Wheel set state detection method and system based on iterative algorithm
By using an iterative algorithm and a weighted value calibration method for wheelset status detection, the environmental interference and reference offset problems of subway train wheel tread detection are solved, achieving high-precision, adaptive intelligent detection and reducing the need for manual calibration.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-04
- Publication Date
- 2026-03-10
AI Technical Summary
Existing methods for detecting the tread surface of subway train wheels are susceptible to environmental interference and prone to reference deviation, making it difficult to accurately quantify and identify defects, resulting in low detection accuracy and the need for frequent manual calibration.
An iterative algorithm-based wheel pair state detection method is adopted. Through iterative data updates and weighted value correction, adaptive adjustment is achieved. Combined with image preprocessing and edge detection, the detection accuracy and robustness are improved.
It significantly improves detection accuracy and system robustness, reduces false alarm and false negative rates, realizes intelligent wheelset status detection, and reduces manual intervention.
Smart Images
Figure CN121639644A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of rail transit, in particular to a wheel set state detection method and system based on an iterative algorithm. BACKGROUND
[0002] With the rapid development of metro vehicles, the state detection of metro train wheels has become a key research field to ensure operation safety. As the part directly contacted by the wheel and the track, the wheel tread surface, the surface scratch, the out-of-roundness and the size deviation of which will significantly increase the vibration amplitude of the train during operation, which not only affects the comfort of passengers, but also may intensify the load on the axle, bogie and other key components due to resonance effect, and thus cause potential traffic safety hazards. Therefore, accurately detecting the size, out-of-roundness and scratch defects of the wheel tread surface is of great significance to ensure the safety of metro operation.
[0003] At present, the detection methods for metro train tread scratches at home and abroad mainly include noise analysis, ultrasonic signal detection and acceleration sensor vibration signal analysis technologies. However, these methods can only qualitatively determine whether the scratch exists, and cannot provide detailed information such as the specific form and size of the scratch, which is not conducive to the accurate judgment of maintenance personnel. In addition, the detection results are easily disturbed by factors such as vehicle speed, vehicle weight, installed vibration and calibration reference, resulting in low accuracy. In recent years, the detection technology based on computer vision has gradually become a research hotspot, which can efficiently and flexibly identify wheel set defects by combining image processing and deep learning algorithms, and significantly improve the detection accuracy and efficiency. However, the existing visual detection scheme still has certain limitations, such as the image being easily disturbed by environmental light, the reference being deviated due to vibration or long-term use of the triggering device, and the need for frequent manual calibration, which limits its reliability in practical application. SUMMARY
[0004] In view of the above-mentioned technical problems that the existing detection methods are easily disturbed, the reference is easily deviated, and it is difficult to accurately and quantitatively identify defects, a wheel set state detection method and system based on an iterative algorithm are provided. The present application mainly updates the data by iteration, increases the iterative algorithm and the weighting value for reference data correction, realizes the self-adaptive adjustment of the wheel set state detection with the disturbance and external environment change, reduces the artificial working time and improves the qualified rate of image detection and judgment.
[0005] The technical means adopted by the present application are as follows: The wheel set state detection method based on the iterative algorithm comprises the following steps: Collecting a wheel set image, wherein the wheel set image includes wheel set defect information; The wheel set image is preprocessed, and the preprocessing includes: detecting noise information in the wheel set image by using an iterative optimization algorithm to obtain a noise threshold distribution map, generating a pixel weight matrix by using a linear weighting method based on the noise threshold distribution map, filtering the wheel set image by using the pixel weight matrix to obtain a preprocessed wheel set image; An edge detection operator is used to extract features from the preprocessed wheel set image to obtain an edge feature map, and the edge feature map is classified and extracted to identify wheel set defects.
