A pantograph millimeter-level visual detection method

Abstract

The application provides a pantograph millimeter-level visual detection method, relates to the technical field of intelligent maintenance of rail transit vehicles and machine vision measurement, and comprises the following steps: firstly, establishing a track reference coordinate system and an intrinsic coordinate system of a pantograph, expanding a carbon slide plate current collection working surface into a millimeter parameter domain, and generating a local millimeter scale field to realize scale unification of the same physical position in multi-view images; then, using a micro-sweep image sequence formed by vehicle passing to perform sub-pixel reconstruction to obtain a carbon slide plate surface height field and pantograph corner and head key profiles; and finally, based on joint inversion of posture, elasticity, abrasion and sudden change defects, fusing images, scales, heights, topologies, time sequences and historical evidence to output pantograph millimeter-level detection results.

Description

Technical Field

[0001] This invention relates to the field of intelligent maintenance and machine vision measurement technology for rail transit vehicles, specifically a millimeter-level visual inspection method for pantographs. Background Technology

[0002] The pantograph is a crucial component for electric locomotives, EMUs, and urban rail transit vehicles to obtain electrical energy from the overhead contact line. Wear, cracks, missing corners, peeling, drooping of the pantograph angle, pantograph head misalignment, and abnormal installation structures on the pantograph's carbon sliding plate can all affect the quality of pantograph-catenary contact. Current pantograph inspection methods mainly include manual inspection from the top of the pantograph, fixed-point photography at the maintenance depot, single-camera image recognition, and fixed-scale image measurement. Manual inspection from the top of the pantograph suffers from long operation times, high operational risks, and reliance on experience for results. Single-camera image recognition is significantly affected by shooting angle, carbon sliding plate reflection, vehicle position deviation, and pantograph elastic deformation, making it difficult to obtain stable millimeter-level measurement results.

[0003] Pantographs are characterized by their elongated shape, multi-depth planes, multiple components obstructing the view, and localized elastic deformation. The central, end, corner, and lateral thickness regions of the carbon pantograph correspond to different spatial depths within the same image. Directly using a uniform pixel scale for size conversion across the entire image results in localized measurement drift. Simultaneously, normal wear of the carbon pantograph, pantograph attitude changes during vehicle passage, elastic drooping of the pantograph head, and cracks or missing corners all manifest as boundary changes or grayscale anomalies in the image. Traditional detection methods based on single-frame image thresholding and contour recognition struggle to distinguish between normal mechanical changes and genuine structural defects, increasing the risk of false alarms and false negatives.

[0004] Existing technologies require a detection method that does not involve contact with the pantograph, the installation of pantograph sensors, or the application of artificial markers. This method would convert multi-view continuous images into a unified millimeter parameter domain and combine local millimeter-scale fields, micro-sweep image reconstruction, and mechanical state inversion to perform millimeter-level detection of the pantograph. This would provide the detection results with clear physical coordinates, scale sources, image sources, and maintenance criteria. Summary of the Invention

[0005] Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a millimeter-level visual inspection method for pantographs, solving the following problems:

[0007] 1. This solves the problems of traditional manual inspections, which are time-consuming, risky, and inconsistent, making it impossible to perform rapid, continuous, and automated millimeter-level inspections of pantographs on vehicles entering the warehouse. Manual inspections rely on the experience of maintenance personnel, and the judgment of local wear, small cracks, end corner defects, and drooping of the carbon plate is easily affected by the observation angle, lighting conditions, and subjective experience, making it difficult to form traceable and quantitative inspection records.

[0008] 2. This solves the problem that existing single-camera image detection uses a fixed pixel scale, which cannot adapt to the multi-depth, multi-angle, and multi-component planar structure of pantographs, resulting in inconsistent measurement accuracy in the central area, end area, bow angle area, and lateral thickness area of ​​the carbon slide plate. When a vehicle passes over the detection gantry, the pantograph undergoes attitude changes and local elastic displacement. Traditional image measurement fails to establish an independent millimeter scale for each detection position, easily leading to dimensional conversion errors.

[0009] 3. This solves the problem that existing visual inspection methods rely on single-frame image recognition, making it difficult to distinguish between normal wear of carbon skateboards, changes in vehicle motion projection, normal elastic deformation of pantographs, and real abrupt defects. Existing methods directly use image grayscale anomalies, edge abrupt changes, or dark lines as candidates for cracks, missing corners, or peeling, and easily misjudge reflections, carbon powder deposition, surface scratches, and oil stains as structural defects.

[0010] 4. This solves the problem of existing detection results lacking a verification chain including image sources, scale sources, 3D height sources, mechanical topology sources, continuous frame sources, and historical trend sources, making it difficult to quickly verify detection alarms on-site. Traditional detection results typically only output abnormal screenshots or single size values, failing to explain the corresponding physical coordinates, camera viewpoint, image frame, and historical trend of that size, thus affecting the application of automatic detection results in actual maintenance processes.

[0011] Technical solution

[0012] To achieve the above objectives, the present invention provides a millimeter-level visual inspection method for pantographs, comprising the following steps:

[0013] Sp1. Establish an external coordinate reference for the detection area, and unify the track centerline, track surface height, detection gantry reference point, top camera coordinates, left oblique camera coordinates, right oblique camera coordinates, and side camera coordinates into the track reference coordinate system;

[0014] Sp2. When the vehicle enters the detection area, acquire the top image sequence, left oblique image sequence, right oblique image sequence and side image sequence, and record the acquisition time, vehicle position trigger value, camera exposure value, supplementary light status value, vehicle number and pantograph number corresponding to each frame of the image.

[0015] Sp3. Extract the carbon slide current receiving zone boundary, carbon slide end boundary, bow angle transition boundary, bow head crossbeam boundary, and mounting base boundary from the top image sequence, left oblique image sequence, right oblique image sequence, and side image sequence, and establish the pantograph intrinsic coordinate system based on the carbon slide current receiving zone boundary, carbon slide end boundary, and bow head crossbeam boundary.

[0016] Sp4. Expand the carbon slide current-collecting working surface in the pantograph intrinsic coordinate system into a millimeter parameter domain composed of length direction parameter coordinates and width direction parameter coordinates, and map the same carbon slide position in the top image sequence, left oblique image sequence, right oblique image sequence and side image sequence to the millimeter parameter domain.

[0017] Sp5. Based on the camera intrinsic parameters, camera extrinsic parameters, camera distortion parameters, orbit reference coordinate system, and pantograph intrinsic coordinate system, generate a local millimeter-scale field in the millimeter parameter domain;

[0018] Sp6. Micro-sweep image frames are selected based on the effective micro-baseline between image frames, gradient of the flow-receiving working surface region, motion blur index, structural alignment index and coverage integrity index, and the selected micro-sweep image frames are registered to the same millimeter parameter domain.

[0019] Sp7. Based on the registered micro-sweep image frames, the surface height field of the carbon slide plate receiving working surface, the carbon slide plate edge profile, the bow corner end profile, and the bow head crossbeam spatial profile are reconstructed.

[0020] Sp8. Decompose the current state of the pantograph into overall attitude variables, elastic deformation variables, continuous wear variables, and abrupt defect variables, and obtain the mechanical state of the pantograph through joint inversion;

[0021] Sp9, within the millimeter parameter domain, calculate the remaining thickness of the carbon slide plate, wear depth, wear volume, wear slope, crack space length, corner boundary size, bow angle droop, bow head center offset, and mounting position deviation;

[0022] Sp10 generates a defect confidence value based on image evidence value, millimeter-scale evidence value, three-dimensional height evidence value, mechanical topology evidence value, continuous frame temporal evidence value, and historical trend evidence value, and outputs the detection conclusion based on the defect confidence value and maintenance threshold.

[0023] Preferably, when establishing the external coordinate reference of the detection area in Sp1, the direction of the track centerline is set as the first coordinate axis of the track reference coordinate system, the direction of the track surface normal is set as the second coordinate axis of the track reference coordinate system, and the direction that is perpendicular to both the first and second coordinate axes is set as the third coordinate axis of the track reference coordinate system. The external parameter matrices of the top camera, the left oblique camera, the right oblique camera, and the side camera relative to the track reference coordinate system are determined using the reference calibration points fixed on the detection gantry.

[0024] Preferably, when establishing the pantograph intrinsic coordinate system in Sp3, the direction of the center line of the carbon slide current receiving strip is taken as the length direction of the pantograph intrinsic coordinate system, the direction of the width of the carbon slide current receiving strip is taken as the width direction of the pantograph intrinsic coordinate system, the normal direction obtained by fitting the plane of the carbon slide current receiving strip is taken as the height direction of the pantograph intrinsic coordinate system, and the center point of the current receiving strip at one end of the carbon slide is taken as the origin of the millimeter parameter domain.

