Nondestructive testing method for axle fracture defects

Abstract

The invention relates to the technical field of defect detection, and discloses an axle fracture defect nondestructive testing method which comprises the following steps: acquiring multiband image data of the surface of an axle under a full-spectrum illumination condition; according to the multiband image data after spectral feature separation, determining the surface phase consistency of the axle to obtain a structural saliency map of the axle; while maintaining the edge integrity in the structural saliency map, performing region segmentation on the structural saliency map to obtain segmented regions of the axle; performing multi-feature voting on the segmented areas to obtain suspected defect areas of the axle; performing credibility evaluation on the suspected defect area according to the image features of the suspected defect area and the acoustic emission features of the axle to obtain a target defect area of the axle; performing depth feature analysis on the target defect area to obtain defect quantitative description and confidence evaluation of the axle; according to the invention, the nondestructive detection efficiency of the axle fracture defect can be improved.

Description

Technical Field

[0001] This invention relates to the field of defect detection technology, and in particular to a non-destructive testing method for vehicle axle fracture defects. Background Technology

[0002] Existing technologies have significant shortcomings in the image acquisition and preprocessing stages of non-destructive testing of vehicle axle fracture defects. They fail to acquire multi-band image data under full-spectrum illumination conditions, relying solely on single-band or ordinary illumination images for detection. This fails to fully utilize the effective information after spectral feature separation, resulting in low differentiation between surface defects and the background. Furthermore, they do not generate structural saliency maps through multi-directional filtering and phase consistency analysis, employing only simple edge detection or grayscale thresholding. This makes it difficult to accurately highlight the structural features of the defect area and is susceptible to noise interference. Consequently, subsequent region segmentation lacks a reliable image foundation, and the segmentation results are prone to edge breakage and region confusion, failing to provide accurate segmentation regions for defect screening.

[0003] Existing technologies have significant shortcomings in defect identification and assessment. They fail to perform multi-feature voting to screen suspected defect areas within segmented regions, relying solely on a single feature to determine defect probability. This results in a large number of non-defect areas being mixed into the suspected defect area, leading to poor screening accuracy. Furthermore, they do not combine image features and acoustic emission features for complementary confidence assessment, relying only on a single type of feature to determine defect authenticity, making it difficult to eliminate false defect signals and resulting in insufficient reliability in target defect area identification. Finally, they lack cross-scale deep feature analysis of target defect areas, only providing simple defect morphology descriptions, failing to generate accurate quantitative results and confidence assessments. This leads to detection results lacking scientific rigor and practicality, making it difficult to meet the needs of accurate detection and quality assessment of vehicle and axle fracture defects. Summary of the Invention

[0004] This invention provides a non-destructive testing method for vehicle axle fracture defects to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a non-destructive testing method for fracture defects in vehicle axles, comprising:

[0006] S1. Acquire multi-band image data of the axle surface under full-spectrum illumination conditions;

[0007] S2. Based on the multi-band image data separated by spectral features, determine the surface phase consistency of the axle to obtain the structural saliency map of the axle;

[0008] S3. While maintaining the integrity of the edges in the structural saliency map, the structural saliency map is segmented to obtain the segmented regions of the vehicle bridge;

[0009] S4. Perform multi-feature voting on the segmented region to obtain the suspected defect region of the axle;

[0010] S5. Based on the image features of the suspected defect area and the acoustic emission features of the axle, the credibility of the suspected defect area is evaluated to obtain the target defect area of ​​the axle.

[0011] S6. Perform in-depth feature analysis on the target defect area to obtain a quantitative description and confidence assessment of the defects of the axle.

[0012] In a preferred embodiment, acquiring multi-band image data of the axle surface under full-spectrum illumination conditions includes:

[0013] Configure full-spectrum illumination parameters to obtain uniform illumination conditions on the axle surface;

[0014] Under the uniform illumination conditions, the original reflected image of the axle surface is acquired to obtain a set of original images of the axle surface;

[0015] The original image set is integrated and standardized to obtain multi-band image data of the vehicle bridge surface.

[0016] In a preferred embodiment, determining the surface phase consistency of the axle based on the multi-band image data separated by spectral features to obtain a structural saliency map of the axle includes:

[0017] Multi-directional filtering is performed on the band image data after spectral feature separation to obtain filtered image data;

[0018] Based on the filtered image data, the phase consistency feature map of the axle is extracted;

[0019] The phase consistency map is normalized and enhanced to obtain the structural saliency map of the axle.

[0020] In a preferred embodiment, the step of segmenting the structural saliency map to obtain the segmented regions of the vehicle axle while maintaining the edge integrity of the structural saliency map includes:

[0021] Edge-guided optimization processing is performed on the structural saliency map to obtain the structural feature map of the axle;

[0022] The structural feature map is subjected to region growth based on predefined dynamic constraints to obtain the initial segmentation region of the vehicle axle;

[0023] Multi-level confidence verification is performed on the initial segmented region to obtain the reliable segmented region of the vehicle axle;

[0024] The topology of the reliable segmentation region is optimized to obtain the segmentation region of the vehicle axle.

[0025] In a preferred embodiment, the step of performing region growing on the structural feature map according to predefined dynamic constraints to obtain the initial segmentation region of the vehicle axle includes:

[0026] In the structural feature map, multiple seed points are selected based on the principle of local maxima;

[0027] Based on the pixel intensity similarity and spatial continuity of the structural feature map, the dynamic growth criterion of the vehicle axle is defined.

[0028] Under the dynamic growth criterion, region growth is performed on the structural feature map to obtain the initial segmentation region of the vehicle bridge.

[0029] In a preferred embodiment, the step of performing multi-level confidence verification on the initial segmented region to obtain the reliable segmented region of the axle includes:

[0030] The texture consistency and grayscale distribution features in the segmented region are evaluated to obtain the preliminary verification results of the initial segmented region;

[0031] Based on the boundary curvature continuity and gradient consistency of the edge pixel sequences in the initial segmentation region, an intermediate verification result for the initial segmentation region is generated.

[0032] Based on the spatial relationship between adjacent regions in the initial segmentation region and the defect shape conformity of the initial segmentation region, the advanced verification result of the initial segmentation region is determined;

[0033] Based on the fusion of the primary verification results, the intermediate verification results, and the advanced verification results, a reliable segmentation region of the vehicle axle in the initial segmentation region is selected.

[0034] In a preferred embodiment, the step of performing multi-feature voting on the segmented region to obtain the suspected defect region of the axle includes:

[0035] Based on the multi-dimensional features of the segmented region, a regional feature description of the segmented region is generated;

[0036] Based on the different feature dimensions in the region feature description, multidimensional anomaly voting is performed on the segmented region to obtain the initial voting result of the segmented region;

[0037] Based on the consistency of the initial voting results, the voting results of different feature dimensions in the initial voting results are weighted and fused to obtain the comprehensive voting result of the segmented region;

[0038] The region screening criteria for the axle are generated based on the comprehensive voting results, and the segmented regions are screened according to the region screening criteria to obtain the suspected defect regions of the axle.

[0039] In a preferred embodiment, the step of assessing the credibility of the suspected defect region based on the image features of the suspected defect region and the acoustic emission features of the axle to obtain the target defect region of the axle includes:

[0040] Based on the texture complexity and shape irregularity measures of the image features in the suspected defect area, a defect correlation analysis is performed on the axle to obtain the credibility of the axle's image features;

[0041] Defect indicative analysis is performed on the acoustic emission characteristics of the axle to obtain the reliability of the acoustic emission characteristics of the axle;

[0042] A complementary evaluation of the image feature confidence level and the acoustic emission feature confidence level is established to obtain the feature complementarity index of the vehicle axle;

[0043] Based on the image feature credibility, the acoustic emission feature credibility, and the feature complementarity index, the comprehensive credibility score of the vehicle axle is calculated.