[0006] Further, the noise information in the wheel set image is detected by using an iterative optimization algorithm to obtain a noise threshold distribution map, which includes: Based on the wheel set image, an initial filtering window with a size of a x a and a pixel set in the window are defined, and the calculation formula of the pixel set in the window is:
[0007] wherein, F m,n is the pixel set in the window, m is the horizontal coordinate of the center point of the initial filtering window, n is the vertical coordinate of the center point of the initial filtering window, x is the horizontal offset of the pixel in the window relative to the center point, y is the vertical offset of the pixel in the window relative to the center point, and a is the size parameter of the window; Based on the pixel set in the window, the pixel average value of the initial filtering window is calculated, and the calculation formula of the pixel average value is:
[0008] wherein, AVG ( F m,n ) is the pixel average value; Based on the pixel average value, the window center point is iteratively updated to obtain an optimized window center point set, and the iteratively updated window pixel average value is calculated according to the optimized window center point set, and the calculation formula of the optimized window center set is:
[0009] wherein, is the optimized window center set, is the horizontal coordinate of the window center point after the kth iteration, is the vertical coordinate of the window center point after the kth iteration, and k is the iteration number, The calculation formula of the iteratively updated window pixel average value is:
[0010] wherein, the average value of the window pixels after iteration; According to the initial average value and the average value after iteration, a noise threshold value of each window region is calculated to form the noise threshold value distribution map, and the calculation formula of the noise threshold value of each window region is:
[0011] wherein, D m,n is the noise threshold value.
[0012] Further, the method of generating a pixel weight matrix based on the noise threshold value distribution map adopts a linear weighting method, including: According to the noise threshold value in the noise threshold value distribution map, the number of noise points in the wheel set image is counted, and the calculation formula of the number of noise points is:
[0013] wherein, NUM F m,n is the number of noise points, and Numnoise(·) is a noise point discrimination function; Based on the total number of noise points, the final size of the filter window is adaptively determined; According to the final size, a weight value corresponding to each pixel position in the window is calculated by using a linear weighting function, and all position weight values together constitute the pixel weight matrix.
[0014] Further, the calculation formula of the linear weighting function is:
[0015] wherein, weight (x,y) is the pixel weight value at position (x,y).
[0016] Further, the mathematical expression of the filter processing is:
[0017] wherein, h m , n is the pixel value of the preprocessed wheel set image at position (x,y), Med is a median filter operator, and * represents a weighted calculation.
[0018] Further, the method of using an edge detection operator to perform feature extraction on the preprocessed wheel set image to obtain an edge feature map, including: A first-order differential Roberts operator is used to perform convolution calculation on the preprocessed wheel set image to extract edge features in the image, and the convolution calculation formula of the Roberts operator is:
[0019] wherein, f ( i , j ) is the gray value of the pre-processed wheel set image at pixel coordinates , f i is the approximation of the first-order partial derivative of the image function in the i direction, f j is the approximation of the first-order partial derivative of the image function in the j direction, f ( i +1, j +1) is the gray value of the pre-processed wheel set image at pixel coordinates , f ( i +1, j ) is the gray value of the pre-processed wheel set image at pixel coordinates , f ( i , j +1) is the gray value of the pre-processed wheel set image at pixel coordinates , f ( i , j ) is the gradient vector of the image at pixel coordinates , g ( i , j ) is the pixel value of the edge feature at coordinates ; The complete edge feature map is obtained by traversing and calculating the entire image.
[0020] The application also includes a wheel set state detection system based on an iterative algorithm, which is used to implement the above-mentioned wheel set state detection method based on an iterative algorithm, and includes: An image acquisition module is configured to acquire a wheel set image containing wheel set defect information. A preprocessing module is connected to the image acquisition module, and the preprocessing module includes: A noise analysis unit is configured to detect noise information in the wheel set image by using an iterative optimization algorithm, and obtain a noise threshold distribution map, A weight matrix generation unit is configured to generate a pixel weight matrix by using a linear weighting method based on the noise threshold distribution map, A filtering processing unit is configured to perform filtering processing on the wheel set image by using the pixel weight matrix, and obtain a pre-processed wheel set image. The feature extraction module is connected with the preprocessing module and is configured to extract features of the preprocessed wheel set image by using an edge detection operator to obtain an edge feature map. The defect identification module is connected with the feature extraction module and is configured to classify and extract the edge feature map to identify wheel set defects.