[0025] Preferably, the millimeter parameter domain in Sp4 consists of length direction parameter coordinates, width direction parameter coordinates, and surface height values. The length direction parameter coordinates are set along the length direction of the carbon slide plate, the width direction parameter coordinates are set along the width direction of the flow receiving zone of the carbon slide plate, and the surface height values ​​are set along the normal direction of the flow receiving working surface of the carbon slide plate. The pixels in the top image sequence, left oblique image sequence, right oblique image sequence, and side image sequence establish a correspondence with the millimeter parameter domain parameter points through a camera projection model.

[0026] Preferably, the local millimeter-scale field in Sp5 is jointly determined by the camera projection Jacobian matrix, the millimeter parameter domain geometric Jacobian matrix, and the pantograph intrinsic coordinate system attitude matrix, and generates local millimeter-scale subfields for the top camera, the left oblique camera, the right oblique camera, and the side camera, respectively.

[0027] Preferably, the screening of micro-sweep image frames in Sp6 includes: calculating the effective micro-baseline length between the current image frame and the reference image frame; calculating the average gradient intensity of the current image frame's current receiving working surface region; calculating the motion blur index of the current image frame; calculating the structural alignment index between the bow head beam boundary, carbon slide end boundary, and mounting base boundary of the current image frame and the pantograph mechanical topology model; calculating the coverage integrity index of the current receiving working surface region in the current image frame; and obtaining an image frame quality score based on the effective micro-baseline length, average gradient intensity, motion blur index, structural alignment index, and coverage integrity index.

[0028] Preferably, when reconstructing the surface height field of the carbon slide plate receiving working surface in Sp7, the selected micro-sweep image frames are registered according to the camera number, frame number and millimeter parameter domain parameter points. The sub-pixel edge position and gray-level gradient extreme value position of the same millimeter parameter domain parameter point in different image frames are extracted, and the surface height field is calculated based on the sub-pixel edge position, the gray-level gradient extreme value position and the local millimeter scale field of multiple frames.

[0029] Preferably, the joint inversion in Sp8 includes: using the observation errors of the top image sequence, left oblique image sequence, right oblique image sequence, and side image sequence as the first constraint term; using the attitude change error of consecutive frames as the second constraint term; using the mechanical topology error between the carbon slide plate, the bow head beam, the bow angle, the mounting base, and the support arm as the third constraint term; using the elastic deformation error of the pantograph as the fourth constraint term; using the continuous wear error of the carbon slide plate as the fifth constraint term; and using the abrupt defect error of the carbon slide plate as the sixth constraint term. The first, second, third, fourth, fifth, and sixth constraint terms are then summed and solved using a weighted summation.

[0030] Preferably, the continuous wear variable and abrupt defect variable of the carbon slide plate in Sp9 are obtained by surface height decomposition. The surface height decomposition is based on the standard surface height of the carbon slide plate, and the current surface height is decomposed into continuous wear height, abrupt defect height and residual error. A parameter domain continuity constraint is applied to the continuous wear height, and a local sparsity constraint is applied to the abrupt defect height.

[0031] Preferably, when generating the defect confidence value in Sp10, the boundary consistency value of the suspected defective object in different camera images is used as the image evidence value, the size stability value of the suspected defective object after local millimeter-scale field conversion is used as the millimeter-scale evidence value, the height mutation value of the suspected defective object in the surface height field is used as the three-dimensional height evidence value, the residual value between the suspected defective object and the pantograph mechanical topology model is used as the mechanical topology evidence value, the parameter domain position stability value of the suspected defective object in consecutive image frames is used as the consecutive frame temporal evidence value, and the trend deviation value between the suspected defective object and historical detection records is used as the historical trend evidence value. The image evidence value, millimeter-scale evidence value, three-dimensional height evidence value, mechanical topology evidence value, consecutive frame temporal evidence value, and historical trend evidence value are fused into the defect confidence value.

[0032] Beneficial effects

[0033] This invention provides a millimeter-level visual inspection method for pantographs. It has the following beneficial effects:

[0034] 1. This invention establishes a track reference coordinate system and a pantograph intrinsic coordinate system, transforming the pantograph detection results from ordinary image coordinates to the millimeter parameter domain of the current receiving working surface. This enables the same carbon slide plate physical position in different cameras, different image frames, and different detection batches to be detected and compared in a unified millimeter coordinate system, thereby improving the traceability of carbon slide plate wear, cracks, missing corners, and bow angle deformation detection.

[0035] 2. This invention generates a local millimeter-scale field within the millimeter parameter domain, enabling each detection position to have an independent pixel-to-millimeter conversion relationship. This avoids the scale drift caused by traditional fixed-scale image measurement at the carbon slide end, bow angle transition area, and lateral thickness region, and improves the consistency of dimensional measurement under multi-depth and multi-angle structural conditions of the pantograph.

[0036] 3. This invention utilizes the continuous minute displacements formed when a vehicle passes through the detection area to perform micro-sweep image frame screening and sub-pixel reconstruction, introducing vehicle motion into the carbon slide plate surface height field recovery process. This improves the reconstruction stability of the surface height, edge contour, crack endpoints, and missing corner boundaries of the flow-receiving working surface, and reduces the reliance on manual gauge verification and contact measurement.

[0037] 4. This invention separates normal posture changes, normal elastic deformation, long-term wear, and real abrupt defects by jointly inverting overall attitude variables, elastic deformation variables, continuous wear variables, and abrupt defect variables. It also generates defect confidence values ​​through image evidence, millimeter-scale evidence, three-dimensional height evidence, mechanical topology evidence, continuous frame time sequence evidence, and historical trend evidence, so that the detection conclusions have clear mechanical interpretation and inspection and repair verification basis. Attached Figure Description

[0038] Figure 1 This is an overall flowchart of the pantograph millimeter-level visual inspection method of the present invention;

[0039] Figure 2 This is a schematic diagram of the deployment structure of the pantograph millimeter-level visual inspection system of the present invention;

[0040] Figure 3 This is a schematic diagram of the construction of the millimeter parameter domain of the flow-receiving working surface according to the present invention;

[0041] Figure 4 This is a schematic diagram of the local millimeter-scale field generation of the present invention;

[0042] Figure 5 This is a schematic diagram of the micro-sweep image frame selection and sub-pixel reconstruction of the present invention;

[0043] Figure 6 This is a schematic diagram of the joint inversion of mechanical state and fusion of defect evidence according to the present invention;

[0044] Figure 7 This is a system framework interface diagram of the present invention;

[0045] Figure 8 This is a system interface diagram illustrating the application scenario of the present invention;

[0046] Figure 9 This is a functional system interface diagram of the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Specific Implementation Example 1:

[0049] like Figures 1 to 9 As shown, a pantograph millimeter-level visual inspection method includes a track reference establishment module, an image synchronous acquisition module, a pantograph intrinsic coordinate construction module, a current-collecting working surface millimeter parameter domain generation module, a local millimeter-scale field generation module, a micro-sweep image frame screening module, a sub-pixel reconstruction module, a mechanical state joint inversion module, a defect evidence fusion module, an inspection conclusion output module, and a historical archive management module.

[0050] The track reference establishment module unifies the track centerline, track surface height, detection gantry reference point, top camera coordinates, left oblique camera coordinates, right oblique camera coordinates, and side camera coordinates into the track reference coordinate system. The track reference coordinate system uses the track centerline direction as the first coordinate axis, the track surface normal direction as the second coordinate axis, and the direction perpendicular to both the first and second coordinate axes as the third coordinate axis. The camera intrinsic parameters, camera extrinsic parameters, distortion parameters, and gantry reference point coordinates are used as fixed external references for subsequent millimeter measurements.

[0051] The image synchronization acquisition module acquires top, left oblique, right oblique, and side image sequences when the vehicle enters the detection area, and records the acquisition time, vehicle position trigger value, camera exposure value, supplementary lighting status value, vehicle number identification value, and pantograph number for each frame. The top image sequence acquires information on the carbon slide plate's current-collecting working surface, current-collecting zone boundary, and surface defects. The left and right oblique image sequences acquire information on the carbon slide plate's end, left and right bow angles, and end corner defects. The side image sequence acquires information on the carbon slide plate's thickness edge, bow head crossbeam, mounting base, and support arm nodes.

[0052] The pantograph intrinsic coordinate construction module extracts the boundaries of the carbon slide current-collecting zone, the carbon slide end boundary, the bow angle transition boundary, the bow head crossbeam boundary, and the mounting base boundary from the top, left oblique, right oblique, and lateral image sequences. It then establishes the pantograph intrinsic coordinate system based on the centerline of the carbon slide current-collecting zone, the position of the carbon slide end, and the relative position of the bow head crossbeam. This pantograph intrinsic coordinate system uses the length direction of the carbon slide as the length direction coordinate, the width direction of the carbon slide current-collecting zone as the width direction coordinate, the normal direction of the carbon slide current-collecting working surface as the height direction coordinate, and the center point of the current-collecting zone at one end of the carbon slide as the origin of the millimeter parameter domain.