[0044] The suspected defect areas are prioritized based on the comprehensive credibility score to obtain the target defect areas of the axle.

[0045] In a preferred embodiment, the formula for calculating the comprehensive credibility score is as follows:

[0046] ;

[0047] In the formula, The overall credibility score is given. The credibility of the image features, The credibility of the acoustic emission characteristics is given. This is the mapping value of the feature complementarity index.

[0048] In a preferred embodiment, the step of performing deep feature analysis on the target defect region to obtain a quantitative description and confidence assessment of the axle defect includes:

[0049] The local and global features of the target defect region are fused across scales to obtain the multi-scale features of the vehicle axle.

[0050] Multi-dimensional feature fusion is performed on the multi-scale features to obtain the fused features of the vehicle axle;

[0051] Based on the fusion features, the set features of curves in the target defect region are determined, and a quantitative description of the vehicle axle defect is generated according to the texture complexity and intensity distribution of the target defect region.

[0052] Based on the fusion features, the defects of the axle are scored for internal and external compliance to obtain the confidence level assessment of the axle.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. This invention lays a high-quality data foundation for vehicle axle fracture defect detection through refined image acquisition and preprocessing. It configures full-spectrum illumination parameters to obtain uniform illumination conditions, acquires multi-band images of the vehicle axle surface, and performs standardization processing to ensure the integrity and consistency of the image data. After spectral separation, the images are subjected to multi-directional filtering, and phase consistency features are extracted and enhanced to generate a saliency map highlighting the defect structure. While maintaining edge integrity, accurate segmentation regions are obtained through edge-guided optimization, dynamic region growing, and multi-level confidence verification, providing a reliable basis for subsequent defect identification.

[0055] 2. This invention significantly improves the accuracy and completeness of axle fracture defect detection by employing multi-dimensional defect screening and comprehensive evaluation. It performs multi-feature voting on segmented regions, integrating anomaly assessment results from different dimensions to screen suspected defect areas, improving the accuracy of initial defect identification. Combining the image features of suspected defect areas with the acoustic emission features of the axle, a comprehensive confidence score is calculated through complementary evaluation to accurately locate the target defect area. Cross-scale feature fusion and in-depth analysis are performed on the target defect area to generate a quantitative description of the defect and an internal and external conformity confidence assessment, providing a comprehensive and scientific basis for axle defect judgment and quality control. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating a non-destructive testing method for axle fracture defects according to an embodiment of the present invention.

[0057] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0058] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0059] This application provides a non-destructive testing method for axle fracture defects. The execution subject of this non-destructive testing method for axle fracture defects includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the non-destructive testing method for axle fracture defects can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0060] Reference Figure 1 The diagram shown is a flowchart illustrating a non-destructive testing method for axle fracture defects according to an embodiment of the present invention. In this embodiment, the non-destructive testing method for axle fracture defects includes:

[0061] S1. Acquire multi-band image data of the axle surface under full-spectrum illumination conditions;

[0062] In this embodiment of the invention, acquiring multi-band image data of the axle surface under full-spectrum illumination conditions includes:

[0063] Configure full-spectrum illumination parameters to obtain uniform illumination conditions on the axle surface;

[0064] Under the uniform illumination conditions, the original reflected image of the axle surface is acquired to obtain a set of original images of the axle surface;

[0065] The original image set is integrated and standardized to obtain multi-band image data of the vehicle bridge surface.

[0066] Based on the wavelength range of the full-spectrum illumination coverage and the reflective properties of the axle surface material, the placement and angle of the lighting equipment were determined to ensure that the light could uniformly cover all areas of the axle surface, including corners and recessed areas. The brightness and light intensity distribution of the lighting equipment were adjusted to avoid localized over-brightness causing glare or localized under-darkness creating shadows. Through multiple adjustments, the light intensity at all points on the axle surface was kept consistent, ultimately achieving uniform illumination conditions for the axle surface.

[0067] Under established uniform illumination, a multi-band image acquisition device is used to capture raw reflected images of different areas of the axle surface in a pre-set shooting order. During the acquisition process, the device parameters and shooting distance are kept constant to ensure uniform imaging specifications for each image. All acquired images are then categorized and organized by region to form a raw image set of the axle surface containing complete information about the axle surface.

[0068] Each image in the original image set of the axle surface undergoes format standardization processing, converting raw images of different formats into a preset standard image format. The image resolution is adjusted to ensure consistent pixel size across all images, eliminating subsequent data processing deviations caused by resolution differences. Pixel values ​​are normalized to unify the pixel brightness range of each image to the same interval, while interference pixels caused by equipment noise are removed. Through this data integration and standardization process, multi-band image data of the axle surface is finally obtained.

[0069] The beneficial effects are that configuring full-spectrum illumination parameters and forming uniform illumination conditions can ensure that the light fully and evenly covers all areas of the vehicle axle surface, including corners, depressions and other easily overlooked areas, avoiding local over-brightness that causes reflections and glare or local under-darkness that forms shadows, allowing the defect features of the vehicle axle surface to be clearly separated from the background information, and providing a stable lighting foundation for subsequent image acquisition.

[0070] Acquiring raw reflection images under uniform illumination and forming a raw image set can completely record the spectral reflection information of different areas on the vehicle axle surface, ensuring the integrity and authenticity of the image data. The image of each area can accurately reflect the surface condition of that part, avoiding the loss or distortion of defect features due to uneven illumination, and providing high-quality raw materials for the generation of multi-band image data.

[0071] Data integration and standardization of the original image set unifies the image format, resolution, and pixel brightness range, eliminating interference caused by minor differences in acquisition conditions between different images, and ensuring consistency and comparability of multi-band image data. Simultaneously, removing device noise pixels further improves image quality, making subsequent spectral feature separation and defect detection based on multi-band image data more accurate and reliable, thus laying a high-quality data foundation for the entire non-destructive testing process for vehicle and axle fracture defects.

[0072] S2. Based on the multi-band image data separated by spectral features, determine the surface phase consistency of the axle to obtain the structural saliency map of the axle;

[0073] In this embodiment of the invention, determining the surface phase consistency of the axle based on the multi-band image data separated by spectral features to obtain the structural saliency map of the axle includes:

[0074] Multi-directional filtering is performed on the band image data after spectral feature separation to obtain filtered image data;

[0075] Based on the filtered image data, the phase consistency feature map of the axle is extracted;

[0076] The phase consistency map is normalized and enhanced to obtain the structural saliency map of the axle.

[0077] A filter template is selected to match the spectral feature separation of the band image data. This template can specifically retain image information related to the surface structure of the vehicle axle while filtering out irrelevant noise. Multiple fixed filtering directions, including horizontal, vertical, and diagonal directions, are set to ensure comprehensive coverage of any possible structural textures on the vehicle axle surface. The filter template is then applied pixel-by-pixel along each set direction to process the band image data. By fusing the template with the pixel information of the corresponding region in the image, the interference from random noise and uneven illumination in the image is weakened, while the structural edges and texture details of the vehicle axle surface are enhanced, ultimately yielding the filtered image data.

[0078] Based on filtered image data, phase analysis is used to extract phase information from the axle surface. This method can accurately capture the phase changes of pixels in different regions of the image, and these changes are directly related to the structural undulations and texture features of the axle surface. The phase information of each pixel is quantized to determine its phase consistency in the image. The higher the phase consistency, the more regular the axle surface structure corresponding to that region. The phase consistency of all pixels is arranged according to their image pixel positions to form a phase consistency feature map of the axle that can intuitively reflect the regularity of the axle surface structure distribution.

