position of the node in the overall graph structure. Second, a feeder direction code can be generated based on whether the node is located upstream or downstream of the feeder. Third, a sequence position code can be generated based on the tower number, tower arrangement order, or segment sequence code.

[0113] The graph structure position encoding, feed direction encoding, and sequence position encoding are concatenated or fused to form a composite position representation of the nodes. This composite position representation, along with the node features, is input into the graph neural network, enabling the model to determine not only whether two nodes are connected, but also their direction and order within the feed lines.

[0114] The upstream anchor, downstream anchor, segment center index, and segment sequence code in the segment anchor representation can also be used to generate composite position codes, and participate in model inference together with information transmission constraints to maintain the continuous association between the image and the topological alignment results and the graph inference process.

[0115] Furthermore, when multiple candidate mapping objects are obtained, legal path gating can be generated based on power supply direction consistency, tower sequence consistency, path reachability, and cross-frame observation stability.

[0116] Legitimate path gating is used to limit which candidate mapping objects and their associated paths can enter the information aggregation process of the graph neural network. For candidate paths with low consistency, isolated by disconnect switches, or that obviously conflict with the tower sequence, the corresponding gating can be turned off or reduced; for candidate paths with high consistency and physical reachability, their information transmission can be preserved.

[0117] Legitimate path gating can be combined with path attenuation: legitimate path gating first eliminates candidate paths that are physically invalid, and path attenuation then determines the transmission degree of the retained paths according to the electrical path length, thereby reducing the propagation of erroneous evidence caused by visual misdetection or obstruction by nearby devices.

[0118] In a preferred embodiment, the graph neural network model can employ a graph Transformer network. The model takes node features and topological relationships as the base graph input, edge attributes to reflect the electrical, spatial, and visual relationships between nodes, information transmission constraints to limit feature aggregation between nodes, and can further combine local region attention, composite position encoding, and legal path gating to perform multi-layer graph feature updates.

[0119] During each layer of graph feature updates, nodes aggregate information based on other nodes with which they have a valid transitive relationship. Visual features corresponding to abnormal visual objects can enter device nodes along their binding relationships; device operating status can propagate along electrical connection edges and power supply direction edges; segment switch status can restrict information propagation across segments. After multi-layer feature aggregation, segment features representing comprehensive fault evidence for the corresponding suspected fault segment can be obtained.

[0120] Based on the segment features, the model outputs the failure probability corresponding to the suspected faulty segment. The failure probability is used to characterize the degree to which visual evidence, equipment status, and topological relationships in the current fault evidence subgraph jointly support the occurrence of a fault in that segment.

[0121] The model can also output fault type labels, such as arcing, cracking, ablation, foreign object bridging, or wire damage.

[0122] When there is only one suspected faulty section, it can be determined whether the fault probability of the suspected faulty section meets the preset fault probability condition; if it does, it is identified as the target faulty section; if it does not, the result to be reviewed can be output.

[0123] When multiple suspected faulty sections exist, they can be sorted from highest to lowest fault probability. The preset fault probability condition can be that the fault probability reaches a preset threshold, or it can be selecting one or more sections ranked highest based on probability. When a unique location result is required, the suspected faulty section with the highest fault probability that meets the threshold can be identified as the target faulty section; when tiered verification is required on-site, multiple priority verification sections can be output.

[0124] When outputting the target fault section, the corresponding fault type label, key evidence nodes, and high-weight electrical paths can also be output simultaneously. Key evidence nodes are abnormal visual object nodes, equipment nodes, or operating status nodes that contribute significantly to the fault probability; high-weight electrical paths are the main information transmission paths that support the current location result during model inference. The above information can be used for maintenance dispatching, on-site verification, or comparison with historical fault cases to improve the interpretability of the location results.

[0125] The graph neural network model can be pre-trained using historical inspection samples labeled with fault sections. Each training sample can include inspection images, collected correlation information, topology operation information of the corresponding feeder, a fault evidence subgraph constructed based on the training sample, and actual fault section labels.