[0021] Compared with the prior art, the present application has the following advantages: The present application combines classical machine vision technology with deep iterative learning algorithm and introduces an iterative detection mechanism of image information interference noise, realizing three core results: first, improving detection accuracy, continuously optimizing feature extraction and noise discrimination model through iterative learning, significantly reducing false positives and false negatives caused by complex background interference; second, enhancing system robustness, the adaptive noise detection mechanism can effectively cope with uncertain factors such as on-site light changes and equipment vibration, ensuring the stability of the detection results; third, realizing intelligent upgrading, combining the rapidity of traditional vision with the self-learning ability of iterative algorithm, promoting the transformation of wheel set detection from relying on artificial experience to data-driven intelligent diagnosis.
[0022] Based on the above reasons, the present application can be widely popularized in the field of rail transportation. BRIEF DESCRIPTION OF DRAWINGS
[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.
[0024] Figure 1 The flowchart of the wheel set state detection method based on the iterative algorithm of the present application.
[0025] Figure 2 The device imaging diagram taken by the device.
[0026] Figure 3 The flowchart of the track defect detection in the prior art. DETAILED DESCRIPTION
[0027] In order to make the personnel in the technical field better understand the present application scheme, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the scope of protection of the present application.
[0028] It is to be understood that the terms "including", "comprising", "having" and their conjugates, as used throughout the specification and in the following Claims, shall not be construed so as to exclude any steps or elements not specifically recited. Thus, the terms "including", "comprising", "having" and their conjugates, as used throughout the specification and in the following Claims, shall be construed as meaning "including, but not limited to".
[0029] The line laser is arranged near the wheel center, and the line laser extension line passes above the wheel center axis; the camera is arranged away from the wheel center, close to the theoretical line laser reflection light path. A precise positioning magnet is arranged at the camera shooting position, and when the physical center of the wheel passes through the center of the magnet, a camera trigger signal is sent to trigger the camera shooting. The device imaging diagram is shown in Figure 2 .
[0030] Using a laser light source, sunlight interference can be effectively avoided, so that the detection device can work normally in the daytime and at night. After the wheel passes through the positioning sensor, the wheel running speed can be calculated, and the wheel can be positioned according to the wheel. When the lowest point of the wheel pair reaches the laser plane, the system sends a shooting command, and the image is shot. Through image recognition processing, we can measure the tread wear, rim thickness, wheel diameter and wheel pair inside distance.
[0031] According to the scratch detected by the tread, all the images of a wheel pair circle can be reflected on the terminal through the video image monitoring module, so that the user can observe and judge the tread scratch condition.
[0032] The wheel pair defect image processing system based on machine vision mainly aims to realize the identification and judgment of the wheel pair defect information, and the core is image processing. Through the wheel pair defect image processing related algorithm, the correct detection of the wheel pair defect image is realized. The track image information detected by the wheel pair defect detection is mainly analyzed and recognized, and then the wheel pair state is analyzed according to the processed image information. The basic flow is shown in Figure 3 .
[0033] The wheel pair image collected on site is easy to produce noise points on the track defect pattern, which interferes with the image information. Image noise information is generally random and uncontrollable image interference information, which may sometimes destroy the details of the original image and cause misjudgment of image recognition. Through image preprocessing, the noise information of the target area can be effectively reduced, the measured image information becomes clear, and feature extraction is facilitated. In order to improve the clarity of the track image, it is necessary to remove and correct the noise and distortion of the image.
[0034] The traditional median filtering processing method is affected greatly by the pixel center point and the size of the filtering window, and when the wheel set image noise points are less, the detection effect is better, and when the noise points are more, the detection effect of the algorithm is very poor.
[0035] The defects in the prior art can be summarized as the following three points: (1) The external factors greatly affect the image information, sometimes destroy the details of the original image, and cause misjudgment of image recognition. (2) The trigger magnetic steel and the data acquisition device are affected by vibration impact and running time, so that the reference data deviates, and the collected and analyzed data has error judgment. (3) It needs to be manually calibrated regularly, and does not have a self-adaptive function.
[0036] The present application adds an iterative algorithm and a window weighting value to the original filtering processing method, and assigns corresponding weights to the wheel set image noise pixel points after adaptive grouping, thereby achieving good noise filtering of the wheel set image.
[0037] As shown in Figure 1 The present application provides a wheel set state detection method based on an iterative algorithm, comprising the following steps: S1. Collecting a wheel set image, wherein the wheel set image includes wheel set defect information.