[0079] First, the maximum and minimum values ​​of phase consistency in the phase consistency feature map are determined. Using these two values ​​as a benchmark, the phase consistency of all pixels in the image is mapped to a fixed numerical range in the same proportion, completing the normalization process and unifying the brightness range of the image for easier subsequent analysis and comparison. Then, an enhancement algorithm is used to process the normalized phase consistency feature map. By increasing the pixel difference between the structural region and the background region, the structural details of the axle surface are made more prominent, while suppressing residual weak noise. After these processes, a structural saliency map of the axle that clearly presents the key structural information of the axle surface is obtained.

[0080] The beneficial effects are that multi-directional filtering of the band image data after spectral feature separation can selectively preserve image information related to the surface structure of the vehicle axle, while effectively filtering out irrelevant noise. By sliding the filter template along multiple fixed directions such as horizontal, vertical, and diagonal, pixel information of corresponding regions is fused pixel by pixel. This can weaken the interference caused by random noise and uneven illumination in the image, enhance the structural edges and texture details of the vehicle axle surface, and the resulting filtered image data can more clearly present the structural features related to defects, providing a high-quality image foundation for subsequent phase consistency feature extraction.

[0081] Phase consistency feature maps of the axle are extracted from filtered image data, accurately capturing phase changes in pixels across different regions of the image. These changes are directly related to the structural undulations and texture features of the axle surface. By quantifying the phase consistency of each pixel, the regularity distribution of the axle surface structure can be intuitively reflected. Regions with high phase consistency correspond to structurally regular areas, while regions with low consistency may have defects. This feature map effectively highlights areas of structural abnormality, providing core feature basis for the subsequent generation of structural saliency maps.

[0082] The phase consistency map is normalized and enhanced. Normalization maps the phase consistency of all pixels in the image to a fixed numerical range in the same proportion, unifying the brightness range of the image and facilitating subsequent comparative analysis of features in different regions. Enhancement further improves the pixel difference between structural regions and background regions, making the structural details of the axle surface more prominent, while suppressing residual weak noise. The structural saliency map of the axle obtained after these two steps clearly presents the key structural information of the axle surface, accurately highlights potential defect areas, and provides reliable image support for subsequent region segmentation, avoiding problems such as edge breakage and region confusion in the segmentation results due to unclear image features.

[0083] S3. While maintaining the integrity of the edges in the structural saliency map, the structural saliency map is segmented to obtain the segmented regions of the vehicle bridge;

[0084] In this embodiment of the invention, the step of segmenting the structural saliency map to obtain the segmented region of the vehicle axle while maintaining the edge integrity of the structural saliency map includes:

[0085] Edge-guided optimization processing is performed on the structural saliency map to obtain the structural feature map of the axle;

[0086] The structural feature map is subjected to region growth based on predefined dynamic constraints to obtain the initial segmentation region of the vehicle axle;

[0087] Multi-level confidence verification is performed on the initial segmented region to obtain the reliable segmented region of the vehicle axle;

[0088] The topology of the reliable segmentation region is optimized to obtain the segmentation region of the vehicle axle.

[0089] The step of performing region growing on the structural feature map according to predefined dynamic constraints to obtain the initial segmentation region of the vehicle axle includes:

[0090] In the structural feature map, multiple seed points are selected based on the principle of local maxima;

[0091] Based on the pixel intensity similarity and spatial continuity of the structural feature map, the dynamic growth criterion of the vehicle axle is defined.

[0092] Under the dynamic growth criterion, region growth is performed on the structural feature map to obtain the initial segmentation region of the vehicle bridge.

[0093] The step of performing multi-level confidence verification on the initial segmented region to obtain the reliable segmented region of the axle includes:

[0094] The texture consistency and grayscale distribution features in the segmented region are evaluated to obtain the preliminary verification results of the initial segmented region;

[0095] Based on the boundary curvature continuity and gradient consistency of the edge pixel sequences in the initial segmentation region, an intermediate verification result for the initial segmentation region is generated.

[0096] Based on the spatial relationship between adjacent regions in the initial segmentation region and the defect shape conformity of the initial segmentation region, the advanced verification result of the initial segmentation region is determined;

[0097] Based on the fusion of the primary verification results, the intermediate verification results, and the advanced verification results, a reliable segmentation region of the vehicle axle in the initial segmentation region is selected.

[0098] Based on the clearly defined edge information in the structural saliency map, an edge-guided model is constructed, using the position and grayscale features of edge pixels as core references. This model optimizes the structural saliency map pixel-by-pixel, enhancing the difference between edge pixels and surrounding background pixels, while repairing edge breaks caused by noise or lighting effects, ensuring the integrity and clarity of critical edges on the axle surface. During the optimization process, the original effective structural information in the structural saliency map is strictly preserved without introducing any additional false features, ultimately yielding the structural feature map of the axle.

[0099] Based on the material properties and structural distribution patterns of the axle surface, predefined dynamic constraints are established. These constraints include core elements such as pixel grayscale similarity, texture consistency, and spatial relationships between adjacent regions. Several representative seed pixels are selected from the structural feature map; these seed pixels must be accurately located at the core positions of different structural regions of the axle. Starting from the seed pixels, and following the predefined dynamic constraints, surrounding pixels that meet the constraints are gradually absorbed, allowing each seed pixel to continuously expand outwards to form an independent region. All the expanded regions together constitute the initial segmentation region of the axle.

[0100] A multi-level confidence verification system was constructed, with each level corresponding to a different verification dimension, including the integrity of the region boundary, the consistency of pixels within the region, and the degree of matching between the region and the actual structure of the vehicle axle. The initial segmented region was sequentially substituted into each level of verification. Lower-level verification first screened out regions with clear boundaries and uniform internal pixels, and then higher-level verification further confirmed the degree of fit between these regions and the actual structure of the vehicle axle. Regions with blurred boundaries, mixed internal pixels, or inconsistencies with the actual structure were marked and removed during the verification process. Regions that passed all levels of verification were retained, resulting in the reliable segmented regions of the vehicle axle.

[0101] The topological structure of the reliably segmented regions is analyzed, with a focus on examining the connectivity, overlap, and boundary connections between regions. For overlapping regions, they are rationally split based on pixel features and structural correlations, ensuring clear boundaries for each region and preventing interference. For regions with unnatural boundary connections or gaps, smooth adjustments are made based on the actual structural morphology of the axle, ensuring the transitions between regions conform to the structural logic of the axle surface. Simultaneously, the shape and contour of the regions are optimized, removing irregular protrusions or depressions caused by segmentation, so that the final segmented regions accurately fit the actual structural distribution of the axle, resulting in the segmented regions of the axle.

[0102] The structural feature map of the axle is scanned pixel by pixel to analyze the grayscale distribution of each pixel and its surrounding neighborhood. The local maximum principle requires that the grayscale value of the selected seed point be higher than that of all its neighboring pixels, and these pixels must correspond to the core locations of different structural regions on the axle surface, ensuring that each seed point represents an independent structural region. Through systematic scanning and grayscale value comparison, all pixels meeting the conditions are selected as seed points. These seed points are evenly distributed in the structural feature map, providing stable starting positions for subsequent region growth.

[0103] A thorough analysis of the grayscale distribution patterns and spatial characteristics of the axle surface structure in the structural feature map was conducted to define dynamic growth criteria. The grayscale intensity similarity criterion stipulates that the difference in grayscale values ​​between the pixel to be grown and the edge pixels of the current region must be controlled within a reasonable range to ensure uniform grayscale within the grown region, conforming to the characteristics of the same structural region. The spatial continuity criterion requires that the pixel to be grown must be directly adjacent to the current region in space, with no obvious structural breaks or abrupt changes between adjacent pixels, ensuring that the region growth process conforms to the actual structural morphology of the axle surface. These two criteria work together to constrain the direction and range of region growth, avoiding region confusion or structural distortion during the growth process.