[0053] The millimeter parameter domain generation module for the current-collecting working surface unfolds the carbon slide current-collecting working surface in the pantograph's intrinsic coordinate system into a millimeter parameter domain composed of length direction parameter coordinates, width direction parameter coordinates, and surface height values. It also maps the physical position of the same carbon slide from different cameras, different image frames, and different detection batches to the same millimeter parameter domain position. The length direction parameter coordinates are set along the length direction of the carbon slide, the width direction parameter coordinates are set along the width direction of the carbon slide current-collecting zone, and the surface height value is set along the normal direction of the carbon slide current-collecting working surface.

[0054] The local millimeter-scale field generation module generates a local millimeter-scale field within the millimeter parameter domain based on camera intrinsic parameters, camera extrinsic parameters, camera distortion parameters, the orbital reference coordinate system, the pantograph intrinsic coordinate system, and the millimeter parameter domain. The local millimeter-scale fields correspond to the top camera, left-angled camera, right-angled camera, and side camera, respectively. Each millimeter parameter domain location has an independent millimeter-scale conversion relationship, ensuring that image pixels in the carbon skateboard's central region, end region, bow angle transition region, and lateral thickness region are converted to millimeter scales according to their spatial location.

[0055] The micro-sweep image frame selection module filters image frames for subpixel reconstruction based on the minute relative displacements formed between consecutive image frames as the vehicle passes through the detection area. The module calculates the effective micro-baseline length between the current image frame and the reference image frame, the average gradient intensity of the current-collecting working surface region, the motion blur index, the pantograph structure alignment index, and the current-collecting working surface coverage integrity index. Based on these indices, it generates an image frame quality score. Image frames whose quality scores reach a set reconstruction threshold enter the subpixel reconstruction process.

[0056] The subpixel reconstruction module registers the selected micro-sweep image frames to the same millimeter parameter domain, and fuses the subpixel edge positions, grayscale gradient extreme positions, and local millimeter-scale fields from different cameras and different image frames at the same millimeter parameter domain position to reconstruct the surface height field of the carbon slide plate's flow-receiving working surface, the carbon slide plate edge contour, the bow corner end contour, and the bow head crossbeam spatial contour.

[0057] The mechanical state joint inversion module decomposes the current state of the pantograph into overall attitude variables, elastic deformation variables, continuous wear variables, and abrupt defect variables, and performs joint inversion by weighted summation of image observation error, attitude smoothing error, mechanical topology error, elastic rationality error, continuous wear error, and abrupt defect error.

[0058] The defect evidence fusion module generates defect confidence values ​​for suspected defect objects, which originate from locally connected regions in the height field of abrupt defects. The module calculates image evidence values, millimeter-scale evidence values, 3D height evidence values, mechanical topology evidence values, continuous frame temporal evidence values, and historical trend evidence values, and then fuses these evidence values ​​into a defect confidence value.

[0059] The test result output module calculates the remaining thickness of the carbon slide plate, maximum wear depth, average wear depth, wear volume, wear slope, crack space length, corner boundary size, bow angle droop, bow head center offset, and mounting position deviation within the millimeter parameter domain. Based on the defect confidence value and maintenance threshold, it outputs four types of test results: qualified, observation warning, time-limited maintenance, and immediate maintenance.

[0060] The historical archive management module writes the vehicle number, pantograph number, inspection time, camera number, effective frame number, millimeter parameter domain coordinates, local millimeter scale field, surface height field, continuous wear height field, abrupt defect height field, key dimension results, defect confidence value, and inspection conclusions into the pantograph's full life inspection archive. It also aligns the current inspection results with historical inspection results in the same millimeter parameter domain to form wear growth curves, uneven wear trend curves, and defect development curves. Specific Implementation Example 2:

[0062] like Figures 1 to 9 As shown, based on the content of the above specific embodiments, the following content is further disclosed:

[0063] In this invention, the pantograph millimeter-level visual inspection method is executed according to a unified data flow. The inspection object is mainly the pantograph carbon slide plate current-collecting working surface, with the pantograph angle, pantograph head beam, mounting base, and support arm key nodes as auxiliary constraint objects. When the system completes an inspection, it first establishes an external coordinate reference for the inspection area, then converts the pantograph image information from multi-camera continuous images into the pantograph intrinsic coordinate system. Subsequently, the carbon slide plate current-collecting working surface is unfolded into a millimeter parameter domain, and a local millimeter-scale field is generated within this millimeter parameter domain. The continuous small displacements formed when the vehicle passes through the inspection area are used to complete micro-sweep image frame selection and sub-pixel reconstruction. Finally, the mechanical state of the pantograph is obtained through the joint inversion of overall attitude variables, elastic deformation variables, continuous wear variables, and abrupt defect variables. The steps of this method are interdependent. The orbital reference coordinate system provides a reference for the camera's extrinsic parameters, the pantograph's intrinsic coordinate system provides a reference for the generation of the millimeter parameter domain, the millimeter parameter domain provides a reference for the alignment of the local millimeter-scale field and historical detection, the local millimeter-scale field provides a reference for the conversion of defect size, the micro-sweep reconstruction results provide surface height input for the joint inversion of mechanical state, and the joint inversion results provide state variable input for the output of detection conclusions.

[0064] In SP1, the establishment of the track reference coordinate system includes reference point acquisition, coordinate axis definition, camera extrinsic parameter solution, and calibration data storage. After the gantry is installed, at least three gantry reference points are fixed on both sides of the track. The relative positions of the gantry reference points with respect to the track centerline and track surface height are measured using a calibration fixture. The track centerline direction is set as the first coordinate axis, the track surface normal direction is set as the second coordinate axis, and the direction perpendicular to the first and second coordinate axes is set as the third coordinate axis. Top, left, right, and side cameras respectively capture images of the gantry reference points and the calibration fixture. The system solves for each camera extrinsic parameter matrix based on the pixel coordinates of the reference points in the image, the three-dimensional coordinates of the reference points in the track reference coordinate system, the camera intrinsic parameter matrix, and the distortion parameters. The camera intrinsic parameter matrix includes focal length parameters, principal point coordinates, and pixel ratio parameters; the camera extrinsic parameter matrix includes rotation matrix and translation vector; and the distortion parameters include radial distortion parameters and tangential distortion parameters. The calibration data is written into the calibration database before testing and is retrieved by the testing server at the start of each test.

[0065] In Sp2, the image synchronization acquisition process is initiated by the vehicle position triggering unit. After detecting a vehicle entering the detection window, the vehicle position triggering unit sends a trigger value to the edge computing unit. The edge computing unit synchronizes the top camera, left oblique camera, right oblique camera, and side camera according to a unified clock. Each camera continuously acquires images at a preset frame rate, and binds each frame to the acquisition time, vehicle position trigger value, camera number, exposure value, gain value, supplementary lighting status value, vehicle number, and pantograph number. The top image sequence is used to acquire the carbon slide plate's current-collecting working surface; the left and right oblique image sequences are used to acquire the carbon slide plate end and the bow angle transition area; and the side image sequence is used to acquire the carbon slide plate thickness edge, bow head crossbeam, mounting base, and support arm nodes. The edge computing unit performs distortion correction, exposure normalization, reflection suppression, and motion blur evaluation on the acquired images, generating a set of valid image frames. Image frames that fail the sharpness and integrity evaluations are not included in the subsequent millimeter parameter domain mapping process.

[0066] In Sp3, the pantograph's intrinsic coordinate system is jointly determined by the boundaries of the carbon slide current collection zone, the carbon slide end boundary, and the pantograph head crossbeam boundary. The system extracts the left and right boundaries and the centerline of the carbon slide current collection zone in the top image, the end boundary and the transition boundary of the pantograph angle in the oblique image, and the pantograph head crossbeam boundary and the mounting base boundary in the side image. During boundary extraction, the system first calculates the grayscale gradient within the pantograph's coarse positioning area, then searches for gradient extrema along a preset search direction, and performs continuity connection and mechanical topological constraint filtering on these gradient extrema. After boundary extraction, the system defines the direction of the carbon slide current collection zone centerline as the length direction of the pantograph's intrinsic coordinate system, the width direction of the current collection zone as the width direction, the normal direction obtained from fitting the local plane of the current collection zone as the height direction, and the center point of the current collection zone at one end of the carbon slide as the origin of the millimeter parameter domain. This pantograph intrinsic coordinate system is generated according to the current actual attitude of the pantograph, ensuring that the subsequent millimeter parameter domain does not depend on fixed image coordinates.