[0038] The wheel set image includes a digital image of a single wheel set or a key area of the wheel set (mainly the wheel tread and the wheel rim). The image should include the texture, geometric contour and possible defect visual features of the wheel set surface, and the defect features include but are not limited to: local texture and color abnormalities caused by surface scratches, linear gray discontinuity caused by cracks, pits or patches formed by peeling, and contour geometric deformation caused by out-of-roundness or wear.
[0039] S2. Preprocessing the wheel set image, the preprocessing including: detecting noise information in the wheel set image by using an iterative optimization algorithm to obtain a noise threshold distribution map, generating a pixel weight matrix based on the noise threshold distribution map by using a linear weighting method, filtering the wheel set image by using the pixel weight matrix to obtain a preprocessed wheel set image.
[0040] Specifically, S2 includes: S21. Detecting noise information in the wheel set image by using an iterative optimization algorithm to obtain a noise threshold distribution map.
[0041] S211. Based on the wheel set image, defining an initial filtering window with a size of a x a and a pixel set in the window.
[0042] Specifically, screening noise points irrelevant to image information from the detection image is the key to image analysis and judgment. For the collected image information, a filter window is selected to process all pixel points, assuming that the pixel (m, n) center of the selected a x a window is and the corresponding gray value is f (m, n).
[0043] The calculation formula of the pixel set in the window is:
[0044] wherein, F m,n is the pixel set in the window, m is the horizontal coordinate of the initial filter window center point, n is the vertical coordinate of the initial filter window center point, x is the horizontal offset of the pixel in the window relative to the center point, y is the vertical offset of the pixel in the window relative to the center point, and a is the size parameter of the window.
[0045] S212. Based on the pixel set in the window, the pixel average value of the initial filter window is calculated, and the calculation formula of the pixel average value is:
[0046] wherein, is AVG ( F m,n ) is the pixel average value.
[0047] S213. Based on the pixel average value, the window center point is iteratively updated to obtain an optimized window center point set, and according to the optimized window center point set, the iterated window pixel average value is calculated, so as to reduce the large change of defect recognition range caused by the pixel center point offset caused by the installed and running vibration. The calculation formula of the optimized window center set is:
[0048] wherein, is the optimized window center set, is the horizontal coordinate of the window center point after the kth iteration, is the vertical coordinate of the window center point after the kth iteration, k is the iteration number, and the iteration number k is selected to converge according to the window a, k = a.
[0049] The calculation formula of the iterated window pixel average value is:
[0050] wherein, is the iterated window pixel average value.
[0051] S214. According to the initial average value and the average value after iteration, the noise threshold value of each window region is calculated to form a noise threshold value distribution map, and the calculation formula of the noise threshold value of each window region is:
[0052] wherein, D m,n is the noise threshold value.
[0053] S22. A pixel weight matrix is generated based on the noise threshold value distribution map using a linear weighting method.
[0054] S221. The number of noise points in the image information is counted by selecting noise points in the wheel pair image. The wheel pair image filtering window size is adaptively determined using the counted number of noise points. The number of noise points in the wheel pair image is counted according to the noise threshold value in the noise threshold value distribution map, and the calculation formula of the number of noise points is:
[0055] wherein, NUM F m,n is the number of noise points, and Numnoise(·) is a noise point discrimination function.
[0056] S222. The final size of the filtering window is adaptively determined based on the total number of noise points.
[0057] S223. The weight value corresponding to each pixel position in the window is calculated using a linear weighting function according to the final size, and the weight values of all positions together constitute a pixel weight matrix.
[0058] The calculation formula of the linear weighting function is:
[0059] wherein, weight (x, y) is the pixel weight value at position (x, y).
[0060] S23. The wheel pair image is filtered using the pixel weight matrix to obtain a preprocessed wheel pair image.
[0061] Specifically, after the noise filtering window size is determined, a certain weight value is assigned to the image noise points to filter out the noise points outside the weight value range. The mathematical expression of the filtering process is:
[0062] wherein, h m , n is the pixel value of the preprocessed wheel pair image at position, and Med is a median filtering operator, and * represents weighted calculation.