[0104] Using multiple selected seed points as the starting points for the growth of their respective regions, the region growth process is initiated according to defined dynamic growth criteria. For the initial region corresponding to each seed point, its adjacent pixels are evaluated one by one. If adjacent pixels simultaneously satisfy the criteria of grayscale intensity similarity and spatial continuity, the pixel is included in the current region, causing the region boundary to expand outward. This evaluation and inclusion process is repeated until no pixels in the current region satisfy the criteria, at which point the region growth stops. After all regions corresponding to all seed points have completed growth, the structural feature map is divided into multiple non-overlapping, clearly defined independent regions, which together constitute the initial segmentation region of the vehicle bridge.

[0105] Texture features are extracted from each independent region within the initial segmentation area. The direction, density, and repetition patterns of the texture within the region are analyzed to determine whether these texture features remain consistent throughout the region, without significant abrupt changes or heterogeneous textures. Simultaneously, the grayscale values ​​of all pixels within the region are statistically analyzed to observe the distribution pattern of the grayscale values, confirming whether the grayscale values ​​are concentrated within a specific range without significant fluctuations or abnormal dispersion. Through a comprehensive evaluation of texture consistency and grayscale distribution characteristics, the internal uniformity and rationality of each initial segmentation region are rated, generating preliminary validation results for the initial segmentation region.

[0106] The edge pixel sequence of each initial segmentation region is extracted, and the boundary curvature between adjacent edge pixels is calculated sequentially. The smoothness of the curvature transition, without abrupt turns or discontinuities, is observed to verify the continuity of the boundary curvature. Simultaneously, the gradient value of each pixel in the edge pixel sequence is calculated, and the consistency of the gradient direction and magnitude is analyzed to ensure a uniform grayscale trend in the edge region, without reversed or chaotic gradient distributions. Combining the evaluation results of boundary curvature continuity and gradient consistency, the edge integrity and rationality of the initial segmentation region are judged, yielding an intermediate-level verification result for the initial segmentation region.

[0107] Analyze the spatial relationship between each initial segmented region and its adjacent regions to confirm whether there is a reasonable connection sequence between the regions, and to ensure there are no overlaps, intersections, or gaps that do not conform to the actual structural logic of the vehicle axle. Simultaneously, compare the typical shape characteristics of common vehicle axle defects to determine whether the outline of the initial segmented region matches the shape of known defects, eliminating irregularly shaped or false regions without corresponding actual defects. By comprehensively considering the rationality of the spatial relationship between adjacent regions and the conformity of defect shapes, the overall structural matching degree of the initial segmented regions is evaluated, forming a high-level verification result for the initial segmented regions.

[0108] We define the weighting of primary, intermediate, and advanced verification results, based on the impact of each verification level on the credibility of the segmented region, ensuring that the fusion result comprehensively reflects the true situation of the region. We then comprehensively calculate the three levels of verification results for each initial segmented region according to the defined weights, resulting in a single fusion score. We set a threshold for the fusion score and select initial segmented regions with scores higher than the threshold. These regions meet the credibility requirements for vehicle-bridge segmentation in terms of internal features, edge morphology, and overall structure, ultimately forming the credible segmented regions for the vehicle-bridge.

[0109] The beneficial effects are that edge-guided optimization processing of the structural saliency map yields the structural feature map of the vehicle axle. It can use the clearly defined edge information in the structural saliency map as the core reference, and strengthen the difference between edge pixels and background pixels through pixel-by-pixel optimization. At the same time, it can repair edge breaks caused by noise or lighting, strictly preserve the original effective structural information and not introduce false features, so that the key edges of the vehicle axle surface are complete and clear, providing accurate structural basis for subsequent region segmentation and avoiding the impact of missing or distorted edge information on segmentation accuracy.

[0110] The initial segmentation regions of the vehicle axle are obtained by performing region growth on the structural feature map based on predefined dynamic constraints. These dynamic constraints combine the surface material properties and structural distribution patterns of the vehicle axle, regulating the region growth process from dimensions such as grayscale similarity, texture consistency, and spatial relationships. Starting with seed pixels located at the core of different structural regions, adjacent pixels that meet the constraints are gradually absorbed, ensuring that the generated initial segmentation regions closely match the actual structural shape of the vehicle axle. This guarantees clear boundaries and uniform internal features for each region, laying the foundation for subsequent verification and selection.

[0111] Multi-level confidence verification is performed on the initial segmented region to obtain a reliable segmented region for the axle. The multi-level verification system can comprehensively evaluate the initial segmented region from different dimensions such as internal features, edge morphology, and overall structural matching degree. First, regions with uniform internal pixels and clear boundaries are screened through low-level verification. Then, high-level verification confirms the fit between the region and the actual structure of the axle, eliminating regions with blurred boundaries, mixed internal structures, or those that do not match the actual structure. This significantly improves the confidence of the segmented region and avoids invalid regions interfering with subsequent defect detection.

[0112] Topological optimization of the reliable segmented regions yields the segmented regions of the vehicle axle. Adjustments can be made to the connectivity, overlap, and boundary connection status between these reliable segmented regions. Overlapping regions are split to ensure no boundary interference, and areas with unnatural connections or gaps are smoothly adjusted to conform to the vehicle axle's structural logic. Simultaneously, the shape and contour of the regions are optimized to remove irregular protrusions or depressions. Ultimately, the segmented regions accurately conform to the actual structural distribution of the vehicle axle, providing a precise and reliable basis for subsequent multi-feature voting to screen for suspected defective regions.

[0113] Based on the principle of local maxima, multiple seed points are selected in the structural feature map. By scanning pixel by pixel and comparing the gray values ​​of each pixel with its surrounding neighborhood, pixels with gray values ​​higher than all adjacent pixels and located at the core of different structural regions of the axle are selected as seed points. These seed points are evenly distributed and can accurately represent independent structural regions, providing stable and accurate starting positions for subsequent region growth, and avoiding deviations in region growth direction or incomplete region coverage due to improper selection of seed points.

[0114] The dynamic growth criteria for the vehicle axle are defined based on the grayscale intensity similarity and spatial continuity of the structural feature map. The grayscale intensity similarity criterion ensures that the difference in grayscale values ​​between the pixels to be grown and the edge pixels of the current region is within a reasonable range, maintaining pixel uniformity within the region. The spatial continuity criterion ensures that the pixels to be grown are directly adjacent to the current region without obvious structural breaks, conforming to the actual structural morphology of the vehicle axle surface. The two criteria work together to regulate the direction and range of region growth, effectively avoiding problems such as region confusion and structural distortion during the growth process.

[0115] Under dynamic growth criteria, region growing is performed on the structural feature map to obtain the initial segmentation region of the vehicle bridge. Starting from various sub-points, adjacent pixels are judged according to criteria one by one, and pixels that meet the conditions are continuously included in the region to expand the boundary until no pixels that meet the conditions can be included. This process allows each seed point to grow into an independent region with clear boundaries, uniform internal features, and no overlap, completely covering different structural parts in the structural feature map. This provides a comprehensive and regular initial region foundation for subsequent multi-level confidence verification, helping to improve the credibility of subsequent segmentation regions.