[0067] In Sp4, the generation process of the millimeter parameter domain for the current-collecting working surface includes parameter domain sampling, image pixel mapping, and historical archive alignment. The millimeter parameter domain consists of length direction parameter coordinates, width direction parameter coordinates, and surface height values. The length direction parameter coordinates cover the effective current-collecting length of the carbon slide, the width direction parameter coordinates cover the effective current-collecting width of the carbon slide, and the surface height value is recorded along the normal direction of the current-collecting working surface. The system generates a set of parameter points within the millimeter parameter domain according to a set sampling interval. Each parameter point has unique length and width direction coordinates. For each parameter point, the system calculates the projected pixel position of the parameter point in the top, left oblique, right oblique, and side images based on the pantograph's intrinsic coordinate system, the track reference coordinate system, and the camera projection model, and establishes a mapping relationship between the parameter point number, 3D coordinates, pixel coordinates, and camera number. The detection results of the same pantograph in the historical archive are also stored according to this millimeter parameter domain, allowing for comparison between the current detection results and historical detection results under the same parameter coordinates.

[0068] In Sp5, the local millimeter-scale field is used to describe the conversion relationship between image pixel changes and actual physical size changes at each millimeter parameter point. The system calculates the local millimeter-scale subfield for each camera. For a parameter point, the system first determines the pixel coordinates of the parameter point in the image based on the camera projection model, and then applies small parameter perturbations along the length and width directions near the parameter point to obtain the pixel coordinate changes and actual physical distance changes of adjacent parameter points, respectively. Subsequently, the corresponding local millimeter scale is calculated. The local millimeter-scale subfield of the top camera is used for the conversion of wear and surface crack size on the flow-receiving working surface; the local millimeter-scale subfields of the left and right oblique cameras are used for the conversion of end corner defects and bow angle sag size; and the local millimeter-scale subfield of the side camera is used for the conversion of the remaining thickness of the carbon slide plate and the position deviation of the bow head beam. The system binds and stores each local millimeter-scale subfield with the parameter point number, so that the size of each subsequent defect comes from the local scale of the corresponding location, rather than from the uniform scale of the entire image.

[0069] In Sp6, the micro-sweep image frame selection process is used to select effective observation frames for subpixel reconstruction from consecutive image frames. When a vehicle passes through the detection window, the pantograph undergoes continuous small displacements relative to the camera. The system sets the image frame where the pantograph is located in the middle of the detection window as the reference image frame and calculates the effective micro-baseline length between the remaining image frames and the reference image frame. The system simultaneously calculates the average gradient intensity, motion blur index, structural alignment index, and coverage integrity index of the current image frame's current working surface region. The average gradient intensity is determined by the mean grayscale gradient within the current working surface region; the motion blur index is determined by the edge spread width and gradient direction consistency; the structural alignment index is determined by the matching residuals between the pantograph head beam boundary, carbon slide end boundary, and mounting base boundary and the mechanical topology model; and the coverage integrity index is determined by the proportion of the effective area of ​​the current working surface in the visible area of ​​the image. The system obtains an image frame quality score based on these indices and writes image frames with scores reaching the reconstruction threshold into the micro-sweep image frame set.

[0070] In Sp7, the subpixel reconstruction process takes a set of micro-sweeped image frames, a local millimeter-scale field, and millimeter-parameter domain mappings as input. The system performs subpixel edge localization on the projection regions of the same parameter point in different cameras and different image frames. The subpixel edge position is determined by the extreme position of the gray-level gradient, the position of the quadratic curve fitting, and the consistency of the edge direction. For the flow-receiving surface of the carbon slide plate, the system fuses surface edge information from the top image, height change information from the oblique image, and thickness edge information from the lateral image to obtain the surface height value of each parameter point. For the bow corner end, the system fuses the end contours from the left and right oblique images to obtain the bow corner end contour and the missing corner boundary. For the bow head beam, the system fuses the straight line boundary from the lateral image and the relative position constraint from the top image to obtain the bow head beam spatial contour. After reconstruction, the system generates the surface height field, the carbon slide plate edge contour, the bow corner end contour, and the bow head beam spatial contour, and binds the above results to millimeter-parameter domain coordinates.

[0071] In Sp8, the joint inversion process of mechanical state is used to decompose the image observation results into overall attitude, elastic deformation, continuous wear, and abrupt defects. The system uses the reconstructed surface height field, edge contour, bow corner end contour, and bow head crossbeam spatial contour as observation inputs, and establishes a mechanical topology model in combination with the pantograph model structural data. The mechanical topology model includes the relative positional relationship between the carbon slide plate and the bow head crossbeam, the connection relationship between the bow corner and the end of the carbon slide plate, the relative positional relationship between the mounting base and the bow head crossbeam, and the connection relationship between the support arm nodes. During the joint inversion, the system first solves the overall attitude variables of the pantograph to reduce the difference between the model projection boundary and the image observation boundary; then it solves the elastic deformation variables to match the current observation within the allowable deformation range of the bow head, bow corner, and support arm; next, it solves the continuous wear variables to ensure that the surface height of the carbon slide plate decreases continuously within the millimeter parameter domain; finally, it solves the abrupt defect variables to include local height abrupt changes that cannot be explained by the overall attitude, elastic deformation, and continuous wear into the defect residual field. Through this process, image anomalies are hierarchically attributed, and mutation defect variables correspond only to local residual regions.

[0072] In Sp9, the calculation of detection indicators is performed within the millimeter parameter domain. The remaining thickness of the carbon slide plate is determined by the distance between the upper and lower edges obtained from the lateral image reconstruction and the height of the flow-receiving working surface; the maximum wear depth is determined by the maximum difference between the standard surface height and the current surface height; the average wear depth is determined by the average continuous wear height within the effective flow-receiving area; the wear volume is obtained by integrating the continuous wear height over the effective flow-receiving area; the wear slope is obtained by the ratio of the surface height difference to the width distance at different sampling positions in the same length direction coordinate; the crack space length is obtained by the centerline length of the slender connected region in the abrupt defect height field; the corner boundary size is obtained by the boundary span of the abrupt defect region at the end of the carbon slide plate; the bow angle sag is obtained by the difference between the height of the left and right bow angle ends and the reference height of the bow head crossbeam; the bow head center offset is obtained by the lateral distance between the centerline of the bow head crossbeam and the center reference line of the track reference coordinate system; the mounting position deviation is obtained by the difference between the reconstructed mounting position and the standard position in the model structure data. All detection indicators are bound to the parameter domain coordinates, camera number, image frame number, and local millimeter-scale field number.

[0073] In Sp10, the defect confidence value generation process is performed on a per-potential defect object basis. Potential defect objects originate from locally connected regions within the abrupt defect height field. For each potential defect object, the system generates image evidence value, millimeter-scale evidence value, 3D height evidence value, mechanical topology evidence value, continuous frame temporal evidence value, and historical trend evidence value. Image evidence value is determined by the boundary consistency of the potential defect object across multiple camera views; millimeter-scale evidence value is determined by the dimensional stability of the potential defect object after conversion through different local millimeter-scale subfields; 3D height evidence value is determined by the height abrupt amplitude and abrupt boundary continuity of the potential defect object in the surface height field; mechanical topology evidence value is determined by the residuals of the potential defect object that cannot be explained by the mechanical topology model and elastic deformation variables; continuous frame temporal evidence value is determined by the stability of the potential defect object falling within the same millimeter parameter domain position in consecutive micro-scanned image frames; and historical trend evidence value is determined by the degree of deviation of the potential defect object from the wear trend in the same pantograph historical inspection archive. The system integrates the above evidence values ​​into a defect confidence value, and outputs conclusions such as qualified, observation warning, time-limited maintenance, and immediate maintenance based on the defect confidence value, dimensional measurement results, and maintenance threshold.

[0074] In a specific application of this invention, the detection server establishes a structured data record for each detection. This structured data record includes vehicle number, pantograph number, detection time, camera number, valid frame number, millimeter parameter domain coordinates, local millimeter-scale field, surface height field, continuous wear height field, abrupt defect height field, key dimension results, defect confidence value, and detection conclusion. When generating the detection report, the system simultaneously outputs an overall pantograph diagram, a unfolded diagram of the carbon slide current-collecting working surface, a wear distribution diagram, an abrupt defect diagram, an anomaly magnification diagram, a key dimension table, and original image lookup records. Maintenance personnel can query the corresponding original camera image, corresponding image frame, and corresponding millimeter-scale conversion result based on the parameter domain coordinates to complete the verification of the detection conclusion. The historical archive management module aligns the current detection result with historical detection results in the same millimeter parameter domain and generates wear growth curves, uneven wear trend curves, and defect development curves for pantograph condition maintenance, periodic verification, and fault tracing. Specific Implementation Example 3:

[0076] like Figures 1 to 9 As shown, based on the content of the above specific embodiments, the following content is further disclosed:

[0077] This implementation takes the pantograph inspection scenario for vehicle depot entry as an example. An inspection gantry is installed on the inspection line. The gantry is equipped with a top camera, a left-angled camera, a right-angled camera, a side camera, a top uniform illumination unit, a low-angle contour illumination unit, a polarization illumination unit, a vehicle position triggering unit, a vehicle number recognition unit, and an edge computing unit. The inspection server is communicatively connected to the edge computing unit. The top camera's field of view covers the carbon slide plate's current-collecting working surface. The left and right angled cameras' fields of view cover the carbon slide plate's end, left and right bow angles, and the bow head's lateral boundary. The side camera's field of view covers the carbon slide plate's thickness edge, the bow head crossbeam, support arm nodes, and mounting base. The vehicle position triggering unit sends a trigger signal to the edge computing unit when a vehicle enters the inspection window. The edge computing unit controls each camera to synchronously acquire a continuous image sequence.

