Urban rail train cable slot cover plate defect detection method and system

Abstract

The application discloses a kind of urban rail train cable slot cover plate defect detection method and system, including identifying structural features in visible light image and establishing assembly topological relationship, generate structure prior graph;Visible light and infrared image cross-modal fusion is carried out, and through the unstable area of inhibiting network prediction inhibits mask to suppress artifact;Fusion features are compared with standard templates, and topological edge geometric scale movement is combined to extract the paired change evidence representing component macroscopic form displacement;In the fusion feature, the local infrared temperature field of adhesive strip node is extracted, and the single-frame structure evidence representing implicit defects is obtained according to the temperature gradient change rate in the direction of adjacent topological edge;Through the state gating decision output of two parallel branches, the explicit and implicit defect results are output, the application effectively eliminates train operation vibration and illumination artifact interference, realizes the collaborative and accurate detection of cable slot cover plate explicit mechanical damage and adhesive inner surface shedding implicit thermal defects.

Description

Technical Field

[0001] This invention relates to the field of rail train component inspection technology, and more specifically to a method and system for detecting defects in cable tray covers of urban rail trains. Background Technology

[0002] During operation, urban rail transit vehicles (such as subways, light rail, and intercity EMUs) have a large number of high-voltage power, control, and communication cables laid in their chassis and roof areas. To provide mechanical and physical protection for these cables, long strip-shaped metal cable tray covers are usually installed on the train chassis and cable routing areas. Because the cable tray covers are subjected to high-frequency vibration, wind swaying, and extreme humidity, heat, and impact from flying rocks for extended periods, the elastic protective rubber strips on the edges of the covers are prone to aging, cracking, and detachment.

[0003] To prevent internal cables from being cut by sharp metal edges due to the failure of protective rubber strips, the industry has widely adopted automated visual inspection systems based on intelligent inspection robots or trackside dynamic image inspection booths. Currently, existing inspection methods fall into two categories. The first category is component condition detection technology based on two-dimensional visible light template comparison. For example, patent number CN118506037B discloses a method for matching components under a rail vehicle. This method uses an inspection robot equipped with an image acquisition unit to capture images of the vehicle's undercarriage and aligns and compares the image to be inspected with a pre-stored defect-free standard template image, based on the texture between the images. The method uses differences in physical characteristics, color differences, or abrupt changes in edges to determine the missing or displaced components. However, this method relies entirely on a one-sided two-dimensional optical perspective. In the actual engineering conditions of cable tray covers, edge protection strips often experience "one-sided latent failure." That is, the inner surface of the strip that adheres to the metal edge has been largely detached due to wear or vibration tearing. However, due to the rigidity of the rubber material itself or the influence of external stress compression, the part of the strip exposed outside the camera still maintains the transient stability of its shape and texture. Since the two-dimensional camera cannot see through and obtain information about the inner surface, such serious hidden dangers score normally in template comparison, leading to a large area of ​​missed detection.

[0004] The second category is geometric shape measurement technology based on three-dimensional laser line scanning or structured light. For example, patent number CN120031866A discloses a dynamic anomaly element adjustment system for key components based on 3D laser scanning. It uses a laser line scanner to capture three-dimensional point clouds on the surface of the component and identifies the bending and tension gaps of the component by calculating the step difference or flatness of the point cloud in the Z-axis (height direction). However, although this method can introduce depth information, it is extremely susceptible to serious interference from "physical thermal and visual artifacts" in the harsh environment under the train. On the one hand, the train runs on the line for a long time, and a thick layer of black heavy matter accumulates on the undercarriage and cable trough cover. When oil, dirt, iron filings, or ice fill the seams of the cover plate or the edges of the rubber strip, the geometric features measured by the 3D point cloud will be completely obliterated, causing the height difference algorithm to fail. On the other hand, oil will form physical thermal resistance on the surface and generate local heat accumulation (specific hot spots) under the heat of the internal power cables. Existing 3D scanning and conventional infrared technology cannot distinguish whether the hot spots are caused by interference from "oil accumulation" or by a real defect caused by "metal exposure and heat dissipation due to missing inner surface of the rubber strip". As a result, the system will generate a very high false alarm rate when facing complex dirt under the vehicle, and cannot meet the needs of automated closed-loop operation and maintenance. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for detecting defects in cable trough covers for urban rail trains. The purpose is to solve the problems in existing technologies where the lack of macroscopic deformation characterization of the inner surface of the rubber strip of the trough cover makes it difficult to achieve accurate identification based on single visible light or simple infrared detection methods. Furthermore, in complex track environments, the trough cover is easily affected by abnormal temperature changes, surface oil contamination, and light and shadow artifacts, resulting in low detection accuracy and high false alarm rate.

[0006] Therefore, this invention provides a method and system for detecting defects in cable trough covers for urban rail trains. By constructing a structural prior diagram that integrates geometric assembly relationships, and based on cross-modal feature fusion and unstable region suppression, adaptive suppression of external artifact interference is achieved. Simultaneously, paired change evidence representing changes in appearance and single-frame structural evidence representing damage to the internal thermal conduction barrier are computed in parallel. Finally, the two sets of evidence are complementaryly fused through state gating decision, which enables accurate detection of latent defects such as detachment of the inner surface of the cable trough cover rubber strip, significantly reducing false alarms caused by environmental noise interference.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for detecting defects in cable tray covers for urban rail trains includes the following steps: The structural topology modeling steps involve identifying structural features in the visible light image of the cable tray cover area to be inspected and marking them as structural nodes in the topology graph. Based on the assembly relationship of each structural feature, topological edges are established, and a structural prior graph is generated through a structural topology coding strategy.

