Train wheel tread whole-circle dense stripping detection method

Abstract

The invention discloses a train wheel tread whole-circle dense stripping detection method, and belongs to the technical field of railway vehicle detection. The relative movement between the static linear array camera and the moving train is utilized to realize the imaging of the wheel pair tread, the imaging view field is large, and the collected original image has the interference of the tread area and the non-tread area, so that the original image needs to be processed, the wheel tread area is extracted, and the accuracy of the wheel pair tread is improved. Panoramic display of the wheel tread is realized through methods of correction stretching, splicing and the like, finally, defects such as whole-circle stripping are detected on the panoramic tread image through a deep learning method, the situation of missed detection of the tread defects is avoided, and the detection result is comprehensive and reliable.

Description

Technical Field

[0001] The invention belongs to the technical field of rail vehicle detection, and relates to a method for detecting the full circle intensive peeling of a train wheel tread. Background Art

[0002] Train wheels are one of the important parts of the running gear. They are in contact with the rails for a long time. In the event of emergency braking, friction and collision, it is easy to cause defects such as peeling and abrasion on the wheel tread. Under long-term continuous friction and collision, such peeling no longer occurs only in a local area of ​​the tread, but expands to an entire circle of the wheel tread. The tread defect of peeling off the entire circle seriously affects the safety of train driving.

[0003] At present, wheel tread defect detection includes contact detection and non-contact non-destructive testing.

[0004] Patent No. CN 108225706A discloses an automated diagnostic method for identifying dense peeling pitting faults on wheels. Based on the existing bearing fault detection device using generalized resonance / resonance demodulation technology, the information output during the operation of the locomotive vehicle containing the impact of the tread fault is detected, and the sample containing the impact information of the tread fault is obtained through the speed tracking sampling technology, and the tread is deeply diagnosed for whether there is a dense peeling pitting fault. This method realizes the impact monitoring and alarm of dense peeling pitting faults during the operation of the wheel, but it cannot completely detect and warn smaller full-circle peeling, and each vehicle needs to be equipped with a set of equipment, which not only requires the modification of existing locomotive equipment, but also increases the equipment cost of the whole vehicle.

[0005] Patent No. CN107146224A discloses an online image detection system and method for train wheel tread defects. It uses an array camera to collect wheel tread images, and uses traditional image processing technology to determine whether there are defects in the tread images. Then, a method combining regional growth and seed point growth is used to separate the tread defect area from the tread image, and the size information of the tread defect part is calculated using the geometric correspondence of image calibration. This method can detect partial tread peeling defects, but the array camera collects the wheel tread in motion, which will cause motion blur, which is not conducive to image analysis. In addition, this method uses traditional image processing technology. Since the tread defect morphology is unknown, there will be tread defect omissions. This method does not mention the detection of full-circle tread peeling. Summary of the invention

[0006] In view of the shortcomings and deficiencies in the prior art, the present invention provides a method for detecting intensive peeling of a whole circle of a train wheel tread.

[0007] The present invention adopts the following technical solutions to solve the above technical problems:

[0008] A method for detecting intensive peeling of a train wheel tread in a whole circle, comprising:

[0009] (1) Tread image acquisition: multiple linear array camera modules are used to collect images of the wheels of a moving train;

[0010] (2) Tread image correction and stitching: locate the tread image to obtain the boundary points on both sides of the tread; perform curve fitting on the boundary points to obtain two boundary lines of the tread; normalize and interpolate the tread within the boundary line according to the row direction to obtain a regular tread image after correction and flattening; tread images collected by different linear array camera modules are fused and stitched after correction and flattening to obtain a 360-degree panoramic image of the tread;

[0011] (3) Defect analysis and sample processing: The 360-degree panoramic image of the tread obtained in the above steps (1) and (2) is used as a model sample, including an image of the tread with a peeling defect and an image of a normal tread;

[0012] (4) Model training and generation: Using the large number of model samples obtained in step (3), the improved RT-DETR deep learning method is used to train and verify the detection model;

[0013] (5) Tread image detection and review: The collected image is processed through steps (1) and (2) and then input into the detection model for peeling defect detection. If an alarm is generated in the detection result, manual review is required.