[0078] Before inspection, the system establishes an external coordinate reference for the inspection area using the track reference establishment module. The track centerline direction is set as the first coordinate axis, the track surface normal direction as the second coordinate axis, and the direction perpendicular to both the first and second coordinate axes as the third coordinate axis. The extrinsic parameter matrices of the top camera, left oblique camera, right oblique camera, and side camera relative to the track reference coordinate system are determined using the inspection gantry reference point. After calibration, the camera intrinsic parameters, extrinsic parameters, and distortion parameters are stored together with the track reference coordinate system in the inspection server.

[0079] When a vehicle passes through the detection area, the vehicle identification unit reads the vehicle number, and the system retrieves the corresponding vehicle model and pantograph model data. This model data includes the standard length and width of the carbon strip, the initial thickness of the carbon strip, the relative position of the pantograph head beam, the dimensions of the pantograph angle structure, the mounting bracket spacing, the relative position of the hinge points, and the maintenance threshold. The image synchronization acquisition module simultaneously acquires top, left-side, right-side, and side image sequences when the vehicle enters the detection window, and binds and stores each frame with the acquisition time, vehicle position trigger value, camera exposure value, supplementary lighting status value, vehicle number, and pantograph number.

[0080] The edge computing unit performs distortion correction, exposure normalization, reflection suppression, motion blur calculation, and coarse localization of the pantograph area on the acquired images. After coarse localization, the pantograph intrinsic coordinate construction module extracts the carbon slide current receiving zone boundary and length direction edge in the top image, the carbon slide end boundary, bow angle transition boundary, and end corner candidate boundary in the left and right oblique images, and the carbon slide thickness edge, bow head crossbeam boundary, and mounting base boundary in the side images. The boundary extraction results are stored according to camera number, image frame number, and pixel coordinate set.

[0081] The pantograph intrinsic coordinate construction module establishes the pantograph intrinsic coordinate system based on the centerline of the carbon slide current-collecting strip, the end boundary of the carbon slide, and the boundary of the pantograph head beam. The pantograph intrinsic coordinate system uses the direction of the carbon slide current-collecting strip centerline as the length direction, the width direction of the carbon slide current-collecting strip as the width direction, the normal direction of the carbon slide current-collecting working surface as the height direction, and the center point of the current-collecting strip at one end of the carbon slide as the origin of the millimeter parameter domain. The current-collecting working surface millimeter parameter domain generation module expands the carbon slide current-collecting working surface into a millimeter parameter domain consisting of length direction parameter coordinates, width direction parameter coordinates, and surface height values, allowing the same physical position of the carbon slide in different cameras and different frames to be mapped to the same parameter point.

[0082] The parameter domain of the flow-receiving working surface is expressed by the following formula:

[0083]

[0084] In formula (1), Indicates the first Parameter points of the working surface during the second test Coordinates in three-dimensional space; Indicates the batch number of the test; Represents the millimeter coordinates along the length of the carbon slide plate; Represents the millimeter parameter coordinates along the width direction of the carbon slide plate; Indicates the first The three-dimensional coordinates of the origin of the parameter domain of the working surface under flow during the second detection; Indicates the first The unit vector of the carbon slide length direction during the second test; Indicates the first The unit vector in the width direction of the carbon slide plate during the second test; Indicates the first The normal unit vector of the carbon slide plate's flow-receiving working surface during the second test; Indicates the first Parameter points during the second detection The surface height value at that location. All the above symbols are represented inline formula format in Word documents.

[0085] The system generates a local millimeter-scale field within the millimeter parameter domain, which is jointly determined by the camera projection model and the parameter domain geometric model. For the ... One camera, parameter points Pixel coordinates are obtained through the camera projection model. The system calculates the Jacobian matrix between the parameter domain coordinate changes and the image pixel coordinate changes, and combines this with the three-dimensional geometric changes of the flow-receiving working surface to obtain the millimeter scale at that parameter point. The local millimeter-scale field does not use a fixed scale for the entire image, but is generated independently at each parameter point, so that the central region, end region, bow angle transition region, and side thickness region of the carbon slide plate all have corresponding scale conversion relationships.

[0086] The local millimeter-scale field is represented by the following formula:

[0087]

[0088]

[0089]

[0090] In formulas (2-1) to (2-3), Indicates the first The camera in the During the second detection, the parameter points were... The resulting local millimeter-scale data is expressed in millimeters per pixel. Indicates the first Parameter points during the second detection The geometric scale of the parameter domain is formed by the geometric changes of the flow-receiving working surface; Indicates the first The camera in the Parameter points during the second detection The projection scale at that location; Indicates the camera number; Indicates the first Parameter points during the second detection 3D coordinates; Represents the three-dimensional coordinates relative to the length direction parameter The partial derivatives; Represents the three-dimensional coordinates relative to the width direction parameter The partial derivatives; Represents the Euclidean norm; Represents parameter points Project to the Pixel coordinates in a camera image; The Jacobian matrix representing the change in pixel coordinates caused by changes in parameter domain coordinates; Denotes the Frobenius norm of a matrix; The above symbols represent square root operations. All of these symbols are represented as inline formulas in Word documents.

[0091] The system performs micro-sweep filtering on consecutive image frames, selecting top, left-angled, right-angled, and side image frames acquired when a vehicle passes through the detection window. The system first selects the image frame where the pantograph is located in the middle of the detection window as the reference image frame. Then, it calculates the effective micro-baseline length, average gradient intensity of the current-collecting working surface area, motion blur index, structure alignment index, and coverage integrity index between the remaining image frames and the reference image frame, and obtains an image frame quality score based on these indices. Image frames with a quality score higher than a preset reconstruction threshold are entered into sub-pixel reconstruction, while image frames with a quality score lower than the preset reconstruction threshold are excluded from the reconstruction input.

[0092] The frame quality score for micro-sweep images is expressed by the following formula:

[0093]

[0094] In formula (3), Indicates the first Micro-sweep reconstruction quality score of frame images; Indicates the frame number in a continuous image sequence; Indicates the first The effective microbaseline length between the frame and the reference image frame; This indicates the reference microbaseline length set by the system; Indicates the first The average gradient intensity of the frame's flow-receiving working surface region; This indicates the reference gradient strength set by the system. Indicates the first Frame motion blur index; This represents the motion fuzzy reference value set by the system; Indicates the first Structural alignment metrics between the pantograph structure and the mechanical topology model; Indicates the first The coverage integrity index of the frame receiving working surface area; Indicates the microbaseline weights; Represents image gradient weights; Indicates the motion blur weights; Indicates the structural alignment weight; This indicates that the weights are fully covered. All of the above symbols are represented as inline formulas in Word documents.

[0095] During subpixel reconstruction, the system registers selected micro-sweep image frames to the same millimeter parameter domain and extracts the subpixel positions of edges and the extreme positions of grayscale gradients near each parameter point. The system fuses observations of the same parameter point from different cameras and different image frames to obtain the surface height field of the carbon slide plate's flow-receiving working surface, and simultaneously obtains the carbon slide plate edge contour, bow corner end contour, and bow head crossbeam spatial contour. For carbon slide plate thickness detection, the system combines the upper and lower edge positions in the lateral image with the working surface height reconstructed from the top image to obtain the remaining thickness of the carbon slide plate; for crack detection, the system combines the elongated abrupt boundary in the top image, the height change in the oblique image, and the defect residual obtained from micro-sweep reconstruction to determine the crack spatial length; for corner defect detection, the system combines the abrupt boundary change at the carbon slide plate end, the bow corner transition boundary, and the parameter domain end residual to determine the corner defect boundary size.

[0096] The system solves for the pantograph's mechanical state through joint inversion. The set of variables in the joint inversion includes overall attitude variables, elastic deformation variables, continuous wear variables, and abrupt defect variables. Overall attitude variables represent the pantograph's rotation and translation relative to the track reference coordinate system; elastic deformation variables represent the allowable deformation of the pantograph head, support arm, pantograph angle, and mounting base within mechanical topology constraints; continuous wear variables represent the smooth height decrease of the carbon slide plate's current-collecting working surface over mileage; and abrupt defect variables represent local height abrupt changes caused by cracks, missing corners, chipped edges, and spalling. The system determines the optimal mechanical state by weighted summation of image observation errors, attitude smoothing errors, mechanical topology errors, elastic rationality errors, continuous wear errors, and abrupt defect errors.