[0008] Furthermore, this step specifically involves extracting edge geometric lines and key deformation corner points from the visible light image to be inspected, determining the coordinates, scale, and orientation of each structural feature as the node attribute vector of the corresponding structural node, then extracting the spatial geometry and assembly relationship between any two structural nodes as the edge type feature vector of the corresponding topological edge, interactively fusing the node attribute vector and the edge type feature vector, and then generating a structural prior map by projecting it into the image feature grid through a Gaussian distribution.

[0009] By upscaling scattered pixel information into a graph topology with rigid physical and mechanical assembly constraints, and projecting the topological features into the image grid through a Gaussian distribution, subsequent feature extraction and residual comparison are restricted to a rigid track locked by the topological prior graph. This physically eliminates false alarms caused by the bumps and swaying of the train at high speed, camera vibration, and slight shifts in shooting angle, thus locking in an absolute spatial reference benchmark for the entire defect detection process.

[0010] The environmental artifact suppression step involves fusing cross-modal features of the visible light image and infrared image to be inspected to obtain multi-scale dual-light fusion features. Based on the surface texture and edge continuity features in the visible light image to be inspected, an unstable region suppression mask is predicted by an unstable region suppression network.

[0011] Furthermore, the cross-modal feature fusion adopts a fusion strategy, including extracting multi-layer high-frequency spatial edge features from the visible light image to be inspected and extracting multi-layer deep semantic features from the infrared image. A guide map transformation weight matrix is ​​constructed based on the multi-layer high-frequency spatial edge features, and local affine transformation calculations are performed on the pixel positions in the multi-layer deep semantic features. The multi-layer deep semantic features after the affine transformation calculation are then concatenated with the corresponding multi-layer high-frequency spatial edge features in the channel dimension to obtain multi-scale dual-light fusion features.

[0012] Guided by the rich high-frequency edge spatial information in visible light images, local affine transformation constraints and high-definition reshaping are performed on infrared features with low spatial resolution and severe thermal radiation edge dispersion at the feature level. This is equivalent to putting a high-definition rigid boundary shell on the blurred infrared local temperature field, correcting and constraining the spatial dispersion of heat, so that the system can extract a high-precision discrete infrared temperature field in an extremely small adhesive strip area, solving the problem of inaccurate localization in small areas in traditional cross-modal detection.

[0013] The dual-evidence parallel computation step compares the residuals of the multi-scale dual-light fusion features with the pre-stored standard template, and extracts paired change evidence representing the macroscopic morphological displacement of the component's appearance by using the geometric scale anomalies of each topological edge in the structural prior image; extracts the local infrared temperature field of the structural node region corresponding to the adhesive strip from the multi-scale dual-light fusion features, and obtains single-frame structural evidence representing the destruction of the internal thermal conduction barrier of the non-visible surface based on the temperature gradient change rate in the direction of the topological edge connected to the structural node. Furthermore, when extracting evidence of paired changes representing macroscopic morphological displacement, a deformable topological alignment strategy and a topological anomaly activation strategy are integrated. This includes extracting local topological subgraph features of multi-scale dual-light fusion features and standard template features at each structural node and its connected topological edges, calculating the cosine similarity matrix, and constructing a two-dimensional deformation offset vector field. Based on this, pixel-level deformable spatial grid alignment based on local interpolation is performed on the multi-scale dual-light fusion features to eliminate non-uniform deformation distortion caused by mechanical parallax. Subsequently, the appearance reconstruction residual after alignment is calculated, and the axial length indentation and projection tilt angle offset along each topological edge between the area to be inspected and the standard template are statistically analyzed as geometric scale anomalies. These are weighted and fused with the appearance reconstruction residual to obtain local macroscopic changes with an initial numerical distribution. Finally, the initial numerical distribution is calculated using nonlinear bending scaling based on a preset bending coefficient and normalized. The output feature probability matrix in the zero-to-one interval serves as evidence of paired changes.

[0014] By using pixel-level deformable mesh stretching and alignment, the pseudo-edge residuals introduced by non-defect factors are forcibly smoothed out. Nonlinear bending and scaling are used to forcibly increase the gradient slope in the medium numerical range, which makes the system's sensitivity to minor loosening, displacement and weak appearance variation of cover plate or rubber strips leap exponentially. At the same time, normalization is used to cut off the energy mutation caused by extreme large-area foreign object obstruction, ensuring the ultra-high robustness of paired change evidence extraction.

[0015] Furthermore, when extracting single-frame structural evidence characterizing the destruction of the internal thermal conduction barrier of the non-visible surface, a topological temperature difference sampling strategy is integrated. This includes extracting the discrete temperature matrix of the local infrared temperature field in the spatial coordinate system, establishing a local sampling axis with the structural node corresponding to the adhesive strip as the origin, and establishing the orthogonal tangent and normal directions of the topological edge connected to it. In the discrete temperature matrix, temperature sampling is performed at equal intervals along the local sampling axis, and then the temperature difference between adjacent sampled pixels in the orthogonal tangent and normal directions is calculated respectively. The change in the temperature difference within a unit pixel distance is used as the corresponding local temperature gradient change rate.

