A railway track crack defect detection method and device based on twin simulation

Abstract

This invention discloses a method and apparatus for detecting cracks and defects in railway tracks based on twin simulation. The invention relates to the technical field of twin simulation, mapping multimodal data of railway tracks to a preset twin simulation space, discretizing the twin simulation object into a grid, and determining the corresponding crack and defect distribution information by combining material parameters and historical crack distribution characteristics. Further, texture rendering is performed using the stress distribution map of the railway track, improving the accuracy of the augmented dataset. Based on the identification of the crack and defect detection vector, multiple critical items of the corresponding crack and defect characteristics are determined; the crack and defect areas corresponding to the railway track are marked, and multi-factor fusion is performed based on the multiple critical items of the crack and defect characteristics. During the fusion process, multiple maintenance nodes are determined, and the graded maintenance content of the railway track is determined by combining the past maintenance history of the railway track, improving the accuracy of the graded maintenance content of the railway track.

Description

Technical Field

[0001] This invention relates to the technical field of twin simulation, and in particular to a method and device for detecting cracks and defects in railway tracks based on twin simulation. Background Technology

[0002] With the large-scale operation of high-speed railways and heavy-haul railway networks, railway tracks are highly susceptible to various cracks and defects under the long-term coupled effects of extreme alternating loads and complex environments. To ensure train operation safety, track inspection systems based on machine vision and sensor technology have been widely used.

[0003] Current data acquisition methods often rely on a single sensor source, such as using only a linear scan camera to acquire apparent images. Furthermore, apparent data and geometric shape and position data are often isolated and lack alignment and fusion mechanisms under a unified spatiotemporal coordinate system, resulting in a single data modality. At the same time, when constructing datasets for model training, existing technologies typically only perform simple affine transformations or brightness adjustments on the acquired real images to expand the samples, lacking a deep representation of the internal mechanical state and topological structure of the track. This affects the accuracy of the augmented dataset, resulting in low accuracy of the graded maintenance content for this railway track. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a method and device for detecting crack defects in railway tracks based on twin simulation.

[0005] This invention provides a method for detecting cracks and defects in railway tracks based on twin simulation, comprising: The apparent data of railway tracks are collected and combined with geometric shape and position data for spatiotemporal alignment to construct multimodal data of railway tracks. The multimodal data of railway tracks is then mapped to a preset twin simulation space to construct a twin simulation object with texture details and topological structure. The twin simulation is discretized into a mesh, and the corresponding crack defect distribution information is determined by combining material parameters and historical crack distribution characteristics. Furthermore, the stress distribution map of the railway track is combined for texture rendering to generate the corresponding enhanced dataset. The augmented dataset is iterated, and multiple crack data combinations are identified during the iteration process. Sliding window inference is performed on each crack data combination to generate a corresponding crack defect detection vector. Based on the identification of the crack defect detection vector, multiple critical items of the corresponding crack defect features are identified. The crack defect area corresponding to the railway track is marked, and multiple critical items with the characteristics of the crack defect are combined to perform multi-factor fusion. During the fusion process, multiple maintenance nodes are identified, and the graded maintenance content of the railway track is determined by combining the previous maintenance history of the railway track.

[0006] This invention provides a railway track crack defect detection device based on twin simulation, which is applied to the aforementioned railway track crack defect detection method based on twin simulation; the railway track crack defect detection device based on twin simulation includes: The twin simulation module is used to collect the appearance data of railway tracks and perform spatiotemporal alignment with geometric shape and position data to construct multimodal data of railway tracks. The multimodal data of railway tracks is then mapped to a preset twin simulation space to construct a twin simulation body with texture details and topological structure. The augmented dataset module is used to mesh and discretize the twin simulation object, and combine the material parameters and historical crack distribution characteristics to determine the corresponding crack defect distribution information. It further combines the stress distribution map of the railway track for texture rendering, thereby generating the corresponding augmented dataset. The defect identification module is used to iterate the augmented dataset and determine multiple crack data combinations during the iteration process. Sliding window inference is performed on each crack data combination to generate the corresponding crack defect detection vector. Based on the identification of the crack defect detection vector, multiple critical items of the corresponding crack defect features are determined. The graded maintenance module is used to mark the crack and defect area corresponding to the railway track, and to integrate multiple critical items with the characteristics of the crack and defect to determine multiple maintenance nodes during the integration process. The graded maintenance content of the railway track is determined by combining the past maintenance history of the railway track.

[0007] Compared with the prior art, the beneficial effects of the present invention are: (1) Collect the apparent data of railway track and perform spatiotemporal alignment with geometric shape and position data to construct multimodal data of railway track. Map the multimodal data of railway track to the preset twin simulation space to construct a twin simulation body with texture details and topological structure. Perform mesh discretization processing on the twin simulation body and determine the corresponding crack defect distribution information by combining material parameters and historical crack distribution characteristics. Further combine the stress distribution map of railway track for texture rendering to generate the corresponding enhanced dataset. The twin simulation body is introduced to further control the crack defect distribution information and improve the accuracy of the enhanced dataset.

[0008] (2) Iterate the enhanced dataset and determine multiple crack data combinations during the iteration process. Perform sliding window inference on each crack data combination to generate the corresponding crack defect detection vector. Based on the identification of the crack defect detection vector, determine multiple critical items of the corresponding crack defect features. Mark the crack defect area corresponding to the railway track and perform multi-factor fusion in combination with multiple critical items of the crack defect features. During the fusion process, determine multiple maintenance nodes and determine the graded maintenance content of the railway track in combination with the previous maintenance history of the railway track. Further control the crack defect detection vector, fully consider the crack defect area and multiple critical items of the crack defect features, and improve the accuracy of the graded maintenance content of the railway track. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the railway track crack defect detection method based on twin simulation in an embodiment of the present invention. Figure 2 This is a flowchart illustrating step S11 in the twin simulation-based railway track crack defect detection method in this embodiment of the invention. Figure 3 This is a flowchart illustrating step S12 in the twin simulation-based railway track crack defect detection method in this embodiment of the invention. Figure 4 This is a flowchart illustrating step S13 in the twin simulation-based railway track crack defect detection method in this embodiment of the invention. Figure 5 This is a flowchart illustrating step S14 in the twin simulation-based railway track crack defect detection method in this embodiment of the invention. Figure 6 This is a schematic diagram of the structural composition of a railway track crack defect detection device based on twin simulation in an embodiment of the present invention. Detailed Implementation

[0010] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0011] Please see Figures 1 to 6 A method for detecting cracks and defects in railway tracks based on twin simulation is presented, applied to a twin simulation scenario. The method includes: Step S11: Collect the appearance data of the railway track and perform spatiotemporal alignment with the geometric shape and position data to construct multimodal data of the railway track. Map the multimodal data of the railway track to the preset twin simulation space to construct a twin simulation body with texture details and topological structure. Step S12: The twin simulation body is meshed and discretized, and the corresponding crack defect distribution information is determined by combining material parameters and historical crack distribution characteristics. The texture is further rendered by combining the stress distribution map of the railway track to generate the corresponding enhanced dataset. Step S13: Iterate through the augmented dataset and determine multiple crack data combinations during the iteration process. Perform sliding window inference on each crack data combination to generate the corresponding crack defect detection vector. Based on the identification of the crack defect detection vector, determine multiple critical items of the corresponding crack defect features. Step S14: Mark the crack defect area corresponding to the railway track, and combine multiple critical items of the crack defect characteristics to perform multi-factor fusion. During the fusion process, determine multiple maintenance nodes, and combine the railway track’s previous maintenance history to determine the graded maintenance content of the railway track.

[0012] refer to Figure 2 In step S11, the specific steps are as follows: S111: Real-time acquisition of railway track appearance data using vehicle-mounted multi-source sensor array, simultaneous acquisition of railway track geometric shape and position data, and joint registration with railway track appearance data, thereby aligning and fusing appearance data and geometric shape and position data in a unified spatiotemporal coordinate system. During the fusion process, multimodal data of railway track is constructed, which includes texture features, geometric shape and thermal features. S112: Obtain a preset twin simulation space, map multimodal data to the preset twin simulation space, and reconstruct the texture details and internal topology of the railway track surface through nonlinear mapping within the twin simulation space, thereby generating a twin simulation of the railway track with consistent physical properties.

[0013] In the embodiments of this application, the apparent data of railway track is collected in real time using an on-board multi-source sensor array, and the geometric shape and position data of railway track are acquired simultaneously. The apparent data of railway track is then combined with the geometric shape and position data for joint registration, thereby aligning and fusing the apparent data and geometric shape and position data in a unified spatiotemporal coordinate system. During the fusion process, multimodal data of railway track is constructed. The multimodal data of railway track includes texture features, geometric shape and thermal features, thus introducing multimodal data.

[0014] At this time, the vehicle-mounted multi-source sensor array integrated into the high-speed inspection vehicle is configured. The array specifically includes a high-resolution linear industrial camera, a high-precision laser profile scanner, a non-contact infrared thermal imager, and an inertial measurement unit (IMU) and odometer encoder combination. As the inspection vehicle travels at high speed along the railway track, the linear industrial camera continuously scans the railway track surface at a preset line frequency to obtain surface image data containing information on rust, oil stains, and texture of the rail surface.

[0015] Simultaneously, a high-precision laser profile scanner is activated to emit high-frequency laser beams onto the track surface and receive reflected beams to generate three-dimensional point cloud data of the railway track cross-section. This three-dimensional point cloud data specifically characterizes the cross-sectional profile and surface undulation information of the track. At the same time, a non-contact infrared thermal imager is used to capture the thermal radiation intensity of the track surface and convert it into thermal map data reflecting the temperature field distribution of the track. During this process, the inertial measurement unit is used to collect the acceleration and angular velocity data of the vehicle in real time, and combined with the pulse signal output by the odometer encoder, to construct high-precision Global Navigation Satellite System / Inertial Navigation System (GNSS / INS) integrated navigation and positioning information. This provides a unified timestamp and spatial position reference for all sensor data, completing the real-time full acquisition of railway track appearance data, geometric shape and position data, and ambient temperature data.

