A method and system for three-dimensional reconstruction of track defects based on magneto-optical sensing imaging

Abstract

The present application relates to the technical field of magneto-optical imaging processing, and particularly relates to a track defect three-dimensional reconstruction method and system based on magneto-optical sensing imaging, wherein the method comprises: performing multi-direction magneto-optical imaging on a steel track, and performing registration processing on the images; after registration, decomposition is performed to obtain a low-frequency subband and a high-frequency subband, gradient edge enhancement is performed on the high-frequency subband, the enhanced high-frequency subband is fused with the low-frequency subband, and a high-quality crack fusion image is generated through reconstruction; multi-scale edge detection is performed, and a crack two-dimensional contour is extracted in combination with morphological processing; the crack is positioned and classified to determine crack positioning information; steel track vibration information when a train passes is obtained; crack texture structure features and time sequence features are extracted, fusion is performed through a cross-modal attention fusion mechanism, and crack quantitative results are generated. Through the present application, full-quantitative evaluation of crack position, length, depth and damage degree can be realized, and the problem of inaccurate crack three-dimensional quantization is solved.

Description

Technical Field

[0001] This invention relates to the field of magneto-optical imaging processing technology, and in particular to a method and system for three-dimensional reconstruction of orbital defects based on magneto-optical sensing imaging. Background Technology

[0002] Current rail crack detection technologies mainly rely on magnetic particle testing, eddy current testing, or traditional image processing methods, but these generally suffer from low detection accuracy, difficulty in capturing microcracks and subsurface cracks, poor crack profile continuity, and insufficient quantitative assessment. Furthermore, with increasing train speeds and more complex track structures, crack detection tasks place higher demands on spatial resolution, temporal response, and quantitative accuracy. Achieving high-precision detection, three-dimensional quantification, and trend prediction of rail cracks has become a key technical challenge for ensuring rail transit safety and has created an urgent need for the development of intelligent detection technologies. Summary of the Invention

[0003] To address one of the aforementioned problems in the prior art, this invention provides a method and system for three-dimensional reconstruction of orbital defects based on magneto-optical sensing imaging.

[0004] To achieve the above objectives, the present invention provides a method for three-dimensional reconstruction of orbital defects based on magneto-optical sensing imaging, comprising: S101, Perform multi-directional magneto-optical imaging on the rail to obtain multi-directional crack observation images, and perform registration processing on the multi-directional crack observation images; S102, the registered multi-directional crack observation image is decomposed to obtain low-frequency sub-band and high-frequency sub-band, gradient edge enhancement is performed on the high-frequency sub-band, and the enhanced high-frequency sub-band is fused and reconstructed with the low-frequency sub-band to generate a high-quality crack fusion image. S103, perform multi-scale edge detection based on the high-quality crack fusion image, and extract the two-dimensional contour of the crack by combining morphological processing; S104, Based on the high-quality crack fusion image, the crack is located and classified to determine the crack location information; S105, acquire rail vibration information when a train passes by, the rail vibration information including at least: acceleration, velocity and vibration mode information; S106, extract crack texture structure features based on the crack two-dimensional contour and crack location information, extract temporal features based on the rail vibration information, and fuse the crack texture structure features and temporal features through a cross-modal attention fusion mechanism to generate crack quantification results.

[0005] Furthermore, the registration process for the multi-directional crack observation images specifically includes: aligning each image of the multi-directional crack observation images in a unified coordinate system; and using brightness normalization and local contrast enhancement to ensure that images in different directions have the same grayscale reference.

[0006] Furthermore, step S102 specifically includes: S1021, each of the registered multi-directional crack observation images is decomposed into multiple scales and multiple directions by non-subsampled shear wave transform to obtain the low-frequency sub-band and the high-frequency sub-band; S1022, Apply the Sobel operator to perform gradient edge enhancement on the high-frequency sub-band to obtain the enhanced high-frequency sub-band; S1023, the enhanced high-frequency sub-band and the low-frequency sub-band are input into the pulse-coupled neural network, and the fused high-frequency features are generated through pulse firing and neighborhood coupling mechanism. The low-frequency sub-band is fused to generate the fused low-frequency features through weighted averaging or maximum value selection strategy. S1024, the high-frequency features and low-frequency features after fusion are reconstructed by non-subsampled shear wave inverse transform to obtain the high-quality crack fusion image.

[0007] Furthermore, step S103 specifically includes: S1031, Based on the high-quality crack fusion image, the weak edge signal of the crack is spatially enhanced by the Gaussian Laplacian operator to obtain multi-scale edge detection results; S1032, based on the morphology, the edge detection results are sequentially subjected to closing operation, opening operation and connected component analysis to extract the two-dimensional contour of the crack.