[0014] Furthermore, the improved RT-DETR deep learning method: S4 and S5 are fused separately to meet the needs of smaller tread peeling detection, where S4 is a feature map extracted by the RT-DETR backbone network according to 16 steps, and S5 is a feature map extracted by the RT-DETR backbone network according to 32 steps.

[0015] Furthermore, the fusion processing is performed on S4 and S5 respectively:

[0016] 1) Perform feature dimensionality reduction processing on S4 and S5 respectively: Q4=K4=V4=Flatten(S4) and Q5=K5=V5=Flatten(S5), where Flatten represents feature dimensionality reduction processing; Q4, K4, V4 and Q5, K5, V5 are processing results respectively;

[0017] 2) Complete the information interaction within the scale: F4 = Reshape (AIFI (Q4, K4, V4)) and F5 = Reshape (AIFI (Q5, K5, V5)), where AIFI is the information interaction network within the scale, Reshape is the feature size change, and F4 and F5 are the results after size transformation respectively;

[0018] 3) Merge the fusion results: F = Concat (F4, F5), where Concat is the feature merging processing module and F is the merging processing result;

[0019] 4) Input the merged processing result F as well as S3 and S4 into the cross-scale fusion network to fuse the features between different scales, that is, O = CCFF({S3, S4, F}), where CCFF is the cross-scale fusion network, O is the fusion processing result, and serves as the input of the subsequent network.

[0020] Furthermore, the defect analysis and sample processing of step (3) are as follows: dense peeling defects appear on the entire circle of the wheel tread, and the number of peeling defects is recorded as N. N contour textures similar to the peeling shape will appear on the panoramic image; if the distance Dis between two adjacent peeling centers on the image is less than the distance threshold Dthresh between the two peeling defect centers on the image, then it is regarded as a whole, and the overall length is greater than the length threshold Lthresh of the peeling defect, and it can be marked as one defect sample; if Dis is not less than Dthresh, and the two peeling lengths are greater than Lthresh, then it is marked as two defect samples; according to this rule, on a panoramic tread image with N dense peelings, at most N defect samples can be marked.

[0021] Furthermore, step (4) model training and generation: the large number of model samples obtained in step (3) are first subjected to image contrast stretching to obtain sample data with more obvious tread peeling defect characteristics, and then the sample data are standardized and then used as training sets and / or verification sets for model training.

[0022] Furthermore, in step (5), the alarm level and defect size are calculated as follows: first, the lateral resolution and longitudinal resolution of the image are pre-calibrated, and then the actual physical length of the tread defect in the lateral and longitudinal directions is calculated by detecting the image pixel size of the tread defect, and finally the alarm level is determined based on the defect length and the number of defects detected.

[0023] Furthermore, 1) the horizontal resolution and vertical resolution of the image are pre-calibrated, which are recorded as Xdpi mm / pixel and Ydpi mm / pixel respectively;

[0024] 2) The pixel sizes of the detected tread defects in the horizontal and vertical directions are recorded as Xp pixels and Yp pixels respectively;

[0025] 3) Calculate the actual physical quantities of the tread defects in the horizontal and vertical directions, record them as Xm mm and Ym mm respectively, that is, Xm = Xp*Xdpi, Ym = Yp*Ydpi;

[0026] 4) The hierarchical alarm is set according to the following steps, namely:

[0027] (a): Determine the sizes of Xm and Ym, and take the maximum value between the two, denoted as L = max(Xm, Ym);

[0028] (b): Determine the number of detected tread defects, denoted as Num;

[0029] (c): Set the grading alarm limit levels as r1, r2, r3 respectively, with the unit being millimeters.

[0030] If r1 ≤ L < r2 and Num < 3, then the system finally displays a level-three alarm;

[0031] If r2 ≤ L ≤ r3 and 3 ≤ Num ≤ 6, then the system finally displays a level-two alarm;

[0032] If L > r3 and Num > 6, then the system finally displays a level-one alarm.

[0033] Furthermore, before the tread image correction and stitching in the step (2), it is necessary to perform image segmentation processing on the train wheel image to obtain the tread image.

[0034] Furthermore, the interpolation processing includes nearest-neighbor interpolation, bilinear interpolation or cubic interpolation.