[0063] S3. The method of edge detection operator is used to extract features from the preprocessed wheel set image, and an edge feature map is obtained. The edge feature map is classified and extracted, and the wheel set defects are identified.
[0064] The part with strong gray scale change in the image is the edge of the image, and the edge often contains most of the main information of the image. Edge detection is completed by convolution using spatial domain differential operator. A first-order differential Roberts operator is used to convolve and calculate the preprocessed wheel set image to extract the edge features of the image. The principle of Roberts operator for edge detection is that the difference between the gray scale values of adjacent pixels on the diagonal is used to replace the gradient solution. The convolution calculation formula of Roberts operator is:
[0065] wherein, f ( i , j ) are the gray scale values of the preprocessed wheel set image at pixel coordinates , f i is the approximate value of the first-order partial derivative of the image function in the i direction, f j is the approximate value of the first-order partial derivative of the image function in the j direction, f ( i +1, j +1) are the gray scale values of the preprocessed wheel set image at pixel coordinates , f ( i +1, j ) are the gray scale values of the preprocessed wheel set image at pixel coordinates , f ( i , j +1) are the gray scale values of the preprocessed wheel set image at pixel coordinates , f ( i , j ) is the gradient vector of the image at pixel coordinates , g ( i , j ) is the pixel value of the edge feature at coordinates . The complete edge feature map is obtained by traversing and calculating the entire image.
[0066] Specifically, the wheel set defects include surface damage type defects and visible defects. The surface damage type defects specifically include scratches, peeling, cracks, and the visible defects include scratches, depressions, and foreign object embedding.
[0067] The application also comprises a wheel set state detection system based on an iterative algorithm, which is used to realize the above-mentioned wheel set state detection method based on an iterative algorithm, comprising: An image acquisition module is configured to acquire a wheel set image containing wheel set defect information.
[0068] A preprocessing module is connected to the image acquisition module, and the preprocessing module comprises: A noise analysis unit is configured to detect noise information in the wheel set image by using an iterative optimization algorithm to obtain a noise threshold distribution map.
[0069] A weight matrix generation unit is configured to generate a pixel weight matrix by using a linear weighting method based on the noise threshold distribution map.
[0070] A filtering processing unit is configured to perform filtering processing on the wheel set image by using the pixel weight matrix to obtain a preprocessed wheel set image.
[0071] A feature extraction module is connected to the preprocessing module and is configured to perform feature extraction on the preprocessed wheel set image by using an edge detection operator method to obtain an edge feature map.
[0072] A defect recognition module is connected to the feature extraction module and is configured to classify and extract the edge feature map to recognize wheel set defects.
[0073] To sum up, the wheel set surface area is extracted by the image preprocessing process, and is smoothed and denoised to obtain a clear image. In order to identify the wheel set surface defects, the wheel set surface defect features are classified and extracted. The optimal features that can best distinguish different types of defects are selected from multiple features to detect and identify the wheel set surface defects.
[0074] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the application, and not to limit them; although the application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the above embodiments, or make equivalent replacement to some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the application.
Claims
1. A wheelset condition detection method based on an iterative algorithm, characterized in that, The method comprises the following steps: Collecting a wheel set image containing wheel set defect information; Pretreating the wheel set image, which comprises: detecting noise information in the wheel set image by using an iterative optimization algorithm to obtain a noise threshold distribution map, generating a pixel weight matrix by using a linear weighting method based on the noise threshold distribution map, and filtering the wheel set image by using the pixel weight matrix to obtain a pretreated wheel set image; Extracting features from the pretreated wheel set image by using an edge detection operator to obtain an edge feature map, classifying and extracting the edge feature map, and identifying wheel set defects.