[0008] Furthermore, the localization and classification of cracks based on the high-quality crack fusion image is specifically implemented using an improved YOLOv12 network. The improved YOLOv12 network introduces a multi-scale feature fusion module in the backbone network, a context information aggregation layer in the intermediate layer, and an adaptive threshold mechanism in the output layer.

[0009] Furthermore, step S106 specifically includes: S1061, Extract the crack texture structure features based on the crack two-dimensional contour and the crack location information using a two-dimensional convolutional network; S1062, the temporal features are extracted based on the rail vibration information using a one-dimensional convolutional network and a temporal convolutional network; S1063, The crack texture structure features and the temporal features are complementaryly integrated through a cross-modal attention fusion layer to obtain integrated features; S1064, The integrated features are input into a residual convolutional network and combined with a graph attention layer to extract deep features; S1065, the crack quantization result is output through a fully connected regression layer. The crack quantization result includes at least the quantized values ​​of crack location, crack length, crack depth, and crack damage degree.

[0010] Furthermore, the method further includes: step S107, inverting the crack quantization result to calculate a high-precision crack quantization result; step S107 specifically includes: S1071, The crack is physically modeled using a magnetic charge model, and the crack is discretized into multiple magnetic charge units in three-dimensional space; S1072, calculate the contribution of each magnetic charge unit to the magneto-optical signal plane based on the two-dimensional profile of the crack and the crack quantization result, and superimpose them to form a magneto-optical simulation image; S1073, Establish a multi-objective optimization function to optimize the crack geometric parameters contained in the crack quantization result, and obtain the high-precision crack quantization result, wherein one of the optimization objectives of the multi-objective optimization function includes: the fitting accuracy between the magneto-optical simulation image and the multi-directional crack observation image.

[0011] In addition, the optimization objectives of the multi-objective optimization function also include: the consistency between the crack depth and the rail vibration response; and / or the continuity of the two-dimensional profile of the crack.

[0012] In addition, the method also includes: S108, introduce a crack dynamic propagation model, establish material mechanics constraints and rail vibration response constraints, simulate the crack under different train operating conditions, predict the crack development trend, and evaluate the rail life based on the prediction results.

[0013] Another aspect of the present invention provides a three-dimensional reconstruction system for orbital defects based on magneto-optical sensing imaging, comprising: An imaging and registration module is used to perform multi-directional magneto-optical imaging on the rail, acquire multi-directional crack observation images, and perform registration processing on the multi-directional crack observation images. A high-quality crack fusion image generation module is used to decompose the registered multi-directional crack observation image to obtain low-frequency sub-bands and high-frequency sub-bands, perform gradient edge enhancement on the high-frequency sub-bands, and fuse and reconstruct the enhanced high-frequency sub-bands with the low-frequency sub-bands to generate a high-quality crack fusion image. The crack two-dimensional contour detection module is used to perform multi-scale edge detection based on the high-quality crack fusion image and extract the crack two-dimensional contour by combining morphological processing. The crack location information detection module is used to locate and classify cracks based on the high-quality crack fusion image and determine crack location information. A rail vibration information acquisition module is used to acquire rail vibration information when a train passes by. The rail vibration information includes at least: acceleration, velocity, and vibration mode information. The crack quantization result generation module is used to extract crack texture structure features based on the crack two-dimensional contour and crack location information, extract temporal features based on the rail vibration information, and fuse the crack texture structure features and temporal features through a cross-modal attention fusion mechanism to generate crack quantization results.

[0014] The beneficial effects of this invention are reflected in the fact that the three-dimensional reconstruction method and system for track defects based on magneto-optical sensing imaging provided by this invention aim to solve the problems of low accuracy, poor continuity, and difficulty in capturing microcracks and subsurface cracks in existing methods in crack contour extraction, microcrack identification, and full quantitative assessment. This invention acquires crack observation images through multi-directional magneto-optical imaging and performs multi-feature fusion on the images to generate high-quality crack fused images, overcoming the shortcomings of traditional methods in capturing small crack features and ensuring continuity. Based on the fused image, this invention extracts and locates the crack contour and fuses magneto-optical image features with track vibration signals to achieve a full quantitative assessment of crack location, length, depth, and damage degree, solving the problem of inaccurate three-dimensional crack quantification. Furthermore, this invention deeply couples magneto-optical signals, three-dimensional crack contours, and vibration responses through a defect magnetic charge model and a crack dynamic diffusion model, and employs a multi-objective adaptive inversion optimization algorithm to achieve high-precision inversion of crack geometric parameters and prediction of future propagation trends, solving the problems of unpredictable crack development trends and the disconnect between data and physical laws. Attached Figure Description

