Railway power supply unmanned aerial vehicle inspection method based on deep learning

Abstract

The application discloses a railway work, electricity and supply integrated unmanned aerial vehicle inspection method based on deep learning, relates to the technical field of railway inspection, and comprises the following steps: acquiring inspection images, mileage, facility account books and a history library, generating inspection basic data; calling historical images and registering with the mileage, generating aligned image pairs; identifying work, electricity and supply facility objects, generating object partition maps; constructing an improved ChangeFormer, generating double-phase Token sequences; generating difference features through anchor point alignment and differential coding; fusing abnormal shapes and stripping false changes, generating attribution results and disposal results and rewriting the history library. The application utilizes computer deep vision and the improved ChangeFormer, realizes work, electricity and supply abnormality identification and disposal, and has the advantages of accurate attribution, low false positives and high closed-loop efficiency.

Description

Technical Field

[0001] This invention relates to the field of railway inspection technology, and in particular to a deep learning-based unmanned aerial vehicle (UAV) inspection method for integrated railway power supply and electrical systems. Background Technology

[0002] With the expansion of railway lines, communication signaling equipment, and traction power supply equipment, integrated inspection of power supply and electrical systems is gradually evolving from manual and track vehicle inspections to drone inspections and intelligent image recognition. Current industry solutions typically utilize drones to collect images of lines, bridges, tunnels, signal lights, overhead contact line supports, insulators, droppers, and power supply cabinets, then identify appearance anomalies using target detection or semantic segmentation models. While these solutions improve inspection coverage efficiency, most still rely on single-image recognition, making it difficult to accurately determine the changing trends of the same facility across different inspection cycles.

[0003] To compensate for the limitations of single-shot identification, existing technologies include schemes that utilize historical image comparison, change detection models, or manual verification to determine facility changes. However, ordinary change detection models are mostly geared towards general images or remote sensing scenarios, typically only outputting the changed area without uniformly linking line mileage, power supply facility ledgers, mileage-based registration, facility object zoning, and historical inspection results. Due to factors such as differences in drone shooting posture, lighting and shadows, vegetation swaying, and temporary obstructions, existing solutions are prone to misjudging non-structural appearance disturbances as anomalies, and struggle to further attribute changed areas to track maintenance anomalies, electrical maintenance anomalies, or power supply anomalies, resulting in inspection results still requiring manual interpretation and processing.

[0004] Therefore, how to provide a deep learning-based method for unmanned aerial vehicle (UAV) inspection of railway power supply systems is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based unmanned aerial vehicle (UAV) inspection method for integrated railway power supply and electrical systems. This invention utilizes computer deep vision and an improved ChangeFormer to achieve anomaly identification and handling in power supply and electrical systems, and has the advantages of accurate attribution, low false alarm rate, and high closed-loop efficiency.

[0006] The railway power supply integrated unmanned aerial vehicle (UAV) inspection method based on deep learning according to an embodiment of the present invention includes the following steps:

[0007] Acquire current drone inspection images, line mileage data, power supply facility ledgers, and historical inspection benchmark databases to generate a basic data set for power supply inspection.

[0008] Based on the basic data set of power supply inspection, historical reference images are retrieved, and the current drone inspection images and historical reference images are registered at the same mileage to generate image pairs aligned at the same mileage.

[0009] Identify power supply facilities in image pairs aligned with the same mileage and generate a partition map of power supply facilities objects.

[0010] An improved ChangeFormer model is constructed, which processes image pairs aligned with the same mileage and partition maps of power supply facilities through a component-aware dual-temporal tokenization layer to generate the current component temporal token sequence and the historical component temporal token sequence.

[0011] The current component temporal token sequence and the historical component temporal token sequence are processed by a stable anchor-guided dual-temporal alignment attention layer and a bidirectional differential interaction Transformer encoding layer to generate dual-temporal component difference features.

[0012] The structural equipment abnormality morphology constraint difference fusion layer, pseudo-change stripping branch and power supply abnormality prototype query decoding layer process the dual-phase component difference characteristics, generate abnormal change mask and power supply abnormality type attribution results, generate inspection and handling results by continuous multi-period trend memory unit, and write back to the historical inspection benchmark library.

[0013] Optionally, the generation of the basic data set for power supply inspection specifically includes:

[0014] The drone collects images of the current inspection along the inspection route of the railway section to be inspected, and attaches the shooting location, shooting time, shooting posture and lens field of view;

[0015] Based on the shooting location, read the route mileage data, classify the current drone inspection image into the corresponding route segment, and obtain the start and end mileage of the route segment;

[0016] Based on the line section and start and end mileage, the corresponding power supply facility objects are extracted from the power supply facility ledger and mapped to the current drone inspection image;

[0017] Based on the power supply facility objects, obtain the historical benchmark images and historical inspection results of the same power supply facility object from the historical inspection benchmark database;

[0018] The current UAV inspection images, line sections, start and end mileage, power supply facilities, historical reference images, and historical inspection results are collected according to the same power supply facility object to generate a basic data set for power supply inspection.

[0019] Optionally, the generation of the same mileage aligned image pairs specifically includes:

[0020] Based on the line sections, start and end mileage and power supply facility objects in the power supply inspection basic data set, select historical benchmark images containing the same power supply facility object from the historical inspection benchmark database;

[0021] Read the shooting position, shooting posture and lens field of view of the current drone inspection image and the historical reference image, and crop the current inspection field of view and the historical reference field of view according to the start and end mileage of the line section;

[0022] Extract the track centerline, bridge railing edge, tunnel portal edge, drainage ditch straight edge, sound barrier post and protective fence post within the current inspection field of view and historical reference field of view to generate a stable registration reference;

[0023] Based on the stable registration benchmark, perform same-mileage registration on the current inspection field of view and the historical benchmark field of view to generate a dual-time relative aligned field of view;

[0024] The current image region and the historical image region in the dual-time relative alignment field of view are cropped at the same image size, and the corresponding power supply facility objects and line sections are attached to generate image pairs aligned with the same mileage.

[0025] Optionally, the generation of the power supply facility object zoning map specifically includes:

[0026] Input the current image block and the historical image block in the image pair aligned with the same mileage into the power supply facility object segmentation network to generate the initial partitioning result of the current object and the initial partitioning result of the historical object.

[0027] Based on the power supply facility ledger, the classification of the segmented areas in the initial partitioning results of the current object and the initial partitioning results of the historical object is corrected to generate the current object classification partitioning result and the historical object classification partitioning result.

[0028] The initial partitioning results of the current object and the initial partitioning results of the historical objects are corrected for dual-phase consistency according to the same power supply facility object, and a consistent partitioning result for the object is generated.

[0029] Perform boundary refinement on the consistent partitioning results of objects to generate object boundary refinement results;

[0030] Map the refined object boundary results back to the same mileage aligned image pair, so that the current image block and the historical image block share the same pixel coordinates for the power supply facility object partition map.

[0031] Optionally, the generation of the current component time phase token sequence and the historical component time phase token sequence specifically includes:

[0032] An improved ChangeFormer model is constructed, with a component-aware dual-temporal tokenization layer set at the model input end to receive image pairs aligned with the same mileage and partition maps of power supply facility objects;

[0033] The current image block and the historical image block are divided into segments of the current image block and segments of the historical image block according to the same pixel grid.

[0034] Based on the partition map of power supply facilities objects, read the power supply facility objects corresponding to each image block segment and generate component attribution results;

[0035] Embed and merge the current image block fragment, historical image block fragments and component attribution results to generate the current component temporal token and the historical component temporal token;

[0036] Arrange the current component phase token and the historical component phase token according to the same power supply facility object and the same pixel grid to generate the current component phase token sequence and the historical component phase token sequence.

[0037] Optionally, the improved ChangeFormer model includes a component-aware dual-temporal tokenization layer, a stable anchor-guided dual-temporal alignment attention layer, a bidirectional differential interaction Transformer encoding layer, a structural equipment abnormal morphology constraint difference fusion layer, a pseudo-change stripping branch, an industrial power supply abnormal prototype query decoding layer, and a continuous multi-period trend memory unit, all connected in sequence.

[0038] Optionally, the generation of the difference features of the dual-phase components specifically includes:

[0039] Based on the zoning map of power supply facilities, stable anchor tokens are extracted from the current component phase token sequence and the historical component phase token sequence.

[0040] By correcting the local misalignment between the current component time phase token sequence and the historical component time phase token sequence through a dual-time phase alignment attention layer guided by stable anchor points, the aligned current component time phase token sequence and historical component time phase token sequence are obtained.

[0041] The aligned current component phase token sequence and historical component phase token sequence are input into the bidirectional differential interactive Transformer encoding layer to extract the current-to-historical change response and the historical-to-current change response;

[0042] The current-to-historical change response, the historical-to-current change response, and the differential residual are fused to generate the dual-temporal component difference characteristics.

[0043] Optionally, the generation of the inspection and handling results specifically includes:

[0044] By fusing the difference features of dual-phase components according to the objects and changes of power supply facilities in the structural equipment anomaly constraint difference fusion layer, candidate difference features of structural equipment anomalies are generated.

[0045] Non-structural appearance perturbations are separated from the candidate difference features of structural equipment anomalies by pseudo-change stripping branch, and a pseudo-change suppression map is generated.

[0046] By querying the abnormal prototypes in the power supply anomaly layer and matching the structural equipment anomaly candidate difference features after pseudo-change suppression, anomaly change masks and power supply anomaly type attribution results are generated.