[0126] During training, following the same process as steps S1 to S4, abnormal visual objects are extracted from the training samples, image and topological alignment is performed, and corresponding fault evidence sub-graphs are constructed. These fault evidence sub-graphs are then input into the graph neural network model to be trained, and the model parameters are updated based on the difference between the model's predicted fault segments and the actual fault segment labels.

[0127] To improve the model's generalization ability under varying operating conditions, topological perturbation enhancements can be applied to the training samples. Topological perturbations can include randomly simulating the closing or opening of interconnecting switches, missing nodes, changes in local edge weights, and segmented reconstruction of the feeder. After each topological perturbation, the image is re-aligned with the topology based on the perturbated topological relationships, and a fault evidence subgraph is constructed, enabling the model to learn to maintain stable localization under conditions of feeder topological changes and information loss.

[0128] Furthermore, visual acquisition perturbations can be applied to the training samples, including shooting viewpoint shift, partial occlusion, and missing adjacent inspection frames. Shooting viewpoint shift is used to simulate changes in UAV flight path or gimbal attitude; partial occlusion is used to simulate foliage, equipment components, or environmental targets obscuring fault areas; missing adjacent inspection frames is used to simulate communication frame drops or incomplete acquisition.

[0129] By jointly introducing topological perturbations and visual acquisition perturbations, the model can simultaneously learn the mapping rules of stable segments under conditions of changes in operating mode and changes in image acquisition, reducing the model's dependence on fixed feeder topology, fixed viewing angle, or complete continuous frames.

[0130] It should be noted that topology disturbances should synchronously update node connections, power supply directions, and switch segment relationships in the fault evidence subgraph to avoid simply changing the input topology while still using the graph labels or path constraints before the disturbance. For disturbances that do not change the semantics of the actual fault segment, the original fault segment labels can be retained; for segment reconstructions that change the original fault segment identifiers, the corresponding labels should be converted synchronously.

[0131] Based on the same inventive concept, this application also provides a power distribution network fault location system for implementing the power distribution network fault location method described above.

[0132] The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more distribution network fault location system embodiments provided below can be found in the limitations of the distribution network fault location method above, and will not be repeated here.

[0133] like Figure 3 As shown in the figure, this application provides a power distribution network fault location system, which includes: The information acquisition module 100 is used to acquire inspection images of the distribution network, the acquisition association information corresponding to the inspection images, and the topology operation information of the corresponding feeders; The grayscale enhancement module 200 is used to perform adaptive grayscale enhancement on the inspection image and extract abnormal visual objects and their visual features from the enhanced inspection image. The suspected fault location module 300 is used to perform image and topology alignment between abnormal visual objects and corresponding topology objects in the feeder based on the collected correlation information and topology operation information, and to determine at least one suspected fault section. The evidence subgraph construction module 400 is used to construct a fault evidence subgraph by associating the visual features, equipment operating status and topological relationships corresponding to the suspected fault section with the suspected fault section as the center. The target fault location module 500 is used to input the fault evidence subgraph of at least one suspected fault section into a pre-trained graph neural network model to obtain the fault probability corresponding to each suspected fault section, and to determine the target fault section based on the fault probability.

[0134] In some embodiments, the grayscale enhancement module 200 is used for: Brightness normalization and noise suppression are performed on the inspection images to obtain the basic grayscale image; Based on the brightness, texture and edge features of different local regions in the base grayscale image, the grayscale enhancement parameters of the corresponding local regions are determined, and the base grayscale image is locally enhanced based on the grayscale enhancement parameters to obtain the enhanced inspection image. Based on the enhanced inspection images, candidate fault regions are extracted, and the abnormal visual objects are identified by combining the correlation between the candidate fault regions and the corresponding equipment targets. Extract texture features, morphological features, brightness distribution features, and contextual features of abnormal visual objects to obtain visual features.