[0035] Furthermore, a plurality of linear array camera modules are arranged on both sides of the track, and the linear array camera module includes a linear array camera and a laser emitter.

[0036] The method for detecting the whole-circle dense peeling of the train wheel tread of the present invention adopts a non-contact detection method based on visual detection technology, analyzes and studies from the aspects of peeling defect data characteristics and deep learning algorithms, and realizes the non-destructive detection of common tread peeling and whole-circle tread peeling defects of the wheel set. By using the relative motion between the stationary linear array camera and the moving train, the imaging of the wheel set tread is realized. Since the imaging field of view is large, the collected original image has interference from both the tread area and the non-tread area. Therefore, it is necessary to process the original image, extract the wheel tread area, and realize the panoramic display of the wheel tread through methods such as correction, stretching and stitching. Finally, the detection of defects such as whole-circle peeling is realized on the panoramic tread image through deep learning, and there will be no missed detection of tread defects, and the detection results are comprehensive and reliable. Brief Description of the Drawings

[0037] Figure 1 is the overall flowchart of the present invention;

[0038] Figure 2 is the original tread diagram of a single wheel collected in the embodiment: (a) The first original tread diagram of a single wheel; (b) The second original tread diagram of a single wheel; (c) The third original tread diagram of a single wheel; (d) The fourth original tread diagram of a single wheel;

[0039] Figure 3 The tread diagrams of individual wheels after correction and flattening in the embodiments are as follows: (a) the tread diagram of the first individual wheel; (b) the tread diagram of the second individual wheel; (c) the tread diagram of the third individual wheel; (d) the tread diagram of the fourth individual wheel;

[0040] Figure 4 is a panoramic mosaic image of a single wheel tread in the embodiment;

[0041] Figure 5 is a detection effect diagram on a panoramic view of the tread in the embodiment; DETAILED DESCRIPTION

[0042] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0043] Embodiment 1:

[0044] from Figure 1 It can be seen that the train wheel tread full circle dense peeling detection method of this embodiment has five main implementation steps, namely tread image acquisition, tread image correction and splicing, defect analysis and sample processing, model training, reasoning and alarm review, which are as follows:

[0045] Step 1: Acquisition of tread image

[0046] In this embodiment, eight linear array camera modules (including linear array cameras and laser transmitters) are placed on both sides of the track according to the rules, and four camera modules are installed on each side at intervals of 1 meter. When a train enters the line, the linear array camera module obtains the incoming signal and starts to collect wheel tread images at the same time. The collected images are arranged according to the module layout and the order of wheel pair entry, and the image data is transmitted through the TCP protocol.

[0047] like Figure 2 The following are four original tread images collected when a single train wheel passes through the linear array camera module. Since the wheel diameter is 1250 mm, the circumference of the wheel tread can be calculated to be approximately 3925 mm. Each linear array camera is set to shoot 1 meter, and the total length of the four linear array cameras can reach 4 meters, which is enough to cover the circumference of the wheel tread.

[0048] Step 2: Tread image correction and stitching

[0049] Since the collected images include not only tread areas but also non-tread areas, image segmentation processing is required (such as semantic segmentation processing technology based on deep learning U-net, three-level positioning segmentation processing technology, etc.) to obtain images of the tread area.

[0050] During the shooting and scanning process of the linear array camera, since the distance between the wheel tread and the linear array camera changes, the image taken by the wheel tread is curved in the vertical direction, and it needs to be corrected and flattened into a regular straight line shape to facilitate subsequent stitching processing.

[0051] Specifically,

[0052] (1) Position the segmented tread area image to obtain the boundary points on both sides of the tread.

[0053] (2) By processing these boundary points through curve fitting technology, two curves can be obtained, which are also the two boundary lines of the tread.

[0054] (3) The treads within the boundary line are normalized in the row direction. In the implementation example, according to the camera scanning resolution and the actual size of the tread width, the tread image width is normalized to 800 pixels, and bilinear interpolation is performed at the same time to obtain the regular tread shape after correction and flattening, as shown in FIG. Figure 3 As shown, there are four tread images after processing. It can be seen from the figure that the curved shape of the tread in the vertical direction has been corrected to a straight line shape.

