A crack dynamic analysis system of a railway track

Abstract

This invention relates to the field of image detection and discloses a dynamic analysis system for railway track cracks, aiming to solve the problems of strong background interference, high false alarm and missed detection rates, and limited dynamic analysis dimensions in existing track crack detection technologies. The system is characterized by comprising: a texture library construction module, an image acquisition module, an image processing module, a texture expansion module, a damage judgment module, and a dynamic analysis module. By constructing a multi-dimensional texture library, combined with image processing and texture expansion modules, accurate crack identification and hierarchical judgment are achieved; through multi-feature index damage judgment and dynamic analysis, damage assessment and development trend prediction of cracks are performed. This application enables accurate, comprehensive, and dynamic monitoring and assessment of railway track cracks, reducing false alarm and missed detection rates, improving maintenance efficiency and accuracy, and ensuring railway transportation safety.

Description

Technical Field

[0001] This invention relates to the field of image detection, and more specifically to a dynamic analysis system for cracks in railway tracks. Background Technology

[0002] As the core load-bearing structure of rail transit, the structural integrity and operational reliability of railway tracks directly affect the safety and efficiency of the entire transportation network. Among these technologies, dynamic crack identification technology for track surfaces is an important branch of condition monitoring. This technology aims to accurately identify and assess cracks and defects on the track surface caused by factors such as fatigue and wear through real-time or near-real-time detection methods, providing a basis for preventative maintenance decisions.

[0003] Current technologies primarily rely on periodic manual inspections or image-based automated detection systems to observe the track surface. However, track maintenance often involves periodic mechanical polishing, resulting in a multi-layered, composite textured structure. The textured layers formed at different maintenance cycles vary in morphology and density. This layered texture creates strong background interference, making it easy for traditional image recognition algorithms to misjudge the boundaries between layers as cracks, or to mask true cracks located in deeper structures due to their texture characteristics. Furthermore, current dynamic analysis methods often focus on the macroscopic geometric parameters of cracks, such as length or depth, using these as the sole criterion for maintenance decisions. This single criterion fails to adequately consider the complexity of the track structure after multiple maintenance cycles, potentially leading to missed detections or misjudgments in complex textured environments, thus creating safety hazards. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a dynamic analysis system for cracks in railway tracks. This application introduces texture library construction, hierarchical identification, and damage texture expansion technologies, and combines damage judgment based on multiple feature indicators with a multi-dimensional dynamic analysis mechanism to achieve accurate identification, damage assessment, and prediction of development trends of cracks on the track surface.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A dynamic analysis system for cracks in railway tracks, characterized in that it includes:

[0007] A texture library construction module, wherein the texture library construction module pre-stores several standard crack information and corresponding texture images into the texture library;

[0008] The image acquisition module is used to collect orbital image information;

[0009] The image processing module acquires the track image information and determines whether there is crack information. If the crack information exists, the track image information is defined as a crack image and sent to the texture expansion module.

[0010] The texture expansion module extracts crack features from the crack image and matches standard texture information and corresponding texture images in the texture library to obtain a damaged texture expansion map after expanding the texture.

[0011] The damage determination module acquires the damage texture expansion map and determines the degree of damage. If the degree of damage exceeds a preset damage threshold, a maintenance signal is generated.

[0012] The dynamic analysis module acquires damage texture expansion maps at different time periods and determines the damage trend. If the damage trend exceeds a preset damage trend threshold, a maintenance signal is generated.

[0013] In this invention, preferably, the texture augmentation module is configured with a texture augmentation strategy, including:

[0014] Crack features are obtained and derived based on the morphological features of the crack, including crack length, crack width, crack direction, and branch distribution in the crack image;

[0015] Calculate the similarity between the crack feature and each of the standard crack information in the texture library. If there is standard crack information with a similarity exceeding a preset similarity threshold, define the standard crack information as similar crack information and define the corresponding texture image as similar texture image.

[0016] The crack image is augmented with texture based on the similar texture image to obtain the damage texture augmentation map.

[0017] In this invention, preferably, the texture expansion module is configured with a damage level judgment strategy, including judging whether there are multiple damage layers based on the texture level of the crack and the embedding depth. If multiple damage layers exist, the texture expansion strategy is executed for each damage layer after identifying multiple damage layers. If there are no multiple damage layers, the texture expansion strategy is executed directly.

[0018] In this invention, preferably, the damage level determination strategy is configured with a crack level identification sub-strategy, including:

[0019] Extract multi-scale texture features of the crack region and its neighborhood, including directionality, statistical texture parameters and deep semantic texture features;

[0020] By combining multi-size texture features and classification models, pixel-level texture layer segmentation is performed on the orbital image;

[0021] A texture hierarchy model is established, which includes the depth and feature parameters of each texture layer. The texture layer represents the polished layer or composite texture structure left by track maintenance at different times.

[0022] The texture of the cracked area is compared with the model, and the layer where the crack is located is determined based on feature similarity and depth information.