[0016] The process of damage to the air insulation layer, which is difficult to observe inside the non-visible surface, is cleverly transformed into a quantitative index of temperature difference faults in the spatial sampling axis by topological edge deadlock. By utilizing the unique bidirectional gradient constraint of "a sharp drop in normal temperature perpendicular to the edge and smooth temperature conduction in tangential temperature parallel to the edge" when the adhesive strip falls off, a precise physical geometric temperature difference filter is constructed. This filter removes large-scale global heat fluctuations caused by uneven external sunlight and overall environmental temperature rise, making the extraction of hidden detachment defects have extremely high physical specificity and accuracy.

[0017] In the state gating decision step, when the paired change evidence determines that the appearance is normal, and the single-frame structural evidence after the unstable region suppression mask determines that the structural node corresponding to the adhesive strip has a local heat leakage anomaly, the latent defect decision result of the adhesive strip detaching is obtained.

[0018] Furthermore, this step deeply integrates thermal mutation differentiation and adaptive adjustment strategies. This includes using the pixel-level instability probability matrix output by the unstable region suppression network (where each pixel value represents the probability of surface oil contamination or light and shadow interference in the corresponding region) as an unstable region suppression mask. The single-frame structural evidence is then transformed into a thermal mutation anomaly response map representing a two-dimensional spatial distribution. The corresponding pixel values ​​in the unstable region suppression mask are inverted to obtain the corresponding effective environmental confidence weight. The effective environmental confidence weight and the thermal mutation anomaly response map are then multiplied and attenuated pixel by pixel to obtain a local heat leakage anomaly response as the suppressed single-frame structural evidence. Simultaneously, the local alignment confidence is determined based on the alignment deviation during cross-modal feature fusion, and the proportion of oil contamination area within the adhesive strip region is determined based on the unstable region suppression mask. The dual-evidence fusion weight is then dynamically adjusted based on the local alignment confidence and the proportion of oil contamination area. When the proportion of oil contamination area is higher than a preset proportion threshold or the local alignment confidence is lower than a preset confidence threshold, the dual-evidence fusion weight is reduced, increasing the weight of the single-frame structural evidence in the decision.

[0019] A perfect adaptive error-proofing closed loop is achieved at the decision-making level. Utilizing pixel-level instability probability reversal and product attenuation, it can spatially "force zero" or deeply suppress false thermal energy caused by surface oil stains and strong reflections at the track site, fully retaining only the heat leakage response truly caused by the detachment of internal adhesive strips. When encountering severe lighting or violent shaking that leads to extremely low cross-modal alignment confidence or excessive oil stain area, the system adaptively reduces the weight of appearance change evidence, which is easily affected by alignment and surface occlusion contamination, and comprehensively increases the decision weight of single-frame structural evidence, which is not restricted by alignment. Finally, only when the external appearance is completely normal (excluding external mechanical damage and false alarms from shaking), and the single-frame structural evidence after pixel attenuation still strongly indicates that there is a local heat leakage anomaly in the adhesive strip, is the hidden detachment result finally determined and output. This constructs an extremely rigorous mutual exclusion verification and false alarm mechanism, completely solving the industry problem of the difficulty in online identification of hidden defects and the large interference of environmental artifacts, minimizing the system's false alarm rate.

[0020] The present invention also provides a defect detection system for cable tray covers of urban rail trains, comprising: Memory, used to store computer programs; The processor is used to execute the computer program to implement the steps of the method for detecting defects in cable trough covers of urban rail trains. Through the hardware combination of high-performance processor and memory, all the above-mentioned high-density rigid topology coding, cross-modal guided fusion, pixel-level product attenuation, dual-sampling axis temperature difference filtering, and state-gated two-dimensional adaptive verification algorithms are converted into real-time computer running instructions with high concurrency. It can be directly deployed on the train's onboard inspection intelligent terminal or track inspection robot, and achieve fully automatic, highly reliable, and no-missed-report online real-time detection and location of hidden defects under the harsh working conditions of high-speed dynamic train operation.

[0021] The beneficial effects of this invention are as follows: 1. This invention uses parallel computation of dual evidence to simultaneously extract paired change evidence and single-frame structure evidence. Through a state gating decision mechanism, under the premise of excluding external mechanical damage and false alarms caused by shaking (no abnormality in appearance), it can accurately locate local heat leakage anomalies caused by internal detachment of the adhesive strip and damage to the air insulation layer. Compared with the traditional visible light image defect detection method that relies heavily on appearance variation to identify defects, this invention achieves high-precision identification and quantification of deep hidden defects that have no external variation but internal changes.

[0022] 2. This invention introduces a structural topology coding strategy at the forefront, which elevates loose pixel features to a graph topology structure with strong physical and mechanical assembly constraints. Combined with a deformable topology alignment strategy, pixel-level spatial grid alignment based on local interpolation is performed through a two-dimensional deformable offset vector field. This adaptive stretching alignment forcibly smooths out pseudo-edge residuals introduced by non-defect factors, locking the entire detection and comparison process within a priori coordinate system with spatial geometric invariance, ensuring zero false negatives and extremely high robustness under high-speed shaking environments.