[0116] The texture consistency and grayscale distribution characteristics of the segmented regions are evaluated to obtain preliminary verification results, which can judge the rationality of the segmented regions from the perspective of internal features. By analyzing the uniformity of texture direction and density, as well as the concentration of grayscale value distribution within the region, regions with uniform internal features and no obvious heterogeneous textures or abnormal grayscale fluctuations can be screened out, and regions with chaotic internal features due to segmentation errors can be eliminated, providing a basic basis for judging the credibility of the initial segmented regions.

[0117] Intermediate verification results are generated based on the boundary curvature continuity and gradient consistency of the edge pixel sequences in the initial segmentation region. This further verifies the effectiveness of the segmented region from the perspective of region edge morphology. By checking the smooth transition of edge curvature and the uniformity of gradient direction and magnitude, it ensures that the edges of the segmented region are complete and regular, conforming to the actual shape of the vehicle axle surface structure. This eliminates false regions that do not conform to the real structure, such as edge breaks and disordered gradients, thereby improving the edge credibility of the segmented region.

[0118] The advanced verification results are determined based on the spatial relationships and defect shape conformity of adjacent areas in the initial segmentation region, which can improve the verification system from the perspective of overall regional structure matching. By confirming the rationality of the connection between adjacent areas, spatial logic contradictions such as overlap and interweaving are avoided. At the same time, by comparing with the typical shapes of common vehicle and axle defects, areas whose contours match the actual defects are selected, and invalid areas with irregular shapes or no corresponding actual defects are further eliminated, so that the verification results are more in line with the actual defect detection needs of vehicle and axle.

[0119] The selection of reliable segmentation regions based on the fusion of verification results from three levels comprehensively assesses the reliability of the initial segmentation region by integrating evaluation conclusions from three dimensions: internal features, edge morphology, and overall structural matching. By setting reasonable weights to integrate the results from each level, the one-sidedness of single-dimensional verification can be avoided, ensuring that the finally selected reliable segmentation regions meet the requirements for vehicle and bridge segmentation in terms of internal features, edge integrity, and overall structure. This provides an accurate and reliable regional basis for subsequent multi-feature voting to select suspected defective regions, reducing the interference of non-defective regions on subsequent detection processes.

[0120] S4. Perform multi-feature voting on the segmented region to obtain the suspected defect region of the axle;

[0121] In this embodiment of the invention, the step of performing multi-feature voting on the segmented region to obtain the suspected defect region of the axle includes:

[0122] Based on the multi-dimensional features of the segmented region, a regional feature description of the segmented region is generated;

[0123] Based on the different feature dimensions in the region feature description, multidimensional anomaly voting is performed on the segmented region to obtain the initial voting result of the segmented region;

[0124] Based on the consistency of the initial voting results, the voting results of different feature dimensions in the initial voting results are weighted and fused to obtain the comprehensive voting result of the segmented region;

[0125] The region screening criteria for the axle are generated based on the comprehensive voting results, and the segmented regions are screened according to the region screening criteria to obtain the suspected defect regions of the axle.

[0126] This analysis delves into the multi-dimensional characteristics of the segmented region, including core features such as texture distribution, grayscale variations, shape contours, boundary morphology, and spatial location. Each feature dimension is described in detail; for example, texture distribution is described in terms of direction, density, and uniformity; grayscale variations are specified in terms of distribution range and fluctuation; shape contours are recorded in terms of geometric form and regularity; boundary morphology is described in terms of smoothness and continuity; and spatial location is defined in terms of its relative relationship with surrounding areas. All feature descriptions are then integrated according to a unified logic to form a comprehensive description of the segmented region's characteristics, fully reflecting its attributes.

[0127] Based on the different feature dimensions defined in the regional feature description, a normal feature benchmark is established for each dimension. This benchmark is determined based on the feature statistics of the defect-free area of ​​the vehicle axle and represents the feature performance of the normal area under each dimension. For each segmented region, its actual performance on each feature dimension is compared with the normal feature benchmark to determine whether there are any abnormalities deviating from the benchmark. If the performance of a certain feature dimension exceeds the normal benchmark range, it is determined that the dimension votes abnormally for the segmented region; otherwise, it votes normally. After all feature dimensions have been voted on, the initial voting results for the segmented regions are summarized.

[0128] Analyzing the consistency of voting results across different feature dimensions in the initial voting results reveals a high degree of consistency. If multiple feature dimensions consistently vote for the same segmented region (either anomalous or normal), this consistency is significant and provides greater reference value for the final comprehensive judgment. Conversely, if significant differences exist between voting results across dimensions, the influence of less reliable dimensions needs to be mitigated. Based on the consistency of voting results across feature dimensions, each dimension is assigned a corresponding weight, with higher consistency resulting in greater weight. The comprehensive voting result for the segmented region is obtained by multiplying the voting results of each dimension by their corresponding weights and summing the results. This comprehensive result more accurately reflects the overall degree of anomaly in the segmented region.

[0129] Based on the distribution of the overall voting results and the actual needs for detecting axle defects, regional screening criteria are established to clarify at what level of overall voting results a segmented area will be judged as a suspected defect area. The screening criteria must balance detection rate and accuracy, ensuring that no real defect areas are missed, nor are too many normal areas misclassified as suspected defects. The overall voting results of all segmented areas are compared with the established regional screening criteria to select segmented areas that meet the standards. These areas exhibit significant anomalies across multiple feature dimensions, and the overall degree of anomaly meets the criteria for judging suspected defects, ultimately forming the suspected defect areas for the axle.

[0130] The beneficial effects are that by generating regional feature descriptions based on the multi-dimensional features of the segmented regions, key attributes such as texture distribution, grayscale changes, shape contours, boundary morphology, and spatial location of the segmented regions can be comprehensively captured. By meticulously characterizing each feature dimension, it ensures that no feature information that may reflect defects is overlooked, allowing the regional feature descriptions to fully and accurately reflect the actual state of the segmented regions. This provides comprehensive feature basis for subsequent anomaly voting and avoids misjudgment or omission of defect features due to a single feature dimension.

[0131] Based on different feature dimensions in the regional feature description, a multi-dimensional anomaly voting process is performed on the segmented regions to obtain initial voting results. This allows for the determination of whether a segmented region exhibits anomalies from multiple independent dimensions. By establishing normal baselines for each feature dimension and comparing the differences between the actual features of the segmented region and the normal baselines, each dimension independently casts an anomaly or normal vote. This approach comprehensively covers potentially defective feature manifestations, reduces the limitations of judging from a single feature dimension, and preliminarily screens out regions with potential defective tendencies.

[0132] By weighting and fusing the voting results of different feature dimensions based on the consistency of the initial voting results, a comprehensive voting result is obtained, which highlights the reference value of highly consistent voting results. When multiple feature dimensions all determine that a certain region is abnormal or normal, it indicates that the feature performance of that region has strong consistency, and it is assigned a higher weight; if there are large differences in voting between dimensions, the influence of low-confidence dimensions is weakened. This weighted fusion method can integrate the judgment advantages of each dimension, allowing the comprehensive voting result to more accurately reflect the overall degree of abnormality of the segmented region and avoid the interference of individual dimension errors on the result.

[0133] Based on the comprehensive voting results, regional screening criteria are generated, and suspected defect areas are identified. Reasonable screening thresholds can be set according to the actual needs of axle defect detection, ensuring that no real defect areas are missed while reducing the possibility of normal areas being misclassified as suspected defects. By comparing the comprehensive voting results of segmented regions with the screening criteria, regions exhibiting significant anomalies across multiple feature dimensions can be accurately identified. These regions have a high probability of being defective, allowing for focused target areas in subsequent credibility assessments and improving the efficiency and accuracy of axle fracture defect detection.

[0134] S5. Based on the image features of the suspected defect area and the acoustic emission features of the axle, the credibility of the suspected defect area is evaluated to obtain the target defect area of ​​the axle.