[0016] The 3D point cloud data acquired by the laser contour scanner is preprocessed to extract key feature points of the railway track section, such as gauge points, rail head arc inflection points, and rail bottom edge points, to construct a geometric skeleton model of the track section. The coordinates of the 2D appearance image acquired by the linear industrial camera are mapped to the 3D laser point cloud coordinate system using the calibration extrinsic parameter matrix between the on-board multi-source sensor arrays.

[0017] Considering the complex vibrations and attitude changes generated by the vehicle during high-speed driving, the "attitude angle information: pitch angle, roll angle, and yaw angle" collected by the inertial measurement unit is introduced to perform geometric distortion correction and motion compensation on the appearance image to eliminate image projection errors caused by vehicle body bumps. The iterative nearest point algorithm is used to finely register the texture features of the corrected appearance image with the geometric features of the 3D point cloud. By continuously minimizing the Euclidean distance error between corresponding points, the pixel-level texture in the appearance data and the spatial coordinates in the geometric shape and position data are accurately matched, thereby strictly aligning the appearance data and the geometric shape and position data in a unified spatiotemporal coordinate system.

[0018] The RGB color channel values ​​of the jointly registered appearance image are mapped to the corresponding 3D point cloud vertices, so that each geometric point carries real surface texture color information, forming a textured 3D color point cloud. At the same time, the thermal image data collected by the infrared thermal imager is mapped to the grayscale value of the 3D point cloud according to the spatiotemporal coordinates, and the temperature value is converted into the thermal attribute field of the point cloud, so that each spatial coordinate point in the 3D model contains the corresponding temperature information.

[0019] Based on this, the track surface roughness and slope information obtained by laser point cloud computing are combined to structurally integrate the above three types of data sources, and abnormal data points caused by sudden changes in illumination or sensor noise are eliminated, thereby constructing multimodal data of railway tracks that includes texture features, geometric shape and thermal features. This multimodal data is stored in the form of structured data packets, and each frame of data contains a complete description of visual appearance, spatial structure and thermodynamic state.

[0020] Specifically, the inspection vehicle travels at 300 km / h on the high-speed railway trunk line. At this time, the onboard multi-source sensor array starts working. The surface data captured by the linear industrial camera not only includes the bright texture of the rail surface, but may also include the shadows of tiny fatigue cracks. At the same time, the laser profile scanner scans and finds a tiny surface depression on the inner side of the rail head caused by long-term wheel-rail rolling. The geometric shape and position data of this depression have changed slightly. Due to high-speed friction, this "tiny depression area, i.e., the potential source of cracks" shows a hot spot feature in the thermal data of the infrared thermal imager that is 2.5°C higher than the temperature of the surrounding rail.

[0021] The system uses IMU data to compensate for the roll angle caused by the centrifugal force generated when the train curves, ensuring that the crack shadow image captured by the camera and the depression geometry scanned by the laser scanner are completely coincident in space, avoiding texture misalignment caused by perspective shift. The system constructs a high-precision multimodal data, which includes "the precise three-dimensional coordinates of the depression: geometry", "the dark texture of the depression surface: texture features", and "the abnormal high temperature thermal value: thermal features". This means that in the multimodal data of the railway track, the detected "anomaly point" is no longer a single image or a depth value, but a comprehensive physical entity description "with a specific shape, specific dark texture, and accompanied by abnormal temperature rise".

[0022] Furthermore, a preset twin simulation space is obtained, multimodal data is mapped to the preset twin simulation space, and the texture details and internal topology of the railway track surface are reconstructed in the twin simulation space through nonlinear mapping, thereby generating a twin simulation of the railway track with consistent physical properties, which fully considers the twin simulation.

[0023] At this point, a pre-constructed twin simulation space is acquired. This simulation space is built based on a high-fidelity physics engine kernel and presets environmental parameters that simulate the real world, including standard atmospheric pressure, gravitational acceleration vector, ambient lighting model, and global material library. Then, a standard geometric template and physical material parameters matching the "railway track type, such as 60N rail" are loaded. These physical material parameters specifically include physical constants such as the density, Young's modulus, Poisson's ratio, and thermal conductivity of the rail material. A dedicated computational region is allocated in the simulation space's memory, a virtual object container with the same scale as the real track is instantiated, and the mesh boundary conditions of the rendering pipeline and physics solver are initialized. During this stage, the system establishes a standard data interface for receiving external data and sets the transformation matrix between the UV coordinate system origin of the texture mapping and the three-dimensional Cartesian coordinate system, thereby providing a virtual reference environment with physical computation capabilities for subsequent reception and processing of multimodal data.

[0024] The system acquires the multimodal railway track data generated in step S111 and injects it into a pre-defined twin simulation space in real time. During this process, nonlinear mapping is performed to resolve perspective distortion and nonlinear projection relationships between the two-dimensional image sensor and the three-dimensional physical space. At this point, a radial basis function neural network or multidimensional scaling analysis algorithm is used to construct a mapping function from local pixel features of the multimodal data to global three-dimensional coordinates in the simulation space. This mapping function can adaptively adjust the projection weights according to the depth information of the data, thereby nonlinearly mapping the appearance texture data and geometric shape data from a unified spatiotemporal coordinate system to the surface of the virtual object in the simulation space. During the mapping process, the system uses bilinear interpolation or trilinear interpolation algorithms to resample and densify the sparse geometric point cloud data, so that the discrete point cloud data can smoothly transition into a continuous simulation mesh surface, while retaining the high-frequency geometric features in the original data, ensuring that the virtual object maintains a high degree of consistency with the real railway track in terms of spatial morphology.

[0025] Optionally, at least 200 control points with distinct geometric features, uniformly distributed on the rail surface, are extracted from the multimodal data of railway tracks. These control points include the apex of the rail head arc, the turning point of the rail bottom edge, and marker points at the rail joint. The corresponding coordinate pairs of these control points in the image pixel coordinate system and the three-dimensional coordinate system of the laser point cloud are recorded. Virtual control points are defined in the twin simulation space using the same spatial distribution rules, and their three-dimensional coordinates are used as the center points of the radial basis functions. Quadratic functions are selected as the basis functions, and their expression is the square root of the sum of the square of the radius and the square of the smoothing parameter, where the smoothing parameter is set to 0.8 times the average distance between control points. The coordinate pairs of the control points extracted from the multimodal data are substituted into the radial basis function equation system. By solving the linear equation system, the weight coefficients of each basis function are obtained, thereby constructing a continuous and reversible nonlinear mapping function from the two-dimensional image pixel coordinates to the three-dimensional simulation space coordinates.

[0026] During the mapping process, for any pixel position of a non-control point, its unique three-dimensional coordinates in the simulation space can be calculated by substituting the distance of that position relative to the center points of all radial basis functions into the basis functions with determined weights and summing them. Based on this, a bicubic interpolation algorithm is used to resample and densify the sparse three-dimensional points after mapping. New points are inserted between adjacent points at equal intervals, so that the final simulation surface point cloud density reaches no less than 25 points per square millimeter. This preserves the high-frequency geometric features in the original data while forming a continuous and smooth simulation mesh surface, completing the nonlinear mapping reconstruction from multimodal data to twin simulation.

[0027] After completing the coordinate mapping, the reconstruction stage of surface texture details and internal topology is entered. Based on the mapped geometric surface, normal mapping and displacement mapping techniques are applied to solve the "texture features, such as rust spots and wear marks" contained in the multimodal data into the normal vector offset and elevation offset of the mesh vertices, thereby reconstructing the texture details of the railway track surface with micron-level unevenness in the simulation space.

[0028] Meanwhile, the geometric shape and position data are transformed into internal voxel meshes using voxelization algorithms. Based on the standard cross-sectional drawings of the rails and historical ultrasonic flaw detection data, the internal structure of the virtual object is reconstructed topologically. Specifically, this includes constructing lattice structure models of the rail head, rail web, and rail bottom, and simulating the distribution of potential internal voids or inclusions. In addition, thermal characteristic data is mapped to the temperature attribute channels of the voxel mesh, forming an integrated data model that is fully bound to the surface texture and internal geometry. This allows for the complete reproduction of the digital structure of the railway track, from its surface micromorphology to its internal macrostructure, within the simulation space.

[0029] The reconstructed digital structure undergoes physical property consistency calibration to generate a railway track twin simulation with consistent physical properties. At this point, the system corrects the thermodynamic boundary conditions of the simulation based on the thermal characteristics contained in the multimodal data, assigns the collected temperature field distribution as the initial thermodynamic state of the simulation, and uses finite element analysis to pre-stress the simulation, ensuring it possesses thermal expansion and contraction characteristics and internal stress distribution consistent with the real track under current environmental temperature and load conditions. Through the real-time collision detection and constraint solving mechanism of the physics engine, the system verifies whether the deformation response of the simulation under virtual external forces conforms to the physical laws of the preset material parameters. If deviations exist, the mesh stiffness is automatically iteratively corrected. After completing the above calibration, the system locks the virtual object containing high-fidelity texture details, a complete internal topology, and accurate physical properties as the final railway track twin simulation. This twin simulation can respond to physical simulation commands in real time, providing an accurate experimental platform for subsequent crack and defect simulation.

[0030] Specifically, the pre-set twin simulation space is loaded with the material parameters of U71Mn steel rails commonly used in high-speed railways, and a physical field simulating the real driving environment is established; the system receives multimodal data about the small concave area on the inner side of the rail head input in stage S11; due to the perspective compression of the images collected by the high-speed movement of the inspection vehicle, the system uses a radial basis function neural network to reverse-map the image features at this location back to three-dimensional space, accurately restoring the real curved shape of the concave area in three-dimensional space, rather than simply pasting a planar image.

[0031] The system converts the "dark texture" identified in S11 into a displacement map of the simulation surface, generating tiny pits visible to the naked eye on the virtual rail head. At the same time, based on historical detection data, the system simulates and generates a slender crack topology structure connected to the pit area and pointing towards the interior of the rail directly below it, and assigns the thermal characteristics of the area, which is 2.5°C higher than the surrounding area, to the voxel of the crack area, so that the crack appears as a "thermal-mechanical" coupled state in the simulation.