[0023] 3. This invention uses a thermal mutation differentiation strategy to invert the predicted pixel-level instability probability matrix into an effective environmental confidence weight, and then multiplies and attenuates it pixel by pixel with the single-frame thermal mutation map. For reflective and oily areas, since their confidence weight is extremely low, their energy is instantly forced to zero or deeply attenuated after multiplication; while the real defect heat leakage signal is fully preserved due to its high confidence. This adaptive multiplication attenuation mechanism cuts off the interference channel of false thermal artifacts from the physical and feature-level. Attached Figure Description

[0024] Figure 1 This is an overall framework diagram of the detection method in this invention; Figure 2 This is a schematic diagram of the detection scene in this invention; Figure 3 This is a thermal image of the inner surface of the adhesive strip detached in this invention; Figure 4 This is a thermal imaging image of the oil stains accumulated on the rubber strip in this invention; Figure 5 This is a diagram illustrating the process of topological temperature difference sampling and single-frame structural evidence extraction in this invention. Detailed Implementation

[0025] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Identical components are denoted by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "upper," and "lower" used in the following description refer to directions in the accompanying drawings, and the terms "bottom surface," "top surface," "inner," and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0026] Example 1 This embodiment, from image input, topology modeling, feature alignment and fusion, bidirectional evidence parallel extraction to state gating decision, is used for fully automated online detection of defects in the internal adhesive strip of the non-visible surface of the cable tray cover.

[0027] like Figure 2As shown, the hardware uses a track inspection robot equipped with a visible light industrial camera and an infrared thermal imaging camera to simultaneously collect images. The robot has a built-in embedded processor and memory. The memory stores standard cable tray cover template images of various vehicle models, topology coding rules, network weights, various judgment thresholds, and a mapping library of structural nodes and topology edges.

[0028] The visible light image to be examined; Simultaneous acquisition of infrared images; : Prior topology diagram of cable tray cover plate structure; : A set of structural nodes; : No. The structural nodes are divided into six categories: cover plate edge, long slot, joint, fastener, adhesive strip, and shielding. : Set of topological edges; :node and Assembly topology edges between; Pixel position Topological coding features at the location; : The coordinates of any pixel within the ROI of the image; The cable tray cover effectively detects the ROI area.

[0029] like Figure 1 , Figure 3-4 As shown, the structural topology modeling steps involve inputting the visible light image of the object to be inspected, which is synchronously acquired by the inspection robot. The cropped image contains only the ROI area of ​​the cable tray cover. After filtering out irrelevant background elements such as undercarriage pipes and bogies, all geometric features within the ROI are extracted based on Canny edge detection, Hough line detection, and contour roundness detection. Node division is then completed: long, parallel boundaries are marked as cover plate edge nodes. Long, repeating dark areas are marked as slot nodes. The straight line markings between adjacent cover plates are the joint nodes. Circular outline markings are used for fastener nodes. The rubber strips on the edge of the cover plate are marked as rubber strip nodes. Large, irregular, dark-colored occlusions are marked as occlusion nodes. All geometric features are uniformly generated into a node set. Each node Generate node attribute vectors, which contain four types of information: coordinates, scale, orientation, and node type.

[0030] Iterate through all nodes and pair them up to construct a set of topological edges. : Spatially adjacent nodes generate adjacent edges, parallel nodes in the same direction generate collinear edges, nodes within the same segment of the cover plate generate edges within the same cover plate segment, and occluding nodes covering the solid structure nodes generate occluding / covering edges; each topological edge Generate edge-type feature vectors to record the spatial geometry and assembly constraints between nodes.

[0031] By interactively fusing node attribute vectors and edge type feature vectors, and mapping the graph topological features to each pixel location in the image through Gaussian distribution projection, pixel-level topological coding features of the entire ROI are generated. Combine all node and edge information to form a structural prior graph. It is stored in memory for subsequent registration and evidence calculation.

[0032] Environmental artifact suppression steps, synchronously reading infrared images from the same location. , respectively , Perform multi-scale feature extraction: extract high-frequency edge features at three levels (1 / 4, 1 / 8, and 1 / 16) using visible light. , Representative feature scale; infrared extraction of three-layer deep temperature semantic features Construct a guided transformation weight matrix based on the visible light edge gradient. Higher gradient magnitudes are assigned higher weights to adhesive strips, slots, and seams; affine transformation is used to correct for spatial dispersion of infrared features. The formula for calculating infrared features after correction is as follows: ,in, For pixels ,scale The visible light edge guides the weights, and the larger the gradient, the higher the weight, constraining the infrared features to fit the high-definition visible light boundary; The infrared features are after boundary correction.

[0033] Channel splicing of visible light high-frequency edges and corrected infrared features generates multi-scale dual-light fusion features. Calculation formula: ,in, For channel-dimensional stitching operations, the output is a dual-modal feature that integrates visible light spatial details and infrared temperature information, providing a unified input for subsequent alignment and evidence calculation.

[0034] Construct a lightweight U-shaped unstable region suppression network, and input dual-light fusion features and topological coding features. The network outputs a pixel-level probability matrix of unstable regions. , A higher value indicates a higher probability that the pixel contains oil stains, glare, shadows, or interference from moving cables; according to Calculate the credibility weight of the effective environment: , Used for all subsequent evidence decay calculations.

[0035] The unstable region suppression network is trained using a dedicated loss function: ,in, This represents the total number of pixels in the cover plate ROI. As an unstable region labeling and pseudo-labeling matrix, this loss-constrained network accurately identifies interfering regions such as oil stains, reflections, cables, and shadows.