[0015] Figure 1 The flowchart shows the three-dimensional reconstruction method for track defects based on magneto-optical sensing imaging provided by this invention. Figure 2 This is a flowchart illustrating the specific method of step S102 provided by the present invention; Figure 3 Here is a flowchart of the specific method for step S106 provided by the present invention; Figure 4 Another flowchart of the three-dimensional reconstruction method for track defects based on magneto-optical sensing imaging provided by the present invention; Figure 5 This is a schematic diagram of the three-dimensional reconstruction system for track defects based on magneto-optical sensing imaging provided by the present invention. Detailed Implementation

[0016] 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.

[0017] Example 1 This embodiment provides a method for three-dimensional reconstruction of orbital defects based on magneto-optical sensing imaging, such as... Figure 1 As shown, it includes: Step S101 involves performing multi-directional magneto-optical imaging on the rail to acquire multi-directional crack observation images, and then registering these images. Specifically, in high-speed rail and subway track inspection, rail cracks, under long-term exposure to train wheels, manifest as surface and subsurface micro-cracks, which appear as linear or mesh-like magnetic disturbance signals in magneto-optical images. Because of the significant differences in crack directionality and scale, unidirectional magneto-optical images often cannot fully represent crack information; therefore, multi-directional magneto-optical images can supplement the signals from cracks in different directions.

[0018] Furthermore, due to differences in brightness, background texture interference, and noise in images from different directions, direct superposition and fusion can easily lead to dilution of crack information or the appearance of pseudo-cracks. Therefore, in an optional implementation, the multi-directional crack observation images are registered, specifically including: aligning each image of the multi-directional crack observation images in a unified coordinate system; and ensuring that images from different directions have the same grayscale reference through brightness normalization and local contrast enhancement. By performing geometric registration on the multi-directionally acquired magneto-optical images, the images are aligned in a unified coordinate system, and brightness normalization and local contrast enhancement ensure that images from different directions have a consistent grayscale basis during the fusion process.

[0019] Step S102: Decompose the registered multi-directional crack observation image to obtain low-frequency sub-band and high-frequency sub-band. Perform gradient edge enhancement on the high-frequency sub-band. Then fuse and reconstruct the enhanced high-frequency sub-band with the low-frequency sub-band to generate a high-quality crack fusion image.

[0020] In one alternative implementation, such as Figure 2 As shown, step S102 may specifically include: Step S1021 involves performing multi-scale and multi-directional decomposition on each registered multi-directional crack observation image using Non-Subsampled Shearlet Transform (NSST) to obtain low-frequency and high-frequency sub-bands. Each registered image is further decomposed using NSST at multiple scales and in multiple directions to obtain low-frequency and high-frequency sub-bands. The low-frequency sub-band mainly contains overall image brightness and large-scale texture information, while the high-frequency sub-band contains crack edge and detail information. The non-subsampling characteristic of NSST ensures that the image resolution is not reduced, while multi-directional decomposition can capture the magnetic disturbance characteristics of the crack in different directions, and multi-scale decomposition can take into account both micro-crack and obvious crack information, thus providing a basis for the complete preservation of crack features.

[0021] Step S1022: Apply the Sobel operator to perform gradient edge enhancement on the high-frequency subband to obtain the enhanced high-frequency subband. Apply the Sobel operator to the high-frequency subband to perform gradient edge enhancement to highlight the crack line structure and suppress the low-frequency background texture. At the same time, normalize the gradient map so that the crack edges of different directions and scales have a consistent response in the subsequent fusion.

[0022] Step S1023: The enhanced high-frequency and low-frequency subbands are input into a Pulse Coupled Neural Network (PCNN). A fused high-frequency feature is generated through pulse firing and neighborhood coupling mechanisms. The low-frequency subband is fused using a weighted average or maximum value selection strategy to generate a fused low-frequency feature. By introducing a PCNN, intelligent fusion of multi-directional crack information is achieved. The Sobel-enhanced high-frequency and low-frequency subbands are used as inputs, and the crack signal is adaptively enhanced by simulating the pulse firing and neighborhood coupling mechanisms of biological neurons. During the iteration process, PCNN can automatically enhance the crack signal while suppressing background interference and noise, thus obtaining a fused high-frequency feature with prominent cracks and good continuity. The low-frequency subband is fused using a weighted average or maximum value selection strategy to maintain the overall brightness and texture consistency of the image.