[0135] In some embodiments, the grayscale enhancement module 200 is used for: Edge response information and hotspot saliency information corresponding to the enhanced inspection image are extracted, and the edge response information and hotspot saliency information are fused with the enhanced inspection image to obtain a fault sensitivity map; Identify equipment targets from the fault sensitivity map and extract fault candidate regions; Based on the relative relationship between the candidate fault regions and the equipment target, the candidate fault regions are filtered, and the remaining candidate fault regions are identified as abnormal visual objects.

[0136] In some embodiments, the suspected fault location module 300 is used for: Based on the collected correlation information, the field of view coverage area corresponding to the inspection image is determined, and based on the topology operation information, candidate topology objects within the field of view coverage area are determined from the corresponding feeder. Based on the position, size, orientation, and spatial relationship between the abnormal visual object and the neighboring device in the inspection image, the abnormal visual object is matched with the candidate topology object to obtain at least one candidate mapping object. Based on the power supply direction of the corresponding feeder, the equipment arrangement sequence, the upstream and downstream relationship of the equipment, and the segment switch boundary, the candidate mapping object is checked for topology consistency. Based on the results of the topology consistency check, the feeder segments corresponding to the candidate mapping objects that meet the topology consistency conditions are identified as suspected fault segments.

[0137] In some embodiments, the evidence subgraph construction module 400 is used for: Centered on the suspected fault section, determine the local topology range associated with the suspected fault section along the upstream and downstream directions of the corresponding feeder; Extract equipment objects, equipment operating status and spatial location information associated with suspected fault sections within the local topology range, and bind abnormal visual objects and their visual features to the corresponding equipment objects or line segments; Based on the binding results, nodes representing device objects, abnormal visual objects, device operating status, and spatial location are constructed, and edges are established between the corresponding nodes according to the electrical connection relationships, spatial relationships, image adjacency relationships, power supply direction relationships, and switch segmentation relationships within the local topology. By combining nodes, edges, and their corresponding visual features, device operating status, and topological relationships, a fault evidence subgraph corresponding to the suspected faulty section is obtained.

[0138] In some embodiments, the target fault location module 500 is used for: For each suspected faulty section, node features, edge attributes, and topological relationships between nodes are extracted from the corresponding fault evidence subgraph. Based on the location of the suspected fault section in the corresponding fault evidence sub-graph, determine the local transmission range associated with the suspected fault section, and determine the information transmission constraints between nodes based on the electrical path length, power supply direction and sectional switch status between nodes within the local transmission range. By inputting node features, edge attributes, topological relationships, and information transmission constraints into a pre-trained graph neural network model, the fault probability corresponding to each suspected faulty section is obtained. Based on the failure probability corresponding to each suspected failure section, the suspected failure sections are compared, and the suspected failure sections that meet the preset failure probability conditions are determined as the target failure sections.

[0139] In some embodiments, the target fault location module 500 is used for: Centered on the nodes associated with the suspected fault section in the corresponding fault evidence subgraph, expand outward according to the preset topological hop count to determine the local node set participating in the reasoning of the suspected fault section, and determine the range corresponding to the local node set as the local propagation range; For node pairs within the local transmission range, the electrical path length between node pairs, the consistency of power supply direction between node pairs, and the status of segment switches on the connection path are determined based on the current topology of the corresponding feeder. For a first node pair with a reachable power supply path and consistent power supply direction, the information transmission weight of the first node pair is determined according to the corresponding electrical path length. For a second node pair with a disconnected sectionalizing switch or inconsistent power supply direction, the information transmission weight of the second node pair is reduced. Information transmission constraints are generated based on the information transmission weights of all node pairs.

[0140] like Figure 4 As shown in the figure, this application provides an electronic device. The electronic device 10 includes a memory 20 and a processor 30. The memory 20 stores a computer program. When the computer program is executed by the processor 30, the processor 30 performs the steps of the power distribution network fault location method as described in the above embodiment.