[0023] In this invention, preferably, the texture augmentation strategy includes:

[0024] Extract the geometric morphological features of the crack and its interaction patterns with the texture layer, including edge gloss, local distortion and surface reflection features;

[0025] A multimodal deep feature matching method is used to calculate the similarity between cracks and standard samples;

[0026] Damage texture augmentation is performed based on similar texture images using a generative model, outputting an augmented image that includes the crack embedding layer and propagation direction.

[0027] In this invention, preferably, the damage judgment module calculates a damage index based on multiple feature indicators. If the damage index exceeds a preset threshold, a maintenance signal is output. The multiple feature indicators include:

[0028] Results of crack location layer identification;

[0029] The length, width, angle, and complexity of the crack;

[0030] The degree to which a crack disturbs the surrounding texture.

[0031] In this invention, preferably, the dynamic analysis module further includes a difference image generation submodule, which is used to obtain image data of the corresponding track area from the damage texture expansion map at different time periods, and generate a difference image by using image registration and difference calculation methods. The difference image is used to characterize the degree of change of crack texture in the same area, thereby calculating the growth rate of crack damage, and determining whether the current crack propagation trend exceeds the preset safe growth rate threshold.

[0032] In this invention, preferably, the dynamic analysis module further includes a surface corrosion identification submodule, which is used to detect abnormal color distribution and texture roughness distribution in the damage texture expansion map; and combine it with the standard rust spot texture image, and determine whether there is a rust area on the track surface through image similarity comparison and color-texture feature analysis. If a rust spot area is detected, the area is marked as a surface corrosion area and a rust warning signal is output.

[0033] In this invention, preferably, the dynamic analysis module includes an internal damage detection submodule. The internal damage detection submodule obtains local texture depressions or abrupt changes in the damage texture expansion map to define them as suspected chipping damage areas, extracts the edge features, brightness gradient changes, and texture direction distribution of the suspected chipping damage areas, and determines whether there are chipping deformation features by combining the pixel intensity fluctuations of the corresponding areas in the difference image.

[0034] In this invention, preferably, the internal damage detection submodule further includes an image morphology change analysis unit, which is used to extract the edge contour, texture density change and grayscale discontinuity features of the suspected chip damage area by comparing the damage texture expansion images at different time points, and combine the area change and shape evolution of the local abrupt region in the difference image to determine whether there is a material defect area caused by internal structural damage, and to mark the damage level.

[0035] The beneficial effects of this invention are:

[0036] This invention solves the problem of traditional methods' inability to distinguish between real cracks and complex maintenance textures by constructing a texture library containing multi-dimensional and multi-scale information, employing high-precision image acquisition, deep learning-assisted image processing, and a key texture expansion module. By introducing crack hierarchy recognition, multi-feature index damage judgment, and a dynamic analysis module combining difference images, surface corrosion, and internal corrosion detection, this invention achieves accurate, comprehensive, and dynamic monitoring and evaluation of railway track cracks, providing a more reliable technical guarantee for railway operation safety. The system proposed in this invention can effectively reduce false alarm rates and missed detection rates, significantly improve the efficiency and accuracy of track maintenance, extend track service life, and ultimately ensure the safe and stable operation of railway transportation. Attached Figure Description

[0037] Figure 1 This is a system block diagram of the present invention;

[0038] Figure 2 This is a flowchart illustrating the texture augmentation strategy in this invention;

[0039] Figure 3 This is a flowchart illustrating the crack level identification sub-strategy in this invention;

[0040] Figure 4 This is a schematic diagram of the dynamic analysis module in this invention;

[0041] Figure 5 This is a schematic diagram of a crack image in an embodiment of the present invention;

[0042] Figure 6 This is a schematic diagram of the damage texture expansion map obtained in an embodiment of the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] It should be noted that when a component is described as "fixed to" another component, it can be directly on the other component or may have a component in between. When a component is considered "connected to" another component, it can be directly connected to the other component or may have a component in between. When a component is considered "set on" another component, it can be directly set on the other component or may have a component in between. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0045] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0046] Please also see Figures 1 to 5 This embodiment provides a module that includes a texture library construction module, an image acquisition module, an image processing module, a texture expansion module, a damage assessment module, and a dynamic analysis module. These modules work together to form a closed-loop detection, evaluation, and prediction system.