[0097] The joint inversion objective function is expressed by the following formula:

[0098]

[0099] In formula (4), Indicates the first The optimal pantograph state set obtained from the second detection; Indicates the first The set of pantograph states to be solved during the next test; This represents the value of the variable that minimizes the objective function. Indicates image observation error; Indicates attitude smoothing error; Indicates mechanical topology error; Indicates the error in the reasonableness of the elasticity; Indicates continuous wear error; Indicates mutation defect error; Indicates the pose smoothing weights; Represents the mechanical topology weights; Indicates the weight of the elastic constraint; Indicates the weight of continuous wear; This represents the weight of mutation defects. All the above symbols are represented as inline formulas in Word documents.

[0100] After calculating the surface height field of the carbon sliding plate, the system decomposes the current surface height into standard surface height, continuous wear height, abrupt defect height, and residual error. The standard surface height is derived from pantograph production line inspection records, initial maintenance records, or vehicle model structure databases; the continuous wear height varies continuously within the parameter domain; the abrupt defect height exhibits a locally concentrated distribution within the parameter domain; and the residual error is formed by image noise, reconstruction errors, and residual surface contamination. The system applies continuous constraints to the continuous wear height within the parameter domain and local sparsity constraints to the abrupt defect height, thus calculating long-term wear and local defects separately.

[0101] The surface height of the carbon skateboard is decomposed using the following formula:

[0102]

[0103] In formula (5), Indicates the first Parameter points during the second detection The actual surface height at that location; Represents the parameter point under standard conditions Reference surface height at the location; Indicates the first Parameter points during the second detection Continuous wear height at the location; Indicates the first Parameter points during the second detection The height of the mutation defect at the location; This represents the residual error caused by image noise, reconstruction error, and surface contamination. This indicates the length direction of the carbon skateboard in millimeters. This represents the width of the carbon skateboard in millimeters. All the symbols mentioned above are represented in inline formulas in the Word document.

[0104] Wear volume is expressed by the following formula:

[0105]

[0106] In formula (6), Indicates the first The continuous wear volume of the carbon slide plate during the second test; This indicates the effective flow-receiving working surface parameter area of ​​the carbon slide plate; Indicates the first Parameter points during the second detection Continuous wear height at the location; Represents a tiny integral unit along the length direction; Represents a tiny integral unit in the width direction; This indicates a double integral over the entire effective flow-receiving working surface parameter region. The system compares the wear volume with the historical wear volume to obtain the wear increase and wear increase rate. All the above symbols are represented in inline formula format in the Word document.

[0107] The eccentric wear slope is expressed by the following formula:

[0108]

[0109] In formula (7), Indicates the first During the second test, the carbon slide plate was on the length coordinate. The wear slope in the width direction at that location; Indicates the first Width coordinate during the second detection Surface height at that location; Indicates the first Width coordinate during the second detection Surface height at that location; This indicates the first sampling position in the width direction of the flow-receiving working surface; This indicates the second sampling position in the width direction of the flow-receiving working surface; This represents the millimeter coordinates along the length of the carbon slide plate. The system compares the wear slope with the inspection threshold to obtain the wear determination result. All the above symbols are represented in inline formula format in the Word document.

[0110] The system generates a defect confidence value for each suspected defect, with the suspected defect object originating from a locally connected region in the height field of abrupt defect changes. For each suspected defect object, the system calculates image evidence value, millimeter-scale evidence value, 3D height evidence value, mechanical topology evidence value, continuous frame temporal evidence value, and historical trend evidence value. The image evidence value is determined by the degree of boundary consistency of the suspected defect object in the top, left oblique, right oblique, and side images; the millimeter-scale evidence value is determined by the dimensional stability of the suspected defect object after conversion using a local millimeter-scale field; the 3D height evidence value is determined by the degree of height abrupt change of the suspected defect object in the surface height field; the mechanical topology evidence value is determined by the residual between the suspected defect object and the pantograph mechanical topology model; the continuous frame temporal evidence value is determined by the stability of the position of the corresponding parameter domain of the suspected defect object in consecutive image frames; and the historical trend evidence value is determined by the degree of trend deviation between the suspected defect object and historical detection records of the same pantograph.

[0111] The defect confidence value is expressed by the following formula:

[0112]

[0113]

[0114]

[0115] In formulas (8-1) to (8-3), Represents a defective object The overall confidence value; The value of combined evidence supporting the establishment of the defect; Indicates the value of reverse evidence; Indicates the suspected defective object number; Indicates the value of image evidence; Indicates evidence values ​​on a millimeter scale; Indicates the three-dimensional height evidence value; Indicates the mechanical topology evidence value; Indicates the temporal evidence value of consecutive frames; Values ​​representing evidence of historical trends; Represents a defective object The set of evidence involved in the fusion; Indicates the evidence number in the evidence set; Indicates the first Confidence value of the evidence; This indicates a multiplication operation. The system compares the defect confidence value with a preset maintenance threshold and outputs four types of inspection conclusions: qualified, observation warning, time-limited maintenance, and immediate maintenance. All the above symbols are represented inline formula format in Word documents.

[0116] In this embodiment, the detection server establishes a structured record for each detection. This structured record includes the vehicle number, pantograph number, detection time, camera number, effective frame number, millimeter parameter domain coordinates, local millimeter-scale field, surface height field, continuous wear height field, abrupt defect height field, key dimension results, defect confidence value, and detection conclusion. The system also generates a visual detection report, which includes an overall pantograph diagram, a unfolded diagram of the carbon slide current-collecting working surface, a wear distribution diagram, an abrupt defect diagram, an anomaly magnification diagram, historical change curves, and an entry point for reverse lookup of the original image. Maintenance personnel can use the parameter domain coordinates of any defective object to reverse look up the corresponding camera, corresponding image frame, and corresponding original boundary position, thereby completing the verification of the detection conclusion.

[0117] In the second embodiment, this method is applied to static verification inspection in a maintenance depot. After the vehicle stops at the maintenance station, a mobile inspection trolley moves a short distance laterally along the pantograph. Top, oblique, and side cameras generate a micro-scan image sequence with a fixed step distance. The system employs the same process as through-feed inspection: parameter domain construction of the current-collecting working surface, local millimeter-scale field generation, sub-pixel reconstruction, joint inversion, and defect confidence value generation. This embodiment is suitable for post-through-feed inspection verification, key vehicle re-inspection, and post-overhaul quality confirmation.

[0118] In the third embodiment, this method is applied to the production line inspection of pantographs. The pantograph is fixed on a standard inspection fixture, and the system acquires images of the carbon sliding plate's current-collecting working surface, the pantograph head beam, the pantograph angle, the mounting base, and the support arm, generating an initial millimeter parameter domain file for the pantograph. During subsequent operational inspections, the system compares the height field of the current-collecting working surface obtained from the current inspection with the initial millimeter parameter domain file in the same coordinate system to obtain the continuous wear height, abrupt defect height, and structural deviation. This embodiment ensures a continuous data link for the same pantograph from the factory to the operation and maintenance stage.

[0119] In summary, this invention forms a complete detection process from image acquisition, millimeter measurement, mechanical state inversion to maintenance conclusion output by combining millimeter parameter domain of the flow-receiving working surface, local millimeter-scale field, micro-sweep sub-pixel reconstruction, joint inversion of attitude elastic wear defects, and defect evidence fusion. It is applicable to through-process inspection, static verification inspection, and production line inspection. Specific Implementation Example 4:

[0121] like Figures 1 to 9 As shown, based on the content of the above specific embodiments, the working principle of the present invention is further disclosed:

[0122] The working principle of this invention is as follows: First, a stable external spatial reference is established through the track reference coordinate system. Then, an intrinsic measurement reference that changes with the current pantograph attitude is generated through the pantograph's own structure. Subsequently, the carbon slide plate current receiving working surface is unfolded into a unified millimeter parameter domain. Then, continuous images from multiple cameras are used to complete local scale conversion, micro-sweep reconstruction, joint inversion of mechanical state, and defect evidence fusion within this millimeter parameter domain. This transforms pixel changes in ordinary images into detection results with clear physical coordinates, millimeter dimensions, and mechanical meaning.