[0135] In this embodiment of the invention, the step of evaluating the credibility of the suspected defect region based on the image features of the suspected defect region and the acoustic emission features of the axle to obtain the target defect region of the axle includes:

[0136] Based on the texture complexity and shape irregularity measures of the image features in the suspected defect area, a defect correlation analysis is performed on the axle to obtain the credibility of the axle's image features;

[0137] Defect indicative analysis is performed on the acoustic emission characteristics of the axle to obtain the reliability of the acoustic emission characteristics of the axle;

[0138] A complementary evaluation of the image feature confidence level and the acoustic emission feature confidence level is established to obtain the feature complementarity index of the vehicle axle;

[0139] Based on the image feature credibility, the acoustic emission feature credibility, and the feature complementarity index, the comprehensive credibility score of the vehicle axle is calculated.

[0140] The suspected defect areas are prioritized based on the comprehensive credibility score to obtain the target defect areas of the axle.

[0141] The formula for calculating the overall credibility score is as follows:

[0142] ;

[0143] In the formula, The overall credibility score is given. The credibility of the image features, The credibility of the acoustic emission characteristics is given. This is the mapping value of the feature complementarity index.

[0144] Image features of each suspected defect area are extracted, with a focus on analyzing texture complexity and shape irregularity. Texture complexity is measured by observing the density, direction variation, and disorder of textures within the area; denser textures and more chaotic directions indicate higher texture complexity. Shape irregularity is judged by comparing the area's contour with conventional geometric shapes; the more obvious the deviation from conventional shapes and the more rugged the edges, the stronger the shape irregularity. Combining the image feature patterns of known axle defects, the texture complexity and shape irregularity of suspected defect areas are determined to match the characteristics of actual defects. A higher degree of matching indicates a stronger defect correlation, and the reliability of the axle's image features is quantitatively evaluated accordingly.

[0145] Acoustic emission signals generated by the axle during operation are collected. These signals contain information related to axle structural changes, defect initiation, or propagation. The acoustic emission characteristics are analyzed, focusing on extracting key features such as signal amplitude, frequency distribution, and duration. These characteristics are compared with those of a normal axle to determine if abnormal fluctuations exist. If the amplitude of the acoustic emission signal exceeds the normal range, the frequency distribution shows abnormal peaks, or the duration deviates from normal patterns, it indicates that the feature is indicative of defects. The stronger the indicative power, the higher the reliability. Based on this analysis, the reliability of the axle's acoustic emission characteristics is determined.

[0146] The correlation and complementarity between image feature reliability and acoustic emission feature reliability are analyzed. If the image feature reliability of a suspected defect area is high but the acoustic emission feature reliability is low, or vice versa, the difference in detection angle between the two types of features needs to be considered. Image features focus on surface visual information, while acoustic emission features focus on internal structural change information; the two can complement and verify each other. By judging whether the two types of features can jointly point to the existence of defects from different dimensions, the one-sidedness of single feature evaluation is avoided. If the abnormal indication direction of the two types of features is consistent or can corroborate each other, the feature complementarity is strong. Based on this, a complementary evaluation mechanism is established to generate a feature complementarity index for the vehicle axle.

[0147] Weights are assigned to image feature reliability, acoustic emission feature reliability, and feature complementarity indicators. These weights are determined based on the importance of each indicator in defect identification, ensuring the comprehensive score fully reflects the true situation of suspected defect areas. The three indicators are then integrated and calculated according to their assigned weights to obtain a single comprehensive reliability score for each suspected defect area. A higher score indicates a greater likelihood that the area is a genuine defect.

[0148] All suspected defect areas are prioritized based on their comprehensive credibility scores, from highest to lowest. Areas with scores above a set threshold are retained first, as these areas undergo dual verification and complementary evaluation using image features and acoustic emission features, indicating a high probability of actual defects. The highest-priority areas are then selected based on the ranking results; these areas represent the most likely locations of actual defects in the axle, ultimately forming the target defect areas for the axle and providing clear guidance for subsequent defect handling and maintenance.

[0149] Image feature reliability is derived from the analysis of image features of suspected defect areas. The texture complexity and shape irregularity features of suspected defect areas are extracted, and combined with the image feature patterns of known defects in the vehicle axle, the degree of consistency between these features and the actual defect features is judged. The image feature reliability is obtained by quantifying the degree of consistency.

[0150] The reliability of acoustic emission characteristics is derived from the analysis of acoustic emission signals collected during the operation of the vehicle axle. Key features such as the amplitude, frequency distribution, and duration of the acoustic emission signals are extracted and compared with the acoustic emission characteristics benchmark of a normal vehicle axle to determine whether there are abnormal fluctuations. The reliability of acoustic emission characteristics is obtained by quantifying the indicative nature of abnormal fluctuations on defects.

[0151] The mapping value of the feature complementarity index comes from the complementarity analysis of the credibility of image features and the credibility of acoustic emission features. It determines whether the two types of features can point to the existence of defects from different dimensions and whether they can complement each other to verify. The corresponding mapping processing is carried out according to the strength of complementarity to obtain the mapping value of the feature complementarity index.

[0152] The formula's significance lies in comprehensively and accurately calculating the overall credibility score of suspected defect areas by integrating the credibility of image features, the credibility of acoustic emission features, and the mapping value of feature complementarity indices. When the mapping value of the feature complementarity index is high, it indicates that image features and acoustic emission features effectively complement each other and jointly corroborate the existence of defects. In this case, the formula multiplies the geometric mean of the two features by this mapping value, highlighting the synergistic verification effect. When the mapping value of the feature complementarity index is low, it indicates weak complementarity between the two types of features. In this case, the formula multiplies the arithmetic mean of the two features by the complement of the complementarity index mapping value, mitigating the limitations of single-feature evaluation. The overall calculation process considers both the fundamental role of the two core features in defect determination and dynamically adjusts the weights of different evaluation dimensions through the complementarity index, ensuring that the overall credibility score truly reflects the probability that a suspected defect area is a real defect, providing a reliable basis for selecting target defect areas.

[0153] The beneficial effects are that by performing defect correlation analysis based on the texture complexity and shape irregularity measures of image features in suspected defect areas to obtain the credibility of image features, the degree of association between suspected areas and defects can be accurately determined from a visual perspective. By comparing the density and disorder of textures and the degree of deviation of shapes from conventional geometric forms with the image feature patterns of known defects in vehicles and axles, the probability that suspected areas are real defects can be quantitatively assessed. This provides a reliable basis for subsequent comprehensive credibility assessment based on surface visual information, avoiding judgment bias caused by relying on only a single feature.

[0154] Defect indicative analysis of the acoustic emission characteristics of axles is performed to obtain the reliability of the acoustic emission characteristics, which can supplement the verification of the authenticity of suspected defect areas from the perspective of internal structure. By extracting key features such as the amplitude, frequency distribution, and duration of the acoustic emission signal, and comparing them with the acoustic emission characteristics of normal axles, it is determined whether there are abnormal fluctuations in the signal reflecting structural changes or defect initiation. The indicative strength of abnormal signals for defects is quantified, adding a judgment dimension of internal structure to the reliability assessment and making up for the limitation that image features can only reflect the surface state.

[0155] Establishing a complementary evaluation of image feature reliability and acoustic emission feature reliability to obtain a feature complementarity index can effectively integrate the detection advantages of both features. Image features focus on surface visual information, while acoustic emission features focus on internal structural change information. By judging whether the two types of features can jointly point to the existence of defects from different dimensions and mutually corroborate each other, the one-sidedness of single feature evaluation can be avoided. If the anomaly indications of the two types of features are consistent, it indicates strong complementarity, which can enhance confidence in the identification of suspected defects and provide key collaborative verification basis for comprehensive reliability scoring.