[0023] Step S1024: The fused high-frequency and low-frequency features are reconstructed using the non-subsampled shear wave inverse transform to obtain a high-quality crack fusion image. The fused high-frequency and low-frequency subbands are then reconstructed using the NSST inverse transform to obtain a final fused image with high contrast, clear crack edges, continuous structure, and weak background interference.

[0024] This invention is a multi-scale crack information fusion method based on NSST, Sobel edge enhancement and PCNN. It extracts multi-scale and multi-directional features of the image through NSST, enhances crack edges with the Sobel operator, and uses the adaptive coupling mechanism of PCNN to achieve intelligent fusion of multi-directional crack information, thereby obtaining a fused image with high contrast and strong continuity.

[0025] Step S103: Perform multi-scale edge detection based on the high-quality crack fusion image, and extract the two-dimensional contour of the crack by combining morphological processing.

[0026] In an optional implementation, step S103 specifically includes: Step S1031 involves spatially enhancing the weak edge signals of the cracks using the Laplacian of Gaussian (LOG) operator based on the high-quality crack fusion image, thereby obtaining multi-scale edge detection results. Specifically, cracks may exhibit multi-scale variations ranging from micro-cracks to obvious cracks, and a single fusion method struggles to capture both micro-cracks and obvious cracks. This step utilizes the high-quality crack fusion image generated in the preceding steps to perform multi-scale edge detection using the Laplacian of Gaussian operator, significantly enhancing the weak edge signals of the cracks spatially. This method addresses the issues of crack breakage and discontinuity that are common in traditional edge detection, and by employing multi-scale processing, it takes into account the characteristics of both micro-cracks and obvious cracks, achieving high precision and high continuity of crack edges.

[0027] Step S1032 involves performing closing, opening, and connected component analysis on the edge detection results based on morphology to extract the two-dimensional contour of the crack. Building upon step S1031, morphological methods can be used to perform closing, opening, and connected component analysis on the edge image to extract a continuous and complete two-dimensional contour of the crack.

[0028] Step S104: Based on the high-quality crack fusion image, the cracks are located and classified to determine the crack location information. To achieve automatic crack detection and full-quantitative evaluation, an improved YOLOv12 network can be used to locate and classify cracks for the crack target detection task. This network introduces a multi-scale feature fusion module in the backbone network to enhance the response capability to microcracks and subsurface cracks; adds a context information aggregation layer in the middle layer to strengthen the expression of crack edge and local structural information; and introduces an adaptive threshold mechanism in the output layer to achieve accurate location and classification of cracks of different sizes and intensities.

[0029] Step S105: Obtain rail vibration information when the train passes by. The rail vibration information includes at least: acceleration, velocity, and vibration mode information. When the train passes by, the crack morphology of the rail under load may be different from that under unloaded conditions. Therefore, the information such as the velocity and acceleration of rail vibration can reflect the dynamic cracking information of the rail and help to assess the actual condition of the rail.

[0030] Step S106 involves extracting crack texture structure features based on the crack's two-dimensional contour and location information, and extracting temporal features based on rail vibration information. The crack texture structure features and temporal features are then fused using a cross-modal attention fusion mechanism to generate crack quantification results. Specifically, a multimodal crack quantification network can be established to achieve preliminary quantitative assessment of crack location, length, depth, and damage degree. Furthermore, a physical constraint and intelligent reconstruction submodule can be added. By establishing a physical constraint model of crack geometry and dynamic response, the track vibration response information and magneto-optical image features are deeply fused to quantify the crack, achieving high-precision reconstruction of the crack's three-dimensional structure and damage degree.

[0031] In one alternative implementation, such as Figure 3 As shown, step S106 specifically includes: Step S1061: Extract crack texture structure features based on crack two-dimensional contour and crack location information using a two-dimensional convolutional network.

[0032] Step S1062: Extract temporal features based on rail vibration information using a one-dimensional convolutional network and a temporal convolutional network.

[0033] Step S1063: The cross-modal attention fusion layer is used to achieve complementary integration of crack texture structural features and temporal features to obtain integrated features.

[0034] Step S1064: Input the integrated features into the residual convolutional network and combine it with the graph attention layer to extract deep features.

[0035] Step S1065: Output crack quantification results through a fully connected regression layer. The crack quantification results include at least the following: crack location, crack length, crack depth, and crack damage degree quantification values.