[0047] The texture library construction module 1 primarily functions to build and continuously update a structured texture database before the system is put into operation. Initial data is collected from artificially damaged or manually labeled samples in the laboratory, including metal lines and their corresponding texture images. The samples cover various damage types, such as cracks, corrosion, and wear, and record the texture characteristics of the damage, including but not limited to roughness, color distribution, and pattern morphology. The database internally stores several sets of standard crack information and their corresponding texture images. Specifically, the design of the standard crack information is not limited to the macroscopic geometric morphology description of the crack; it delves further into the microstructure level, recording detailed microstructural characteristics of the crack, such as grain boundary corrosion, fatigue striations, or brittle fracture surface characteristics. Simultaneously, it includes the edge characteristics of the crack, such as edge sharpness, smoothness, serration degree, and the presence of micro-scraping phenomena. More importantly, the standard crack information also covers the interaction mode between the crack and the surrounding matrix material, such as whether the crack extends along grain boundaries, whether it passes through inclusions, and the plastic deformation characteristics caused by localized stress concentration in the crack tip region. Furthermore, under specific track materials and different maintenance backgrounds, such as periodic polishing and emergency repairs, the surface reflection characteristics exhibited by cracks are also recorded in detail, including their reflectivity under visible light and anisotropic reflection characteristics. These characteristics help distinguish the fresh metal exposed inside the crack from the surrounding oxide layer or contaminants. The texture images are highly diverse and representative, covering real crack image samples collected under different lighting conditions, such as direct sunlight, shadows, nighttime supplemental lighting, different wear levels, from newly laid tracks to severe wear, different polishing cycles, such as new polishing, intermediate wear, and near-repolishing, and different contamination conditions, such as oil stains, rust, and dust adhesion. At the same time, the database also contains corresponding crack-free background texture samples, as well as representative multi-level mechanical polishing texture samples formed at different maintenance stages. These samples are crucial for subsequent crack level identification.

[0048] Each data item in the texture library is accompanied by detailed metadata, such as, but not limited to: track type, steel grade, service life, last maintenance date, environmental parameters during image acquisition, actual crack depth (accurate to the micrometer level if obtainable through non-destructive testing methods such as ultrasound or eddy current), and crack evolution history, including crack initiation time and early growth rate. The texture library construction module uses a series of high-resolution imaging devices to perform refined image acquisition on representative track samples. These devices include high-resolution optical microscopes for microscopic morphology observation; laser confocal microscopes for acquiring three-dimensional morphology and depth information of cracks; and scanning electron microscopes combined with energy dispersive spectroscopy (EDS) for analyzing the microstructure, chemical composition, and fracture morphology of cracks. Simultaneously, metallographic analysis, fatigue testing, fracture toughness testing, and other materials science methods are used to gain in-depth understanding of the internal structure and development mechanism of cracks. All this refined data, after annotation and classification, is systematically organized and stored in the texture database, forming a knowledge graph of track damage and texture containing multi-dimensional and multi-scale information.

[0049] The core function of the image acquisition module is to acquire high-resolution image information of the railway track surface in real-time or near real-time. Specifically, the image acquisition module consists of three parts: a high-speed industrial camera array, a high-brightness and uniformly distributed LED lighting system, and a high-precision positioning and attitude sensor. The camera array is typically installed below a track inspection vehicle or dedicated inspection platform, and its imaging field of view covers the entire top surface and both sides of the track. The image acquisition module transmits the acquired raw image data to the back-end processing unit at a rate of several gigabytes per second via a high-speed data bus. The raw image data has at least 12 or 16 bits of pixel depth to retain rich grayscale information, avoid quantization loss, and provide high dynamic range raw data for subsequent image processing and analysis.

[0050] The image processing module is primarily responsible for receiving the track image information acquired by the image acquisition module and performing preprocessing, noise suppression, and preliminary crack candidate region identification. Specifically, after receiving the raw track image information, the image processing module first performs a series of refined preprocessing operations. This includes geometric correction, such as correcting lens distortion using pre-calibrated camera intrinsic parameters, and correcting image deformation and projection distortion caused by camera pose changes using perspective transformation and homography matrix correction. Simultaneously, radiometric correction is performed, such as eliminating the light field inhomogeneity of the illumination system using planar correction, and enhancing the local contrast of the image through adaptive histogram equalization, especially in areas of insufficient or overexposed illumination. After preprocessing, the image processing module 3 employs advanced image enhancement and denoising algorithms to effectively suppress random noise and background clutter on the track surface generated during image acquisition while preserving crack details. Subsequently, the image processing module utilizes a deep learning-based semantic segmentation network to classify each pixel in the image to identify potential crack regions. The semantic segmentation network can employ variants of the U-Net, DeepLabV3+, or HRNet architectures. The network has been pre-trained on a large-scale labeled dataset containing pixel-level annotations of different types of cracks under various complex background textures, such as transverse cracks, longitudinal cracks, diagonal cracks, star-shaped cracks, and hairline cracks, covering various environmental conditions and damage levels. During training, a cross-entropy loss function combined with the Dice coefficient is used as an evaluation metric to optimize the network's segmentation performance. The network outputs a pixel-level crack probability map or binary mask, where the value of each pixel represents its probability of being a crack, thus indicating the presence of crack information in the image. If the area of ​​the crack pixel region or its confidence level in the crack probability map exceeds a preset crack presence threshold, for example, if the crack pixel area exceeds 100 pixels, or the average crack probability exceeds 0.7, then the track image information is defined as a crack image, and it is sent to the texture augmentation module along with its initially identified crack region mask. The threshold for the presence of cracks is not a fixed value, but is determined through comprehensive statistical analysis and optimization of a large amount of historical false alarm rate, false negative rate data, and sensitivity requirements for micro-cracks.