[0141] This application provides a computer-readable storage medium storing a computer program thereon, which, when executed, implements the steps of the power distribution network fault location method as described in the above embodiments.

[0142] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, electronic devices, and computer storage media described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0143] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0144] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0145] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0146] In the embodiments provided by this invention, it should be understood that the disclosed systems, electronic devices, computer storage media, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between devices or units through some interfaces, and may be electrical, mechanical, or other forms.

[0147] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0148] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0149] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for executing all or part of the steps of the methods described in the various embodiments of the present invention through a computer device (which may be a personal computer, a server, or a network device, etc.). The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0150] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for locating faults in a power distribution network, characterized in that, The method includes: Acquire inspection images of the power distribution network, the acquisition association information corresponding to the inspection images, and the topology operation information of the corresponding feeders; The inspection image is subjected to grayscale adaptive enhancement, and abnormal visual objects and their visual features are extracted from the enhanced inspection image. Based on the collected association information and the topology operation information, the abnormal visual object is matched with the topology object in the corresponding feeder to determine at least one suspected faulty section. Centered on the suspected faulty section, the visual features, equipment operating status and topological relationships corresponding to the suspected faulty section are associated to construct a fault evidence subgraph; The fault evidence subgraph of at least one of the suspected fault sections is input into a pre-trained graph neural network model to obtain the fault probability corresponding to each of the suspected fault sections, and the target fault section is determined based on the fault probability.

2. The method for locating faults in a power distribution network according to claim 1, characterized in that, The step of performing grayscale adaptive enhancement on the inspection image and extracting abnormal visual objects and their visual features from the enhanced inspection image includes: The inspection image is normalized for brightness and suppressed for noise to obtain a basic grayscale image; Based on the brightness, texture and edge features of different local regions in the base grayscale image, grayscale enhancement parameters for the corresponding local regions are determined, and the base grayscale image is locally enhanced based on the grayscale enhancement parameters to obtain the enhanced inspection image. Based on the enhanced inspection image, a fault candidate region is extracted, and the abnormal visual object is determined by combining the correlation between the fault candidate region and the corresponding equipment target. The texture features, morphological features, brightness distribution features, and contextual features of the abnormal visual object are extracted to obtain the visual features.

3. The method for locating faults in a power distribution network according to claim 2, characterized in that, The step of extracting fault candidate regions from the enhanced inspection image and determining the abnormal visual object by combining the correlation between the fault candidate regions and the corresponding equipment targets includes: The edge response information and hotspot saliency information corresponding to the enhanced inspection image are extracted, and the edge response information and hotspot saliency information are fused with the enhanced inspection image to obtain a fault sensitivity map; Identify equipment targets from the fault sensitivity map and extract fault candidate regions; Based on the relative relationship between the fault candidate region and the device target, the fault candidate region is filtered, and the retained fault candidate region is determined as the abnormal visual object.

4. The method for locating faults in a power distribution network according to claim 1, characterized in that, The step of performing image-topology alignment between the abnormal visual object and the corresponding topology object in the feeder, based on the collected correlation information and the topology operation information, to determine at least one suspected faulty section includes: The field of view coverage area corresponding to the inspection image is determined based on the collected association information, and candidate topology objects within the field of view coverage area are determined from the corresponding feeder based on the topology operation information. Based on the position, size, orientation, and spatial relationship between the abnormal visual object and neighboring devices in the inspection image, the abnormal visual object is matched with the candidate topology object to obtain at least one candidate mapping object. Based on the power supply direction of the corresponding feeder, the equipment arrangement order, the upstream and downstream relationship of the equipment, and the segmented switch boundary, the candidate mapping object is checked for topology consistency. Based on the results of the topology consistency check, the feeder segments corresponding to the candidate mapping objects that meet the topology consistency conditions are identified as the suspected fault segments.