[0036] Based on step S106, this embodiment can establish a multimodal crack quantization network. The network input consists of two types of information: one is the aforementioned high-quality crack fusion image features, and the other is the rail vibration signal features, which contain crack information when a train passes by. The rail vibration signal extracts temporal features through one-dimensional convolution and temporal convolution networks, and the magneto-optical image extracts spatial texture and crack structure features through two-dimensional convolution. Subsequently, a cross-modal attention fusion layer is used to achieve complementary integration of image and vibration features. Based on the fused features, a residual convolution network combined with a graph attention module is used to extract deep features, and a fully connected regression layer is used to calculate the crack length, depth, and damage degree, achieving end-to-end closed-loop processing from two-dimensional contour to preliminary quantization.

[0037] Specifically, this embodiment first uses the generated high-quality crack fusion image as spatial feature input, and combines it with a multimodal quantization network to extract the two-dimensional contour and location features of the crack. At the same time, the acceleration, velocity, and vibration modal information generated by the train passing over the rail are used as time series input. Through a pre-constructed physical constraint model, the amplitude, frequency, and modal features of the vibration response are mapped to the crack depth and damage degree, realizing the physical coupling of spatial and dynamic information.

[0038] In terms of network structure, the physical constraint model adopts a multi-branch fusion design: one branch processes the spatial features of cracks in high-quality crack fusion images, extracting crack texture and contour information through two-dimensional convolution and residual modules; the other branch processes the temporal features of vibration signals, extracting dynamic response patterns through one-dimensional convolution and temporal convolution networks. Subsequently, the two types of features are integrated through a cross-modal attention fusion layer, enabling image features and vibration response features to complement each other in space and time, thereby enhancing the representation capability of microcracks and subsurface cracks.

[0039] Furthermore, in an optional implementation, to ensure the output conforms to physical constraints, physical consistency regularization can be introduced after the fusion layer to constrain the crack depth prediction, ensuring that the model output is highly consistent with the actual physical response of the vibration signal, thereby avoiding non-physical interpretations that may occur in purely data-driven methods. Finally, the fused features are processed through a graph attention module and a fully connected regression layer to generate crack location heatmaps, length, depth, and damage degree quantification values, achieving an end-to-end closed loop from two-dimensional contours to three-dimensional reconstruction.

[0040] This embodiment provides a three-dimensional reconstruction method for track defects based on magneto-optical sensing imaging. It aims to address the problems of low accuracy, poor continuity, and difficulty in capturing microcracks and subsurface cracks in existing methods for crack contour extraction, microcrack identification, and comprehensive quantitative assessment. By acquiring crack observation images through multi-directional magneto-optical imaging and fusing multiple features into these images, a high-quality crack fused image is generated, overcoming the shortcomings of traditional methods in capturing minute crack features and ensuring continuity. This embodiment extracts and locates crack contours based on the fused image and integrates magneto-optical image features with track vibration signals to achieve a comprehensive quantitative assessment of crack location, length, depth, and damage degree, thus solving the problem of inaccurate three-dimensional crack quantification.

[0041] In one alternative implementation, such as Figure 4 As shown, the method in this embodiment further includes: step S107, inverting the crack quantization results to calculate high-precision crack quantization results. Step S107 aims to further improve the quantization accuracy of crack geometric parameters and is used for subsequent prediction and simulation verification of crack propagation trends. Step S107 may specifically include the following steps: Step S1071: Physical modeling of the crack is performed using a magnetic charge model, and the crack is discretized into multiple magnetic charge units in three-dimensional space. In specific implementation, the rail crack is first physically modeled using a defect magnetic charge model, and the crack is discretized into several magnetic charge units in three-dimensional space. Each unit generates a local magnetic flux anomaly under an external magnetization field.

[0042] Step S1072: Calculate the contribution of each magnetic charge unit to the magneto-optical signal plane based on the two-dimensional crack profile and crack quantization results, and superimpose them to form a magneto-optical simulation image; determine the geometric features such as the length, width, depth and location of the crack based on the two-dimensional crack profile and preliminary quantization results provided in the previous steps, and calculate the contribution of each magnetic charge unit to the magneto-optical signal plane, and superimpose them to form the overall magneto-optical response simulation result, i.e., the magneto-optical simulation image.