[0032] Based on the provided thermal data, the simulation engine calculates the thermal stress concentration caused by local high temperature in the crack area. This makes the virtual twin simulation not only look like a real track in terms of mechanical properties, but also shows a stress concentration and propagation trend that is highly consistent with the real physical environment when subjected to virtual train wheelset loads. This generates a twin simulation that can accurately reflect the physical state of the crack defect under the railway track.

[0033] refer to Figure 3 In step S12, the specific steps are as follows: S121: Mesh the twin simulation of the railway track and perform discretization processing using an adaptive mechanism to construct the corresponding finite element mesh body, which contains geometric information and material properties; S122: Synchronously collect the material parameters of the railway track, and perform multiple calculations in the finite element mesh body in combination with the corresponding historical crack distribution characteristics to further determine the corresponding crack defect distribution information, which covers the crack occurrence location and propagation path; S123: Obtain the stress distribution map of the railway track and combine it with the crack defect distribution information for superposition and fusion. In the process of fusion, the corresponding crack extension trajectory is determined, and the corresponding stress concentration degree is combined for differentiated texture rendering, thereby generating an enhanced dataset containing stress shading features and multi-scale crack morphology.

[0034] In the embodiments of this application, the twin simulation of the railway track is meshed and discretized using an adaptive mechanism to construct a corresponding finite element mesh body. This finite element mesh body contains geometric information and material properties, thus introducing the finite element mesh body.

[0035] At this point, the railway track twin simulation generated in step S112 is acquired, and its three-dimensional geometric boundary representation data and internal topological connection relationship are extracted. Based on this boundary representation data, the octree spatial indexing algorithm is used to recursively partition the three-dimensional bounding box where the twin simulation is located, thereby generating an initial regular hexahedral mesh unit covering the entire track volume. In this process, a basic mesh size threshold is set, and the "overall contour of the track, such as the large-scale geometric features of the rail head, rail web, and rail bottom" is divided into coarse-grained spatial parts to form an initial meshing framework. At the same time, the high-frequency geometric features carried by the surface displacement map and normal map of the twin simulation are extracted. By calculating the rate of curvature change between adjacent mesh nodes, the key geometric boundaries of curvature abrupt changes are identified, providing a spatial position reference for subsequent local mesh refinement, thereby completing the initial transformation from continuous topological surface to discrete mesh unit.

[0036] Based on the initial meshing framework, an adaptive refinement mechanism is introduced to process the non-uniform discretization of the mesh. An adaptive evaluation index function is then constructed, with local geometric curvature and estimated stress gradient as independent variables. For regions on the track surface with abrupt curvature changes, such as the transition arc between the rail base and rail waist, local surface depressions, and high stress concentration areas estimated based on historical stress models, the adaptive evaluation index function outputs a higher refinement weight coefficient. The system dynamically reduces the hexahedral mesh size threshold for the corresponding region based on this weight coefficient, performing local mesh refinement operations, i.e., multi-level sub-splitting of the coarse-grained parent mesh to generate fine-grained hexahedral or degenerate tetrahedral meshes. Conversely, for the main track region with gentle curvature and uniform stress distribution, such as the portion of the rail body far from the wheel-rail contact area, its coarse-grained mesh state is retained. Through this adaptive mechanism, the discretization error is controlled within a preset tolerance range without significantly increasing computational resource consumption, thus completing the non-uniform mesh discretization dependent on track structure characteristics.

[0037] After adaptive discretization, topology reconstruction and node fusion operations are performed on the generated mesh elements at all levels. The vertex coordinates of each mesh element are extracted as geometric information, and a mapping relationship between local node numbers and the global coordinate system is established. To address the step size phenomenon between adjacent meshes caused by adaptive refinement, transition mesh elements are created at the interface between coarse and fine meshes. Multi-point constraint technology is used to map the degrees of freedom of the fine mesh to the boundary nodes of the coarse mesh through shape function interpolation. By traversing all mesh interfaces, displacement inconsistency or stress artifacts caused by inconsistent mesh sizes are eliminated, ensuring that the discretized finite element mesh body maintains spatial continuity and smoothness in geometric shape, thereby constructing a finite element mesh body framework with high-fidelity geometric features.

[0038] The material parameters of the twin simulation in step S112 are mapped to the aforementioned mesh framework to endow it with physical and mechanical properties. At this time, according to the physical characteristics of different parts of the railway track, the "isotropic elastic material parameters, including Young's modulus, Poisson's ratio, density, and yield strength" corresponding to the track body are assigned to the basic mesh elements of the main body regions of the rail head, rail web, and rail bottom. At the same time, for the internal topology reconstructed in step S112, the corresponding heterogeneous material parameters or defect material parameters are mapped to the mesh elements occupied by the topology. Then, the "geometric information: node coordinates, element topology connection relationship" and "material properties: elastic matrix, mass matrix" of each mesh element are structurally bound, and finally a finite element mesh body containing accurate geometric information and complete material properties is generated, providing a discretized calculation model that conforms to the finite element analysis standard for subsequent mechanical solution and crack simulation.

[0039] Specifically, for the twin simulation generated in S112, which includes the small indentation on the inner side of the rail head and potential internal cracks, the system uses the octree algorithm to perform initial meshing of the entire 60N rail model. At this time, the rail head, rail web, and rail bottom are all divided into coarse-grained hexahedral meshes with a side length of about 5 mm.

[0040] The system calculations revealed that the surface curvature of the recessed area on the inner side of the rail head is much greater than that of the surrounding flat area. At the same time, based on the high-speed train wheel-rail contact mechanics model, it was estimated that there is an extremely high contact stress gradient in this area. Therefore, the adaptive mechanism was triggered to subdivide the grid in the recessed area and the surrounding 2 cm range twice, dynamically reducing the grid side length to about 1.25 mm, forming a high-density fine-grained grid group; while for the gently stressed rail bottom area, the 5 mm coarse-grained grid was kept unchanged.

[0041] The system generates a transition mesh at the boundary between the 1.25 mm fine mesh and the 5 mm coarse mesh, and employs a multi-point constraint algorithm to ensure that when wheel-rail loads are applied, the small deformations in the fine mesh region can be smoothly transferred to the coarse mesh region, avoiding false stress concentrations at mesh abrupt changes. The system batch-assigns the Young's modulus of U71Mn steel (210 GPa) and Poisson's ratio (0.3) to the main track mesh. Crucially, for the reconstructed elongated crack topology, the system weakens the stiffness matrix of the mesh elements it occupies, reducing the local stiffness to near zero, making it exhibit discontinuous microcrack surfaces in mechanical behavior. The resulting finite element mesh not only accurately encompasses the tiny surface depressions under the railway track in terms of geometry, but also perfectly reproduces the weakening characteristics of potential cracks in terms of internal structure and mechanical properties, providing an extremely accurate finite element calculation basis for the next step of applying loads and generating stress distribution maps.

[0042] Furthermore, the material parameters of the railway track are collected synchronously, and multiple calculations are performed in the finite element mesh body in combination with the corresponding historical crack distribution characteristics to further determine the corresponding crack defect distribution information. This crack defect distribution information covers the crack occurrence location and propagation path, and fully considers the crack defect distribution information.

[0043] At this time, real-time material parameters of the target railway track are collected simultaneously. These material parameters include not only the macroscopic elastic modulus, Poisson's ratio, and density obtained through non-destructive testing, but also microscopic mechanical parameters reflecting local material degradation, such as fatigue limit, fracture toughness, and crack propagation threshold. The collected material parameters are mapped to the finite element mesh generated in step S121 according to their spatial location. At this time, for the work-hardened areas on the track surface caused by long-term wheel-rail friction, or the local rusted areas caused by environmental erosion, a gradient assignment algorithm is used to reduce and correct the yield strength and fatigue limit of the corresponding mesh elements. This makes the finite element mesh not only have macroscopic geometric features, but also accurately carry the spatial heterogeneous mechanical properties reflecting the actual service state, providing a physical basis for subsequent crack evolution calculations.

[0044] A pre-defined database of historical defects in railway tracks is retrieved, and historical crack distribution characteristics matching the current track type and service conditions are extracted. These historical crack distribution characteristics include the statistically dominant orientation, dip angle, depth ratio, and spatial clustering density of historical cracks. Based on the aforementioned historical crack distribution characteristics, a random seeding algorithm is used to generate discrete initial micro-defect groups in the stress concentration sensitive areas on and inside the track surface within the finite element mesh. These micro-defect groups are characterized as initial crack surfaces with extremely weakened stiffness within the mesh elements. Simultaneously, a historical residual stress field is introduced as an initial boundary condition. The finite element solver is used to perform static equilibrium solutions on the finite element mesh containing the initial micro-defect groups to obtain the initial stress field distribution state of the track before the application of external dynamic loads, thereby completing the initial initialization of the crack evolution model.

[0045] After initialization, dynamic boundary conditions simulating real train loads are applied, initiating a multi-operation process. This multi-operation process involves alternating finite element stress solutions for external loads and internal material damage evolution calculations. In each load step iteration, the stress-strain field distribution of the entire track is solved using the finite element method, extracting the stress intensity factor or J integral value of each initial micro-defect tip and high stress gradient region. The calculated stress intensity factor is substituted into the constitutive equation based on Paris's law or continuous damage mechanics evolution to calculate the micro-defect propagation increment under the current load step. If this propagation increment causes the damage accumulation of adjacent mesh elements to exceed their fracture toughness threshold, mesh separation or further stiffness degradation is triggered, simulating the micro-propagation of cracks. Through the alternating multi-iterative calculations of the above mechanical solutions and damage evolution, the physical process of material from micro-defect initiation to macro-crack propagation under alternating loads is realistically reduced.

[0046] The calculation terminates when the multi-operation reaches the preset fatigue life cycle or the crack extends to the critical size, and the spatial topology of the calculation results is extracted. At this time, all element sets that have undergone stiffness degradation and mesh separation in the finite element mesh are traversed, and the coordinate sequence and connectivity of these elements in three-dimensional space are extracted. Based on the coordinate sequence and connectivity, a continuous spatial crack surface or curve is fitted to determine the corresponding crack defect distribution information. This crack defect distribution information not only covers the "crack occurrence location, i.e., the three-dimensional spatial coordinates of the initial occurrence point", but also accurately includes the "crack propagation path, i.e., the trajectory vector of the crack tip extending under multiple load steps and its depth extension shape", providing an accurate geometric input source for the subsequent high-fidelity crack rendering of the twin simulation.