[0047] The continuous multi-period trend memory unit compares the abnormal change mask and the attribution results of power supply anomaly types with the historical inspection results for the same power supply facility object, and generates inspection and handling results.

[0048] The inspection and handling results will be written back to the historical inspection benchmark database, and the current drone inspection image after the handling is completed will be used as the historical benchmark image of the corresponding power supply facility.

[0049] Optionally, the abnormal prototype in the power supply anomaly prototype query decoding layer is an anomaly query template used to retrieve the structural equipment anomaly candidate difference features after pseudo-change suppression. It includes engineering anomaly prototypes, electrical anomaly prototypes, and power supply anomaly prototypes. Each anomaly prototype includes the applicable power supply facility object, the corresponding change form, the pixel scale of the change area, the change direction, and the mask decoding content. The power supply anomaly prototype query decoding layer restores the image block fragments that have completed prototype matching to the pixel coordinates of the image pairs aligned with the same mileage, forming an anomaly change mask, and forms the power supply anomaly type attribution result with the matched anomaly prototype and the corresponding power supply facility object.

[0050] The beneficial effects of this invention are:

[0051] This invention unifies the organization of current UAV inspection images, line mileage data, power supply facility ledgers, and historical inspection benchmark databases. This allows UAV images to move beyond single-shot identification and instead establish correspondences with specific line sections, specific power supply facilities, and historical inspection results. By generating aligned image pairs through mileage registration and further generating a partitioned map of power supply facility objects, current and historical images can be compared under the same facility object, line location, and pixel coordinates. This reduces misjudgments caused by differences in shooting location, field of view, and unclear object attribution, improving the comparability and accuracy of integrated power supply inspection data.

[0052] The improved ChangeFormer model constructed in this invention introduces a component-aware dual-temporal tokenization layer, a stable anchor-guided dual-temporal alignment attention layer, and a bidirectional differential interactive Transformer encoding layer on the basis of ordinary dual-temporal change detection. This enables the binding of power supply facility objects during the feature extraction stage, the correction of local misalignments using stable structural boundaries during the feature alignment stage, and the simultaneous extraction of change responses from the current to the past and from the past to the current stage during the feature encoding stage. Therefore, compared to conventional change detection methods that only output the changed area, this invention can more stably distinguish between real structural equipment anomalies and pseudo-changes such as lighting, shadows, vegetation swaying, temporary occlusion, and residual deviations in shooting angles, reducing false alarms during inspections and improving the detection reliability of the anomaly mask.

[0053] This invention further attributes the changed areas to engineering anomalies, electrical anomalies, and power supply anomalies through a structural equipment anomaly morphology constraint difference fusion layer, pseudo-change stripping branch, and power supply anomaly prototype query decoding layer. The resulting anomalies are then attributed to engineering, electrical, and power supply anomalies by a continuous multi-period trend memory unit combined with historical inspection results. Thus, the inspection results not only identify the location of the anomaly but also provide the anomaly type, corresponding facility object, trend marker, and handling level. This reduces the workload of manual review and interpretation, improves the efficiency of collaborative inspections of engineering, electrical, and power supply systems, anomaly trend judgment, and on-site handling decisions, and allows the historical inspection benchmark database to be continuously updated with the handling results, forming a traceable closed-loop inspection mechanism. Attached Figure Description

[0054] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0055] Figure 1 The flowchart shows the railway power supply integrated unmanned aerial vehicle inspection method based on deep learning proposed in this invention.

[0056] Figure 2 This is a flowchart illustrating the component token generation process of the deep learning-based UAV inspection method for integrated railway power supply and electrical systems proposed in this invention.

[0057] Figure 3 This is a flowchart illustrating the anomaly prototype decoding process of the railway power supply integrated UAV inspection method based on deep learning proposed in this invention. Detailed Implementation

[0058] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0059] refer to Figures 1-3A deep learning-based method for unmanned aerial vehicle (UAV) inspection of integrated railway power supply and electrical systems includes the following steps:

[0060] Acquire current drone inspection images, line mileage data, power supply facility ledgers, and historical inspection benchmark databases to generate a basic data set for power supply inspection.

[0061] Based on the basic data set of power supply inspection, historical reference images are retrieved, and the current drone inspection images and historical reference images are registered at the same mileage to generate image pairs aligned at the same mileage.

[0062] Identify power supply facilities in image pairs aligned with the same mileage and generate a partition map of power supply facilities objects.

[0063] An improved ChangeFormer model is constructed, which processes image pairs aligned with the same mileage and partition maps of power supply facilities through a component-aware dual-temporal tokenization layer to generate the current component temporal token sequence and the historical component temporal token sequence.

[0064] The current component temporal token sequence and the historical component temporal token sequence are processed by a stable anchor-guided dual-temporal alignment attention layer and a bidirectional differential interaction Transformer encoding layer to generate dual-temporal component difference features.

[0065] The structural equipment abnormality morphology constraint difference fusion layer, pseudo-change stripping branch and power supply abnormality prototype query decoding layer process the dual-phase component difference characteristics, generate abnormal change mask and power supply abnormality type attribution results, generate inspection and handling results by continuous multi-period trend memory unit, and write back to the historical inspection benchmark library.

[0066] In this embodiment, the generation of the basic data set for power supply inspection specifically includes:

[0067] The drone collects images of the current inspection along the inspection route of the railway section to be inspected, and attaches the shooting location, shooting time, shooting posture and lens field of view;

[0068] Based on the shooting location, read the route mileage data, classify the current drone inspection image into the corresponding route segment, and obtain the start and end mileage of the route segment;

[0069] Based on the line section and start and end mileage, the corresponding power supply facility objects are extracted from the power supply facility ledger and mapped to the current drone inspection image;

[0070] The engineering and electrical facilities include track maintenance facilities, electrical maintenance facilities, and power supply facilities. Track maintenance facilities include lines, bridges, tunnels, roadbeds, slopes, protective fences, drainage ditches, sound barriers, retaining walls, culverts, track slabs, fasteners, ballast beds, and clearance perimeter facilities. Electrical maintenance facilities include signal lights, switch machine external structures, trackside signal boxes, transponders, communication towers, communication boxes, cable troughs, and cable identification signs. Power supply facilities include overhead contact line supports, cantilever arms, positioners, insulators, droppers, contact wires, catenary wires, power supply cabinets, substation external equipment, cable trenches, and power supply perimeter.

[0071] Based on the power supply facility objects, obtain the historical benchmark images and historical inspection results of the same power supply facility object from the historical inspection benchmark database;

[0072] The current UAV inspection images, line sections, start and end mileage, power supply facilities, historical reference images, and historical inspection results are collected according to the same power supply facility object to generate a basic data set for power supply inspection.

[0073] In this embodiment, the generation of the image pairs aligned with the same mileage specifically includes:

[0074] Based on the line sections, start and end mileage and power supply facility objects in the power supply inspection basic data set, select historical benchmark images containing the same power supply facility object from the historical inspection benchmark database;

[0075] The selection process for the historical reference images is as follows: Using the current UAV inspection image as the center, extend 50m towards the smaller mileage and 50m towards the larger mileage, forming a 100m long line segment; when the power supply facility is a bridge, tunnel, sound barrier, protective fence, drainage ditch, or slope, use the start and end mileages of that power supply facility in the power supply facility register as the line segment; when the power supply facility is a signal, trackside signal box, transponder, contact wire support, cantilever arm, insulator, power supply cabinet, or substation external equipment, use the center mileage of that power supply facility in the power supply facility register as the center, extend 25m towards the smaller mileage and 25m towards the larger mileage, forming a 50m long line segment; retrieve historical reference images from the historical inspection reference database that overlap in line segments and contain the same power supply facility, sort them by shooting posture difference from smallest to largest, and select the historical reference image with the highest sorting value;

[0076] Read the shooting position, shooting posture and lens field of view of the current drone inspection image and the historical reference image, and crop the current inspection field of view and the historical reference field of view according to the start and end mileage of the line section;

[0077] The process of cropping the current inspection field of view and the historical reference field of view is as follows: the image area covering the line section in the current UAV inspection image is taken as the current inspection field of view, and the image area covering the same line section in the historical reference image is taken as the historical reference field of view; when the line section length is 100m, both the current inspection field of view and the historical reference field of view retain the image area within 20m on both sides of the line centerline; when the line section adopts the start and end mileage or the center mileage of the power supply facility, both the current inspection field of view and the historical reference field of view retain the image area within 5m extending outward from the outer contour of the power supply facility.

[0078] Extract the track centerline, bridge railing edge, tunnel portal edge, drainage ditch straight edge, sound barrier post and protective fence post within the current inspection field of view and historical reference field of view to generate a stable registration reference;

[0079] The process of generating the stable registration benchmark is as follows: First, edge extraction is performed on the current inspection field of view and the historical benchmark field of view to obtain linear edges and column edges; then, broken edges with a continuous length of less than 30 pixels in the linear edges are deleted, and short edges with a height of less than 40 pixels in the column edges are deleted; according to the location of the power supply facilities, paired linear edges located in the track area are assigned to the track centerline, continuous linear edges located on both sides of the bridge are assigned to the bridge railing edges, closed edges located in the tunnel entrance outline are assigned to the tunnel portal edges, and closed edges located in the drainage area are assigned to the tunnel portal edges. Parallel edges on both sides of the ditch are classified as straight edges of the drainage ditch, and vertical edges that appear at intervals along the line direction are classified as sound barrier posts or protective fence posts. Edges corresponding to vehicles, personnel, temporary construction objects, swaying vegetation, shadows, and water surface reflections are deleted, and edges corresponding to the same power supply facility object in both periods of the field of view are retained to generate stable registration benchmarks. When the current inspection field of view and the historical benchmark field of view contain catenary support posts, signal posts, or communication towers, the edges of the catenary support posts, signal posts, and communication towers are used as auxiliary stable registration benchmarks.