[0043] Step S1073: A multi-objective optimization function is established to optimize the crack geometric parameters contained in the crack quantization results, obtaining high-precision crack quantization results. One of the optimization objectives of the multi-objective optimization function is the fitting accuracy between the magneto-optical simulation image and the multi-directional crack observation image. The multi-objective adaptive inversion optimization algorithm improves the accuracy of crack geometric parameters and achieves high-precision crack size inversion by establishing a multi-objective optimization function and using the fitting accuracy between the simulated signal and the actual observation data as one of the optimization objectives. In the specific implementation process, the multi-objective adaptive inversion optimization algorithm first uses the crack quantization results obtained in the previous steps as the initial input to generate a candidate population of crack geometric parameters. By initializing the probability distribution of the crack profile and two-dimensional features, the initial search range of parameters such as crack length, width, and depth is determined, and the search space is dynamically reduced according to the preliminary quantization results to improve inversion efficiency and accuracy. On this basis, a multi-objective optimization framework is constructed for optimization, one of the indicators being the fitting accuracy between the simulated magneto-optical signal and the observed signal to ensure that the inversion results are highly consistent with the actual observations.

[0044] Furthermore, in an optional implementation, the optimization objectives of the multi-objective optimization function also include: consistency between crack depth and rail vibration response; and / or continuity of the crack's two-dimensional profile. The consistency objective between crack depth and rail vibration response ensures the physical rationality of crack quantification and dynamic characteristics; the continuity objective of crack geometry avoids the influence of local anomalous parameters on overall quantification. Through the design of multi-objective optimization, the algorithm can achieve a globally accurate search for crack parameters while satisfying physical constraints.

[0045] During the iterative update process, the multi-objective adaptive inversion optimization algorithm can employ an adaptive evolutionary strategy, combining genetic algorithms and simulated annealing. The genetic algorithm dynamically adjusts the crossover and mutation probabilities based on the fitness of the candidate population, improving the convergence speed of microcrack and subsurface crack parameters. The simulated annealing mechanism introduces random jump capabilities to prevent getting trapped in local optima and guides the search process through a physical constraint energy function, ensuring that the inversion results conform to both observational data and physical laws. Simultaneously, the multi-objective adaptive inversion optimization algorithm can also introduce a confidence-weighted mechanism, assigning higher search weights to high-confidence regions in the initial quantification results and increasing the exploration intensity in low-confidence regions, achieving a balance between local fine-tuning and global optimization.

[0046] To fully utilize multimodal information, the multi-objective adaptive inversion optimization algorithm in this embodiment can also introduce a cross-modal feature fusion strategy in each iteration, jointly evaluating the errors of the magneto-optical simulation image and the track vibration signal, and dynamically adjusting their importance through an adaptive weighting mechanism. This strategy can enhance the complementarity of spatial and temporal features, making the inversion of the depth and width of microcracks and subsurface cracks more accurate.

[0047] The multi-objective adaptive inversion optimization algorithm outputs the final inversion results after reaching the preset convergence conditions, including crack length, width, depth and local damage degree, and generates an error distribution map for visualization verification and simulation analysis.

[0048] In alternative implementations, such as Figure 4 As shown, the method in this embodiment further includes: Step S108 introduces a crack dynamic propagation model, establishes material mechanics constraints and rail vibration response constraints, simulates cracks under different train operating conditions, predicts crack development trends, and assesses rail lifespan based on the prediction results. Step S108 introduces a time dimension, extrapolating the current state to the future through the crack dynamic propagation model, extending from precise inversion quantification of the present to future trend prediction. In this step, a crack dynamic propagation model in the time dimension is established, and rail vibration response and material mechanics constraints are integrated into crack development prediction, combining physical constraints with dynamic prediction. By simulating crack propagation behavior under different train operating conditions, crack length, depth, and damage degree are linked to changes in magneto-optical image signals, enabling prediction of future crack development trends and lifespan assessment.

[0049] In the modular structure design of the algorithm model, a multi-branch feature processing and physical constraint fusion strategy can be adopted: one branch inputs the 3D reconstruction features (including crack quantization results) and 2D contour information output from the previous steps, while the other branch inputs the vibration signal features generated by the train passing over the rails and material mechanical constraints. These two types of features are deeply fused through a physical constraint coupling layer and an inversion optimization algorithm, ensuring that crack size and depth predictions conform to both observational data and actual physical laws. Ultimately, while outputting high-precision crack quantization results (including crack width, length, depth, and local damage degree), it also provides crack propagation trend prediction, offering decision support for track crack maintenance and risk assessment.