[0055] (4) The tread images collected by different linear array camera modules are rectified and flattened in the above steps and then fused and stitched to obtain a 360-degree panoramic image of the wheel. Any abnormality on the wheel tread can be viewed from the panoramic image. In the implementation example, the stitched tread panoramic image is uniformly standardized to 13,000 pixels according to the circumference of the wheel tread, such as Figure 4 Shown is an image of the tread of a single wheel.

[0056] Step 3: Defect analysis and sample processing

[0057] Through step 2, a panoramic image of the wheel tread is obtained. Under the action of the linear array camera and the laser, the tread panoramic image is uniformly imaged, is not affected by ambient light, has no blind spots caused by interference from wheel components, has high image clarity, and can truly reflect the condition of the wheel tread.

[0058] If a real peeling defect occurs on the wheel tread, a peeling-like outline will appear on the tread panoramic image, and the outline texture will be more prominent than the surrounding textures under the action of the laser. If its length is greater than the set threshold Lthresh (stripping defect length threshold), the defective texture will be marked as a sample. In this embodiment, Lthresh is set to 60 pixels.

[0059] If there are two real peeling defects on the wheel tread, two contour textures similar to the peeling shape will appear on the panoramic image. The distance between the two peeling defect centers on the image is recorded as Dis. If Dis is less than the set threshold Dthresh (the distance threshold between the two peeling defect centers on the image), then it is taken as a whole, and the overall length of the two peeling defects is required to be greater than Lthresh (the length threshold of the peeling defect), then it can be marked as a defect sample; if Dis is not less than Dthresh, and the two peeling lengths are greater than Lthresh, then it is marked as two defect samples for subsequent deep learning detection model training. In this embodiment, Dthresh is set to 10 pixels.

[0060] If there are dense peeling defects on the entire wheel tread, and the number of peeling defects is N, then N contour textures similar to peeling shapes will appear on the panoramic image. A single peeling texture is slender in the direction perpendicular to the track, and the pixel distance between adjacent peeling defect textures is close. If the distance Dis between two adjacent peeling centers on the image is less than Dthresh (the distance threshold between the two peeling defect centers on the image), then it is regarded as a whole, and the overall length is greater than Lthresh (the length threshold of the peeling defect), and it can be marked as one defect sample; if Dis is not less than Dthresh, and the two peeling lengths are greater than Lthresh, then it is marked as two defect samples; according to this rule, on a panoramic tread image with N dense peelings, at most N defect samples can be marked. Such data samples are more conducive to annotation processing and obtaining a large amount of peeling sample data. The number of defect samples is crucial to subsequent model training. The more the number, the more objectively it can reflect the distribution of peeling defects on the wheel tread, and it can also further prompt the model to learn more accurate tread peeling texture features during training.

[0061] Step 4: Model training and generation

[0062] In order to verify the effectiveness of the tread dense peeling defect data sample processing method, this embodiment adopts a supervised deep learning method based on the Transformer framework, namely Real-Time Detection Transformer (hereinafter referred to as RT-DETR) to realize the training and reasoning of the tread defect detection model. RT-DETR is an end-to-end target object detector that can maintain a balance between detection accuracy and reasoning speed. Based on the Transformer ideological network framework, it also introduces an image convolution-based backbone network and an efficient hybrid encoder to obtain a higher reasoning speed. By decoupling intra-scale interactions and cross-scale fusion, it can efficiently process multi-scale features.

[0063] RT-DETR extracts three scales of output from the backbone network, with step sizes of 8, 16, and 32, respectively, denoted as S3, S4, and S5. Since only S5 is fused, the model has a higher detection rate for large targets and a lower detection rate for small targets; however, in the detection of dense peeling defects on the tread, there are a large number of small target peeling defects. In order to detect the peeling defects of these smaller targets, RT-DETR needs to be improved. The present invention fuses S4 (S4 is a feature map extracted by the RT-DETR backbone network according to a step size of 16) and S5 (S5 is a feature map extracted by the RT-DETR backbone network according to a step size of 32) to meet the needs of smaller tread peeling detection.