[0080] Based on the stable registration benchmark, perform same-mileage registration on the current inspection field of view and the historical benchmark field of view to generate a dual-time relative aligned field of view;

[0081] The same-mileage registration first uses the track centerline in the current inspection field of view and the track centerline in the historical reference field of view as the primary alignment objects, ensuring that the positional deviation of the two track centerlines in the image width direction is less than 8 pixels; then, it uses the edges of bridge railings, tunnel portals, drainage ditch lines, sound barrier posts, protective fence posts, contact wire support posts, signal post edges, and communication towers as auxiliary alignment objects, ensuring that the edge direction deviation of the same auxiliary alignment object in the current inspection field of view and the historical reference field of view is less than 3 degrees and the edge endpoint deviation is less than 12 pixels; after alignment is completed, the current inspection field of view and the historical reference field of view are cropped to the same coverage area to generate a dual-time relative alignment field of view;

[0082] The current image region and the historical image region in the dual-time relative alignment field of view are cropped according to the same image size, and the corresponding power supply facility objects and line sections are attached to generate image pairs aligned with the same mileage.

[0083] The same image size cropping process is as follows: the current image area and the historical image area in the dual-time relative alignment field of view are uniformly cropped into 1024×1024 pixel image blocks; when the power supply facility object is a bridge, tunnel, sound barrier, protective fence, drainage ditch, slope, track bed, contact wire or catenary, it is continuously cropped along the line direction with a step size of 512 pixels; when the power supply facility object is a fastener, a local area of ​​track slab, culvert opening, a local area of ​​retaining wall, signal, switch machine external structure, trackside signal box, transponder, communication box, cable identification plate, contact wire support, cantilever arm, positioner, insulator, dropper, power supply cabinet or substation external equipment, a 512×512 pixel image block is generated with the center position of the power supply facility object as the cropping center; the cropped current image block and the corresponding historical image block are paired according to the same power supply facility object and the same line section to generate image pairs aligned with the same mileage.

[0084] In this embodiment, the generation of the power supply facility object zoning map specifically includes:

[0085] Input the current image block and the historical image block in the image pair aligned with the same mileage into the power supply facility object segmentation network to generate the initial partitioning result of the current object and the initial partitioning result of the historical object.

[0086] The power supply facility object segmentation network comprises a backbone feature extraction layer, a multi-scale context fusion layer, an upsampling decoding layer, and a boundary refinement layer connected in sequence. The current image block and historical image block in the same mileage-aligned image pair are adjusted to 1024×1024 pixels and then input into the backbone feature extraction layer. The backbone feature extraction layer extracts track texture, bridge and tunnel contours, slope surfaces, drainage ditch boundaries, sound barrier panels, protective fence grids, local appearance of fasteners, signal contours, external structural contours of switch machines, contours of trackside signal boxes, external contours of transponders, edges of contact wire supports, edges of cantilever arms, edges of insulator skirts, dropper segments, contact wire and catenary wire shapes, power supply cabinet contours, and substation external equipment contours. The multi-scale context fusion layer fuses shallow edge features with deep facility semantic features to generate a facility category response map. The upsampling decoding layer restores the facility category response map to the input image size and outputs the facility category corresponding to each pixel. The boundary refinement layer refines features based on the track centerline, bridge railing edges, tunnel contours, and other parameters. The boundary pixels of adjacent facility categories are assigned to the following parameters: tunnel portal edge, drainage ditch straight edge, sound barrier post, protective fence post, fastener boundary, signal light position boundary, trackside signal box boundary, transponder outer contour, contact wire support edge, cantilever arm edge, insulator skirt edge, dropper segment, contact wire and catenary cable alignment, and power supply cabinet boundary. This process corrects the initial partitioning results for the current object and the initial partitioning results for historical objects. The facility categories include engineering facilities, electrical facilities, and power supply facilities. The categories of engineering facilities include tracks, bridges, tunnels, roadbeds, slopes, protective fences, drainage ditches, sound barriers, retaining walls, culverts, track slabs, fasteners, ballast beds, and clearance perimeter facilities; the categories of electrical facilities include signal lights, switch machine external structures, trackside signal boxes, transponders, communication towers, communication boxes, cable troughs, and cable identification signs; and the categories of power supply facilities include overhead contact line supports, cantilever arms, positioners, insulators, droppers, contact wires, catenary wires, power supply cabinets, substation external equipment, cable trenches, and power supply perimeter fences.

[0087] Based on the power supply facility ledger, the classification of the segmented areas in the initial partitioning results of the current object and the initial partitioning results of the historical object is corrected to generate the current object classification partitioning result and the historical object classification partitioning result.

[0088] The power supply facility object attribution correction involves comparing each segmented region in the current object's initial partitioning result and the historical object's initial partitioning result with the outer contour of the power supply facility object within the same line section in the power supply facility ledger. When the overlap area between a segmented region and the outer contour of a power supply facility object reaches 60% of the segmented region's area, the segmented region is assigned to that power supply facility object. When a segmented region simultaneously covers the outer contours of two power supply facility objects, the power supply facility object with the larger overlap area is assigned to the segmented region. When the area of ​​a segmented region is less than 0.2% of the area of ​​a 1024×1024 pixel image block and does not cover fasteners, drainage ditch edges, sound barrier posts, protective fence posts, signal light positions, transponders, insulators, droppers, locators, or cable identification plates, the segmented region is deleted.

[0089] The initial partitioning results of the current object and the initial partitioning results of the historical objects are corrected for dual-phase consistency according to the same power supply facility object, and a consistent partitioning result for the object is generated.

[0090] The dual-temporal consistency correction first compares the overlap of each segmented region in the initial partitioning result of the current object with the segmented regions of the same facility category in the initial partitioning result of the historical object; when the overlap area of ​​two segmented regions of the same facility category reaches 70% of their combined area, the overlapping area is retained and a gap of less than 16 pixels is filled between their boundaries; when a segmented region that has not appeared in the historical image block appears in the current image block, and the segmented region is within the scope of the power supply facilities registered in the power supply facilities ledger, the segmented region is retained; when a segmented region appears only in a single temporal phase and does not fall within the scope of the power supply facilities registered in the power supply facilities ledger, the segmented region is deleted.

[0091] Perform boundary refinement on the consistent partitioning results of objects to generate object boundary refinement results;

[0092] The boundary refinement process integrates the boundaries between adjacent facility categories into the outer contour of the corresponding power supply facility object: A 5-pixel-width boundary band is used to correct the edges of bridge railings, tunnel portals, drainage ditches, sound barrier posts, protective fence posts, and retaining walls; when the boundaries of adjacent facility categories overlap, 2 pixels on each side of the boundary band's center line are assigned to the larger power supply facility object; when the fastener area overlaps with the track slab area, the fastener area is retained, and the corresponding pixels are deducted from the track slab area; when the drainage ditch area overlaps with the slope area, the area within the straight edge of the drainage ditch is retained as the drainage ditch area; when the signal light area overlaps with the communication tower area, the signal light position area is retained, and the corresponding pixels are deducted from the communication tower area; when the contact wire support and the cantilever area overlap, the cantilever connection area is retained, and the corresponding pixels are deducted from the contact wire support area; when the insulator overlaps with the cantilever or catenary cable area, the insulator skirt area is retained, and the corresponding pixels are deducted from the adjacent linear facility area.

[0093] Map the refined object boundary results back to the same mileage aligned image pair, so that the current image block and the historical image block share the same pixel coordinates for the power supply facility object partition map.

[0094] In this embodiment, the generation of the current component time phase token sequence and the historical component time phase token sequence specifically includes:

[0095] An improved ChangeFormer model is constructed, with a component-aware dual-temporal tokenization layer set at the model input end to receive image pairs aligned with the same mileage and partition maps of power supply facility objects;

[0096] The component-sensing dual-temporal tokenization layer reads the current image block, historical image block, and power supply facility object partition map from the same mileage aligned image pair. The current image block, historical image block, and power supply facility object partition map have the same pixel size and the same pixel coordinates, so that the same pixel position corresponds to the current image content, historical image content, and power supply facility object at the same time.

[0097] The current image block and the historical image block are divided into segments of the current image block and segments of the historical image block according to the same pixel grid.

[0098] The same pixel grid is divided as follows: when the current image block and the historical image block are 1024×1024 pixels, a 16×16 pixel non-overlapping grid is used for segmentation, generating 4096 current image block segments and 4096 historical image block segments respectively; when the current image block and the historical image block are 512×512 pixels, a 16×16 pixel non-overlapping grid is used for segmentation, generating 1024 current image block segments and 1024 historical image block segments respectively; the current image block segments and the historical image block segments maintain consistency in segmentation order, pixel coordinates, and arrangement position.