[0123] Specifically, before the vehicle enters the detection area, the system has completed the calibration between the track reference coordinate system, the detection gantry reference point, and the coordinate systems of each camera. The track centerline, track surface height, and detection gantry reference point form an external fixed reference. The top camera, left oblique camera, right oblique camera, and side camera are all unified to this external fixed reference through camera intrinsic parameters, camera extrinsic parameters, and distortion parameters. When the vehicle enters the detection area, the vehicle position trigger unit sends a trigger signal to the edge computing unit. The edge computing unit synchronously controls multiple cameras to acquire a continuous image sequence and binds each frame of image with the acquisition time, vehicle position trigger value, camera number, exposure value, supplementary lighting status value, vehicle number, and pantograph number. This process ensures that subsequent detection not only has image data but also a temporal and spatial index corresponding to the vehicle position and camera viewpoint.

[0124] After image acquisition, the system first performs distortion correction, exposure normalization, reflection suppression, and motion blur evaluation on the images. Then, it extracts the carbon slide current-collecting zone boundary and centerline from the top image, the carbon slide end boundary and bow angle transition boundary from the oblique image, and the carbon slide thickness edge, bow head crossbeam boundary, mounting base boundary, and support arm node boundary from the side image. These boundaries are not isolated image contours but rather stable geometric features within the pantograph's mechanical structure. The system uses the carbon slide current-collecting zone centerline to determine the length direction of the pantograph's intrinsic coordinate system, the width direction of the carbon slide current-collecting zone to determine the width direction, and the height direction of the pantograph's intrinsic coordinate system using the fitting normal of the current-collecting working surface. The center point of the current-collecting zone at one end of the carbon slide is used as the origin of the millimeter parameter domain. Therefore, the detection process no longer depends on the absolute position of the pantograph in the image but on the local mechanical reference formed by the pantograph's own structure.

[0125] After establishing the pantograph's intrinsic coordinate system, the system expands the carbon strip's current-collecting working surface into a millimeter parameter domain. This millimeter parameter domain consists of length direction parameter coordinates, width direction parameter coordinates, and surface height values. The length direction parameter coordinates correspond to the position of the carbon strip extending laterally along the vehicle, the width direction parameter coordinates correspond to the position within the width range of the carbon strip's current-collecting zone, and the surface height value corresponds to the actual height of the carbon strip after current wear. The system maps the same physical position of the carbon strip in the top image, left oblique image, right oblique image, and side image to the same millimeter parameter point, aligning data from different cameras, different image frames, and different inspection batches under the same physical coordinates. This principle allows historical inspection results, current inspection results, and subsequent inspection results to be compared around the same current-collecting working surface coordinates, thus forming a millimeter-level inspection archive for the entire pantograph's lifespan.

[0126] Because the pantograph has a long, narrow shape, multiple depth planes, and multiple angles, the system does not use a uniform pixel scale for the entire image. Instead, it generates a local millimeter-scale field within the millimeter parameter domain. For each millimeter parameter point, the system calculates the pixel position of that point in each camera image based on the camera projection relationship. Then, it applies small parameter perturbations along the length and width directions to obtain the pixel change and actual physical distance change corresponding to adjacent parameter points, and then calculates the local millimeter scale at that location. The central area of ​​the carbon slide plate mainly uses the local millimeter-scale subfield from the top camera, the end corner and bow angle areas mainly use the local millimeter-scale subfields from the left and right oblique cameras, and the remaining thickness of the carbon slide plate and the position of the bow head beam mainly use the local millimeter-scale subfields from the side cameras. This principle ensures that crack length, corner size, bow angle sag, wear slope, and remaining thickness are all converted according to the actual spatial scale of the location, avoiding measurement drift caused by using the same scale at different depth locations.

[0127] When a vehicle passes through the detection gantry, the pantograph experiences continuous minute displacements relative to the camera. Traditional methods treat this displacement as detection interference, while this invention treats it as a source of micro-sweep observations. The system uses the image frame where the pantograph is located in the middle of the detection window as the reference frame, calculates the effective micro-baseline length between other image frames and the reference frame, and scores the image frames based on the average gradient intensity of the current-collecting working surface area, motion blur index, structural alignment index, and coverage integrity index. Image frames whose quality scores reach the reconstruction threshold are selected into the micro-sweep image frame set. Since these image frames are observations from the same pantograph under continuous minute viewing angle changes, after registering them to the same millimeter parameter domain, the system can fuse the sub-pixel edge positions, grayscale gradient extreme values, and local millimeter scales of multiple cameras and multiple frames at the same parameter point, thereby reconstructing the surface height field of the carbon slide current-collecting working surface, the carbon slide edge contour, the bow angle end contour, and the bow head crossbeam spatial contour.

[0128] After the surface height field is reconstructed, the system does not directly classify image anomalies as faults. Instead, it uses a joint mechanical state inversion to interpret the sources of anomalies in a hierarchical manner. The current state of the pantograph is decomposed into overall attitude variables, elastic deformation variables, continuous wear variables, and abrupt defect variables. The overall attitude variables are used to explain the overall rotation and translation of the pantograph relative to the track reference coordinate system when the vehicle passes through the detection area; the elastic deformation variables are used to explain the normal deformation of the pantograph head, pantograph angle, support arm, and mounting base within the mechanically permissible range; the continuous wear variables are used to explain the smooth height decrease caused by long-term flow friction on the carbon slide plate; and the abrupt defect variables are used to explain local height abrupt changes caused by cracks, missing corners, chipped edges, and peeling. The system uses a weighted summation of image observation errors, attitude smoothing errors, mechanical topology errors, elastic rationality errors, continuous wear errors, and abrupt defect errors to jointly solve the problem. This ensures that each image change is first explained by attitude, elasticity, and continuous wear, and only the remaining unexplained local abrupt changes are included in the defect residual field.

[0129] On the pantograph's receiving surface, normal wear typically exhibits continuous variation, while cracks, chipped corners, spalling, and edge breakage manifest as localized abrupt changes. Based on this service pattern, the system decomposes the current surface height into standard surface height, continuous wear height, abrupt defect height, and residual error. The standard surface height is derived from the pantograph's initial profile or vehicle structure database; the continuous wear height is constrained by continuity within the millimeter parameter domain; the abrupt defect height is constrained by localized sparsity within the millimeter parameter domain; and the residual error corresponds to image noise, surface contamination, and reconstruction error. Through this decomposition process, the system separates normal wear caused by long-term operation from abrupt defects caused by localized damage, avoiding the direct identification of carbon deposits, surface reflections, shallow scratches, and oil stains as structural faults.

[0130] The calculation of detection indicators is also completed within the millimeter parameter domain. The remaining thickness of the carbon slide plate is determined by the distance between the upper and lower edges obtained from the lateral image reconstruction and the height of the flow-receiving working surface. The maximum wear depth is determined by the maximum difference between the standard surface height and the current surface height. The average wear depth is determined by the average continuous wear height within the effective flow-receiving area. The wear volume is obtained by integrating the continuous wear height over the effective flow-receiving area. The wear slope is obtained by the ratio of the height difference to the width distance at different width sampling positions under the same length direction coordinate. The crack space length is obtained by the centerline length of the slender connected region in the abrupt defect height field. The corner boundary size is obtained by the boundary span of the abrupt defect region at the end of the carbon slide plate. The bow angle sag is obtained by the difference between the height of the bow angle end and the reference height of the bow head beam. The bow head center offset is obtained by the lateral distance between the centerline of the bow head beam and the track reference centerline. The mounting position deviation is obtained by the difference between the reconstructed mounting position and the standard structural position. All detection indicators are bound to the corresponding millimeter parameter domain coordinates, camera number, image frame number, and local millimeter-scale field number.

[0131] Finally, the system generates an evidence chain for each suspected defect. Suspected defects originate from locally connected regions within the height field of abrupt defect changes. The system calculates image evidence values, millimeter-scale evidence values, 3D height evidence values, mechanical topology evidence values, continuous frame temporal evidence values, and historical trend evidence values. Image evidence values ​​reflect the degree of boundary consistency of the suspected defect across multiple camera images; millimeter-scale evidence values ​​reflect the dimensional stability of the suspected defect after conversion through different local millimeter-scale subfields; 3D height evidence values ​​reflect the degree of height abrupt change of the suspected defect in the surface height field; mechanical topology evidence values ​​reflect the degree to which the suspected defect cannot be explained by the normal mechanical topology and elastic deformation of the pantograph; continuous frame temporal evidence values ​​reflect the stability of the suspected defect falling into the same millimeter parameter position in consecutive micro-scanned image frames; and historical trend evidence values ​​reflect the degree of deviation of the suspected defect from the historical wear trend of the same pantograph. The system integrates these evidence values ​​into a defect confidence value and, combined with maintenance thresholds, outputs conclusions of compliance, observation warning, time-limited maintenance, and immediate maintenance.