[0050] This embodiment also provides a three-dimensional reconstruction system for track defects based on magneto-optical sensing imaging, used to execute the aforementioned method. The details already described in the method section will not be repeated here; only the system structure will be explained. Figure 5 As shown, the orbital defect three-dimensional reconstruction system based on magneto-optical sensing imaging in this embodiment includes: The imaging and registration module 501 is used to perform multi-directional magneto-optical imaging of the rail, acquire multi-directional crack observation images, and perform registration processing on the multi-directional crack observation images. The high-quality crack fusion image generation module 502 is used to decompose the registered multi-directional crack observation image to obtain low-frequency sub-bands and high-frequency sub-bands, perform gradient edge enhancement on the high-frequency sub-bands, and fuse and reconstruct the enhanced high-frequency sub-bands with the low-frequency sub-bands to generate a high-quality crack fusion image. The crack two-dimensional contour detection module 503 is used to perform multi-scale edge detection based on high-quality crack fusion images and extract the crack two-dimensional contour by combining morphological processing. The crack location information detection module 504 is used to locate and classify cracks based on high-quality crack fusion images and determine crack location information. The rail vibration information acquisition module 505 is used to acquire rail vibration information when a train passes by. The rail vibration information includes at least: acceleration, velocity and vibration mode information. The crack quantization result generation module 506 is used to extract crack texture structure features based on crack two-dimensional contour and crack location information, extract temporal features based on rail vibration information, and fuse crack texture structure features and temporal features through a cross-modal attention fusion mechanism to generate crack quantization results.

[0051] This embodiment provides a three-dimensional reconstruction system for track defects based on magneto-optical sensing imaging. It aims to address the problems of low accuracy, poor continuity, and difficulty in capturing microcracks and subsurface cracks in existing methods for crack contour extraction, microcrack identification, and comprehensive quantitative assessment. By acquiring crack observation images through multi-directional magneto-optical imaging and fusing multiple features into these images, a high-quality crack fused image is generated, overcoming the shortcomings of traditional methods in capturing minute crack features and ensuring continuity. This embodiment extracts and locates crack contours based on the fused image and integrates magneto-optical image features with track vibration signals to achieve a comprehensive quantitative assessment of crack location, length, depth, and damage degree, thus solving the problem of inaccurate three-dimensional crack quantification.

[0052] In the description of the embodiments of the present invention, it should be understood that the terms "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "center," "top," "bottom," "top," "bottom," "inner," "outer," "inner side," and "outer side," etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the accompanying drawings and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. "Inner side" refers to the interior or enclosed area or space. "Outer perimeter" refers to the area surrounding a specific component or specific area.

[0053] In the description of embodiments of the present invention, the terms "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first," "second," "third," or "fourth" may explicitly or implicitly include one or more of that feature. In the description of the present invention, unless otherwise stated, "a plurality of" means two or more.

[0054] In the description of the embodiments of the present invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," "joining," and "assembly" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0055] In the description of embodiments of the present invention, specific features, structures, materials or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0056] In the description of the embodiments of the present invention, it should be understood that "-" and "~" represent a range between two numerical values, and this range includes the endpoints. For example, "AB" represents a range greater than or equal to A and less than or equal to B. "A~B" represents a range greater than or equal to A and less than or equal to B.

[0057] In the description of embodiments of the present invention, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0058] 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 method for three-dimensional reconstruction of track defects based on magneto-optical sensing imaging, characterized in that, include: S101, Perform multi-directional magneto-optical imaging on the rail to obtain multi-directional crack observation images, and perform registration processing on the multi-directional crack observation images; S102, the registered multi-directional crack observation image is decomposed to obtain low-frequency sub-band and high-frequency sub-band, gradient edge enhancement is performed on the high-frequency sub-band, and the enhanced high-frequency sub-band is fused and reconstructed with the low-frequency sub-band to generate a high-quality crack fusion image. S103, perform multi-scale edge detection based on the high-quality crack fusion image, and extract the two-dimensional contour of the crack by combining morphological processing; S104, Based on the high-quality crack fusion image, the crack is located and classified to determine the crack location information; S105, acquire rail vibration information when a train passes by, the rail vibration information including at least: acceleration, velocity and vibration mode information; S106, extract crack texture structure features based on the crack two-dimensional contour and crack location information, extract temporal features based on the rail vibration information, and fuse the crack texture structure features and temporal features through a cross-modal attention fusion mechanism to generate crack quantification results.

2. The method of three-dimensional reconstruction of rail defects according to claim 1, characterized in that, The registration process for the multi-directional crack observation images specifically includes: Align the images of the multi-directional crack observation in a unified coordinate system; By normalizing brightness and enhancing local contrast, images from different directions can have the same grayscale reference.