[0099] Based on the partition map of power supply facilities objects, read the power supply facility objects corresponding to each image block segment and generate component attribution results;

[0100] The process for generating the component attribution result is as follows: For each current image block segment and each historical image block segment, read the pixels of the power supply facility objects within the same pixel range in the power supply facility object partition map; when a single power supply facility object covers 70% of the pixel area of ​​the image block segment, the image block segment is assigned to that power supply facility object; when multiple power supply facility objects cover the image block segment, the power supply facility object with the largest covering pixel area is taken as the attribution object of the image block segment; when the image block segment covers both fasteners and track slabs, the image block segment is assigned to fasteners; when the image block segment covers both drainage ditches and slopes, the image block segment is assigned to drainage ditches; when the image block segment does not contain any power supply facility object pixels, the image block segment is marked as a background segment.

[0101] Embed and merge the current image block fragment, historical image block fragments and component attribution results to generate the current component temporal token and the historical component temporal token;

[0102] The embedding and merging process expands each current image block fragment and each historical image block fragment into a fragment pixel sequence, which includes image grayscale values, color values, and texture gradients within a 16×16 pixel range. The fragment pixel sequence is mapped to a 256-dimensional visual feature, the power supply facility object in the component attribution result is mapped to a 64-dimensional component embedding, the line segment is mapped to a 32-dimensional mileage embedding, and the row and column positions of the image block fragment in the image are mapped to a 32-dimensional position embedding. The visual features, component embedding, mileage embedding, and position embedding are sequentially concatenated and then linearly compressed to obtain a 384-dimensional current component time phase token and a 384-dimensional historical component time phase token.

[0103] Arrange the current component phase token and the historical component phase token according to the same power supply facility object and the same pixel grid to generate the current component phase token sequence and the historical component phase token sequence;

[0104] The arrangement first arranges each image block fragment in the direction from small mileage to large mileage within the line section, and then arranges them in the row and column positions of the image block fragments within the same mileage position; each current component time phase token in the current component time phase token sequence corresponds to the same pixel grid and the same power supply facility object as the historical component time phase token in the historical component time phase token sequence with the same arrangement position; the current component time phase token and the historical component time phase token corresponding to the background fragment are retained at the end of the sequence and do not participate in the subsequent power supply anomaly type attribution result generation.

[0105] In this embodiment, the improved ChangeFormer model includes a component-aware dual-temporal tokenization layer, a stable anchor-guided dual-temporal alignment attention layer, a bidirectional differential interaction Transformer encoding layer, a structural equipment abnormal morphology constraint difference fusion layer, a pseudo-change stripping branch, an industrial power supply abnormal prototype query decoding layer, and a continuous multi-period trend memory unit connected in sequence.

[0106] Among them, the component perception dual-temporal tokenization layer is used to bind image blocks with power supply facilities objects; the stable anchor point guided dual-temporal alignment attention layer is used to correct local misalignment of component features in dual-temporal phases; the bidirectional differential interaction Transformer encoding layer is used to extract change responses bidirectionally between two phase images; the structural equipment abnormality morphology constraint difference fusion layer is used to organize difference features according to the abnormal morphology of engineering, electrical, and power supply; the pseudo-change stripping branch is used to separate non-structural appearance disturbances; the power supply abnormality prototype query decoding layer is used to match different power supply abnormality prototypes; and the continuous multi-phase trend memory unit is used to compare the trends of multiple phase changes of the same power supply facility object.

[0107] The improved ChangeFormer model, compared to existing ChangeFormer models which primarily address the processing of change regions in ordinary dual-temporal image output, introduces a partition map of railway power supply facilities at the input. This transforms current and historical image blocks into current and historical component temporal token sequences, each containing the component's affiliation, track section location, and image block fragment location. This imbues the model with railway power supply facility semantics even before feature extraction. In the feature alignment stage, a stable anchor-guided dual-temporal alignment attention layer is introduced. This layer utilizes track centerlines, bridge railing edges, tunnel portal edges, drainage ditch straight edges, sound barrier pillars, and protective fence pillars to locally correct misalignment of component features between the two phases, reducing pseudo-changes caused by differences in drone shooting angles and residual field-of-view bias. In the feature encoding stage, a bidirectional differential interactive Transformer encoding layer simultaneously extracts current-to-historical and historical-to-current change responses, enabling the extraction of additions, expansions, offsets, occlusions, continuations, and deletions. Loss, fracture, and deformation can all be included in the dual-phase component difference characteristics. In the result decoding stage, through the structural equipment abnormal morphology constraint difference fusion layer, pseudo-change stripping branch, and power supply abnormal prototype query decoding layer, ordinary change areas are further attributed to power supply abnormality types such as crack expansion, slope slippage, drainage blockage, protective facility damage, sound barrier detachment, foreign object intrusion, local changes in track bed, signal tilting and obstruction, switch machine external abnormality, trackside signal box damage, transponder missing and offset, contact wire support tilting, cantilever positioning device deformation, insulator damage and contamination, dropper loosening and breakage, and contact wire catenary morphology abnormality. Thus, this invention not only improves the detection stability of real structural equipment abnormal changes in railway UAV inspection, reduces false alarms caused by lighting, shadows, vegetation swaying, temporary obstruction, and residual deviation of shooting angle, but also enables the inspection results to directly correspond to specific power supply facilities, abnormal change masks, power supply abnormality type attribution results, and inspection handling results, improving the accuracy of multi-phase inspection comparison, abnormal trend judgment, and on-site handling decision-making.

[0108] The training samples for the improved ChangeFormer model are generated from the historical inspection benchmark library. First, images from two adjacent UAV inspections are extracted for the same line section and the same power supply facility object, with an inspection interval of 15 to 180 days. The previous image is used as a historical image block, and the subsequent image as the current image block, and then cropped into 1024×1024 pixel or 512×512 pixel image blocks. For image pairs confirmed as genuine anomalies by manual review, annotators delineate the abnormal change areas pixel by pixel on the current image block and label the corresponding power supply anomaly type, forming genuine anomaly training samples. For image pairs of the same power supply facility object that do not exhibit anomalies but show changes in lighting, shadows, vegetation movement, wet marks after rain, vehicle obstruction, personnel obstruction, or changes in shooting angle, annotators delineate non-abnormal change areas, forming pseudo-change training samples. For image pairs of the same power supply facility object that show continuous appearance in both images and do not exhibit abnormal changes, normal training samples are formed. The training samples consist of real abnormal training samples, pseudo-change training samples, and normal training samples in a 4:3:3 ratio, forming training batches. Each training batch contains 16 pairs of dual-temporal images, with each of the track maintenance, electrical maintenance, and power supply abnormality types appearing at least once in every 10 consecutive training batches. Simultaneously, the annotators mark the stable anchor pixel positions corresponding to the track centerline, bridge railing edge, tunnel portal edge, drainage ditch straight edge, sound barrier column, protective fence column, signal column, contact wire support, and communication tower on the track maintenance, electrical maintenance, and power supply facility object partition map. The stable anchor pixel positions are represented by a 3-pixel wide line strip or the outer contour area of ​​the column. During training, the network parameters outside the track maintenance, electrical maintenance, and power supply abnormality prototype query decoding layer are fixed. Real abnormal training samples of each type of track maintenance, electrical maintenance, and power supply abnormality are input to the structural equipment abnormality morphology constraint difference fusion layer. The 512-dimensional structural equipment abnormality candidate difference features covered by the abnormal change area are extracted and averaged, serving as the initial query vector for the corresponding abnormality prototype.After initializing the abnormal prototype, the following layers are continuously trained: a component-aware dual-temporal tokenization layer, a stable anchor-guided dual-temporal alignment attention layer, a bidirectional differential interactive Transformer encoding layer, a structural equipment abnormal morphology constraint difference fusion layer, a pseudo-change stripping branch, and an industrial power supply abnormal prototype query decoding layer. The training runs for 120 rounds. In the first 20 rounds, only the stable anchor-guided dual-temporal alignment attention layer and the pseudo-change stripping branch are updated to ensure the stable anchor alignment result matches the stable anchor pixel position and the pseudo-change suppression map covers non-abnormal change areas. From round 21 to round 80, the bidirectional differential interactive Transformer encoding layer, the structural equipment abnormal morphology constraint difference fusion layer, and the industrial power supply abnormal prototype query decoding layer are updated simultaneously to ensure the output abnormal change mask matches the manually labeled abnormal changes. The network parameters are updated jointly from round 81 to round 120 to ensure that the attribution results for power supply anomalies match the manually labeled anomaly types. After each training round, validation samples not used in training are used for verification. The improved ChangeFormer model is considered complete when the validation results for eight consecutive rounds simultaneously satisfy the following conditions: average intersection-over-union ratio (IoU) of the anomaly mask is not less than 0.78, the accuracy of power supply anomaly type attribution is not less than 0.90, the false positive rate is not higher than 0.08, and the average positional deviation after Token alignment is not more than 6 pixels. If the above validation results are not simultaneously met after 120 rounds of training, the model parameters from the round with the highest average IoU of the anomaly mask and a false positive rate not higher than 0.10 are selected as the model parameters after training.

[0109] In this embodiment, the generation of the difference features of the dual-phase components specifically includes:

[0110] Based on the zoning map of power supply facilities, stable anchor tokens are extracted from the current component phase token sequence and the historical component phase token sequence.

[0111] The extraction process of the stable anchor point token is as follows: the current component phase token and the historical component phase token of the pixel positions corresponding to the track centerline, bridge railing edge, tunnel portal edge, drainage ditch straight edge, sound barrier column, protective fence column, signal column, contact wire support column and communication tower are used as candidate anchor point tokens; when the stable structure pixel area in the image block segment corresponding to the candidate anchor point token reaches 50% of the pixel area of ​​the image block segment, the candidate anchor point token is retained as a stable anchor point token; when there are fewer than 4 consecutively retained stable anchor point tokens in the same power supply facility object, the component phase tokens that are the same as the power supply facility object and contain linear edges or column edges are added from the adjacent image block segments.