[0051] As a preferred embodiment of the present invention, the texture augmentation module is described in reference to the appendix. Figure 2 As shown, its core function is to perform deep feature extraction on the crack image output by the image processing module, perform similarity matching with a texture library, and generate a damaged texture augmentation map based on the matching results. The texture augmentation module is configured with a damaged texture augmentation strategy, which includes the following specific steps:

[0052] First, the system acquires and obtains crack features based on the morphological features of the crack and its interaction with the surrounding texture. The morphological features encompass the geometric description of the crack in the crack image, specifically including: crack length, which is obtained by extracting the crack centerline and calculating its path length using a skeletonization algorithm; crack width, which is obtained by distance transformation or by sampling at equal intervals on the crack skeleton and performing orthogonal cross-sectional analysis to obtain the maximum and average crack width; crack orientation, which is determined by principal component analysis or Hough transform to identify the main and local orientation changes of the crack, and represented by the angle with the longitudinal axis of the crack path; and branch distribution, which is quantified using graph theory methods to determine the number of crack branches, branch lengths, and branch angles, for example, by statistically analyzing the average length of all branches and the average angle with the main trunk. Furthermore, the crack features further include micro-texture features of the crack region, such as local binary pattern histogram features, which are used to describe the contrast and structure of local textures; gray-level co-occurrence matrix features, such as multiple statistical quantities such as contrast, correlation, energy, entropy, and homogeneity, to comprehensively characterize the subtle texture changes of the crack region and its adjacent regions; and multi-scale texture descriptors based on wavelet transform, which are used to capture texture information in different frequency bands.

[0053] Furthermore, to capture the interaction pattern between cracks and the track background texture, this invention also extracts the gloss variation characteristics of crack edges. For example, by analyzing the specular reflection intensity and distribution at the crack edges, the openness and depth of the crack can be determined by combining the incident light angle; the degree of disturbance of the local texture by the crack can be quantified, for example, by comparing the difference in texture entropy or local structural similarity index between the inner and outer regions of the crack; and the surface reflection difference characteristics between cracked and non-cracked regions. The surface reflection difference characteristics are manifested in three aspects: the difference between specular and diffuse reflection components, anisotropic reflection characteristics, and the scattering effect at the crack edges. The surface of a normal track after mechanical polishing is relatively flat, and its reflected light contains a strong specular reflection component, forming a bright highlight area under a certain illumination angle. However, the interior of a crack is rough, irregular, and often accompanied by oxides or contaminants, and its reflected light is mainly diffuse reflection, with the specular reflection component significantly weakened or even disappearing. Therefore, by analyzing the specular reflection intensity of specific areas in the image, cracks can be effectively distinguished from the background. The polished texture of the track surface is directional, resulting in anisotropic reflective properties, meaning that brightness varies significantly depending on the viewing direction. In contrast, the reflective properties inside the crack are more isotropic, and its brightness is less affected by changes in the viewing direction. By analyzing the difference in the degree of anisotropy, the crack region can be further identified. At the interface between the crack and the substrate, due to the presence of microscopic steps, spalling, or stress concentration areas, the incident light is strongly scattered, which may appear as a darker or brighter edge line than the surrounding area in the image.

[0054] Next, the similarity between the crack feature and the standard crack information stored in the texture library is calculated. The similarity calculation employs a multimodal deep feature matching method. Specifically, the crack feature is first mapped to a high-dimensional embedding space using a pre-trained deep neural network. Simultaneously, the standard crack information in the texture library undergoes the same feature extraction and mapping process to generate corresponding feature vectors. Subsequently, in the high-dimensional embedding space, the similarity is quantified by calculating the cosine similarity, Euclidean distance, or Mahalanobis distance between the feature vectors. To improve the robustness of the matching, especially for variations in illumination, angle, and noise, the similarity calculation can also employ metric learning methods. For example, the feature extraction network can be trained using a Siamese network or triplet loss, making similar crack features closer together in the embedding space and dissimilar crack features farther apart, thereby enhancing the discriminative power of the features. If one or more standard crack information pieces have a similarity exceeding a preset similarity threshold with the current crack feature, then the standard crack information is defined as similar crack information, and its corresponding texture image is defined as a similar texture image. The similarity threshold was determined by cross-validating a large number of real and fake crack samples and optimizing it through ROC curve analysis, F1 score maximization, and other methods, aiming to balance the accuracy and recall of the matching.