2. The wheelset condition detection method based on iterative algorithm according to claim 1, characterized in that, The method of detecting noise information in the wheel set image by using an iterative optimization algorithm to obtain a noise threshold distribution map comprises: Based on the wheel set image, an initial filtering window with a size of a x a and a pixel set in the window are defined, and the calculation formula of the pixel set in the window is: wherein, F m,n is a set of pixels within a window, m is the horizontal coordinate of the center point of the initial filter window, n is the vertical coordinate of the center point of the initial filter window, x is the horizontal offset of a pixel within the window relative to the center point, y is the vertical offset of a pixel within the window relative to the center point, and a is a size parameter of the window. Based on the pixel set in the window, the pixel average value of the initial filtering window is calculated, and the calculation formula of the pixel average value is: wherein, for AVG ( F m,n ) is the pixel average value; Based on the pixel average value, the window center point is iteratively updated to obtain an optimized window center point set, and the iterative window pixel average value is calculated according to the optimized window center point set, and the calculation formula of the optimized window center set is: wherein, is the optimized window center set, is the horizontal coordinate of the window center point after the kth iteration, is the vertical coordinate of the window center point after the kth iteration, k is the iteration number, The calculation formula of the iterative window pixel average value is: wherein, is the average value of the window pixels after iteration; According to the initial average value and the iterative average value, the noise threshold of each window region is calculated to form the noise threshold distribution map, and the calculation formula of the noise threshold of each window region is: wherein D m,n is a noise threshold.
3. The wheelset condition detection method based on iterative algorithm according to claim 1, characterized in that, The method of generating a pixel weight matrix by using a linear weighting method based on a noise threshold distribution map comprises: According to the noise threshold in the noise threshold distribution map, the number of noise points in the wheel set image is counted, and the calculation formula of the number of noise points is: wherein NUM F m,n is the number of noise points, Numnoise(·) is a noise point discrimination function. Based on the total number of noise points, the final size of the filtering window is adaptively determined; According to the final size, the weight value corresponding to each pixel position in the window is calculated by using a linear weighting function, and all position weight values jointly constitute the pixel weight matrix.
4. The wheelset condition detection method based on iterative algorithm according to claim 3, characterized in that, The calculation formula of the linear weighting function is: wherein, weight (x,y) is the pixel weight at position (x,y).
5. The wheelset condition detection method based on iterative algorithm according to claim 1, characterized in that, The mathematical expression of the filtering processing is: wherein h m n are pixel values of the pre-processed wheelset image at position, Med is a median filter operator, and * denotes a weighted calculation. 6. The wheelset condition detection method based on iterative algorithm according to claim 1, characterized in that, The method of extracting features from the pretreated wheel set image by using an edge detection operator to obtain an edge feature map comprises: A first-order differential Roberts operator is used to perform convolution calculation on the pretreated wheel set image to extract edge features in the image, and the convolution calculation formula of the Roberts operator is: in, f ( i , j () represents the preprocessed wheelset image in pixel coordinates grayscale value at that location f i This is an approximation of the first-order partial derivative of the image function in the i-direction. f j This is an approximation of the first-order partial derivative of the image function in the j-direction. f ( i +1, j +1) represents the preprocessed wheelset image in pixel coordinates. grayscale value at that location f ( i +1, j () represents the preprocessed wheelset image in pixel coordinates grayscale value at that location f ( i , j +1) represents the preprocessed wheelset image in pixel coordinates. The grayscale value at that location, ▽ f ( i , j () represents the image in pixel coordinates The gradient vector at that point, g ( i , j () represents edge features in coordinates Pixel value at; A complete edge feature map is obtained by traversing and calculating the entire image.
7. A wheelset condition detection system based on an iterative algorithm for implementing the wheelset condition detection method based on an iterative algorithm according to any one of claims 1 to 6, characterized in that It comprises: An image acquisition module is configured to acquire a wheel set image containing wheel set defect information; A pretreatment module is connected with the image acquisition module, and the pretreatment module comprises: A noise analysis unit is configured to detect noise information in the wheel set image by using an iterative optimization algorithm to obtain a noise threshold distribution map, A weight matrix generation unit is configured to generate a pixel weight matrix by using a linear weighting method based on the noise threshold distribution map, A filtering processing unit is configured to perform filtering processing on the wheel set image by using the pixel weight matrix to obtain a preprocessed wheel set image; A feature extraction module is connected with the preprocessing module and configured to perform feature extraction on the preprocessed wheel set image by using an edge detection operator to obtain an edge feature map; A defect identification module is connected with the feature extraction module and configured to classify and extract the edge feature map to identify wheel set defects.