[0112] By correcting the local misalignment between the current component time phase token sequence and the historical component time phase token sequence through a dual-time phase alignment attention layer guided by stable anchor points, the aligned current component time phase token sequence and historical component time phase token sequence are obtained.

[0113] The correction is based on the stable anchor token in the current component phase token sequence. It searches for historical stable anchor tokens within the same power supply facility object's corresponding two rows above and two columns to the left and right in the historical component phase token sequence. For each candidate historical stable anchor token, its 256-dimensional visual features are subtracted dimension-by-dimensional from the current stable anchor token's 256-dimensional visual features, and the absolute value is taken. All absolute values ​​are then added together, and the historical stable anchor token with the smallest sum is selected as the corresponding anchor. Based on the offset of the corresponding anchor point in the row and column positions, the positions of historical component phase tokens within the same power supply facility object are moved, so that the current component phase token and the historical component phase token are re-corresponding within the same pixel grid.

[0114] The aligned current component phase token sequence and historical component phase token sequence are input into the bidirectional differential interactive Transformer encoding layer to extract the current-to-historical change response and the historical-to-current change response;

[0115] The bidirectional differential interactive Transformer coding layer includes four consecutively connected differential interactive coding blocks. Each differential interactive coding block sequentially performs current-to-history interaction, history-to-current interaction, and differential residual preservation. The current-to-history interaction uses the current component phase token as the retrieval initiation object and extracts the addition, expansion, offset, and occlusion responses of the current image relative to the historical image from the historical component phase token of the same power supply facility object. The history-to-current interaction uses the historical component phase token as the retrieval initiation object and extracts the continuation, disappearance, breakage, and deformation responses of the historical structure in the current image from the current component phase token of the same power supply facility object.

[0116] The current-to-historical change response, the historical-to-current change response, and the differential residual are fused to generate dual-temporal component difference characteristics;

[0117] The differential residual is obtained by subtracting the aligned current component phase token and the historical component phase token dimension by dimension and taking the absolute value. The 384-dimensional current-to-historical change response, the 384-dimensional historical-to-current change response, and the 384-dimensional differential residual are sequentially concatenated into 1152-dimensional difference features, which are then linearly compressed to obtain 512-dimensional dual-temporal component difference features. The linearly compressed dual-temporal component difference features retain the corresponding power supply facility objects, line sections, and image block fragment row and column positions, so that each dual-temporal component difference feature can correspond to the specific power supply facility object area in the image pair aligned with the same mileage. The difference features corresponding to the background fragments are deleted, and the dual-temporal component difference features corresponding to the power supply facility objects are retained and arranged in the direction from small mileage to large mileage within the line section.

[0118] In this embodiment, the generation of the inspection and handling results specifically includes:

[0119] By fusing the difference features of dual-phase components according to the objects and changes of power supply facilities in the structural equipment anomaly constraint difference fusion layer, candidate difference features of structural equipment anomalies are generated.

[0120] The structural equipment abnormal morphology constraint difference fusion layer reads the power supply facility object, line section, and image block segment row and column positions corresponding to the difference features of each dual-temporal component. It performs a continuity check on the difference features of dual-temporal components belonging to the same power supply facility object in adjacent image block segments. When three or more consecutive image block segments exhibit elongated enhancement responses in the same direction, they are classified as linear expansion morphology. When four or more consecutive adjacent image block segments within the same power supply facility object form a connected enhancement region, they are classified as planar expansion morphology. When the region corresponding to the historical component temporal token has a complete facility boundary and the region corresponding to the current component temporal token lacks that boundary boundary... When the boundary is defined, it is classified as a block-shaped missing shape; when the edge of the same power supply facility object is offset by more than 12 pixels from the historical image block in the current image block, it is classified as a boundary displacement shape; when a newly added connected area in the facility area around the boundary extends more than 20 pixels towards the track centerline, it is classified as an occlusion intrusion shape; when the area of ​​the dark connected area in the drainage ditch area or the roadbed depression area is larger than the historical image block by more than 15%, it is classified as a water body expansion shape; the structural equipment abnormal shape constraint difference fusion layer rearranges the 512-dimensional dual-temporal component difference features that have completed shape classification according to the same power supply facility object to obtain structural equipment abnormal candidate difference features.

[0121] Non-structural appearance perturbations are separated from the candidate difference features of structural equipment anomalies by pseudo-change stripping branch, and a pseudo-change suppression map is generated.

[0122] The pseudo-change stripping branch performs appearance perturbation checks on the current and historical image block segments corresponding to the structural equipment anomaly candidate difference features. When the overall brightness difference between two image block segments exceeds 35 gray levels and the edge position offset is less than 4 pixels, the difference is marked as an illumination change response. When a dark area covers more than 6 adjacent image block segments along the same direction and the boundary of the power supply facility remains continuous, the difference is marked as a shadow change response. When the area of ​​green pixels in the area surrounding the slope or clearance facilities changes by more than 20% and does not extend into the 3m range on both sides of the track centerline, the difference is marked as a vegetation sway response. When the outline of a vehicle, personnel, or temporary construction object appears in a single phase and the outline does not overlap with the outer outline of the power supply facility object in the power supply facility ledger, the difference is marked as a temporary occlusion response. When there is still an overall translation of less than 8 pixels in the same direction after the stable anchor point Token is corrected, the difference is marked as a residual deviation response of the shooting angle. The image block segments containing the above responses are written into the pseudo-change suppression map, and the intensity of the structural equipment anomaly candidate difference features at the corresponding positions in the pseudo-change suppression map is reduced to 30% of the original intensity.

[0123] By querying the abnormal prototypes in the power supply anomaly layer and matching the structural equipment anomaly candidate difference features after pseudo-change suppression, anomaly change masks and power supply anomaly type attribution results are generated.

[0124] The continuous multi-period trend memory unit compares the abnormal change mask and the attribution results of power supply anomaly types with the historical inspection results for the same power supply facility object, and generates inspection and handling results.

[0125] The trend comparison reads the three most recent historical inspection results of the same power supply facility object from the historical inspection benchmark library. The current anomaly mask is overlapped with the anomaly mask in the historical inspection results according to the same line segment and the same pixel coordinates. When the area of ​​the current anomaly mask increases by more than 10% compared to the most recent historical inspection result, the trend is marked as continuously expanding. When the center position of the current anomaly mask moves more than 15 pixels towards the track centerline relative to the most recent historical inspection result, the trend is marked as developing towards the line side. When the same power supply anomaly type attribution result appears in all three most recent historical inspection results, the trend is marked as recurring. When the anomaly mask has been dealt with in the historical inspection results and still overlaps with the position before the treatment by more than 50%, the trend is marked as recurring after treatment. Continuous multi-period trend memory is also used. The unit generates inspection and handling results based on the attribution results of power supply anomalies, trend markers, and power supply facility objects. The inspection and handling results include power supply facility objects, line sections, anomaly change masks, attribution results of power supply anomaly types, trend markers, and handling levels. The handling levels are divided into on-site verification, time-limited repair, and immediate handling. When the crack extension length exceeds 120 pixels, the slope slippage develops towards the line side, the foreign object encroaches towards the track centerline, the drainage blockage covers 80% of the drainage ditch width, the signal light position area is obscured by more than 30%, the top of the contact wire support is offset by more than 15 pixels, the lower end connection point of the dropper is offset by more than 10 pixels, or the dark staining area of ​​the insulator covers more than 15% of the insulator partition area, the anomalies that continue to expand or recur are identified as time-limited repairs, and the remaining anomalies are identified as on-site verification.

[0126] The inspection and handling results will be written back to the historical inspection benchmark database, and the current drone inspection image after the handling is completed will be used as the historical benchmark image of the corresponding power supply facility object.

[0127] The write-back process involves writing the power supply facility objects, line sections, abnormal change masks, power supply anomaly type attribution results, trend markers, and handling levels from the inspection and handling results into the corresponding power supply facility object name in the historical inspection benchmark database. When manual verification confirms that the abnormal change mask is a real anomaly, the abnormal change mask is retained as a historical inspection result for subsequent trend comparison. When manual verification confirms that it is a change in lighting, a change in shadow, vegetation swaying, temporary obstruction, or residual deviation in shooting angle, the abnormal change mask is deleted from the position corresponding to the real anomaly and retained as a pseudo-change sample of the pseudo-change stripping branch. When the power supply facility object corresponding to the inspection and handling result has been repaired and the current drone inspection image after handling is obtained by drone re-photographing, the image is replaced with the historical benchmark image of the corresponding power supply facility object.

[0128] In this embodiment, the abnormal prototype in the power supply anomaly prototype query decoding layer is an anomaly query template used to retrieve the structural equipment anomaly candidate difference features after pseudo-change suppression. It includes engineering anomaly prototypes, electrical anomaly prototypes, and power supply anomaly prototypes. Each anomaly prototype includes the applicable power supply facility object, the corresponding change form, the pixel scale of the change area, the change direction, and the mask decoding content. The power supply anomaly prototype query decoding layer restores the image block fragments that have completed prototype matching to the pixel coordinates of the image pairs aligned with the same mileage, forming an anomaly change mask, and forms the power supply anomaly type attribution result with the matched anomaly prototype and the corresponding power supply facility object.