[0055] Finally, the crack image is texture-enlarged based on the similar texture image to obtain the damage texture augmentation map. This texture augmentation is not a simple image overlay or pixel mixing, but rather achieved using advanced generative models such as generative adversarial networks or variational autoencoders. Specifically, the generative model takes the crack image and the similar texture image as input. The generative model has been pre-trained on massive amounts of labeled data, learning the embedding, growth, and evolution patterns of cracks under different background textures. The training data includes real crack images, artificially synthesized cracks, and various background textures selected from a texture library. The model can seamlessly blend the crack region in the crack image with the surrounding texture background based on contextual information provided by the similar texture image, such as the type of rail steel, polishing level, and wear degree, generating a highly realistic, visually coherent, and physically plausible damage texture augmentation map. The damage texture augmentation map not only clearly shows the geometric shape of the crack, but more importantly, it places the crack within its true texture context, clearly displaying the crack's embedding level. For example, it can accurately identify whether a crack is located in the first polished layer, penetrates the first polished layer into the second polished layer, or is located at the interface between layers. Simultaneously, the expanded image also shows the interaction pattern between the crack and the surrounding polished texture, such as whether the crack extends along the texture direction, cuts across the texture, causes local texture distortion or peeling, or whether local material deformation occurs due to crack propagation. During the expansion process, the generative model can also infer the possible propagation direction and potential expansion trend of the crack by analyzing its local features, and implicitly incorporate this information into the generated expanded image, for example, by generating slight texture deformation along the predicted propagation path. This damage texture expansion image provides richer, more accurate, and contextually semantic visual information for subsequent damage assessment and dynamic analysis, effectively reducing misjudgments and missed detections caused by complex background textures. The acquired crack image is shown below. Figure 5 The generated damage texture expansion map can be referred to as follows. Figure 6 .

[0056] In a preferred embodiment of the present invention, the texture expansion module is further configured with a damage level judgment strategy to address the challenge of multi-level composite texture structures on railway track surfaces, such as layered textures formed by grinding or repair at different times. The damage level judgment strategy includes: first, determining whether multiple damage layers exist based on the texture level of the crack and its embedding depth in the image. Specifically, the judgment process relies on a crack level identification sub-strategy. If the judgment result indicates the presence of multiple damage layers, for example, a crack spanning two or more polishing layers, then for each damage layer, based on the texture features of that level and the crack's behavior at that level, the texture expansion strategy is executed separately, i.e., texture expansion is performed on the crack's behavior at different damage levels, generating multi-level damage texture expansion sub-images; if the judgment result indicates the absence of multiple damage layers, such as a crack entirely within a single texture layer and not crossing level boundaries, then the texture expansion strategy is directly executed, generating a single damage texture expansion image.

[0057] In a preferred embodiment of the present invention, the damage level determination strategy is configured with a crack level identification sub-strategy, the specific execution steps of which are described in the appendix. Figure 3 As shown below:

[0058] The first step involves extracting multi-scale texture features from the crack region and its neighborhood. These multi-scale texture features include directional features, statistical texture parameters, and deep semantic texture features. Specifically, directional features are extracted at multiple scales and directions using Gabor filter banks or histograms of oriented gradients to capture the periodicity and directionality of the track polishing texture, such as the direction of the wear marks. Statistical texture parameters are calculated using methods such as gray-level co-occurrence matrix and gray-level run-length matrix to calculate various features such as contrast, entropy, energy, correlation, and long-run emphasis, quantifying the roughness, uniformity, and complexity of the texture. Deep semantic texture features are extracted from the crack image using a pre-trained deep convolutional neural network. This network has been trained on a dataset containing a large number of track texture images under different maintenance conditions, enabling it to learn the visual patterns and material properties specific to different polishing levels.

[0059] The second step involves the system combining multi-scale texture features and a classification model to perform pixel-level texture layer segmentation on the track image. Then, a pixel-level classification model, such as a fully convolutional network, a variant of the U-Net architecture, or Mask R-CNN, is used to classify each pixel in the track image, assigning it to a different texture layer. This classification model has been pre-trained on a large dataset of labeled multi-level polished texture regions. The model outputs a probability map of each pixel belonging to a specific texture layer, thus achieving fine-grained texture layer segmentation of the track surface, accurate to the micrometer level.

[0060] The third step involves establishing a texture hierarchy model. This model is a structured database that stores the characteristic parameters of each texture layer formed during different polishing and maintenance cycles, along with their corresponding typical depth information. For example, different types of rail steel, after being processed by specific polishing equipment and techniques, will form a surface layer with specific microstructures, surface roughness, optical properties, and average thickness. The model defines a set of feature vectors, depth ranges, and typical connection methods with other layers for each level. This model is built based on extensive laboratory data, field-collected data, and expert experience, and is regularly updated to adapt to new maintenance processes and materials.

[0061] The fourth step involves comparing the texture of the crack region with the texture hierarchy model. Based on the similarity between the features extracted from the crack region and the features of each texture layer in the model, and combined with the depth information from the pixel-level texture layer segmentation results obtained in the second step, the system determines the layer in which the crack resides. Specifically, the feature vector of the crack region is matched with the feature vectors of each texture layer in the texture hierarchy model, for example, by calculating cosine similarity or using a pre-trained support vector machine classifier. Simultaneously, combined with the pixel-level texture layer segmentation results obtained in the second step, statistical analysis is performed on the pixels in the crack region to determine the proportion of pixels from different texture layers within that region. If the crack is mainly located within a specific texture layer, for example, if more than 90% of the crack pixels are classified into the same layer, then the crack is determined to belong to that layer. If the crack spans the boundaries of multiple texture layers and exhibits significant features in different layer regions, for example, the crack head is located in the first layer and the middle extends into the second layer, then the crack is determined to be a cross-layer crack, and the layer sequence and depth information involved are recorded, for example, starting from 0 micrometers on the surface layer and extending to 100 micrometers in the second layer. For example, when the gloss variation at the crack edge, the local texture perturbation pattern, and the typical features of a specific layer in the texture hierarchy model are highly consistent, it can be determined that the crack is mainly embedded in that layer.