[0132] The working principle of this invention is embodied in a step-by-step transformation process from image to mechanical state: First, the external spatial reference problem is solved using the orbital reference coordinate system; second, the self-reference problem under the current posture is solved using the pantograph's intrinsic coordinate system; third, the alignment problem of multiple cameras, multiple frames, and historical data is solved using the millimeter parameter domain; fourth, the inconsistency of pixel scale at different spatial locations is solved using a local millimeter-scale field; fifth, the insufficient height information of a single-frame image is solved using micro-sweep sub-pixel reconstruction; sixth, the confusion between normal posture, normal elasticity, continuous wear, and abrupt defects is solved using joint inversion of mechanical state; and seventh, the problem of verification of detection conclusions and application in maintenance is solved using a defect evidence chain. Through the above continuous processing, this invention elevates the pantograph visual inspection results from ordinary image anomalies to detection conclusions with millimeter-scale dimensions, mechanical constraints, and historical traceability.

[0133] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising a reference structure" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0134] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A millimeter-level visual inspection method for pantographs, characterized in that, Includes the following steps: Sp1. Establish an external coordinate reference for the detection area, and unify the track centerline, track surface height, detection gantry reference point, top camera coordinates, left oblique camera coordinates, right oblique camera coordinates, and side camera coordinates into the track reference coordinate system; Sp2. When the vehicle enters the detection area, acquire the top image sequence, left oblique image sequence, right oblique image sequence and side image sequence, and record the acquisition time, vehicle position trigger value, camera exposure value, supplementary light status value, vehicle number and pantograph number corresponding to each frame of the image. Sp3. Extract the carbon slide current receiving zone boundary, carbon slide end boundary, bow angle transition boundary, bow head crossbeam boundary, and mounting base boundary from the top image sequence, left oblique image sequence, right oblique image sequence, and side image sequence, and establish the pantograph intrinsic coordinate system based on the carbon slide current receiving zone boundary, carbon slide end boundary, and bow head crossbeam boundary. Sp4. Expand the carbon slide current-collecting working surface in the pantograph intrinsic coordinate system into a millimeter parameter domain composed of length direction parameter coordinates and width direction parameter coordinates, and map the same carbon slide position in the top image sequence, left oblique image sequence, right oblique image sequence and side image sequence to the millimeter parameter domain. Sp5. Based on the camera intrinsic parameters, camera extrinsic parameters, camera distortion parameters, orbit reference coordinate system, and pantograph intrinsic coordinate system, generate a local millimeter-scale field in the millimeter parameter domain; Sp6. Micro-sweep image frames are selected based on the effective micro-baseline between image frames, gradient of the flow-receiving working surface region, motion blur index, structural alignment index and coverage integrity index, and the selected micro-sweep image frames are registered to the same millimeter parameter domain. Sp7. Based on the registered micro-sweep image frames, the surface height field of the carbon slide plate receiving working surface, the carbon slide plate edge profile, the bow corner end profile, and the bow head crossbeam spatial profile are reconstructed. Sp8. Decompose the current state of the pantograph into overall attitude variables, elastic deformation variables, continuous wear variables, and abrupt defect variables, and obtain the mechanical state of the pantograph through joint inversion; Sp9, within the millimeter parameter domain, calculate the remaining thickness of the carbon slide plate, wear depth, wear volume, wear slope, crack space length, corner boundary size, bow angle droop, bow head center offset, and mounting position deviation; Sp10 generates a defect confidence value based on image evidence value, millimeter-scale evidence value, three-dimensional height evidence value, mechanical topology evidence value, continuous frame temporal evidence value, and historical trend evidence value, and outputs the detection conclusion based on the defect confidence value and maintenance threshold.

2. The pantograph millimeter-level visual inspection method according to claim 1, characterized in that: When establishing the external coordinate reference of the detection area in Sp1, the direction of the track centerline is set as the first coordinate axis of the track reference coordinate system, the direction of the track surface normal is set as the second coordinate axis of the track reference coordinate system, and the direction that is perpendicular to both the first and second coordinate axes is set as the third coordinate axis of the track reference coordinate system. The external parameter matrices of the top camera, the left oblique camera, the right oblique camera, and the side camera relative to the track reference coordinate system are determined using the reference calibration points fixed on the detection gantry.

3. The pantograph millimeter-level visual inspection method according to claim 1, characterized in that: When establishing the pantograph intrinsic coordinate system in Sp3, the direction of the center line of the carbon slide current receiving strip is taken as the length direction of the pantograph intrinsic coordinate system, the direction of the width of the carbon slide current receiving strip is taken as the width direction of the pantograph intrinsic coordinate system, the normal direction obtained by fitting the plane of the carbon slide current receiving strip is taken as the height direction of the pantograph intrinsic coordinate system, and the center point of the current receiving strip at one end of the carbon slide is taken as the origin of the millimeter parameter domain.

4. The pantograph millimeter-level visual inspection method according to claim 1, characterized in that: The millimeter parameter domain in Sp4 consists of length direction parameter coordinates, width direction parameter coordinates, and surface height values. The length direction parameter coordinates are set along the length direction of the carbon slide plate, the width direction parameter coordinates are set along the width direction of the flow receiving zone of the carbon slide plate, and the surface height values ​​are set along the normal direction of the flow receiving working surface of the carbon slide plate. The pixels in the top image sequence, left oblique image sequence, right oblique image sequence, and side image sequence are correlated with the millimeter parameter domain parameter points through a camera projection model.

5. The pantograph millimeter-level visual inspection method according to claim 1, characterized in that: The local millimeter-scale field in Sp5 is jointly determined by the camera projection Jacobian matrix, the millimeter parameter domain geometric Jacobian matrix, and the pantograph intrinsic coordinate system attitude matrix, and generates local millimeter-scale subfields for the top camera, the left oblique camera, the right oblique camera, and the side camera, respectively.

6. The pantograph millimeter-level visual inspection method according to claim 1, characterized in that: The selection of micro-sweep image frames in Sp6 includes: calculating the effective micro-baseline length between the current image frame and the reference image frame; calculating the average gradient intensity of the current image frame's current working surface region; calculating the motion blur index of the current image frame; calculating the structural alignment index between the bow head beam boundary, carbon slide end boundary, and mounting base boundary of the current image frame and the pantograph mechanical topology model; calculating the coverage integrity index of the current working surface region in the current image frame; and obtaining an image frame quality score based on the effective micro-baseline length, average gradient intensity, motion blur index, structural alignment index, and coverage integrity index.

7. The pantograph millimeter-level visual inspection method according to claim 1, characterized in that: When reconstructing the surface height field of the carbon slide plate's flow-receiving working surface in Sp7, the selected micro-sweep image frames are registered according to camera number, frame number, and millimeter parameter domain parameter points. The sub-pixel edge position and gray-level gradient extreme value position of the same millimeter parameter domain parameter point in different image frames are extracted, and the surface height field is calculated based on the sub-pixel edge position, the gray-level gradient extreme value position, and the local millimeter-scale field.

8. The pantograph millimeter-level visual inspection method according to claim 1, characterized in that: The joint inversion in Sp8 includes: using the observation errors of the top image sequence, left oblique image sequence, right oblique image sequence, and side image sequence as the first constraint term; using the attitude change error of consecutive frames as the second constraint term; using the mechanical topology error between the carbon slide plate, the bow head beam, the bow angle, the mounting base, and the support arm as the third constraint term; using the elastic deformation error of the pantograph as the fourth constraint term; using the continuous wear error of the carbon slide plate as the fifth constraint term; and using the abrupt defect error of the carbon slide plate as the sixth constraint term. The first, second, third, fourth, fifth, and sixth constraint terms are then summed and solved using a weighted summation.

9. The pantograph millimeter-level visual inspection method according to claim 1, characterized in that: The continuous wear variable and abrupt defect variable of the carbon slide plate in Sp9 are obtained through surface height decomposition. The surface height decomposition is based on the standard surface height of the carbon slide plate, and the current surface height is decomposed into continuous wear height, abrupt defect height and residual error. A parameter domain continuity constraint is applied to the continuous wear height, and a local sparsity constraint is applied to the abrupt defect height.

10. The pantograph millimeter-level visual inspection method according to claim 1, characterized in that: When generating a defect confidence value in Sp10, the boundary consistency value of the suspected defect object in different camera images is used as the image evidence value; the size stability value of the suspected defect object after local millimeter-scale field conversion is used as the millimeter-scale evidence value; the height mutation value of the suspected defect object in the surface height field is used as the three-dimensional height evidence value; the residual value between the suspected defect object and the pantograph mechanical topology model is used as the mechanical topology evidence value; the parameter domain position stability value of the suspected defect object in consecutive image frames is used as the consecutive frame temporal evidence value; and the trend deviation value between the suspected defect object and historical detection records is used as the historical trend evidence value. The image evidence value, millimeter-scale evidence value, three-dimensional height evidence value, mechanical topology evidence value, consecutive frame temporal evidence value, and historical trend evidence value are fused into a defect confidence value.

Images