[0047] Specifically, the system retrieves typical defects of this type of rail from the database—inner side core damage of the rail head, i.e. transverse fatigue cracks. Statistical analysis shows that these cracks mostly initiate 3-5 mm below the rail head fillet, and the propagation path tends to extend inward at an angle of 20 to 30 degrees with the tread surface. Based on this, the system pre-sets 10 micrometer-level initial defect parts in the dangerous area inside the rail head of the finite element mesh body according to the Poisson distribution law.

[0048] The system applied a wheel-rail contact load equivalent to that of a high-speed train operating at 300 km / h to the rail head, driving tens of millions of high-cycle fatigue calculations. When the iteration reached the 2 millionth iteration, the equivalent stress intensity factor of the crack tip of seed defect No. 7, located 4 mm below the fillet of the rail head, exceeded the fatigue crack propagation threshold value corresponding to the material for the first time, marking that the micro-defect entered the propagation stage from the stable state. Based on the Paris fatigue crack propagation law and combined with the propagation coefficient and exponent of U71Mn rail material, it was calculated that the defect produced a micro-propagation increment of 0.01 mm under this load step. As the number of iterations increased further, defect No. 7 continued to extend under the continuous cyclic load, and its propagation path gradually approached the adjacent seed defect No. 5. When the distance between the two defect tips decreased to a point where it was insufficient to maintain independent propagation, the defects merged, and the two originally separate micro-defects connected to form a continuous macro-crack surface, completing the complete evolution process from micro-defect initiation to macro-crack formation.

[0049] The system extracts all mesh elements that have undergone penetration and stiffness degradation to determine the distribution information of the crack defect: the crack is precisely located 4.2 mm below the gauge angle, and its expansion path shows an oblique development from the initiation point towards the tread surface at a 25-degree angle, and it branches at a depth of 15 mm. This crack distribution information, which contains precise location and three-dimensional expansion trajectory, provides extremely accurate bionic data support for the next step of realistic texture rendering and stress field visualization on the twin simulation body.

[0050] Therefore, by acquiring the stress distribution map of the railway track and combining it with the crack defect distribution information for superposition and fusion, the corresponding crack extension trajectory can be determined during the fusion process. Differentiated texture rendering is then performed based on the corresponding stress concentration level, thereby generating an enhanced dataset containing stress shading features and multi-scale crack morphology. A twin simulation is introduced to further control the crack defect distribution information and improve the accuracy of the enhanced dataset.

[0051] At this point, a railway track stress distribution map generated by solving the finite element mesh under simulated conditions is obtained. This stress distribution map represents the equivalent stress distribution state inside and on the surface of the track in the form of a cloud map and is stored as a discrete nodal stress tensor matrix. The three-dimensional spatial coordinates corresponding to each data point in the stress distribution map are extracted and mapped to the surface and internal topological space of the twin simulation object based on the global coordinate system transformation matrix constructed in step S121. During the mapping process, a bicubic interpolation algorithm is used to resample the sparse stress node data to achieve pixel-level alignment with the high-resolution geometric mesh of the twin simulation object, thereby constructing a continuous and high-fidelity three-dimensional stress field distribution model in the twin simulation space, providing an accurate physical field reference for subsequent crack-stress superposition.

[0052] The crack defect distribution information, including the crack occurrence location and propagation path, determined in step S122, is spatially superimposed and fused with the aforementioned three-dimensional stress field distribution model. At this point, using the crack occurrence location as an anchor point, the three-dimensional geometric curve of the crack propagation path is embedded into a high-gradient stress concentration region. During this fusion process, a trajectory correction mechanism driven by the stress intensity factor is introduced, namely, extracting the stress gradient direction vector at the leading edge of the crack tip, and weighting and fusing this direction vector with the tangent vector of the original crack propagation path to calculate the corrected crack dynamic extension trend. Based on this extension trend, a spline curve interpolation algorithm is used to extend and expand the original crack path, thereby determining the corresponding crack extension trajectory under the constraint of the current real stress field. This extension trajectory not only reflects historical evolution characteristics but also accurately predicts the dynamic propagation trend of the crack under the current extreme stress field.

[0053] After determining the crack extension trajectory, the physically-based differential texture rendering pipeline is activated. A local coordinate system is established along the crack extension trajectory, and differential parameters for texture rendering are calculated based on the stress concentration in the crack area. At this point, the stress concentration coefficient is divided into multiple gradient threshold intervals. For "high stress concentration areas, such as crack tips and origin areas," high-frequency normal perturbation and high-contrast displacement mapping techniques are used to render macroscopic crack textures with sharp edges, wide openings, and obvious plastic deformation tearing characteristics. High-gloss reflection properties are superimposed to simulate the metallic luster of a real rail fracture surface. For "medium-low stress concentration areas, such as crack extension tails or micro-crack areas," low-frequency normal smoothing transition and low-contrast color masking techniques are used to render micro-crack textures with extremely narrow widths, blunt edges, and rust and oxidation filling. Simultaneously, the stress values ​​in the stress distribution map are linearly mapped to a preset stress color spectrum, such as transitioning from cool-toned blue low-stress areas to warm-toned red high-stress areas. This is used as the base color and overlaid on the surface of the twin simulation to achieve a visually penetrating expression of the physical stress field.

[0054] Based on the rendered twin simulation, multi-view virtual sampling and data encapsulation operations are performed. By setting different lighting conditions, such as front lighting, backlighting, and shadow occlusion, and using a multi-view virtual camera array in the virtual space, two-dimensional projection slicing is performed on the track surface containing stress shading features and differentiated crack textures. During the acquisition process, the semantic segmentation mask, bounding box coordinates, and corresponding stress level labels of cracks in each frame are extracted simultaneously and formatted into a structured annotation file. By traversing all crack extension trajectories and stress gradient combinations, the system outputs a batch of image sample sets covering macroscopic large cracks, microscopic fine cracks, different stress background colors, and complex lighting interference, thereby generating an enhanced dataset containing stress shading features and multi-scale crack morphology. This dataset is directly used for the training and iteration of subsequent deep learning detection models.

[0055] Specifically, the system superimposes the oblique crack path generated by historical evolution with the current dark red high-stress zone; calculations show that the direction of the maximum principal stress in front of the crack tip has deviated by 5 degrees compared with the historical propagation path. Based on this, the system corrects and determines a new crack extension trajectory—that is, the crack will accelerate its extension in the direction biased towards the inside of the tread surface in the future. This dynamic prediction trajectory is precisely encoded into the twin simulation.

[0056] The system visually reconstructs this composite trajectory: at the "crack tip with the highest stress concentration coefficient: deep red area", the rendering engine generates a macroscopic open crack texture with a width of 0.5 mm and a jagged tearing shape at the edge, accompanied by metallic reflection; while at the "crack tail with lower stress: yellow or green area", a microscopic mesh-like fine crack texture with a width of only 0.05 mm and filled with black rust is rendered; at the same time, the entire track surface is covered with a semi-transparent stress shading base map from blue to red, so that the originally invisible internal stress field and surface crack shape are visually deeply integrated.

[0057] The virtual camera takes multi-angle photos of the railhead, which features a red and blue stress background and has both wide tear gaps and fine rust lines, under various complex lighting conditions commonly found on railway tracks, such as twilight backlighting and artificial night lighting. It automatically generates pixel-level mask labels, thus producing a large number of high-quality augmented datasets. The detection model trained on this dataset can not only identify tiny surface rust gaps in real inspections, but also, by generalizing its understanding of "stress coloring features," can keenly capture deep nuclear damage manifestations accompanied by specific deformation features, significantly reducing the false negative rate.

[0058] refer to Figure 4 In step S13, the specific steps are as follows: S131: The augmented dataset is used as input data and trained iteratively multiple times. During the iteration process, multiple crack data combinations are determined through an adaptive sampling strategy. These multiple crack data combinations include different lighting conditions, occlusion levels, and crack types. S132: A dynamic-scale sliding window is used to extract and infer features from each combination of crack data, thereby capturing the local texture gradient and global morphological context of the crack, and generating a high-dimensional crack defect detection vector, which includes crack length, width, direction and depth. S133: Dynamically identify the crack defect detection vector, and combine cluster analysis mechanism to determine the corresponding crack defect characteristics during the identification process. Further combine the defect response mapping relationship of railway track to screen out the key factors that lead to track failure risk, thereby determining multiple critical items of the corresponding crack defect characteristics. Multiple critical items include but are not limited to rail head peeling, rail web vertical cracks and rail bottom core damage.

[0059] In the embodiments of this application, the augmented dataset is used as input data and trained iteratively multiple times. During the iteration process, multiple crack data combinations are determined through an adaptive sampling strategy. These multiple crack data combinations include different lighting conditions, occlusion levels, and crack types, thus introducing multiple crack data combinations.

[0060] At this point, the augmented dataset containing stress coloring features and multi-scale crack morphology generated in step S123 is obtained and used as the input data stream for the deep neural network architecture. The image samples in the augmented dataset and their corresponding "structured annotation files, including semantic masks, bounding boxes, and stress level labels" are loaded into the video memory buffer and grouped according to the preset batch size. At the same time, the defect detection network model based on attention mechanism and multi-scale feature fusion is initialized, and the initial learning rate, momentum parameter, and weight decay coefficient are configured. The parameter matrix of each convolutional layer of the network is assigned values ​​using the pre-trained weights or He normal distribution initialization strategy, thereby constructing an initial model framework with feature extraction and classification regression capabilities, providing a computational carrier for subsequent iterative optimization.

[0061] The iterative training process is initiated. In each iteration, the image samples of the current batch are input into the aforementioned deep neural network. Multi-level feature maps are extracted through the backbone network, and the predicted category probability, bounding box offset, and mask contour of the crack defect are output through the feature pyramid network and the detection head. The network prediction results and the real labeled data are input into a composite loss function, which is composed of a weighted sum of focal classification loss, complete intersection-union regression loss, and binary cross-entropy segmentation loss. Based on this composite loss function, the forward propagation error of the current batch is calculated, and the gradient matrix of the loss function with respect to the parameters of each layer of the network is calculated using the backpropagation algorithm. The network weight parameters are updated through stochastic gradient descent or adaptive moment estimation optimizer, thereby continuously reducing the model's fitting residuals to the crack features in multiple iterations.