[0062] In a preferred embodiment of the present invention, the damage judgment module is configured to acquire the damaged texture expansion map output by the texture expansion module and calculate a damage index based on multiple feature indicators. If the damage index exceeds a preset damage threshold, a maintenance signal is generated. The multiple feature indicators specifically include:

[0063] The first indicator is the crack level identification result. This result comes directly from the damage level judgment strategy in the texture expansion module. For example, cracks located on the outermost layer may require immediate attention because they are directly exposed to the environment and loads; while cracks located in deeper layers, such as the base metal layer, with a depth exceeding 150 micrometers, even if the length is similar, may indicate deeper structural problems or material defects, and their damage index weight is usually higher because they may involve more severe fatigue damage.

[0064] The second metric includes the crack's length, width, angle, and complexity. The length is calculated by multiplying the number of pixels along the crack's centerline after skeletonization by the actual size of a single pixel, for example, in millimeters. The width is obtained by sampling at equal intervals along the crack's centerline and calculating the average distance perpendicular to the crack direction. The angle refers to the angle between the crack's principal axis and the longitudinal axis of the crack path. The complexity is quantified by calculating the Euler number of the crack skeleton, the number of branch nodes, and the curvature or fractal dimension of the crack path. These geometric parameters carry significant weight in the damage index calculation; generally, the longer, wider, and more complex the crack, the higher its damage index.

[0065] The third indicator is the degree of disturbance caused by the crack to the surrounding texture. This indicator quantifies the extent of damage or deformation caused by the crack to the normal polished texture structure of the track surface. For example, it is calculated by determining the difference in texture entropy between the cracked region and the adjacent non-cracked region, the decrease in the Local Structural Similarity Index (SSIM), or the degree of breakage, distortion, or misalignment of the polished texture lines at the crack edge. The greater the degree of disturbance, the more severe the damage to the track structure, thus assigning a higher damage index weight. For example, if the SSIM value decreases by more than 0.2, it indicates a significant disturbance.

[0066] In addition, the multi-feature indicators may also include the crack opening degree, the sharpness of the crack tip, and the color or spectral anomaly of the crack region.

[0067] The damage assessment module uses a pre-trained multi-feature fusion model, such as a regression model based on support vector regression, multilayer perceptron, or gradient boosting decision tree, to take the aforementioned feature indicators as input and output a continuous numerical value, which is the damage index. During training, the regression model utilizes a large number of track crack samples with known damage levels and corresponding multi-feature indicators. The damage threshold is set according to railway industry safety regulations and historical fault data; it can be a fixed value or an adaptive threshold dynamically adjusted based on track type, operating speed level, environmental conditions, etc. For example, for high-speed railways, a 0.8 mm surface crack may already reach the maintenance threshold; while for ordinary freight lines, a 1.5 mm surface crack triggers maintenance. If the calculated damage index exceeds the preset damage threshold, the damage assessment module immediately generates a maintenance signal. The maintenance signal carries detailed information on the crack location, damage index value, the contribution of the most important feature indicators, and the suggested maintenance priority, and is sent to the railway operation management system or dispatch center via an API interface for timely intervention.

[0068] As a preferred embodiment of the present invention, the dynamic analysis module is described in reference to the appendix. Figure 4 As shown, its core function is to acquire damage texture expansion maps at different time periods and determine the damage trend based on the analysis of these images. If the damage trend, such as the crack growth rate, exceeds a preset damage trend threshold, a maintenance signal is generated. The dynamic analysis module further includes a difference image generation submodule, a surface corrosion recognition submodule, and an internal damage detection submodule.

[0069] The difference image generation submodule is configured to acquire image data from damage texture expansion maps of the same track area acquired at different time periods. Specifically, the difference image generation submodule first performs high-precision image registration on the damage texture expansion maps of the same track area acquired at different time periods. The image registration employs a global homography transformation or a local non-rigid deformation model based on a combination of feature point matching and the RANSAC algorithm to ensure that the pixel-level alignment accuracy of the two images reaches the sub-pixel level, thereby eliminating image offset, rotation, and scale changes caused by vehicle vibration, slight changes in camera attitude, or deviations in the position of the track detection platform. After accurate registration, the difference image generation submodule uses pixel-level difference calculation methods, such as absolute difference, Gaussian difference, or variants of structural similarity index, to generate the difference image. The pixel intensity values ​​or structural similarity deviations in the difference image are used to characterize the degree of change in crack texture in the same area between different time periods, calculate the crack damage growth rate, such as the growth or regression of crack length, width, branch structure, and the evolution of texture features within the crack area. Subsequently, the difference image generation submodule analyzes the changes in the crack region within the difference image. This includes, for example, the area of ​​newly appearing crack regions, the increase in the area of ​​existing crack regions, the extension of crack length, the increase in crack width, or the increase in crack branches. The growth rate can be expressed as the area growth rate per unit time. Finally, the difference image generation submodule compares the current crack damage growth rate with a preset safe growth rate threshold. This safe growth rate threshold is preset based on the characteristics of the track material, service environment, traffic load level, and industry maintenance standards. For example, for a specific track steel, a warning is triggered if the surface crack length growth rate exceeds 0.1 mm / month. If the crack damage growth rate exceeds the safe growth rate threshold, the dynamic analysis module generates a maintenance signal containing the crack location, current growth rate, estimated time to reach critical size, and recommended maintenance measures. This maintenance signal emphasizes the dynamic evolution trend of the crack, providing crucial time window information for preventative maintenance.