[0129] The anomaly prototypes include crack propagation prototypes, slope slippage prototypes, drainage blockage prototypes, protective facility damage prototypes, sound barrier detachment prototypes, foreign object intrusion prototypes, and track bed local changes prototypes. The crack propagation prototype corresponds to linear propagation patterns in bridges, tunnels, retaining walls, and track slabs, extracting continuous length, endpoint extension direction, and bifurcation position. When the linear propagation pattern is located within the corresponding engineering facility object of a bridge, tunnel, retaining wall, or track slab, and its continuous length exceeds 48 pixels, it is matched with the crack propagation prototype. The slope slippage prototype corresponds to planar propagation patterns and edge... For boundary displacement patterns, extract the exposed area area, the downward direction of the boundary, and the distance moved towards the track. When the planar expansion pattern is located within the engineering facility object corresponding to the slope, and the boundary moves more than 20 pixels towards the track, match the slope slip prototype. For drainage blockage prototypes, extract the water body expansion pattern and planar filling pattern in the drainage ditch. Extract the ditch coverage width, the area of ​​the dark connected area, and the water body boundary expansion direction. When the water body expansion pattern or planar filling pattern is located within the engineering facility object corresponding to the drainage ditch, and the coverage of the drainage ditch width reaches 50%, match the drainage blockage prototype. The prototype for damaged protective facilities corresponds to the block-shaped missing shape and boundary displacement shape in the protective fence. It extracts the gaps in the mesh, the offset of the posts, and the position of the continuous boundary interruption. When the block-shaped missing shape or the boundary displacement shape is located within the engineering facility object corresponding to the protective fence, it matches the prototype for damaged protective facilities. The prototype for detached sound barriers corresponds to the block-shaped missing shape in the sound barrier. It extracts the height of the missing panel, the width of the missing panel, and the offset position of the supporting components. When the block-shaped missing shape is located within the engineering facility object corresponding to the sound barrier, and the height of the missing area exceeds 80 pixels, it matches the prototype for detached sound barriers. The foreign object intrusion prototype corresponds to the obstruction and intrusion patterns in the perimeter facilities of the clearance. It extracts the distance and movement direction between the newly added obstruction area and the track centerline. When the obstruction and intrusion pattern is located in the area corresponding to the perimeter facilities of the clearance and approaches the track centerline by more than 20 pixels, it matches the foreign object intrusion prototype. The track bed local change prototype corresponds to the planar expansion pattern and blocky missing pattern in the track bed. It extracts the ballast texture abrupt change area, local missing area and newly added accumulation area. When the planar expansion pattern or blocky missing pattern is located within the track bed corresponding to the track bed engineering facility object, it matches the track bed local change prototype.

[0130] The electrical anomaly prototypes include signal tilt and obstruction prototypes, switch machine external anomaly prototypes, trackside signal box damage prototypes, transponder missing and offset prototypes, and cable trough cover plate anomaly prototypes. The signal tilt and obstruction prototype corresponds to the boundary displacement and obstruction intrusion patterns within the signal, extracting the signal outline tilt direction, the light position obstruction area, and the outer outline offset distance. When the signal's outer outline offsets more than 12 pixels relative to historical image blocks, or when the light position area is covered by more than 30% of the newly added obstruction area, the signal tilt and obstruction prototype is matched. The switch machine external anomaly prototype corresponds to the block-like missing patterns, planar expansion patterns, and water expansion patterns in the switch machine's external structure, extracting box cover gaps, shell damage areas, dark water-filled connected areas, and foreign object-covered areas. When the area of ​​the missing area in the switch machine's external structure exceeds 8% of the area of ​​the corresponding segmented area, or when the dark water-filled connected area covers more than 20% of the bottom of the box, the switch machine external anomaly prototype is matched. The trackside signal box damage prototype corresponds to the trackside signal box. The block-shaped missing patterns and boundary displacement patterns in the image are used to extract the door opening area, the damaged area of ​​the box body, and the direction of door gap expansion. When the door boundary is offset by more than 10 pixels relative to the historical image block or the newly added damaged connected area on the box surface is more than 200 pixels, the trackside signal box damage prototype is matched. The transponder missing offset prototype corresponds to the block-shaped missing patterns and boundary displacement patterns in the transponder. The missing area of ​​the transponder's outer contour, the offset distance of the installation position, and the change of the relative position of the sleeper are extracted. When the outer contour of the transponder exists in the historical image block and the corresponding position in the current image block is missing the outer contour, or when the center position of the current transponder is offset by more than 10 pixels relative to the historical image block, the transponder missing offset prototype is matched. The cable trough cover plate abnormal prototype corresponds to the linear fracture pattern, block-shaped missing pattern, and boundary displacement pattern in the cable trough. The cover plate fracture, the missing area of ​​the cover plate, and the misalignment of the trough edge are extracted. When the edge interruption of two or more consecutive cover plates or the missing area has a continuous length of more than 64 pixels along the line direction, the cable trough cover plate abnormal prototype is matched.

[0131] The power supply anomaly prototypes include contact wire support tilting prototypes, cantilever positioning device deformation prototypes, insulator damage and contamination prototypes, dropper loosening and breakage prototypes, contact wire catenary abnormality prototypes, power supply cabinet appearance abnormality prototypes, and substation external equipment abnormality prototypes. The contact wire support tilting prototype corresponds to the boundary displacement shape of the contact wire support, extracting the tilting direction of the support centerline, the position of the support bottom, and the offset distance of the support top. When the support centerline in the current image block deviates more than 3 degrees relative to the support centerline in historical image blocks, or the support top offset exceeds 15 pixels, the contact wire support tilting prototype is matched. The cantilever positioning device deformation prototype corresponds to the linear expansion shape and boundary displacement shape of the cantilever and positioner. The prototype for insulator positioning device deformation is matched when the insulator is found to have blocky and blocky missing morphologies, extracting the position of the cantilever arm endpoint, the position of the locator connection point, and the bending area of ​​the rod. For insulator damage and contamination prototypes, the prototypes correspond to blocky missing morphologies and planar expansion morphologies in the insulator, extracting the skirt gaps, dark contamination areas on the surface, and string continuity. When there are more than two missing edges on the insulator skirt or the dark contamination area covers more than 15% of the insulator partition area, the prototypes for insulator damage and contamination are matched. For dropper loosening and breakage prototypes, the prototypes correspond to linear breakage morphologies and boundary displacement morphologies in the dropper, extracting the continuity of the dropper segment, the position of the upper and lower connection points, and the vertical offset distance. When a continuous dropper line segment in a historical image block breaks in the current image block, or when the lower end connection point of the dropper line shifts by more than 10 pixels, a dropper line loosening and breaking prototype is matched; the contact wire and catenary morphology abnormal prototype corresponds to the linear expansion and boundary displacement patterns in the contact wire and catenary, extracting the linear offset between the center line, sag direction, and adjacent support points. When the center line of the contact wire or catenary shifts by more than 12 pixels relative to the historical image block and more than 3 consecutive image block segments in the same interval show the same direction of shift, a contact wire and catenary morphology abnormal prototype is matched; the power supply cabinet appearance abnormal prototype corresponds to the block-like missing patterns, boundary displacement patterns, and occlusion intrusion patterns in the power supply cabinet, extracting the door opening... For areas, damaged enclosure areas, areas obstructed by external foreign objects, and enclosure boundary offsets, when the enclosure door boundary offset exceeds 10 pixels, the damaged enclosure area exceeds 200 pixels, or the newly added obstructed area covers more than 20% of the enclosure partition area, the abnormal prototype of the power supply enclosure appearance is matched; for abnormal prototypes of external equipment in substations, block-like missing shapes, boundary displacement shapes, and obstruction intrusion shapes in external equipment of substations, the missing areas of the equipment outer contour, the offset position of the equipment bracket, and the newly added foreign object area within the enclosure are extracted. When the missing area of ​​the equipment outer contour exceeds 8% of the area of ​​the corresponding segmented area of ​​the external equipment of the substation or the newly added foreign object area approaches the outer contour of the live equipment by more than 20 pixels, the abnormal prototype of external equipment in substations is matched.

[0132] Example 1: To verify the feasibility of this invention in practice, it was applied to an integrated unmanned aerial vehicle (UAV) inspection scenario for an integrated power supply and electrical engineering section of a railway in a mountainous area under the jurisdiction of a railway bureau. This section spans from K132+400 to K138+900, with a total length of 6.5 km. Along the line are 2 bridges, 1 tunnel, 8 slopes, 14 drainage ditches, continuous sections of protective fencing, 6 signal lights, 11 trackside signal boxes, 18 transponders, 126 overhead contact line supports, 126 sets of cantilever arms, 252 strings of insulators, numerous droppers, and 4 power supply cabinets. Manual inspection of this section presents challenges, including separate records from track maintenance, electrical maintenance, and power supply professionals, making it difficult to compare images uniformly for the same facility. This is particularly problematic at tunnel entrances, bridge transition sections, and areas with dense overhead contact line supports. Slightly different drone angles during two separate images make it easy for ordinary image change detection to misinterpret shadows, vegetation movement, and temporary obstructions as anomalies, resulting in a significant workload for on-site verification.