[0062] During the iterative training process, an adaptive sampling strategy based on loss feedback and feature mining is introduced to dynamically adjust the composition ratio of input samples in subsequent iterations. At this time, after each preset interval of iteration, the distribution of classification loss and regression loss of the current model on "samples of different subclasses, such as different lighting, different degrees of occlusion, and different crack types" is statistically analyzed. For the difficult sample subset with loss values ​​higher than a preset threshold, that is, samples that the model cannot accurately identify at present, such as micro-cracks under strong backlight or hidden cracks under heavy occlusion, the adaptive sampling strategy exponentially increases the corresponding sampling weights. For the simple sample subset with loss values ​​lower than a preset threshold, the sampling weights are proportionally reduced. Through this dynamic weight adjustment mechanism, in the data loading stage of subsequent iterations, the system selectively extracts samples from the augmented dataset according to the updated sampling probability distribution, so that the network can focus on difficult samples with complex feature distributions in the later stage of training, avoiding the model from getting trapped in local optima and overfitting.

[0063] Based on the execution results of the above adaptive sampling strategy, multiple crack data combinations are dynamically determined during iterative training. At this time, the samples in the augmented dataset are divided into multiple orthogonal dimensions according to the data domain labels, namely, "lighting condition dimension, including front lighting, backlighting, side lighting and shadow occlusion", "occlusion degree dimension, including no occlusion, light dirt occlusion, heavy corrosion or foreign object occlusion", and "crack type dimension, including macroscopic transverse fracture, microscopic oblique fatigue crack, network peeling and core damage surface micro-pits". After each adaptive sampling, the system extracts the domain labels corresponding to the current high sampling weight samples and cross-combines the domain labels with strong correlation, thereby dynamically determining the multiple extreme crack data combinations that the current network needs to optimize the most, such as the combination of "strong backlight + heavy corrosion occlusion + microscopic oblique fatigue crack", "side shadow + no occlusion + macroscopic transverse fracture", etc. These multi-dimensional cross-crack data combinations are continuously input into the network for targeted training, forcing the model to learn the robust expression of the essential characteristics of cracks under various extreme interference conditions, significantly improving the model's generalization detection ability under complex working conditions.

[0064] Specifically, the system loads the enhanced dataset generated by S123, which has red and blue stress background colors and multi-scale crack textures, into the detection network based on YOLOv8 and starts iterative training. When it reaches the 50th round, the overall loss value of the network has converged, but the system finds through the loss feedback mechanism that the regression loss of the model on certain types of samples is still high: one type is cracks under the condition of "local brightening caused by reflection of the track surface under twilight backlight", and the other type is micro-cracks under the condition of "severe oil and rust covering the bottom of the track".

[0065] Based on this, the adaptive sampling strategy was triggered, and the system dynamically adjusted the data loading probability from round 51 to round 100: the sampling weights of the two dimensions "dusk backlight" and "heavy oil stains obscuring" were increased from the original 5% to 25% and 20% respectively, while the weight of the originally easy-to-detect "front lighting without obstruction" sample was reduced to 2%.

[0066] Guided by weight adjustments, the system automatically combines and identifies several highly challenging crack data combinations for railway tracks in subsequent iterations. For example, the system identifies a challenging data combination: "dusk backlighting + severe sludge obscuring + micro-pit crack type on the surface of nuclear damage," and another combination: "lateral shadowing + light rust coverage + transverse fatigue crack type on the rail base." In the following thousands of iterations, the network intensively extracts features and calculates loss penalties for these extreme data combinations, forcing the network's attention mechanism to learn to penetrate the interference of reflections and oil stains, accurately capture the displacement characteristics of crack edges and the distribution patterns of stress coloring maps, thereby ensuring that the final trained model maintains an extremely high defect detection rate and positioning accuracy when facing complex surface contamination and variable lighting conditions on high-speed railway lines.

[0067] Furthermore, a sliding window with a dynamic scale is used to extract and infer features from each combination of crack data, thereby capturing the local texture gradient and global morphological context of the crack, thus generating a high-dimensional crack defect detection vector. This crack defect detection vector includes crack length, width, direction, and depth, fully considering the crack defect detection vector.

[0068] At this point, multiple crack data combinations determined in step S131 are acquired, and a dynamic-scale sliding window mechanism is configured for the image sample characteristics contained in each crack data combination. Based on prior metadata in the current crack data combination, such as illumination conditions, occlusion degree, and crack type labels, an adaptive mapping relationship between window size and prior features is established. For data combinations containing "macroscopic long strip cracks, such as longitudinal cracks along the rail waist," a rectangular sliding window with a "large aspect ratio, such as a long side covering 512 pixels and a short side covering 128 pixels" is dynamically generated to cover the global extension range of the crack. For data combinations containing microscopic short cracks or high occlusion, a "small-scale square window, such as 128 pixels × 128 pixels," is dynamically generated to focus on the extraction of local high-resolution texture features. The dynamically generated sliding window is deployed bidirectionally on the image samples with a "preset overlap rate, such as a step size of 25% of the window size," ensuring sufficient contextual information and boundary details are obtained in crack feature regions of any scale.

[0069] During the sliding window scanning process, local image patches captured by the window are input into a pre-trained feature extraction network for inference. At this time, a feature pyramid architecture with multi-level receptive fields is adopted to perform layer-by-layer convolution and downsampling on the local image patches. In the shallow layers of the network, small-scale convolutional kernels are used to extract the micro-texture gradient information of the crack edge. This texture gradient information represents the degree of abrupt change in gray value and normal direction between the crack and the surrounding rail substrate. Especially under high reflectivity or shadow occlusion conditions, this local texture gradient is a key feature to distinguish real cracks from surface oil artifacts. At the same time, in the deep layers of the network, the receptive field is expanded through large-scale convolutional kernels and self-attention mechanism to capture the global morphological context of the crack. This global morphological context represents the topological continuity, bifurcation law, and relative spatial position relationship of the crack curve with geometric structures such as rail head fillet or rail bottom slope. By cross-channel splicing and fusion of shallow texture gradient features and deep global context features, the model can not only identify isolated micro-damage, but also understand the overall direction and extension intention of the crack in a complex background.

[0070] After multi-scale feature fusion is completed, the fused feature map is input into the multi-task decoding head for regression decoding. At this time, the decoding head includes a classification branch and a geometric parameter regression branch. The classification branch outputs the confidence score of the presence of crack defects in the current window to filter out background interference. The geometric parameter regression branch quantitatively calculates the morphological parameters of the crack based on the fused feature map. For crack length and direction, the regression target is modeled as the center point coordinates and deflection angle of a rotated bounding box. The crack direction vector is determined by calculating the arctangent value of the deflection angle. The pixel length of the crack is calculated based on the long side scale of the rotated box and the downsampling ratio of the feature map. Then, it is converted into a physical length value by combining the physical resolution parameters of the imaging system. For crack width, the spacing between the extreme points of the texture gradient along the crack direction normal is extracted, and its maximum width and average width are calculated. For crack depth, the depth-grayscale mapping inversion model is constructed using the stress coloring features and shadow gradient distribution given in step S123. The equivalent depth information of the crack inside the rail is estimated by statistically analyzing the stress color gradation level and edge shadow attenuation coefficient in the crack area.

[0071] For all candidate regions covered by the sliding window on a single image, a non-maximum suppression algorithm is executed to eliminate overlapping and redundant detection results, and the best remaining detection results are structured and encoded. Then, the calculated physical length, maximum / average physical width, orientation angle, and equivalent depth values ​​are encapsulated according to a preset data structure, along with the confidence score and spatial coordinate information of the detection result, thereby generating a high-dimensional crack defect detection vector. The specific dimension of this crack defect detection vector is V=[Conf,x,y,Length,Widthmax,Widthavg,θdir,Depth], where Conf is the confidence score, (x,y) are the crack center coordinates, Length is the length, Width is the width, θdir is the orientation angle, and Depth is the equivalent depth. By traversing the inference results of all crack data combinations, the system outputs a set of high-dimensional crack defect detection vectors that closely approximate the actual physical dimensions in batches, providing quantitative and multi-dimensional data support for subsequent critical project assessments.

[0072] Specifically, for the common data combination of "hidden oblique fatigue cracks on the inner side of the rail head" in high-speed rail rails, since such cracks are easily confused with surface sludge under high-speed backlight, the system dynamically generates a slender sliding window with an aspect ratio of 4:1, such as 512×128 pixels. The long side of this window just matches the oblique direction in which the crack may extend, avoiding the introduction of too much irrelevant background interference by the conventional square window.

[0073] When the sliding window scans to the core region of the crack, the shallow layer of the network keenly captures the 0.05 mm micro-texture step gradient generated at the crack edge under backlight, which is a high-frequency fracture signal invisible from half a meter away. The deep network, combined with the arc geometry boundary of the rail head, captures the global morphological context of the crack extending downwards along the tread at a 30-degree angle, making the model confident that this is a real source of expanding fatigue, rather than a surface scratch.

[0074] The regression branch accurately solves the problem based on fusion features: the deflection angle of the rotated bounding box is extracted as 32 degrees, which is converted to a physical length of 18 mm; the spacing between the extreme points of the texture gradient is measured along the normal direction, and the maximum width is calculated to be 0.15 mm; more importantly, based on the deep red stress shading background map inside the crack and the edge shadow attenuation coefficient, the system inversely calculates that the equivalent depth of the crack has reached 4.2 mm, which touches the internal dangerous area of ​​the rail.

[0075] The system generates a high-dimensional detection vector V=[0.95,1250.5,340.2,18.0,0.15,0.08,32] for the crack at this location. [4.2] This vector not only tells the system "there is a crack here", but also precisely describes "at mileage 1250.5, there is an extremely dangerous fatigue crack that is 18 mm long, 0.15 mm wide, 4.2 mm deep and develops at a 32-degree angle", providing an unambiguous quantitative basis for the next step in S133 to assess whether it constitutes a "nuclear injury critical item".