[0156] The comprehensive credibility score calculated based on three types of indicators can fully integrate information from surface visuals, internal structure, and feature co-verification to form an overall quantitative judgment on the authenticity of suspected defect areas. By reasonably allocating the weights of each indicator, it takes into account both the basic judgment role of the two core features and dynamically adjusts the degree of influence of different dimensions through complementary indicators, so that the scoring results can truly reflect the probability that the suspected area is a real defect, avoiding misjudgments caused by single indicators or non-coordinated evaluation.

[0157] Suspected defect areas are prioritized based on comprehensive credibility scores to identify target defect areas, enabling precise targeting of high-credibility potential defect regions. Areas with high priority, verified across multiple dimensions, and high comprehensive credibility are selected from highest to lowest score. These areas are the most likely locations in the axle to contain genuine fracture defects, eliminating a large number of low-credibility false positives. This provides clear and precise targets for subsequent in-depth feature analysis and defect processing, significantly improving the accuracy and efficiency of axle fracture defect detection and reducing the investment of ineffective detection resources.

[0158] This comprehensive credibility scoring formula effectively integrates the mapping values ​​of image feature credibility, acoustic emission feature credibility, and feature complementarity index, enabling precise quantitative assessment of the authenticity of suspected defect areas. When the mapping value of the feature complementarity index is high, it indicates that image features and acoustic emission features mutually corroborate each other from both surface visual and internal structural dimensions, collaboratively pointing to the existence of defects. In this case, the formula multiplies the geometric mean of the two feature credibility values ​​by this mapping value, highlighting the effect of their collaborative verification and allowing the comprehensive score to better reflect the high credibility of the two types of features consistently pointing to defects. When the mapping value of the feature complementarity index is low, it indicates weak complementarity between the two types of features, potentially indicating insufficient information from a single feature dimension. In this case, the formula multiplies the arithmetic mean of the two feature credibility values ​​by the complement of the mapping value, weakening the limitations of single-feature evaluation and avoiding scoring bias due to inconsistencies between the two types of features.

[0159] The formula takes into account the detection advantages and synergistic relationship of different feature dimensions. It not only emphasizes the reflection of surface defect morphology by image features, but also does not ignore the indication of internal structural anomalies by acoustic emission features. At the same time, it dynamically adjusts the weight ratio of different evaluation dimensions through feature complementarity index to ensure that the comprehensive credibility score can truly and comprehensively reflect the probability that the suspected defect area is a real defect.

[0160] The score calculated based on this formula can provide an objective and accurate quantitative basis for prioritizing suspected defect areas, helping to screen out the target defect areas with the highest authenticity, effectively reducing the interference of false defect areas on subsequent inspection processes, and improving the accuracy and reliability of vehicle axle fracture defect detection.

[0161] S6. Perform in-depth feature analysis on the target defect area to obtain a quantitative description and confidence assessment of the defects of the axle.

[0162] In this embodiment of the invention, the step of performing deep feature analysis on the target defect region to obtain a quantitative description and confidence assessment of the axle defect includes:

[0163] The local and global features of the target defect region are fused across scales to obtain the multi-scale features of the vehicle axle.

[0164] Multi-dimensional feature fusion is performed on the multi-scale features to obtain the fused features of the vehicle axle;

[0165] Based on the fusion features, the set features of curves in the target defect region are determined, and a quantitative description of the vehicle axle defect is generated according to the texture complexity and intensity distribution of the target defect region.

[0166] Based on the fusion features, the defects of the axle are scored for internal and external compliance to obtain the confidence level assessment of the axle.

[0167] Local and global features of the target defect area are extracted. Local features focus on detailed information within the defect area, including subtle texture variations, local edge undulations, and local grayscale distribution differences. Local feature extraction is achieved by meticulously scanning and recording these details pixel by pixel. Global features focus on the overall morphology of the defect area, its spatial location on the axle surface, and its overall relative relationship with surrounding normal structures. Global feature extraction is achieved by outlining and spatially locating the overall contour of the defect area. The extracted local and global features are categorized by feature type and fused across scales using feature concatenation. Local detailed features are embedded into the corresponding spatial locations of global features, ensuring that the fused features encompass both the subtle details of the defect and overall distribution information, ultimately yielding multi-scale features of the axle.

[0168] The multi-scale features of the vehicle axle are classified into different dimensions, including texture features, shape features, grayscale features, and spatial location features. For each dimension, redundant or repetitive information is removed, retaining only the key features that accurately reflect the defect attributes. Then, a fixed fusion weight is assigned to each feature dimension according to its importance in describing the defect, with higher-importance dimensions receiving larger weights. After combining the key features of each dimension with their corresponding weights, they are integrated according to a unified feature format, allowing the features from different dimensions to complement each other and form a fusion feature that comprehensively covers all aspects of the vehicle axle's defect attributes.

[0169] Based on the fusion features of the axle, edge curves and internal texture curves of the target defect region are extracted. The set of these curves is analyzed to determine their overall direction, curvature, and connectivity. Simultaneously, texture complexity and intensity distribution features are separated from the fusion features. Texture complexity is determined by statistically analyzing the density and frequency of texture changes within the defect region, while intensity distribution is obtained by analyzing the range and concentration of grayscale values ​​within the region. Combining the fusion features of the curves, texture complexity, and intensity distribution, the size, shape, texture density, and intensity level of the defect are quantified, forming a quantitative description of the axle defect.

[0170] Internal and external features of defects are extracted from the fusion features of the axle. Internal features include texture uniformity, grayscale consistency, and structural integrity within the defect. External features include the clarity of the defect boundary, its connection with surrounding structures, and the regularity of its overall shape. A conformity assessment standard for internal and external features is established, based on the correspondence between typical internal and external features of various known defects in the axle. The extracted defect internal and external features are compared with the assessment standard to determine the degree of conformity between internal features, between external features, and between internal and external features. A corresponding score is given based on the degree of conformity; this score is the confidence assessment of the axle. A higher score indicates a higher degree of agreement between the quantitative description of the defect and the actual defect.

[0171] The beneficial effects are that cross-scale fusion of local and global features of the target defect area yields multi-scale features of the vehicle axle, enabling the simultaneous capture of subtle details and overall distribution information of the defect. Local features can accurately extract minute texture changes, local edge undulations, and local grayscale distribution differences within the defect area, while global features can clearly present the overall shape of the defect, its spatial location on the axle surface, and its relative relationship with surrounding normal structures. By classifying the two according to feature type and fusing them in a stitching manner, embedding local detail features into the corresponding spatial locations of global features, the loss of details or insufficient overall understanding caused by single-scale analysis can be avoided, providing comprehensive and rich feature support for subsequent deep feature analysis.

[0172] Multi-scale features of the vehicle axle are fused using multi-dimensional features to obtain fused features, which can integrate key defect information from different dimensions. First, the multi-scale features are divided into dimensions such as texture features, shape features, grayscale features, and spatial location features. For each dimension, redundant information is removed, retaining the core content that accurately reflects the defect attributes. Then, fixed fusion weights are assigned according to the importance of each dimension to the defect description. The key features of each dimension are combined with their corresponding weights and integrated in a unified format. This ensures that the fused features cover all aspects of the key attributes of the defect while avoiding interference from irrelevant information, providing accurate and focused feature basis for subsequent quantitative description and confidence assessment of defects.

[0173] Based on the fusion features of the vehicle axle, the set features of curves in the target defect region are determined, and the texture complexity and intensity distribution of the target defect region are combined to generate a quantitative description of the vehicle axle defects, which can transform defect attributes into specific quantifiable indicators.