[0133] In this embodiment, the implementer uses a drone to collect current drone inspection images along a preset route. Each image is accompanied by the shooting location, shooting time, shooting posture, and lens field of view. The system also reads line mileage data, power supply facility ledgers, and historical inspection benchmark databases to form a basic data set for power supply facility inspections. For objects such as bridge railings near K134+120, signal lights near K135+760, catenary supports near K136+280, and drainage ditches near K137+040, the method retrieves historical benchmark images of the same power supply facility object from the historical inspection benchmark database. It then performs same-mileage registration by using the track centerline, bridge railing edges, drainage ditch straight edges, signal light supports, and catenary support edges to generate same-mileage aligned image pairs between the current image block and historical image blocks. Subsequently, the power supply facility object segmentation network divides objects such as bridges, drainage ditches, signal lights, trackside signal boxes, contact wire supports, cantilever arms, insulators, droppers, and power supply cabinets in the image into corresponding regions, generating a power supply facility object zoning map. This allows each image segment to be associated with a specific facility object, rather than simply participating in recognition as an ordinary image region.

[0134] The improved ChangeFormer model was trained using historical inspection benchmarks from the section and adjacent sections, resulting in 3260 sets of dual-temporal image pairs. These included 1280 sets of real anomaly training samples, 970 sets of pseudo-change training samples, and 1010 sets of normal training samples. The real anomaly training samples were manually reviewed and annotated pixel-by-pixel with anomaly regions, including types of electrical and power supply anomalies such as crack propagation, drainage blockage, signal obstruction, transponder offset, contact wire support tilting, insulator contamination, dropper loosening, and abnormal appearance of power supply boxes and cabinets. The pseudo-change training samples were annotated with regions corresponding to changes in lighting, shadows, vegetation swaying, vehicle obstruction, personnel obstruction, and residual deviations in shooting angles. The normal training samples were image pairs of the same facility object with continuous appearance and no abnormal changes. During training, the loss function consists of the anomaly mask cross-union ratio (CUIR) loss, the power supply anomaly type cross-entropy loss, the spurious change suppression loss, and the stable anchor point alignment deviation loss. The anomaly mask CUIR loss constrains the mask's fit to manually labeled areas; the power supply anomaly type cross-entropy loss constrains the anomaly attribution category; the spurious change suppression loss reduces responses to illumination, shadows, and temporary occlusion; and the stable anchor point alignment deviation loss constrains the positional correspondence of stable objects such as the track centerline, signal post, and overhead contact line support in the two image periods. By the 96th training iteration, the average CUIR of the anomaly mask in the validation samples reached 0.812, the power supply anomaly type attribution accuracy reached 0.918, the spurious change false alarm rate decreased to 0.067, and the average positional deviation of the aligned stable anchor points (tokens) was 4.9 pixels, satisfying the convergence condition.

[0135] In five consecutive UAV inspection verifications, the method of this invention identified 31 track maintenance anomalies, 12 electrical anomalies, and 19 power supply anomalies. Of these, 57 were confirmed as genuine anomalies after manual verification, 5 were false alarms, and 3 were missed. Compared with the method of directly detecting dual-phase changes using the ordinary ChangeFormer, which generates 18 false alarms in the same batch of images, mainly concentrated in the shadows under bridges, swaying vegetation on slopes, and temporary obstruction areas next to the catenary support pillars, this invention reduces such false alarms to 5 by using a dual-phase alignment attention layer guided by stable anchor points and a pseudo-change stripping branch. Taking the catenary support pillar near K136+280 as an example, there is a slight shooting offset between the pillar centerline in the historical image and the current image. The ordinary method identifies the entire edge of the pillar as a large-area change. This invention uses the edge of the catenary support pillar as a stable anchor point to complete local alignment, retaining only the real change area of ​​the pillar top offset by 16.8 pixels relative to the historical image, and generates an immediate handling level by matching the power supply anomaly prototype with the catenary support pillar tilt prototype. Taking the drainage ditch near K137+040 as an example, the dark area expanded after rain, covering 82% of the ditch width. The method attributes this to a drainage blockage prototype and, combined with the results of the three most recent historical inspections, determines that it is continuously expanding. Implementation results show that this invention can directly convert UAV image change detection results into power supply facility objects, abnormal change masks, power supply anomaly type attribution results, trend markers, and handling levels. This reduces the time for manual cross-disciplinary verification and interpretation, shortening the time for processing inspection results for a single 6.5km section from 4.6 hours to 1.7 hours. The consensus rate of reviewers regarding handling suggestions reached 91.3%, indicating that this invention has good feasibility and application value in integrated power supply UAV inspections.

[0136] Table 1 Comparison of Anomaly Identification and Handling Performance of Integrated Power Supply Unmanned Aerial Vehicles in Inspection

[0137] YOLOv8-seg single-phase segmentation recognition method 0.638 72.4 24.7 18.6 13.8 4.4 FC-Siam-Diff Dual-Temporal Change Detection Method 0.697 78.1 18.9 13.5 10.6 3.6 ChangeFormer Two-Phase Change Detection Method 0.754 84.6 14.2 9.4 8.7 2.8 Method of the present invention 0.812 91.8 6.7 5.2 4.9 1.7

[0138] This comparison used UAV inspection data from the same railway section for verification. The test scope covered power supply facilities such as tracks, bridges, tunnels, drainage ditches, signal lights, trackside signal boxes, transponders, catenary supports, cantilever arms, insulators, droppers, contact wires, catenary cables, and power supply cabinets. The test data included 620 pairs of dual-phase images. After manual verification, 57 real anomalies were confirmed. The spurious changes mainly came from shadows under bridges, swaying vegetation on slopes, wet marks after rain, vehicle obstruction, personnel obstruction, and differences in UAV shooting posture. The data in Table 1 do not show an equidistant increasing relationship. The variation range of each method on different indicators is different, which can reflect the actual interference of lighting, obstruction, posture deviation, and differences in facilities in the power supply UAV inspection images on the experimental results.

[0139] From the perspective of the average intersection-over-union (IoU) ratio of the anomaly masks, the YOLOv8-seg single-phase segmentation and recognition method only judges the abnormal region based on the current image, lacking comparison with the same mileage changes in historical images. The average IoU ratio of the anomaly masks is 0.638, which easily includes normal appearance changes in the abnormal region in scenarios such as shadows under bridges and offset edges of catenary supports. The FC-Siam-Diff dual-phase change detection method introduces the difference between the two images, improving the average IoU ratio of the anomaly masks to 0.697. However, it mainly relies on pixel-level or feature-level differences, which is insufficient for utilizing the categories and component affiliations of railway power supply facilities. When encountering slender structures such as contact wires, catenary cables, and droppers, the mask boundaries are prone to breakage. The ChangeFormer dual-phase change detection method achieves an average IoU ratio of 0.754, indicating that the Transformer structure has a stronger expressive ability for dual-phase semantic changes. However, it still mainly outputs ordinary change regions, does not bind power supply facility objects at the input end, and does not use stable anchor points for local misalignment correction. In comparison, the average cross-union ratio of the anomaly mask in the method of the present invention reaches 0.812, which is 0.058 higher than the ChangeFormer dual-phase change detection method. This is because the same mileage registration first maps the current image and the historical image to the same line section, and then restricts the anomalous area attribution through the power supply facility object zoning map, so that the anomaly mask can fit the boundary of the specific facility object.

[0140] In terms of the accuracy of attribution for power and electrical anomalies, the YOLOv8-seg single-phase segmentation identification method achieves 72.4%, the FC-Siam-Diff dual-phase change detection method achieves 78.1%, the ChangeFormer dual-phase change detection method achieves 84.6%, and the method of this invention reaches 91.8%. While the aforementioned comparative methods can output target areas or change areas, they still rely on post-processing rules when distinguishing between engineering anomalies, electrical anomalies, and power supply anomalies. In particular, anomalies such as transponder missing / offset, signal tilt / obstruction, insulator damage / contamination, and loose / broken suspension wires are prone to attribution bias due to small target size, slender structures, or overlap with the background. The method of this invention, through a prototype query decoding layer for power and electrical anomalies, matches the candidate difference features of structural equipment anomalies after pseudo-change suppression to the prototypes of engineering anomalies, electrical anomalies, and power supply anomalies, respectively. This allows change areas to not only be detected but also directly correlated with anomaly types and facility objects, thus significantly improving attribution accuracy.

[0141] In terms of false positive rate and false negative rate, the false positive rate of the method in this invention is 6.7%, lower than the 14.2% of the ChangeFormer dual-temporal change detection method; the false negative rate is 5.2%, lower than the 9.4% of the ChangeFormer dual-temporal change detection method. This difference mainly comes from the stable anchor-guided dual-temporal alignment attention layer and the false change stripping branch. The stable anchor-guided dual-temporal alignment attention layer uses long-term stable structures such as track centerlines, bridge railing edges, sound barrier pillars, signal pillars, contact wire supports, and communication towers to correct local misalignment between the two phases of images, reducing the average edge deviation after registration to 4.9 pixels, while the average edge deviation of the ChangeFormer dual-temporal change detection method is 8.7 pixels. The false change stripping branch further identifies overall brightness changes, shadow changes, vegetation swaying, temporary occlusion, and residual deviations in shooting angle, and reduces the intensity of abnormal response in related areas, thus reducing the possibility of misjudging non-structural appearance disturbances as real anomalies.