[0076] Therefore, the crack defect detection vector is dynamically identified, and the corresponding crack defect characteristics are determined by combining the cluster analysis mechanism during the identification process. Furthermore, the defect response mapping relationship of the railway track is combined to screen out the key factors that lead to track failure risk, thereby determining multiple critical items of the corresponding crack defect characteristics. These critical items include, but are not limited to, rail head peeling, rail web vertical cracks and rail bottom core damage.

[0077] At this point, the high-dimensional crack defect detection vector generated in step S132, containing crack length, width, orientation, and depth, is obtained, and the dynamic recognition mechanism is activated. This high-dimensional detection vector is then input into a pre-constructed nonlinear feature projection space, which is obtained through dimensionality reduction training on massive historical defect data using a manifold learning algorithm. During the projection process, an adaptive weight allocation strategy is adopted, dynamically adjusting the projection weights based on the dispersion of values ​​in each dimension of the vector. The depth and orientation dimensions, which significantly affect the integrity of the track structure, are assigned higher-order projection coefficients, while the length and width dimensions are assigned corresponding coefficients as morphology correction factors. Through this dynamic projection calculation, the detection vectors originally scattered in the high-dimensional space are mapped to feature coordinate points in the low-dimensional manifold space, thereby achieving the separability measurement and dynamic recognition of different crack morphologies in the feature space, providing standardized input for subsequent clustering analysis.

[0078] Within the aforementioned low-dimensional manifold feature space, a clustering analysis mechanism based on density and distance constraints is introduced to divide the mapped feature coordinate points into clusters. Then, DBSCAN or ordered nearest neighbor graph clustering algorithms are used, with preset neighborhood radius and minimum sample number as constraints, to traverse all coordinate points in the feature space. For any coordinate point, if there are other coordinate points within its neighborhood radius with a number of samples not less than the minimum, this region is identified as a high-density core area, and adjacent high-density core areas are connected and merged into the same cluster. During this process, crack vectors with similar length-to-depth ratios, directional dip angles, and width-to-depth ratios will be clustered in the same cluster, while isolated and deviating anomalous vectors are marked as noise points or rare defects. By extracting the boundary contours and centroids of the clusters, the typical crack defect features represented by each cluster are determined. These defect features not only include the numerical range of a single dimension but also encompass the coupling relationships between multi-dimensional parameters.

[0079] After determining the characteristics of the crack defect, a pre-set railway track defect response mapping database is retrieved. This database, built based on finite element ultimate bearing capacity simulation and fracture mechanics theory, stores mapping functions between various combinations of crack characteristic parameters and the decay of track structure mechanical response. The crack defect characteristics obtained from the current cluster analysis are input into this database, and forward inference calculation is performed. At this time, the stress intensity factor, section bending stiffness reduction rate, and fatigue crack propagation rate of the current crack characteristic under the design wheel load are calculated. If the above mechanical response indicators exceed the preset safety threshold, the crack characteristic is determined to be a key factor leading to track failure risk. On this basis, through a reverse tracing mechanism, the crack characteristic dimensions corresponding to exceeding the safety threshold are extracted, such as depth exceeding the critical value and orientation perpendicular to the main force direction, and these are marked as risk-dominant factors, thereby achieving accurate screening from superficial defect characteristics to deep failure mechanisms.

[0080] Based on the key factors identified above that lead to track failure risk, and combined with the preset track defect taxonomy standards, a mapping and matching operation for critical items is performed. At this point, a decision tree classification model is constructed between the combination of key factors and the labels of critical items. The node splitting basis of this decision tree is the threshold range of risk-dominant factors such as depth, orientation, and location. The current combination of key factors is input into the decision tree, and the tree traverses downwards along the branch path corresponding to the feature value, eventually landing on a leaf node. This leaf node corresponds to the identified critical item. Through this mapping and matching, the abstract risk factors are transformed into engineering-identifiable critical item categories, thereby identifying multiple critical items with corresponding crack defect characteristics. These critical items include, but are not limited to, railhead peeling, vertical cracks in the rail web, and core damage at the rail base, completing the closed-loop judgment from quantitative detection vectors to qualitative hazard levels and categories.

[0081] Specifically, the system projects the above vectors onto the manifold space and finds that V1 has the characteristics of being shallow and long with a small inclination angle, and is clustered in the same category as the "fatigue microcrack cluster" in the history database; while V2 has the characteristics of being short and wide, with an approximately vertical inclination angle and a very large depth, and is highly clustered with the "internal brittle fracture cluster", thus identifying two completely different defect characteristics.

[0082] The system inputs these two types of defect features into the defect response mapping database: For the shallow and long feature represented by V1, the calculated stress intensity factor is still lower than the crack propagation threshold, and the cross-sectional stiffness reduction rate is only 1.5%, which does not reach the failure risk threshold, so the key factor is eliminated; while for the short, wide, deep and vertical feature represented by V2, the calculated stress intensity factor is close to the critical value of rail fracture toughness, and the cross-sectional bending stiffness reduction rate is as high as 12%, which is very easy to cause instantaneous brittle fracture under the impact of high-speed train wheel load, so "large depth-to-width ratio + near vertical orientation + deep propagation" is selected as the key factor leading to track failure risk.

[0083] The decision tree model makes its judgment based on the above key factors: the defect depth exceeds 8 mm, reaching the danger zone inside the rail head, the angle between its direction and the longitudinal axis of the rail is greater than 75 degrees, it is perpendicular to the direction of the principal tensile stress, and the surface opening width is significant. These features accurately hit the leaf node in the decision tree corresponding to the extremely high risk of derailment, and the system ultimately identifies it as the "rail head nuclear damage" critical item. At the same time, if another key factor selected is "located at the neutral axis of the rail web, vertically oriented, and deeply penetrating", it is identified as the "rail web vertical crack" critical item. Thus, the system achieves accurate qualitative and hazard classification of hidden and fatal defects under railway tracks.

[0084] refer to Figure 5 In step S14, the specific steps are as follows: S141: Collect the crack defect area corresponding to the railway track, and perform spatiotemporal alignment with multiple critical items of the crack defect characteristics. In this way, time decay factors are introduced during the alignment process, and multi-factor fusion evaluation of the crack defect area is carried out. Multiple maintenance nodes are dynamically determined during the fusion process. S142: Based on the railway track database, retrieve the past maintenance history of the railway track, determine the historical maintenance content based on the identification of the past maintenance history, and determine the graded maintenance content of the railway track by combining the node location of multiple maintenance nodes and the corresponding crack defect content. The graded maintenance content covers maintenance plans for ordinary cracks and maintenance plans for critical cracks.

[0085] In the embodiments of this application, the crack defect area corresponding to the railway track is collected, and multiple critical items of the crack defect characteristics are spatiotemporally aligned. In this way, a time decay factor is introduced during the alignment process, and then a multi-factor fusion evaluation of the crack defect area is carried out. During the fusion process, multiple maintenance nodes are dynamically determined and multiple maintenance nodes are introduced.

[0086] At this point, based on the output of step S133, the spatial location information of the crack defect area corresponding to the railway track is collected. At this point, the absolute mileage coordinates, lateral offset, and three-dimensional boundary contour data of the defect center point are extracted from the high-dimensional crack defect detection vector to form a spatial bounding box of the defect area. At the same time, multiple critical items corresponding to the crack defect characteristics determined in step S133 are obtained, such as discretized risk labels for rail head peeling, rail web vertical cracks, and rail bottom core damage. The above critical items are attached to the corresponding spatial bounding box as risk attribute fields. By establishing a mapping relationship between spatial geometric data and qualitative risk labels, a set of track defect spatial entities containing exact geographical locations and specific critical categories is constructed, providing a spatial index benchmark for subsequent multi-dimensional data fusion.

[0087] Historical inspection records and maintenance logs from the track lifecycle management database are retrieved and spatiotemporally aligned with the aforementioned defect spatial entity set. Then, using absolute mileage coordinates as the primary key, the defect spatial entity at the current inspection moment is sequentially spliced ​​with historical defect evolution records at the same mileage location to construct the temporal sequence state space of the defect at that location. During this alignment process, a time decay factor based on the material fatigue cumulative damage theory is introduced. This time decay factor is specifically manifested as a time decay function, with its independent variable being the time interval Δt between the current inspection moment and the last effective maintenance moment at the same location. Different decay indices are set for different critical projects. For example, a high-exponential decay model is used for rapidly developing railhead peeling, while a medium-exponential decay model is used for relatively slowly developing railbed core damage. By multiplying the severity vector of historical defects by this time decay function value, a nonlinear weighting of the influence of historical states on the current state is achieved. This ensures that the further away the historical damage record is from the current moment, the lower its contribution to the current fusion assessment, truly reflecting the objective law of performance degradation of track materials under natural aging and continuous dynamic loads.

[0088] After completing spatiotemporal alignment and time decay processing, a multi-factor fusion assessment is performed. At this point, a multi-factor fusion assessment model is constructed with the current defect morphology characteristics, the risk level of critical items, and the historical damage state after decay as input vectors. The "geometric dimension parameters, such as depth and area" and "mechanical response parameters, such as stress concentration factor" of the current defect are weighted and summed with the corresponding critical item weights to calculate the immediate risk index of the current defect. At the same time, the historical damage vector after decay is concatenated with the immediate risk index at the feature level and input into a nonlinear fusion layer composed of a multilayer perceptron or a long short-term memory network. This fusion layer comprehensively considers the current destructive force of the defect and the historical cumulative damage effect, and outputs a probability value of track local structural instability between 0 and 1. This probability value is used as the comprehensive risk quantification index of the multi-factor fusion assessment, thereby accurately quantifying the real risk level of the crack defect area causing overall track failure in the current and short term.

[0089] Optionally, the comprehensive risk quantification index is a dimensionless value between 0 and 1, used to characterize the comprehensive probability that the crack defect area will cause local instability of the track structure in the current state. This index needs to be correlated with measurable mechanical parameters in order to predict the risk evolution.