[0070] The surface corrosion identification submodule is configured to detect abnormal color distribution and texture roughness distribution in the damage texture expansion map. Specifically, the surface corrosion identification submodule first analyzes the color features of each pixel in the damage texture expansion map. For example, it extracts the absolute values ​​and ratios of the red, green, and blue channels in the RGB color space, or converts to the HSV / Lab color space to extract hue, saturation, and brightness components to identify areas that match typical rust colors. Simultaneously, the module quantifies the texture roughness of the track surface by calculating the gray-level co-occurrence matrix features, local binary modes, and wavelet transform coefficients of local areas to identify surface unevenness and granular textures caused by corrosion. For example, rusted areas typically have high GLCM contrast but low energy. Further, the surface corrosion identification submodule compares the results with standard rust texture images pre-stored in a texture library. These standard rust texture images cover different types and severity of rust samples, including standard rust texture types such as pitting corrosion, flaky corrosion, uniform corrosion, and crevice corrosion. The system calculates the similarity between detected abnormal regions and standard rust samples using image similarity comparison algorithms, such as structural similarity index, feature vector matching, or deep learning-based similarity measurement networks. Furthermore, the system analyzes hue and saturation components in the color space and combines this with texture roughness indices to further improve the accuracy of rust identification. If a region matching rust characteristics and with a similarity exceeding a preset threshold is detected, the surface rust identification submodule marks this region as a surface rust area and outputs a rust location, type, severity, and rust warning signal, which is then sent to the railway operation management system. The rust warning signal can be categorized into different levels based on the area, depth, and growth trend of the rust.

[0071] The internal damage detection submodule is configured to acquire local texture depressions or abrupt changes in the damage texture expansion map, defining these as suspected chipping damage areas. This module aims to indirectly detect material damage inside the track, such as material loss due to fatigue or corrosion, through surface deformation features. Specifically, the internal damage detection submodule first uses high-pass filtering, Laplacian operators, or deep learning-based anomaly detection models to scan the damage texture expansion map and identify areas exhibiting local surface depressions, pits, or abnormal interruptions or breaks in texture arrangement. These areas typically show localized dimming, sharp edges, or significant changes in local curvature. For these suspected chipping damage areas, the module further extracts detailed edge features, brightness gradient changes, and anomalies in texture direction distribution. Furthermore, the internal damage detection submodule combines the pixel intensity fluctuations in the corresponding areas of the difference image generated by the difference image generation submodule to determine whether chipping-type defect features exist. If the difference image shows a consistent decrease in pixel intensity in the same area, for example, if the pixel value of the center point of a recessed area continues to show a negative value in successive difference images over a period of time, this may indicate a continuous loss of surface material over time, thus increasing suspicion of chipping damage.

[0072] In a preferred embodiment of the present invention, the internal damage detection submodule further includes an image morphology change analysis unit, used to extract the edge contours, texture density changes, and grayscale discontinuities of the chipped area by comparing damage texture expansion maps at different time points. Specifically, the image morphology change analysis unit uses advanced image segmentation algorithms to accurately delineate the edge contours of the chipped area in the damage texture expansion maps at multiple time points and tracks the evolution of these contours over time. By comparing the contours at different time points, the area growth rate, shape change trend, and depth evolution of the chipped area can be quantified. Simultaneously, the unit also analyzes the texture density changes within the chipped area, for example, by calculating the pixel density or texture unit number of local areas. Chipping leads to the loss of surface material, resulting in the loss or disruption of texture information, contrasting with the texture of the surrounding normal areas. Furthermore, the unit detects grayscale discontinuities between the chipped area and the surrounding normal areas, i.e., significant brightness or color jumps appearing at the chipped boundary, indicating a sudden interruption of the surface material. The image morphology change analysis unit further combines the area changes and shape evolution of local abrupt changes in the difference image to determine whether there are material loss areas caused by internal structural damage such as fatigue or corrosion. For example, if the area of ​​a local depression in the difference image continues to increase, and its shape evolution is consistent with the expansion pattern of the detached area, it strongly supports the judgment that internal damage caused material detachment. Finally, the unit marks the damage level based on the size, depth, growth rate of the detached area, and its potential impact on the surrounding structure. The damage level marking serves as a maintenance signal, indicating the severity of internal structural damage, and is sent to the railway operation management system to trigger corresponding non-destructive detection or repair measures.