3. The method of claim 1, wherein, Step S102 specifically includes: S1021, each of the registered multi-directional crack observation images is decomposed into multiple scales and multiple directions by non-subsampled shear wave transform to obtain the low-frequency sub-band and the high-frequency sub-band; S1022, Apply the Sobel operator to perform gradient edge enhancement on the high-frequency sub-band to obtain the enhanced high-frequency sub-band; S1023, the enhanced high-frequency sub-band and the low-frequency sub-band are input into the pulse-coupled neural network, and the fused high-frequency features are generated through pulse firing and neighborhood coupling mechanism. The low-frequency sub-band is fused to generate the fused low-frequency features through weighted averaging or maximum value selection strategy. S1024, the high-frequency features and low-frequency features after fusion are reconstructed by non-subsampled shear wave inverse transform to obtain the high-quality crack fusion image.

4. The method of claim 1, wherein, Step S103 specifically includes: S1031, Based on the high-quality crack fusion image, the weak edge signal of the crack is spatially enhanced by the Gaussian Laplacian operator to obtain multi-scale edge detection results; S1032, based on the morphology, the edge detection results are sequentially subjected to closing operation, opening operation and connected component analysis to extract the two-dimensional contour of the crack.

5. The method of claim 1, wherein, The localization and classification of cracks based on the high-quality crack fusion image is specifically implemented using an improved YOLOv12 network. The improved YOLOv12 network introduces a multi-scale feature fusion module in the backbone network, a context information aggregation layer in the intermediate layer, and an adaptive threshold mechanism in the output layer.

6. The method of claim 1, wherein, Step S106 specifically includes: S1061, Extract the crack texture structure features based on the crack two-dimensional contour and the crack location information using a two-dimensional convolutional network; S1062, the temporal features are extracted based on the rail vibration information using a one-dimensional convolutional network and a temporal convolutional network; S1063, The crack texture structure features and the temporal features are complementaryly integrated through a cross-modal attention fusion layer to obtain integrated features; S1064, The integrated features are input into a residual convolutional network and combined with a graph attention layer to extract deep features; S1065, the crack quantification results are output through a fully connected regression layer. The crack quantification results include at least the quantified values ​​of crack location, crack length, crack depth, and crack damage degree.

7. The method for three-dimensional reconstruction of track defects according to claim 1, characterized in that, The method further includes: step S107, inverting the crack quantization result to calculate a high-precision crack quantization result; step S107 specifically includes: S1071, The crack is physically modeled using a magnetic charge model, and the crack is discretized into multiple magnetic charge units in three-dimensional space; S1072, calculate the contribution of each magnetic charge unit to the magneto-optical signal plane based on the two-dimensional profile of the crack and the crack quantization result, and superimpose them to form a magneto-optical simulation image; S1073, Establish a multi-objective optimization function to optimize the crack geometric parameters contained in the crack quantization result, and obtain the high-precision crack quantization result, wherein one of the optimization objectives of the multi-objective optimization function includes: the fitting accuracy between the magneto-optical simulation image and the multi-directional crack observation image.

8. The method for three-dimensional reconstruction of track defects according to claim 7, characterized in that, The optimization objectives of the multi-objective optimization function also include: The consistency between the crack depth and the rail vibration response; and / or the continuity of the two-dimensional profile of the crack.

9. The method for three-dimensional reconstruction of track defects according to claim 7, characterized in that, The method further includes: S108, introduce a crack dynamic propagation model, establish material mechanics constraints and rail vibration response constraints, simulate the crack under different train operating conditions, predict the crack development trend, and evaluate the rail life based on the prediction results.

10. A three-dimensional reconstruction system for orbital defects based on magneto-optical sensing imaging, characterized in that, include: An imaging and registration module is used to perform multi-directional magneto-optical imaging on the rail, acquire multi-directional crack observation images, and perform registration processing on the multi-directional crack observation images. A high-quality crack fusion image generation module is used to decompose the registered multi-directional crack observation image to obtain low-frequency sub-bands and high-frequency sub-bands, perform gradient edge enhancement on the high-frequency sub-bands, and fuse and reconstruct the enhanced high-frequency sub-bands with the low-frequency sub-bands to generate a high-quality crack fusion image. The crack two-dimensional contour detection module is used to perform multi-scale edge detection based on the high-quality crack fusion image and extract the crack two-dimensional contour by combining morphological processing. The crack location information detection module is used to locate and classify cracks based on the high-quality crack fusion image and determine crack location information. A rail vibration information acquisition module is used to acquire rail vibration information when a train passes by. The rail vibration information includes at least: acceleration, velocity, and vibration mode information. The crack quantization result generation module is used to extract crack texture structure features based on the crack two-dimensional contour and crack location information, extract temporal features based on the rail vibration information, and fuse the crack texture structure features and temporal features through a cross-modal attention fusion mechanism to generate crack quantization results.

Images