[0036] The dual-evidence parallel computation steps, deformable topology alignment and paired change evidence extraction, firstly, read the dual-light fusion features of the standard template corresponding to the vehicle model and the cover plate segment from the memory. Extract local topological subgraphs of all structural nodes in the template and the image to be inspected, calculate the cosine similarity matrix of the subgraphs, and construct a two-dimensional deformation offset vector field. Offset field prediction formula: ,in, This is an offset prediction convolutional network that takes fused features, template features, and topological prior codes as input and outputs pixel-level horizontal and vertical offsets.

[0037] According to the offset field Perform bilinear resampling alignment on the template features to obtain aligned template features: ,in, This represents bilinear interpolation resampling, which completes local pixel-level alignment between the template and the image to be inspected, eliminating spatial misalignment caused by shooting shake, perspective, and assembly deviation.

[0038] Topological edge The geometrical variation is obtained by statistically analyzing the axial length indentation and projection tilt offset between the image to be inspected and the template. ; Calculate the fusion feature residuals between the aligned template and the feature to be detected The initial macroscopic change is obtained by weighted fusion of the two: ,in, The structural node weights are determined by the fact that the cover plate solid area has a high weight and the occluded area has a low weight.

[0039] Nonlinear scaling and normalization are performed using a gated bending activation function: ,in, A preset bending coefficient is used to amplify the gradient of moderately weak defects; the final paired change evidence matrix is ​​as follows: ,in, A higher value indicates that the cover plate has abnormal shape or size relative to the template.

[0040] Topological temperature difference sampling and single-frame structural evidence extraction, such as Figure 5 As shown, in the dual-light fusion feature Center positioning adhesive strip node The corresponding local infrared discrete temperature matrix, with Using the origin as the reference point, construct orthogonal sampling axes along the connected topological edges: a normal sampling axis (perpendicular to the edge of the adhesive strip) and a tangent sampling axis (parallel to the edge of the adhesive strip), to characterize the topological edges representing the direction of the adhesive strip at the nodes. The tangent direction at the point is used as the tangent sampling axis (parallel to the edge of the adhesive strip), and the direction perpendicular to the tangent and pointing to both the inner and outer sides of the cable tray cover is used as the normal sampling axis (perpendicular to the edge of the adhesive strip). Pixel temperature sequences are extracted at equal intervals along the two axes. , , The formula for calculating the gradient change rate per unit pixel, representing the pixel spacing between adjacent sampling points: The sampling points were calculated. After obtaining the one-dimensional temperature gradient, the one-dimensional temperature gradient value is mapped back to the two-dimensional pixel coordinates corresponding to each sampling point in the image RIO, thus constructing a thermal two-dimensional catastrophe response map: When the temperature difference along the normal is much greater than that along the tangent, it indicates damage to the internal insulation layer of the adhesive strip; evidence of single-frame structure after attenuation by an unstable mask: , A higher value indicates that the non-visible inner surface of the adhesive strip has detached, causing a hidden defect of heat leakage.

[0041] The state-gated decision step assigns a confidence weight. , and thermal mutation response diagram Pixel-by-pixel multiplication attenuation filters out false hot spots caused by oil stains and reflections, retaining only the true heat leakage signal for generation. ; Calculating local alignment confidence in feature alignment bias during cross-modal feature fusion ; Traverse all pixels in the adhesive strip node region and count The percentage of pixels exceeding a threshold is used to determine the percentage of oil-stained area. Preset oil contamination threshold Confidence threshold Basic integration weight Adaptive fusion weight calculation formula: Automatically reduce when there is a lot of oil or the alignment accuracy is low. Increase the weight of single-frame infrared evidence.

[0042] The final defect probability matrix is ​​generated by fusing two pieces of evidence: ,right Connectivity partitioning is performed to obtain candidate defect regions; if the average value of the candidate regions is... Below the threshold for no abnormal appearance (no visible damage, warping, or missing parts), and the area is average. When the value exceeds the heat leakage judgment threshold, the output will determine the hidden defect of the inner surface of the adhesive strip falling off; if the average value of the candidate area is higher... The value is higher than the threshold for no abnormal appearance (or there is a significant offset on both sides of the edge), and the average value of the area is higher than the threshold for no abnormal appearance (or there is a significant offset on both sides of the edge). When the temperature exceeds the heat leakage threshold, the output shows a clear defect judgment result indicating large-area detachment of the adhesive strip or warping of the cover plate. If the candidate area... If the mean is too high, a manual review prompt will be issued simultaneously.

[0043] Defect output: The processor outputs complete defect information, including the location of the cover plate project to which the defect belongs, the pixel coordinates of the defect area, the corresponding adhesive strip structure node, the average value of paired change evidence, the average value of single-frame thermal evidence, the proportion of oil contamination in the area, and the manual review mark. At the same time, it outputs a visual overlay layer (structural node topology diagram, unstable mask, defect heat map) and stores it in the memory for maintenance personnel to retrieve and review.

[0044] Example 2 This embodiment, based on Embodiment 1, replaces the unified two-dimensional offset vector field of the entire map with independent offset prediction of sub-topological subgraphs. Complete cover plate topology diagram Cut into independent sub-graphs according to the segmented cover plate boundaries and joint topological edges. Each sub-diagram contains a set of continuous holes, slots, adhesive strips, and fastener nodes.