[0073] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A dynamic analysis system for cracks in railway tracks, characterized in that, include A texture library construction module, wherein the texture library construction module pre-stores several standard crack information and corresponding texture images into the texture library; The image acquisition module is used to collect orbital image information; The image processing module acquires the track image information and determines whether there is crack information. If the crack information exists, the track image information is defined as a crack image and sent to the texture expansion module. The texture expansion module extracts crack features from the crack image and matches standard texture information and corresponding texture images in the texture library to obtain a damaged texture expansion map after expanding the texture. The damage determination module acquires the damage texture expansion map and determines the degree of damage. If the degree of damage exceeds a preset damage threshold, a maintenance signal is generated. The dynamic analysis module acquires damage texture expansion maps at different time periods and determines the damage trend. If the damage trend exceeds a preset damage trend threshold, a maintenance signal is generated.

2. The dynamic crack analysis system for railway tracks according to claim 1, characterized in that, The texture augmentation module is configured with a texture augmentation strategy, including: Crack features are obtained and derived based on the morphological features of the crack, including crack length, crack width, crack direction, and branch distribution in the crack image; Calculate the similarity between the crack feature and each of the standard crack information in the texture library. If there is standard crack information with a similarity exceeding a preset similarity threshold, define the standard crack information as similar crack information and define the corresponding texture image as similar texture image. The crack image is augmented with texture based on the similar texture image to obtain the damage texture augmentation map.

3. The dynamic crack analysis system for railway tracks according to claim 2, characterized in that, The texture augmentation module is configured with a damage level judgment strategy, including determining whether there are multiple damage layers based on the texture level of the crack and the embedding depth. If multiple damage layers exist, the texture augmentation strategy is executed for each damage layer separately after identifying multiple damage layers. If there are no multiple damage layers, the texture augmentation strategy is executed directly.

4. The dynamic crack analysis system for railway tracks according to claim 3, characterized in that, The damage level determination strategy is configured with a crack level identification sub-strategy, including: Extract multi-scale texture features of the crack region and its neighborhood, including directionality, statistical texture parameters and deep semantic texture features; By combining multi-scale texture features and classification models, pixel-level texture layer segmentation is performed on the orbital image; A texture hierarchy model is established, including the depth and feature parameters of each texture layer. The texture layer represents the polishing layer or composite texture structure left by track maintenance at different times. The texture of the cracked area is compared with the model, and the layer where the crack is located is determined based on feature similarity and depth information.

5. The dynamic crack analysis system for railway tracks according to claim 4, characterized in that, The texture augmentation strategy includes: Extract the geometric morphological features of the crack and its interaction pattern with the texture layer. The geometric morphological features include edge gloss, local distortion and surface reflection features. The interaction pattern represents the interpenetration relationship between the crack and the texture layer. A multimodal deep feature matching method is used to calculate the similarity between cracks and standard samples; Damage texture augmentation is performed based on similar texture images using a generative model, outputting an augmented image that includes the crack embedding layer and propagation direction.

6. The dynamic crack analysis system for railway tracks according to claim 3, characterized in that, The damage assessment module calculates a damage index based on multiple feature indicators. If the damage index exceeds a preset threshold, a maintenance signal is output. The multiple feature indicators include: Results of crack location layer identification; The length, width, angle, and complexity of the crack; The degree to which a crack disturbs the surrounding texture.

7. The dynamic crack analysis system for railway tracks according to claim 1, characterized in that, The dynamic analysis module further includes a difference image generation submodule, which is used to obtain image data of the corresponding track area from the damage texture expansion map at different time periods, and generate a difference image by using image registration and difference calculation methods. The difference image is used to characterize the degree of change of crack texture in the same area, thereby calculating the growth rate of crack damage and determining whether the current crack propagation trend exceeds the preset safe growth rate threshold.

8. The dynamic crack analysis system for railway tracks according to claim 1, characterized in that, The dynamic analysis module also includes a surface corrosion recognition submodule, which is used to detect abnormal color distribution and texture roughness distribution in the damage texture expansion map; and combined with the standard rust spot texture image, it determines whether there is a rust area on the track surface through image similarity comparison and color-texture feature analysis. If a rust spot area is detected, the area is marked as a surface corrosion area and a rust warning signal is output.

9. The dynamic crack analysis system for railway tracks according to claim 7, characterized in that, The dynamic analysis module includes an internal damage detection submodule. The internal damage detection submodule obtains local texture depressions or abrupt changes in the damage texture expansion map to define suspected chipping damage areas, extracts the edge features, brightness gradient changes, and texture direction distribution of the suspected chipping damage areas, and determines whether there are chipping deformation features by combining the pixel intensity fluctuations of the corresponding areas in the difference image.

10. The dynamic crack analysis system for railway tracks according to claim 9, characterized in that, The internal damage detection submodule also includes an image morphology change analysis unit, which is used to extract the edge contour, texture density change and grayscale discontinuity features of the suspected chip damage area by comparing the damage texture expansion maps at different time points. Combined with the area change and shape evolution of the local abrupt region in the difference image, it determines whether there is a material defect area caused by internal structural damage and marks the damage level.

Images