[0090] At this point, the comprehensive risk quantification index is linearly mapped to the ratio of the equivalent stress intensity factor to the fracture toughness of the material at the crack tip. Specifically, when the comprehensive risk quantification index is 0, the ratio of the equivalent stress intensity factor to the fracture toughness is 0, indicating a defect-free state; when the comprehensive risk quantification index is 1, the equivalent stress intensity factor equals the fracture toughness of the material, indicating a critical instability state. For 60N type U71Mn rail material, its plane strain fracture toughness is measured to be 38 MPa multiplied by the square root of m through a standard compact tensile test. Therefore, the equivalent stress intensity factor corresponding to any comprehensive risk quantification index value is equal to that index value multiplied by 38 MPa multiplied by the square root of m. Simultaneously, the correspondence between the equivalent stress intensity factor and the characteristic depth of the crack under this crack geometry is established through finite element analysis. That is, under a fixed wheel load, the equivalent stress intensity factor is proportional to the square root of the crack depth. The proportionality coefficient is obtained through single-parameter sensitivity analysis of this specific crack configuration, specifically taking a value of 9.5 MPa.

[0091] Based on the comprehensive risk quantification index output from the multi-factor fusion assessment, combined with the track instability risk threshold array, multiple maintenance nodes are dynamically extrapolated and determined. At this point, using the time axis as the horizontal axis and the comprehensive risk quantification index as the vertical axis, the future risk evolution curve of the defect area is extrapolated based on the fatigue crack propagation rate model and the current risk value. Multi-level risk intervention thresholds corresponding to different critical projects are preset in the track operation and maintenance specifications, such as yellow warning thresholds, orange speed limit thresholds, and red lockdown thresholds. The time of the first intersection between the risk evolution curve and the above multi-level thresholds is calculated, and this intersection time is mapped to the actual railway operation calendar, thereby dynamically determining the corresponding maintenance nodes. For example, the time node reaching the yellow warning threshold is determined as the "recent observation node," the time node reaching the orange speed limit threshold is determined as the "planned maintenance node," and the time node reaching the red lockdown threshold is determined as the "emergency intervention node." Through this dynamic extrapolation mechanism, a leap from static defect identification to dynamic life prediction is achieved, providing precise time window guidance for the subsequent formulation of graded maintenance strategies.

[0092] Specifically, the system collects the spatial coordinates of the crack defect area, such as the absolute mileage K115+200, inside the rail head, and attaches the critical item "rail head core damage" determined in step S133 to this area, forming a defect entity with a risk label. The system searches the historical database and finds that the mileage was where rail grinding and repair were carried out 3 years ago, with historical records of surface microcracks. The system introduces a time decay factor. Since "rail head core damage" is a high-risk defect that expands rapidly, its decay index is set to 0.8. After the time decay function is calculated, the weight of the historical microcrack damage from 3 years ago on the current state has decayed to less than 10%, while the currently detected core damage features occupy an absolutely dominant position, avoiding interference from historical minor defects in the judgment of the current serious crisis.

[0093] The system adds the weights of "current nuclear damage depth: 12 mm" and "high stress concentration coefficient" to the critical item "railhead nuclear damage" to obtain an immediate risk index of 0.75. At the same time, it overlays the historical damage status after attenuation and calculates the current comprehensive risk quantification index of the area as 0.78 through nonlinear fusion layer calculation, which is already in the high danger range.

[0094] The system extrapolates the risk evolution curve of this core damage based on Paris' crack propagation law. The preset yellow warning threshold for high-speed rail rails is 0.80, and the orange speed limit threshold is 0.90, at which point the core damage area accounts for 15% of the rail head area. Calculations show that the risk curve will reach the 0.80 threshold in 15 days and the 0.90 threshold in 45 days. Based on this, the system dynamically determines two key maintenance nodes: the time point 15 days later is marked as the "recent observation node," requiring increased inspection frequency; the time point 45 days later is marked as the "planned maintenance node," requiring rail replacement or reinforcement to be completed before this time, otherwise a speed limit intervention scenario for high-speed rail will be triggered.

[0095] Furthermore, based on the railway track database, the historical maintenance history of the railway track is retrieved. The historical maintenance content is determined by identifying the past maintenance history, and the graded maintenance content of the railway track is determined by combining the node locations of multiple maintenance nodes and the corresponding crack defects. The graded maintenance content covers maintenance plans for ordinary cracks and maintenance plans for critical cracks. The crack defect detection vector is further controlled, and multiple critical items of crack defect area and crack defect characteristics are fully considered, which improves the accuracy of the graded maintenance content of the railway track.

[0096] At this point, based on the distributed relational database of the railway track, using the absolute mileage coordinates of the defect area and the track component number determined in step S141 as the primary key, the system retrieves the past maintenance history of the railway track within a preset historical period. The database stores log records of all inspections, repairs, and overhauls. The system extracts historical maintenance work order streams matching the current defect mileage through the data access interface. The system performs serialization and parsing of these work order streams, extracting the operation type code, operation depth parameters, material consumption records, and post-operation re-inspection results corresponding to each historical time node. By removing invalid and interrupted operation records, the discrete work order data is reconstructed into a historical maintenance event chain arranged in a time sequence, thereby obtaining the complete set of intervention measures experienced by the mileage location in the past time period, providing a structured data source for subsequent historical maintenance content identification.

[0097] Semantic recognition and feature extraction are performed on the aforementioned historical maintenance event chain to determine the historical maintenance content. Then, based on natural language processing technology or a rule matching engine, the unstructured text descriptions in the work orders are transformed into standardized "maintenance action labels," such as "rail grinding," "rail welding," and "rail replacement," along with corresponding geometric parameter labels, such as grinding depth 0.5mm and welding length 30mm. On this basis, an intervention effectiveness attenuation assessment mechanism is introduced. Combining the time difference between the historical operation timestamp and the current timestamp, and the cumulative total weight borne by the track after the historical operation, the effectiveness attenuation coefficient of the historical maintenance content is calculated. For example, for historical "rail grinding" operations, due to the wear and fatigue accumulation of the surface hardened layer after grinding, its effectiveness in eliminating surface stress concentration decreases exponentially with the total weight borne. Using this effectiveness attenuation coefficient, the system quantifies and reduces the actual residual protection effect of the historical maintenance content, identifying whether the historical intervention measures have completely failed or still have partial load-bearing capacity, thereby clarifying the current true foundation state of the track.

[0098] After clarifying the historical baseline, multiple maintenance nodes dynamically determined in step S141 are obtained, and the distribution location of each maintenance node on the time axis and its corresponding hazard level label are extracted. Simultaneously, the crack defect content associated with the maintenance node in step S133 is retrieved. This defect content includes the "geometric morphological parameters of the crack: length, width, and depth", spatial orientation, and "the hazard project category, such as rail head peeling, rail web vertical crack, and rail bottom core damage". A spatiotemporal constraint-based correlation mapping model is established to couple the mechanical damage mechanism of the current crack defect with the intervention failure mechanism of the historical maintenance content. At this point, it is determined whether the current crack defect is located in the edge overlap area or stress overlap area of ​​the historical maintenance area. If the current "rail web vertical crack" happens to be located in the heat-affected zone of the historical "welding repair area", it indicates that the historical maintenance content not only failed to suppress the defect, but also became the cause of the new defect due to residual stress. Through this spatiotemporal constraint correlation, the system clarifies the evolution cause of the crack defect under the background of historical intervention and the current stress boundary, providing accurate boundary conditions for formulating targeted maintenance plans.

[0099] The effectiveness evaluation results of comprehensive historical maintenance content, the time margin of maintenance nodes, and the urgency of crack defects are input into a pre-set multi-dimensional constrained decision tree model to determine the graded maintenance content of the railway track. This multi-dimensional constrained decision tree branches along a three-dimensional coordinate axis of "time margin - defect urgency - historical intervention effectiveness." For defects with "sufficient time margin: at a recent observation node," "low defect urgency: ordinary micro-cracks with no propagation trend," and for which historical maintenance effectiveness still exists, the decision tree outputs a maintenance plan for ordinary cracks. This plan specifically includes low-intensity, non-invasive methods such as preventative rail grinding, applying rust-proof and crack-inhibiting coatings, or routine monitoring and tracking. Interventional maintenance instructions; conversely, for defects with "tight time margin: at the planned maintenance or emergency intervention node" and "high defect urgency: belonging to critical items such as rail head peeling, rail web vertical cracks and rail bottom core damage" and where the effectiveness of historical maintenance has been completely ineffective or historical intervention was the cause of cracking, the decision tree outputs maintenance plans for critical cracks. These plans specifically include high-intensity, interventional repair instructions such as rail cutting and replacement in the defective area, in-situ aluminothermic welding repair, implementation of stress relief hole drilling or emergency speed limit closure. Through the above hierarchical logic, the system automatically generates hierarchical maintenance content that is highly matched with the actual condition and evolution risk of the track, realizing precise allocation of maintenance resources and closed-loop risk control.

[0100] Specifically, regarding the "rail head damage" and a "fine surface fatigue line" detected at the K115+200 mileage of the high-speed railway, the system retrieved past maintenance records from the engineering system database using this mileage as the primary key. Analysis revealed that "rail contour grinding" had been carried out at this location two years ago, with a grinding depth of 0.3mm.

[0101] The system determined that the "rail grinding" was aimed at surface corrugation and micro-cracks. However, considering that it had accumulated a total weight of 150 million tons of high-speed rail passing through it over the past two years, the system calculated that its intervention effectiveness attenuation coefficient had reached 0.85, meaning that the stress concentration eliminated by the original grinding had been basically restored and deepened.

[0102] The system correlates the current defects with the nodes: it finds that the current "fine fatigue lines on the surface" are located exactly in the center of the grinding zone from 2 years ago, which is a normal fatigue recurrence; while the fatal "railhead core damage" originated in the hardened transition zone inside the grinding zone from 2 years ago, indicating that the hardness gradient difference formed by historical grinding is one of the potential causes of this railhead core damage.

[0103] The decision tree model makes a judgment based on the above comprehensive information: for "fine fatigue marks on the surface", its urgency is low and it is at the "recent observation node", the decision tree classifies it as a common crack and outputs the maintenance plan for common cracks as follows: arrange preventive rail grinding during the next maintenance window, such as cutting depth of 0.2mm, to eliminate surface micro-marks and apply long-lasting rust inhibitor.