[0045] For each independent topological subgraph Calculate the fusion features to be detected within the subgraph separately. Features of template subgraph The cosine similarity matrix is ​​used to generate a local offset vector field specific to the subgraph. Subgraph offset prediction formula: Different cover plate segments and different rubber strip areas use independent offset parameters to avoid mutual interference from the perspective differences at both ends of the long cover plate.

[0046] Offset field constraint rule optimization: at the adhesive strip node Hole and slot edge nodes , joint nodes Add offset abrupt change constraints to the corresponding pixel positions; occluded nodes. The upper limit of the offset amplitude is forcibly limited in the region to prevent excessive stretching of the occluded area and smoothing out the real defects.

[0047] Block offset field stitching and fusion: Separating the independent offset fields of each sub-image Perform a smooth interpolation transition at the seam topology edge to generate a complete global offset field. Subsequent template feature resampling formula Residual calculation, paired change evidence The extraction process is exactly the same as in Example 1, and the meanings of all formulas and variables remain unchanged.

[0048] This embodiment completely solves the technical pain point of mutual traction and interference of perspective distortion at both ends of narrow components by cutting the complete topology map into independent sub-maps according to the cover plate seam and assigning independent parameters to the local offset vector field of each sub-map. This enables cross-modal alignment accuracy to reach the pixel level. In addition, by adding offset abrupt change constraints to the strongly assembled constraint areas (adhesive strips, slots, seams) and limiting the upper limit of offset amplitude in the occluded areas, it prevents the network from treating "real component misalignment, warping, or crack residuals" as defect-free images and forcibly stretching and smoothing them during forced registration, thus ensuring evidence of pairing changes. Authenticity and keen insight.

[0049] Example 3 This embodiment is based on Embodiment 1, and optimizes the unstable region suppression network branch. The unstable network in Embodiment 1 only inputs dual-light fusion features. + Topological coding features This embodiment adds three auxiliary input channels: visible light saturation channel. Infrared average temperature channel Edge continuity feature channel The four types of features are spliced ​​together and then input into a U-shaped decoding network to improve the accuracy of distinguishing oil stains, glare, cables, and shadows.

[0050] Network training pseudo-annotation generation optimization: Glare pseudo-annotation: Pixels with visible light saturation higher than the saturation threshold and brightness higher than the brightness threshold are automatically marked as unstable samples; Shadow pseudo-annotation: Regions with visible light brightness lower than the dark light threshold and edge continuity lower than the texture threshold are marked; Active cable pseudo-annotation: Connected regions with long, continuous dark color and no periodic hole or groove texture are marked; Oil stain pseudo-annotation: Large areas with low saturation, smooth texture and no clear edges are marked; The loss function is improved by adding a topological node weighted constraint term, resulting in the following loss formula for unstable regions: ,in, For structural node weights, occluded nodes The region increases the loss weight, and the cover plate solid structure node Reduce the weight of unstable prediction loss to prevent the network from misjudging normal cover plate structures as interference areas.

[0051] The network output remains a pixel-level unstable probability matrix. Credibility weight Cross-modal fusion formula, dual evidence calculation, adaptive fusion weights, defect probability All calculation formulas are completely consistent with those in Example 1, with only the unstable network input and training constraints being optimized.

[0052] This embodiment introduces visible light saturation, infrared average temperature, and edge continuity features as three auxiliary inputs, enabling the network to accurately distinguish between "temperature rises caused by oil buildup and heat dissipation due to adhesive strip detachment," thus improving the accuracy of environmental interference discrimination. Furthermore, a topology node weighting constraint term is added to the loss function. The prediction loss weights of the cover plate's physical structural nodes (such as edges, holes, and rubber strips) are reduced, thus establishing a hard defense at the bottom layer of algorithm training. This prevents the network from misjudging normal or slightly defective core structures of the cover plate as "oil stains or shadow interference areas" and being forced to zero, thus balancing high noise filtering and zero missed detections.

[0053] Example 4 This embodiment, based on the visible light and infrared dual-modality of Embodiment 1, adds project location pairing and anchoring, ROI structured normalization, three-stage network training, morphological defect verification, multi-level interpretable output, and structure-weighted loss.

[0054] For project location matching and ROI illumination normalization, firstly, the vehicle number, left and right side markings, cover plate segment number, and trackside acquisition point markings carried in the visible light and infrared image file names are read to establish the mapping relationship between the image pairs to be inspected and the corresponding standard template images; if the image markings are missing or the template and target cover plate positions do not match, a manual review label is directly marked.

[0055] Secondly, based on the structural prior diagram Extract the long side of the cover plate, the edge of the rubber strip, and the arrangement direction of the holes and slots, and fit a local coordinate system specific to the cover plate; simultaneously perform scale, rotation, and perspective coarse normalization on the visible light image to be inspected and the standard template image; decompose the RGB visible light image into three channels: brightness, chromaticity, and edge, and use large-scale guided filtering to estimate the non-uniform illumination field to complete pixel-level local brightness normalization, thereby reducing the interference caused by uneven lighting under the vehicle.

[0056] in, The structure is weighted by the Dice cross-entropy segmentation loss; This represents the loss due to structural topology consistency. To smooth the registration loss; Losses due to unstable regions; Weights for structural consistency loss; To register and smooth the loss weights; Weights are lost for unstable regions; This is the defect real annotation matrix; For structural boundary probabilities; This is a candidate connected region for defects; This is the threshold for defect determination; This is the threshold for unstable verification.