[0174] The edge curves and internal texture curves of defects are extracted from the fusion features. The overall direction, curvature and connection relationship between curves are analyzed to determine the set features. At the same time, the density and direction change frequency of textures in the defect area are statistically analyzed to obtain the texture complexity. The intensity distribution range and concentration of gray values ​​in the area are analyzed to obtain the intensity distribution. Combining this information, the size, shape, texture density and intensity level of the defects are specifically quantified, so that the defect description can move away from the limitations of qualitative judgment and provide an objective and accurate quantitative reference for the assessment and treatment of vehicle and axle defects.

[0175] Based on the fusion characteristics of the axle, internal and external conformity scores are scored to obtain the confidence level assessment of the axle, which can verify the reliability of the quantitative description of defects. Internal and external features of defects are extracted from the fusion characteristics. Internal features include texture uniformity, grayscale consistency, and structural integrity within the defect; external features include boundary clarity, connection with surrounding structures, and overall morphological regularity. Evaluation criteria are established based on the correspondence between typical internal and external features of known axle defects. The extracted internal and external features are compared with the criteria to determine the degree of conformity between internal features, between external features, and between internal and external features, and a score is given. A higher score indicates a higher degree of conformity between the quantitative description of the defect and the actual defect, providing a clear judgment standard for the credibility of the detection results and enhancing the scientific rigor and practicality of axle fracture defect detection results.

[0176] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0177] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0178] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A non-destructive testing method for fracture defects in vehicle axles, characterized in that, The method includes: S1. Acquire multi-band image data of the axle surface under full-spectrum illumination conditions; S2. Based on the multi-band image data separated by spectral features, determine the surface phase consistency of the axle to obtain the structural saliency map of the axle; S3. While maintaining the edge integrity of the structural saliency map, perform region segmentation on the structural saliency map to obtain the segmented region of the vehicle bridge; S4. Perform multi-feature voting on the segmented region to obtain the suspected defect region of the axle; S5. Based on the image features of the suspected defect area and the acoustic emission features of the axle, the credibility of the suspected defect area is evaluated to obtain the target defect area of ​​the axle. S6. Perform in-depth feature analysis on the target defect area to obtain a quantitative description and confidence assessment of the defects of the axle.

2. The non-destructive testing method for axle fracture defects as described in claim 1, characterized in that, The acquisition of multi-band image data of the axle surface under full-spectrum illumination conditions includes: Configure full-spectrum illumination parameters to obtain uniform illumination conditions on the axle surface; Under the uniform illumination conditions, the original reflected image of the axle surface is acquired to obtain a set of original images of the axle surface; The original image set is integrated and standardized to obtain multi-band image data of the vehicle bridge surface.

3. The non-destructive testing method for axle fracture defects as described in claim 1, characterized in that, The process of determining the surface phase consistency of the axle based on the multi-band image data separated according to spectral features to obtain the structural saliency map of the axle includes: Multi-directional filtering is performed on the band image data after spectral feature separation to obtain filtered image data; Based on the filtered image data, the phase consistency feature map of the axle is extracted; The phase consistency map is normalized and enhanced to obtain the structural saliency map of the axle.

4. The non-destructive testing method for axle fracture defects as described in claim 1, characterized in that, The step of segmenting the structural saliency map to obtain the segmented regions of the vehicle axle while maintaining the edge integrity of the structural saliency map includes: Edge-guided optimization processing is performed on the structural saliency map to obtain the structural feature map of the axle; The structural feature map is subjected to region growing according to predefined dynamic constraints to obtain the initial segmentation region of the vehicle bridge; Multi-level confidence verification is performed on the initial segmented region to obtain the reliable segmented region of the vehicle axle; The topology of the reliable segmentation region is optimized to obtain the segmentation region of the vehicle axle.

5. The non-destructive testing method for axle fracture defects as described in claim 4, characterized in that, The step of performing region growing on the structural feature map according to predefined dynamic constraints to obtain the initial segmentation region of the vehicle axle includes: In the structural feature map, multiple seed points are selected based on the principle of local maxima; Based on the pixel intensity similarity and spatial continuity of the structural feature map, the dynamic growth criterion of the vehicle axle is defined. Under the dynamic growth criterion, region growth is performed on the structural feature map to obtain the initial segmentation region of the vehicle bridge.

6. The non-destructive testing method for axle fracture defects as described in claim 5, characterized in that, The step of performing multi-level confidence verification on the initial segmented region to obtain the reliable segmented region of the axle includes: The texture consistency and grayscale distribution features in the segmented region are evaluated to obtain the preliminary verification results of the initial segmented region; Based on the boundary curvature continuity and gradient consistency of the edge pixel sequences in the initial segmentation region, an intermediate verification result for the initial segmentation region is generated. Based on the spatial relationship between adjacent regions in the initial segmentation region and the defect shape conformity of the initial segmentation region, the advanced verification result of the initial segmentation region is determined; Based on the fusion of the primary verification results, the intermediate verification results, and the advanced verification results, a reliable segmentation region of the vehicle axle in the initial segmentation region is selected.

7. The non-destructive testing method for axle fracture defects as described in claim 1, characterized in that, The step of performing multi-feature voting on the segmented region to obtain the suspected defect region of the axle includes: Based on the multi-dimensional features of the segmented region, a regional feature description of the segmented region is generated; Based on the different feature dimensions in the region feature description, multidimensional anomaly voting is performed on the segmented region to obtain the initial voting result of the segmented region; Based on the consistency of the initial voting results, the voting results of different feature dimensions in the initial voting results are weighted and fused to obtain the comprehensive voting result of the segmented region; The region screening criteria for the axle are generated based on the comprehensive voting results, and the segmented regions are screened according to the region screening criteria to obtain the suspected defect regions of the axle.

8. The non-destructive testing method for axle fracture defects as described in claim 1, characterized in that, The step of assessing the credibility of the suspected defect area based on the image features of the suspected defect area and the acoustic emission features of the axle to obtain the target defect area of ​​the axle includes: Based on the texture complexity and shape irregularity measures of the image features in the suspected defect area, a defect correlation analysis is performed on the axle to obtain the credibility of the axle's image features; Defect indicative analysis is performed on the acoustic emission characteristics of the axle to obtain the reliability of the acoustic emission characteristics of the axle; A complementary evaluation of the image feature confidence level and the acoustic emission feature confidence level is established to obtain the feature complementarity index of the vehicle axle; Based on the image feature credibility, the acoustic emission feature credibility, and the feature complementarity index, the comprehensive credibility score of the vehicle axle is calculated. The suspected defect areas are prioritized based on the comprehensive credibility score to obtain the target defect areas of the axle.

9. The non-destructive testing method for axle fracture defects as described in claim 8, characterized in that, The formula for calculating the overall credibility score is as follows: ； In the formula, The overall credibility score is given. The credibility of the image features, The credibility of the acoustic emission characteristics is given. This is the mapping value of the feature complementarity index.

10. The non-destructive testing method for axle fracture defects as described in claim 1, characterized in that, The step of performing in-depth feature analysis on the target defect area to obtain a quantitative description and confidence assessment of the axle defect includes: The local and global features of the target defect region are fused across scales to obtain the multi-scale features of the vehicle axle. Multi-dimensional feature fusion is performed on the multi-scale features to obtain the fused features of the vehicle axle; Based on the fusion features, the set features of curves in the target defect region are determined, and a quantitative description of the vehicle axle defect is generated according to the texture complexity and intensity distribution of the target defect region. Based on the fusion features, the defects of the axle are scored for internal and external compliance to obtain the confidence level assessment of the axle.

Images