[0142] In terms of single-result processing time, the YOLOv8-seg single-phase segmentation and identification method requires 4.4 hours, the FC-Siam-Diff dual-phase change detection method requires 3.6 hours, and the ChangeFormer dual-phase change detection method requires 2.8 hours. The method of this invention shortens this to 1.7 hours. This is because the first three methods typically require manual re-assignment of the changed areas to line sections, facility objects, anomaly types, and handling suggestions. This is especially time-consuming when track maintenance, electrical maintenance, and power supply objects are mixed, requiring extensive manual verification and cross-disciplinary interpretation. The method of this invention simultaneously generates anomaly change masks, anomaly type attribution results for track maintenance, electrical maintenance, and power supply, trend markers, and handling levels at the output end. It also writes the inspection and handling results back to the historical inspection benchmark database, enabling subsequent inspections to directly read historical anomaly locations and handling statuses, reducing redundant interpretation and processing.

[0143] In summary, the method of this invention achieves closed-loop processing from UAV image acquisition, dual-temporal change recognition, anomaly type attribution, to inspection and handling write-back through the continuous coordination of the basic data set of power supply and electrical engineering inspection, image pairs aligned with the same mileage, partition map of power supply and electrical engineering facility objects, improved ChangeFormer model, structural equipment anomaly morphology constraint difference fusion layer, pseudo-change stripping branch, and power supply and electrical engineering anomaly prototype query decoding layer. Experimental data show that this method is superior to the YOLOv8-seg single-temporal segmentation and recognition method, FC-Siam-Diff dual-temporal change detection method, and ChangeFormer dual-temporal change detection method in terms of anomaly mask accuracy, anomaly type attribution, pseudo-change suppression, missed detection control, and result processing efficiency. It is more adaptable to inspection scenarios with a wide variety of railway power supply and electrical engineering facility objects, large differences in structural morphology, numerous shooting interferences, and high requirements for closed-loop handling.

[0144] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based unmanned aerial vehicle (UAV) inspection method for integrated railway power supply and electrical systems, characterized in that, Includes the following steps: Acquire current drone inspection images, line mileage data, power supply facility ledgers, and historical inspection benchmark databases to generate a basic data set for power supply inspection. Based on the basic data set of power supply inspection, historical reference images are retrieved, and the current drone inspection images and historical reference images are registered at the same mileage to generate image pairs aligned at the same mileage. Identify power supply facilities in image pairs aligned with the same mileage and generate a partition map of power supply facilities objects. An improved ChangeFormer model is constructed, which processes image pairs aligned with the same mileage and partition maps of power supply facilities through a component-aware dual-temporal tokenization layer to generate the current component temporal token sequence and the historical component temporal token sequence. The current component temporal token sequence and the historical component temporal token sequence are processed by a stable anchor-guided dual-temporal alignment attention layer and a bidirectional differential interaction Transformer encoding layer to generate dual-temporal component difference features. The structural equipment abnormal morphology constraint difference fusion layer, pseudo change stripping branch and power supply abnormal prototype query decoding layer process the dual-phase component difference characteristics, generate abnormal change mask and power supply abnormality type attribution results, generate inspection and handling results by continuous multi-period trend memory unit, and write back to the historical inspection benchmark library. The improved ChangeFormer model includes a component-aware dual-temporal tokenization layer, a stable anchor-guided dual-temporal alignment attention layer, a bidirectional differential interaction Transformer encoding layer, a structural equipment abnormal morphology constraint difference fusion layer, a pseudo-change stripping branch, an industrial power supply abnormal prototype query decoding layer, and a continuous multi-period trend memory unit. The generation of the inspection and handling results specifically includes: By fusing the difference features of dual-phase components according to the objects and changes of power supply facilities in the structural equipment anomaly constraint difference fusion layer, candidate difference features of structural equipment anomalies are generated. Non-structural appearance perturbations are separated from the candidate difference features of structural equipment anomalies by pseudo-change stripping branch, and a pseudo-change suppression map is generated. By querying the abnormal prototypes in the power supply anomaly layer and matching the structural equipment anomaly candidate difference features after pseudo-change suppression, anomaly change masks and power supply anomaly type attribution results are generated. The continuous multi-period trend memory unit compares the abnormal change mask and the attribution results of power supply anomaly types with the historical inspection results for the same power supply facility object, and generates inspection and handling results. The inspection and handling results will be written back to the historical inspection benchmark database, and the current drone inspection image after the handling is completed will be used as the historical benchmark image of the corresponding power supply facility.

2. The method for unmanned aerial vehicle (UAV) inspection of railway power supply and electrical systems based on deep learning according to claim 1, characterized in that, The generation of the basic data set for power supply inspection specifically includes: The drone collects images of the current inspection along the inspection route of the railway section to be inspected, and attaches the shooting location, shooting time, shooting posture and lens field of view; Based on the shooting location, read the route mileage data, classify the current drone inspection image into the corresponding route segment, and obtain the start and end mileage of the route segment; Based on the line section and start and end mileage, the corresponding power supply facility objects are extracted from the power supply facility ledger and mapped to the current drone inspection image; Based on the power supply facility objects, obtain the historical benchmark images and historical inspection results of the same power supply facility object from the historical inspection benchmark database; The current UAV inspection images, line sections, start and end mileage, power supply facilities, historical reference images, and historical inspection results are collected according to the same power supply facility object to generate a basic data set for power supply inspection.

3. The method for railway power supply and electrical system integrated UAV inspection based on deep learning according to claim 1, characterized in that, The generation of the same mileage aligned image pairs specifically includes: Based on the line sections, start and end mileage and power supply facility objects in the power supply inspection basic data set, select historical benchmark images containing the same power supply facility object from the historical inspection benchmark database; Read the shooting position, shooting posture and lens field of view of the current drone inspection image and the historical reference image, and crop the current inspection field of view and the historical reference field of view according to the start and end mileage of the line section; Extract the track centerline, bridge railing edge, tunnel portal edge, drainage ditch straight edge, sound barrier post and protective fence post within the current inspection field of view and historical reference field of view to generate a stable registration reference; Based on the stable registration benchmark, perform same-mileage registration on the current inspection field of view and the historical benchmark field of view to generate a dual-time relative aligned field of view; The current image region and the historical image region in the dual-time relative alignment field of view are cropped at the same image size, and the corresponding power supply facility objects and line sections are attached to generate image pairs aligned with the same mileage.

4. The method for railway power supply and electrical system integrated UAV inspection based on deep learning according to claim 1, characterized in that, The generation of the power supply facility object zoning map specifically includes: Input the current image block and the historical image block in the image pair aligned with the same mileage into the power supply facility object segmentation network to generate the initial partitioning result of the current object and the initial partitioning result of the historical object. Based on the power supply facility ledger, the classification of the segmented areas in the initial partitioning results of the current object and the initial partitioning results of the historical object is corrected to generate the current object classification partitioning result and the historical object classification partitioning result. The initial partitioning results of the current object and the initial partitioning results of the historical objects are corrected for dual-phase consistency according to the same power supply facility object, and a consistent partitioning result for the object is generated. Perform boundary refinement on the consistent partitioning results of objects to generate object boundary refinement results; Map the refined object boundary results back to the same mileage aligned image pair, so that the current image block and the historical image block share the same pixel coordinates for the power supply facility object partition map.

5. The method for railway power supply and electrical system integrated UAV inspection based on deep learning according to claim 1, characterized in that, The generation of the current component time phase token sequence and the historical component time phase token sequence specifically includes: An improved ChangeFormer model is constructed, with a component-aware dual-temporal tokenization layer set at the model input end to receive image pairs aligned with the same mileage and partition maps of power supply facility objects; The current image block and the historical image block are divided into segments of the current image block and segments of the historical image block according to the same pixel grid. Based on the partition map of power supply facilities objects, read the power supply facility objects corresponding to each image block segment and generate component attribution results; Embed and merge the current image block fragment, historical image block fragments and component attribution results to generate the current component temporal token and the historical component temporal token; Arrange the current component phase token and the historical component phase token according to the same power supply facility object and the same pixel grid to generate the current component phase token sequence and the historical component phase token sequence.

6. The method for unmanned aerial vehicle (UAV) inspection of railway power supply and electrical systems based on deep learning according to claim 1, characterized in that, The generation of the difference features of the dual-phase components specifically includes: Based on the zoning map of power supply facilities, stable anchor tokens are extracted from the current component phase token sequence and the historical component phase token sequence. By correcting the local misalignment between the current component time phase token sequence and the historical component time phase token sequence through a dual-time phase alignment attention layer guided by stable anchor points, the aligned current component time phase token sequence and historical component time phase token sequence are obtained. The aligned current component phase token sequence and historical component phase token sequence are input into the bidirectional differential interactive Transformer encoding layer to extract the current-to-historical change response and the historical-to-current change response; The current-to-historical change response, the historical-to-current change response, and the differential residual are fused to generate the dual-temporal component difference characteristics.

7. The method for railway power supply and electrical system integrated UAV inspection based on deep learning according to claim 1, characterized in that, The abnormal prototypes in the power supply anomaly prototype query decoding layer are abnormal query templates used to retrieve the structural equipment anomaly candidate difference features after pseudo-change suppression. They include engineering anomaly prototypes, electrical anomaly prototypes and power supply anomaly prototypes. Each anomaly prototype includes the applicable power supply facility object, the corresponding change form, the pixel scale of the change area, the change direction and the mask decoding content. The power supply anomaly prototype query decoding layer restores the image block fragments that have completed prototype matching to the pixel coordinates of the image pairs aligned with the same mileage, forming an anomaly change mask, and uses the matched anomaly prototypes and corresponding power supply facility objects to form the power supply anomaly type attribution result.

Images