5. The method for locating faults in a power distribution network according to claim 1, characterized in that, The step involves using the suspected faulty section as the center, associating the visual features, equipment operating status, and topological relationships corresponding to the suspected faulty section to construct a fault evidence sub-graph, including: Centered on the suspected fault section, determine the local topology range associated with the suspected fault section along the upstream and downstream directions of the corresponding feeder; Extract the equipment objects, equipment operating status and spatial location information associated with the suspected fault section from the local topology range, and bind the abnormal visual objects and their visual features to the corresponding equipment objects or line segments; Based on the binding results, nodes representing the device object, abnormal visual object, device operating status, and spatial location are constructed, and edges are established between the corresponding nodes according to the electrical connection relationship, spatial relationship, image adjacency relationship, power supply direction relationship, and switch segmentation relationship within the local topology range; By combining the nodes, edges, and corresponding visual features, device operating status, and topological relationships, a fault evidence subgraph corresponding to the suspected faulty section is obtained.

6. The method for locating faults in a power distribution network according to claim 1, characterized in that, The step of inputting the fault evidence subgraph of at least one of the suspected fault sections into a pre-trained graph neural network model to obtain the fault probability corresponding to each of the suspected fault sections, and determining the target fault section based on the fault probability, includes: For each suspected faulty section, node features, edge attributes, and topological relationships between nodes are extracted from the corresponding fault evidence subgraph. Based on the location of the suspected fault section in the corresponding fault evidence sub-graph, the local transmission range associated with the suspected fault section is determined, and based on the electrical path length, power supply direction and segment switch status between nodes within the local transmission range, the information transmission constraints between nodes are determined. The node features, edge attributes, topological relationships, and information transmission constraints are input into the pre-trained graph neural network model to obtain the fault probability corresponding to each suspected fault segment. Based on the fault probability corresponding to each of the suspected fault sections, the suspected fault sections are compared, and the suspected fault sections that meet the preset fault probability conditions are determined as the target fault sections.

7. The method for locating faults in a power distribution network according to claim 6, characterized in that, The step of determining the local transmission range associated with the suspected fault section based on its position in the corresponding fault evidence sub-graph, and determining the information transmission constraints between nodes based on the electrical path length, power supply direction, and sectionalizing switch status within the local transmission range, includes: Centered on the node associated with the suspected fault segment in the corresponding fault evidence subgraph, expand outward according to a preset topology hop count to determine the local node set participating in the reasoning of the suspected fault segment, and determine the range corresponding to the local node set as the local propagation range; For the node pairs within the local transmission range, the electrical path length between the node pairs, the consistency of the power supply direction between the node pairs, and the segment switch status on the connection path are determined according to the current topology of the corresponding feeder. For a first node pair where the power supply path is reachable and the power supply direction is consistent, the information transmission weight of the first node pair is determined according to the corresponding electrical path length. For a second node pair where the connection path is interrupted by a sectionalizing switch or the power supply direction is inconsistent, the information transmission weight of the second node pair is reduced. The information transmission constraints are generated based on the information transmission weights of all the node pairs.

8. A power distribution network fault location system, characterized in that, The system includes: The information acquisition module is used to acquire inspection images of the distribution network, the acquisition association information corresponding to the inspection images, and the topology operation information of the corresponding feeders; The grayscale enhancement module is used to perform adaptive grayscale enhancement on the inspection image and extract abnormal visual objects and their visual features from the enhanced inspection image. The suspected fault location module is used to perform image and topology alignment between the abnormal visual object and the corresponding topology object in the feeder based on the collected association information and the topology operation information, and to determine at least one suspected fault section. The evidence subgraph construction module is used to construct a fault evidence subgraph by associating the visual features, equipment operating status and topological relationships corresponding to the suspected fault section with the suspected fault section as the center. The target fault location module is used to input the fault evidence subgraph of at least one of the suspected fault sections into a pre-trained graph neural network model to obtain the fault probability corresponding to each of the suspected fault sections, and to determine the target fault section based on the fault probability.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the power distribution network fault location method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the steps of the power distribution network fault location method as described in any one of claims 1-7.

Images