[0057] The three-stage hierarchical network training strategy employed in this invention involves a phased, progressive training model for the constructed deep learning network. The first stage involves pre-training the basic feature branches, separately training the unstable region suppression network and the structural topology encoding head, using only... As parameters for loss function optimization, the registration offset head and dual evidence feature extraction branches are not updated; the training objective is to enable the network to recognize visual distinguishing features of cover plate physical structure, oil stains, reflections, moving cables, and shadows.

[0058] Phase 2: Registration and cross-modal fusion branch training. The topology encoding and unstable network weights completed in Phase 1 are frozen. Only the deformable topology shift prediction head, visible light, and infrared cross-modal fusion modules are trained, and the registration smoothing loss is optimized. Eliminate artifacts caused by image capture shake and perspective deviation; Registration smoothing loss calculation formula: ,in, For pixels The structural boundary probability of the location is not forced to smooth the offset field in high-probability areas such as cover plate joints, hole and groove edges, and rubber strip boundaries, so as to avoid the network from smoothing out real cracks and cover plate misalignment defects.

[0059] Phase 3: Joint optimization training of the entire network, unlocking all network weights, simultaneously optimizing the four types of loss terms, and the comprehensive formula for the total loss: ,in, These are the weighting coefficients for structural consistency, registration smoothing, and unstable region losses, respectively; structural node weights are incorporated into the calculation of all loss terms. To enhance the learning constraint strength of defects in the solid area of ​​the cover plate. The segmentation loss is a combination of structure-weighted Dice cross-entropy, based on the defect annotation matrix. To improve the accuracy of defective pixel prediction, specifically in the third stage, the paired change evidence obtained from parallel computation of dual evidence will be used. Evidence of Single-Frame Structure Adaptive state-gated fusion is performed to output the final defect probability map. Defects are accurately labeled using manual work or pre-annotated matrix. To monitor the target, calculate and The sum of Dice and cross-entropy loss between them is used as This enables end-to-end joint gradient backpropagation optimization of all network parameters; To mitigate the structural topology consistency loss, the defect prediction probabilities of adjacent cover plate edge nodes in the same row of holes and slots are kept consistent, thus filtering out isolated noise and false defects.

[0060] This embodiment fundamentally solves the "gradient friction and oscillation" problem that often occurs in the joint training of multi-task complex networks by first pre-training to identify basic visual distinguishing features, second focusing on registration and using boundary probability to protect real cracks, and third jointly optimizing the entire network. This ensures the efficient convergence and stable output of the multi-scale dual-light feature network.

[0061] Dual evidence fusion output defect probability map Subsequently, candidate connected regions with defects are obtained through connected component segmentation. Layered morphological verification was performed based on the node types in the structural prior diagram, along the nodes of the adhesive strip. Hole and slot edge nodes Extending the generated elongated connected domain increases the defect confidence weight in that region; cross-joint topological edges Furthermore, connected regions with obvious offsets on both sides are identified as high-risk defects such as cover plate warping and large-area detachment of adhesive strips; connected regions that are not attached to any cover plate solid structure node and have loose textures without a fixed shape are identified as oil stains and shadow interference, automatically raising the level of manual review. The defect judgment of connected areas follows the disclosure standard criteria: Defect judgment conditions: Manual review trigger conditions: .

[0062] In this embodiment, after the dual evidence fusion output, the system does not simply perform black-box classification. Instead, it combines the structural prior graph nodes to perform hierarchical morphological verification. By verifying whether the connected domain extends along the adhesive strip node and whether it crosses the topological edge of the seam, the system directly filters out random false alarms caused by irregular external environmental patches, making the detection results have extremely high physical credibility.

[0063] In addition to the basic output fields of Example 1, the system also outputs vehicle number, acquisition point location, image acquisition time, standard template matching status, overall image quality score, percentage of oil stain pixels on a single section of the cover plate, and average registration confidence level for the entire section. The system includes: overall heat leakage risk classification of the cover plate; defect attribution to topology node type, mean of paired change evidence area, mean of infrared single-frame thermal evidence, proportion of unstable pixels in the region, defect pixel area, defect severity classification, and maintenance manual review suggestions; cover plate ROI mask layer, structural node thermal layer, local offset vector direction layer, instability suppression mask layer, defect segmentation contour layer, and infrared temperature difference gradient thermal layer; all layers and text results are uniformly stored in local storage, supporting online retrieval and offline review by maintenance personnel.