[0104] For "rail head core damage", the urgency is extremely high and it is at the "planned maintenance node". There are only 45 days left before the risk gets out of control. At the same time, the hardened layer caused by historical grinding is the cause. Conventional grinding can no longer repair it and will instead accelerate its expansion. The decision tree classifies it as a critical crack and outputs the maintenance plan for critical cracks as follows: immediately issue a speed limit dispatch order, such as reducing the speed from 250km / h to 160km / h, and arrange for large track maintenance machinery to enter the site within 7 days to carry out the "rail removal and aluminothermic welding replacement of the core damage section" operation.

[0105] Please see Figure 6 The twin simulation-based railway track crack defect detection device is applied to the aforementioned twin simulation-based railway track crack defect detection method; the twin simulation-based railway track crack defect detection device includes: The twin simulation module 21 is used to collect the appearance data of the railway track and perform spatiotemporal alignment with the geometric shape and position data to construct multimodal data of the railway track. The multimodal data of the railway track is mapped to the preset twin simulation space to construct a twin simulation body with texture details and topological structure. The augmented dataset module 22 is used to perform mesh discretization on the twin simulation object, and combine the material parameters and historical crack distribution characteristics to determine the corresponding crack defect distribution information. It further combines the stress distribution map of the railway track to perform texture rendering, thereby generating the corresponding augmented dataset. The defect identification module 23 is used to iterate the enhanced dataset and determine multiple crack data combinations during the iteration process. It performs sliding window inference on each crack data combination to generate a corresponding crack defect detection vector. Based on the identification of the crack defect detection vector, it determines multiple critical items of the corresponding crack defect features. The graded maintenance module 24 is used to mark the crack defect area corresponding to the railway track, and to integrate multiple critical items with the characteristics of the crack defect to determine multiple maintenance nodes during the integration process, and to determine the graded maintenance content of the railway track by combining the previous maintenance history of the railway track.

[0106] It should be noted that although multiple modules are mentioned in the detailed description above, this division is not mandatory; in fact, according to the embodiments of this disclosure, the features and functions of two or more modules or described above can be embodied in one module; conversely, the features and functions of one module described above can be further divided into multiple modules to be embodied.

[0107] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein; this application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein; the specification and embodiments are to be considered exemplary only.

[0108] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for detecting crack defects in railway tracks based on twin simulation, characterized in that, include: The apparent data of railway tracks are collected and combined with geometric shape and position data for spatiotemporal alignment to construct multimodal data of railway tracks. The multimodal data of railway tracks is then mapped to a preset twin simulation space to construct a twin simulation object with texture details and topological structure. The twin simulation is discretized into a mesh, and the corresponding crack defect distribution information is determined by combining material parameters and historical crack distribution characteristics. Furthermore, the stress distribution map of the railway track is combined for texture rendering to generate the corresponding enhanced dataset. The augmented dataset is iterated, and multiple crack data combinations are identified during the iteration process. Sliding window inference is performed on each crack data combination to generate a corresponding crack defect detection vector. Based on the identification of the crack defect detection vector, multiple critical items of the corresponding crack defect features are identified. The crack defect area corresponding to the railway track is marked, and multiple critical items with the characteristics of the crack defect are combined to perform multi-factor fusion. During the fusion process, multiple maintenance nodes are identified, and the graded maintenance content of the railway track is determined by combining the previous maintenance history of the railway track.

2. The railway track crack defect detection method based on twin simulation according to claim 1, characterized in that, The process involves collecting the apparent data of the railway track and combining it with geometric shape and position data for spatiotemporal alignment to construct multimodal data of the railway track. This multimodal data is then mapped to a pre-defined twin simulation space to construct a twin simulation object with texture details and topological structure, including: The apparent data of railway tracks is collected in real time using an on-board multi-source sensor array, and the geometric shape and position data of the railway tracks are acquired simultaneously. The apparent data and geometric shape and position data are then jointly registered and aligned in a unified spatiotemporal coordinate system. During the fusion process, multimodal data of the railway tracks is constructed, which includes texture features, geometric shape and thermal features.

3. The railway track crack defect detection method based on twin simulation according to claim 2, characterized in that, The process of collecting apparent data of railway tracks and combining it with geometric shape and position data for spatiotemporal alignment to construct multimodal data of railway tracks, mapping the multimodal data of railway tracks to a preset twin simulation space, and constructing a twin simulation object with texture details and topological structure, also includes: A preset twin simulation space is obtained, multimodal data is mapped to the preset twin simulation space, and the texture details and internal topology of the railway track surface are reconstructed in the twin simulation space through nonlinear mapping, thereby generating a twin simulation of the railway track with consistent physical properties.

4. The railway track crack defect detection method based on twin simulation according to claim 1, characterized in that, The process involves discretizing the twin simulation object into a mesh, determining the corresponding crack defect distribution information by combining material parameters and historical crack distribution characteristics, and further combining this with the stress distribution map of the railway track for texture rendering, thereby generating a corresponding enhanced dataset, including: The twin simulation of the railway track is meshed and discretized using an adaptive mechanism to construct the corresponding finite element mesh body, which contains geometric information and material properties. The material parameters of the railway track are collected synchronously, and multiple calculations are performed in the finite element mesh body in combination with the corresponding historical crack distribution characteristics to further determine the corresponding crack defect distribution information, which covers the crack occurrence location and propagation path.

5. The railway track crack defect detection method based on twin simulation according to claim 4, characterized in that, The process of meshing and discretizing the twin simulation object, determining the corresponding crack defect distribution information by combining material parameters and historical crack distribution characteristics, and further combining the stress distribution map of the railway track for texture rendering to generate the corresponding enhanced dataset also includes: The stress distribution map of the railway track is obtained and superimposed and fused with the crack defect distribution information. During the fusion process, the corresponding crack extension trajectory is determined, and the corresponding stress concentration degree is combined to perform differentiated texture rendering, thereby generating an enhanced dataset containing stress shading features and multi-scale crack morphology.

6. The railway track crack defect detection method based on twin simulation according to claim 1, characterized in that, The process involves iterating over the enhanced dataset, determining multiple crack data combinations during the iteration, and performing sliding window inference on each crack data combination to generate a corresponding crack defect detection vector. Based on the identification of this crack defect detection vector, several critical items are determined to identify the corresponding crack defect characteristics, including: The augmented dataset is used as input data and trained iteratively multiple times. During the iteration process, an adaptive sampling strategy is used to determine multiple crack data combinations, which include different lighting conditions, occlusion levels and crack types. A dynamic-scale sliding window is used to extract and infer features from each combination of crack data, thereby capturing the local texture gradient and global morphological context of the crack, and generating a high-dimensional crack defect detection vector, which includes crack length, width, direction and depth.

7. The railway track crack defect detection method based on twin simulation according to claim 6, characterized in that, The process involves iterating over the enhanced dataset, determining multiple crack data combinations during the iteration, and performing sliding window inference on each crack data combination to generate a corresponding crack defect detection vector. Several critical projects that determine the corresponding crack defect characteristics based on the identification of this crack defect detection vector also include: The crack defect detection vector is dynamically identified, and the corresponding crack defect characteristics are determined by combining the cluster analysis mechanism during the identification process. Furthermore, the defect response mapping relationship of the railway track is combined to screen out the key factors that lead to track failure risk, thereby identifying multiple critical items of the corresponding crack defect characteristics. These critical items include, but are not limited to, rail head peeling, rail web vertical cracks, and rail bottom core damage.

8. The railway track crack defect detection method based on twin simulation according to claim 1, characterized in that, The marked crack defect area of ​​the railway track is used to integrate multiple critical items related to the crack defect characteristics. During this integration process, multiple maintenance nodes are identified, and the graded maintenance content of the railway track is determined based on its past maintenance history, including: The crack defect area corresponding to the railway track is collected, and multiple critical projects with the characteristics of the crack defect are spatiotemporally aligned. The time decay factor is introduced in the alignment process, and then a multi-factor fusion assessment of the crack defect area is carried out. Multiple maintenance nodes are dynamically determined in the fusion process.

9. The railway track crack defect detection method based on twin simulation according to claim 8, characterized in that, The method involves marking the crack defect area corresponding to the railway track, and combining multiple critical items related to the crack defect characteristics for multi-factor fusion. During the fusion process, multiple maintenance nodes are identified, and the graded maintenance content of the railway track is determined based on its past maintenance history. This also includes: Based on the railway track database, the historical maintenance history of the railway track is retrieved. The historical maintenance content is determined by identifying the past maintenance history. Combined with the node locations of multiple maintenance nodes and the corresponding crack defects, the graded maintenance content of the railway track is determined. The graded maintenance content covers maintenance plans for ordinary cracks and maintenance plans for critical cracks.

10. A railway track crack defect detection device based on twin simulation, characterized in that, The twin simulation-based railway track crack defect detection device is applied to the twin simulation-based railway track crack defect detection method as described in any one of claims 1-9; The railway track crack defect detection device based on twin simulation includes: The twin simulation module is used to collect the appearance data of railway tracks and perform spatiotemporal alignment with geometric shape and position data to construct multimodal data of railway tracks. The multimodal data of railway tracks is then mapped to a preset twin simulation space to construct a twin simulation body with texture details and topological structure. The augmented dataset module is used to mesh and discretize the twin simulation object, and combine the material parameters and historical crack distribution characteristics to determine the corresponding crack defect distribution information. It further combines the stress distribution map of the railway track for texture rendering, thereby generating the corresponding augmented dataset. The defect identification module is used to iterate the augmented dataset and determine multiple crack data combinations during the iteration process. Sliding window inference is performed on each crack data combination to generate the corresponding crack defect detection vector. Based on the identification of the crack defect detection vector, multiple critical items of the corresponding crack defect features are determined. The graded maintenance module is used to mark the crack and defect area corresponding to the railway track, and to integrate multiple critical items with the characteristics of the crack and defect to determine multiple maintenance nodes during the integration process. The graded maintenance content of the railway track is determined by combining the past maintenance history of the railway track.

Images