[0064] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting defects in cable tray covers for urban rail trains, characterized in that: Includes the following steps: The structural topology modeling steps involve identifying each structural feature in the visible light image of the cable trough cover area to be inspected and marking them as structural nodes in the topology graph. Based on the assembly relationship of each structural feature, topology edges are established to generate a structural prior graph. The environmental artifact suppression step involves fusing cross-modal features of the visible light image and infrared image to be inspected to obtain multi-scale dual-light fusion features. Based on the surface texture and edge continuity features in the visible light image to be inspected, an unstable region suppression mask is predicted by an unstable region suppression network. The dual-evidence parallel computation step compares the multi-scale dual-light fusion features with the pre-stored standard templates for residual comparison, and extracts paired change evidence representing the macroscopic morphological displacement of the component's appearance by using the geometric scale anomalies of each topological edge in the structural prior graph. The local infrared temperature field of the structural node region corresponding to the adhesive strip is extracted from the multi-scale dual-light fusion features. Based on the temperature gradient change rate in the direction of the topological edge connected to the structural node, single-frame structural evidence characterizing the destruction of the internal thermal conduction barrier of the non-visible surface is obtained. In the state gating decision step, when the paired change evidence determines that the appearance is normal, and the single-frame structure evidence after the unstable region suppression mask determines that the structure node corresponding to the adhesive strip has a local heat leakage anomaly, the latent defect decision result of the adhesive strip falling off the inner surface is obtained. The dual-evidence parallel computing step includes a topological temperature difference sampling strategy. The topological temperature difference sampling strategy includes extracting the discrete temperature matrix of the local infrared temperature field in the spatial coordinate system, taking the structural node corresponding to the adhesive strip as the origin, establishing a local sampling axis along the orthogonal tangent direction and normal direction of the topological edge connected to it, and performing temperature sampling of equally spaced pixels along the local sampling axis in the discrete temperature matrix. Then, the temperature difference between adjacent sampling pixels in the orthogonal tangent direction and normal direction is calculated respectively, and the change of the temperature difference within a unit pixel distance is used as the corresponding local temperature gradient change rate. The dual-evidence parallel computing step also includes a topology anomaly activation strategy. The topology anomaly activation strategy includes calculating the appearance reconstruction residuals of the multi-scale dual-light fusion features and the standard template in the local areas involved by each topology edge, and statistically calculating the axial length indentation and projection tilt angle offset between the cable tray cover area to be inspected and the standard template along each topology edge as the corresponding geometric scale anomaly amount. The dual-evidence parallel computation step also includes a deformable topology alignment strategy. The deformable topology alignment strategy includes extracting local topological subgraph features of multi-scale dual-light fusion features and standard template features at each structural node and its connected topological edges, calculating the cosine similarity matrix of the local topological subgraph features at the feature level, constructing a two-dimensional deformable offset vector field based on the cosine similarity matrix, and performing pixel-level deformable spatial mesh alignment of the multi-scale dual-light fusion features based on local interpolation according to the two-dimensional deformable offset vector field.

2. The method for detecting defects in cable tray covers for urban rail trains according to claim 1, characterized in that: The state gating decision step includes a hot mutation differentiation strategy. The thermal mutation differentiation strategy includes using the pixel-level instability probability matrix output by the unstable region suppression network as an unstable region suppression mask, wherein the position value of each pixel in the matrix represents the probability that the corresponding region is surface oil stains or light and shadow interference. The single-frame structural evidence is transformed into a thermal mutation anomaly response map representing a two-dimensional spatial distribution, and the values ​​of each pixel in the unstable region suppression mask are reversed to obtain the corresponding effective environmental confidence weight. The effective environmental confidence weight and the thermal mutation anomaly response map are then multiplied and attenuated pixel by pixel to obtain the local thermal leakage anomaly response as evidence of the suppressed single-frame structure.

3. The method for detecting defects in cable tray covers for urban rail trains according to claim 2, characterized in that: The state gating decision step also includes an adaptive adjustment strategy. The adaptive adjustment strategy includes determining the local alignment confidence based on the alignment deviation during the cross-modal feature fusion process, determining the proportion of oil stain area within the adhesive strip area based on the unstable region suppression mask, and then dynamically adjusting the dual evidence fusion weights based on the local alignment confidence and the proportion of oil stain area. When the proportion of the oil stain area is higher than the preset proportion threshold or the confidence of the local alignment is lower than the preset confidence threshold, the weight of the dual evidence fusion is reduced and the weight of the single-frame structure evidence in the judgment is increased.

4. The method for detecting defects in cable tray covers for urban rail trains according to claim 3, characterized in that: The topological anomaly activation strategy further includes weighted fusion of appearance reconstruction residuals and geometric scale anomalies to obtain local macroscopic changes with initial numerical distributions. Then, the initial numerical distributions are subjected to nonlinear bending scaling calculations based on preset bending coefficients, and the scaled data is normalized to output a feature probability matrix in the zero-to-one interval as evidence of paired changes.

5. The method for detecting defects in cable trough covers for urban rail trains according to claim 1, characterized in that: The environmental artifact suppression step includes a fusion strategy. The fusion strategy includes extracting multi-layer high-frequency spatial edge features from the visible light image to be inspected and extracting multi-layer deep semantic features from the infrared image. A guide map transformation weight matrix is ​​constructed based on the multi-layer high-frequency spatial edge features. Local affine transformation calculations are performed on each pixel position in the multi-layer deep semantic features. The multi-layer deep semantic features after the affine transformation calculation are then concatenated with the corresponding multi-layer high-frequency spatial edge features in the channel dimension to obtain multi-scale dual-light fusion features.

6. The method for detecting defects in cable trough covers for urban rail trains according to claim 1, characterized in that: The structural topology modeling step includes a structural topology encoding strategy. The structural topology coding strategy includes extracting edge geometric lines and key deformation corners from the visible light image to be inspected, determining the coordinates, scale, and orientation of each structural feature as the node attribute vector of the corresponding structural node, extracting the spatial geometry and assembly relationship between any two structural nodes as the edge type feature vector of the corresponding topological edge, interactively fusing the node attribute vector and the edge type feature vector, and then generating a structural prior map by projecting the Gaussian distribution into the image feature grid.

7. A defect detection system for cable trough covers of urban rail trains, characterized in that: include: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the steps of a method for detecting defects in cable trough covers for urban rail trains as described in any one of claims 1-6.

Images