[0064] Specifically:

[0065] (1) Perform feature dimensionality reduction processing on S4 and S5 respectively, that is, Q4=K4=V4=Flatten(S4) and Q5=K5=V5=Flatten(S5), where Flatten represents the feature dimensionality reduction processing module, and Q4, K4, V4 and Q5, K5, V5 are the processing results respectively;

[0066] (2) Complete the information interaction within the scale, F4 = Reshape (AIFI (Q4, K4, V4)) and F5 = Reshape (AIFI (Q5, K5, V5)), where AIFI is the information interaction network module within the scale, Reshape is the feature size change module, and F4 and F5 are the results after size transformation respectively;

[0067] (3) Merge the fusion results, F = Concat (F4, F5), where Concat is the feature merging processing module and F is the merging processing result;

[0068] (4) The above-mentioned combined processing results F and S3, S4 are input into the cross-scale fusion network to fuse the features between different scales, that is, O = CCFF({S3, S4, F}), where CCFF is the cross-scale fusion network module and O is the fusion processing result, which serves as the input of the subsequent network.

[0069] Through step three, a large amount of labeled tread peeling sample data can be obtained. First, image contrast stretching processing is performed to obtain sample data with more obvious tread peeling defect features; then the sample data is standardized to meet the data interface of the improved RT-DETR deep learning method during training and reasoning. In this embodiment, when using the improved RT-DETR method to train the detection model, positive sample data and negative sample data are first required, that is, tread images with peeling defects and normal tread images, and then they are divided into training sets and verification sets according to a certain ratio (in this embodiment, the positive and negative sample ratio is 2:1), the training set is specifically used for model training, and the verification set is used to test the performance of model training online; secondly, the model hyperparameters are pre-set and trained in the built server. When the model is trained, a model for detecting tread peeling defects will be generated every 20 times. When the number of training times reaches the preset parameter value, that is, 160 times, the model training is completed, and 8 detection models will be generated at the same time; finally, the detection model is tested with the verification set, and a model with higher accuracy and recall rate is selected. In this embodiment, a model trained 120 times is selected for subsequent model deployment and reasoning.

[0070] Step 5: Tread image detection and review

[0071] After the above steps, a detection model with better performance can be obtained; after the collected tread image is processed by steps one and two, it can be input into the model for peeling defect detection.

[0072] If an alarm is detected in the test results, the owner needs to be notified to manually check to determine the alarm level and conduct an on-site inspection to verify the size of the defect to confirm whether the alarm is valid.

[0073] In this embodiment, the alarm level and defect size are calculated as follows:

[0074] First, the horizontal and vertical resolutions of the image are pre-calibrated. Then, the image pixel size of the tread defect is detected by the system, and the actual physical length of the tread defect in the horizontal and vertical directions is calculated. Finally, the alarm level is determined according to the defect length and the number of defects detected. Specifically:

[0075] (1) Pre-calibrate the horizontal resolution and vertical resolution of the image, which are recorded as Xdpi mm / pixel and Ydpi mm / pixel respectively. In this embodiment, Xdpi=Ydpi=0.15 mm / pixel;

[0076] (2) The pixel sizes of the tread defects detected by the system in the horizontal and vertical directions are recorded as Xp pixels and Yp pixels respectively;

[0077] (3) Calculate the actual physical quantities of the tread defects in the horizontal and vertical directions, record them as Xm mm and Ym mm respectively, that is, Xm = Xp*Xdpi, Ym = Yp*Ydpi.

[0078] (4) The hierarchical alarm is set according to the following steps, that is

[0079] (a): Determine the sizes of Xm and Ym, and take the maximum value between the two, denoted as L = max(Xm, Ym);

[0080] (b): Determine the number of detected tread defects in the system, denoted as Num;

[0081] (c): Set the hierarchical alarm limit levels as r1, r2, r3 respectively, with the unit being millimeters.

[0082] If r1 ≤ L < r2 and Num < 3, then the system finally displays a level-three alarm;

[0083] If r2 ≤ L ≤ r3 and 3 ≤ Num ≤ 6, then the system finally displays a level-two alarm;

[0084] If L > r3 and Num > 6, then the system finally displays a level-one alarm;

[0085] In this embodiment, r1 = 15, r2 = 30, r3 = 60, and the hierarchical alarm limit range will be adjusted according to different customer requirements.

[0086] If there is no alarm in the detection result, then the wheel tread of this train is normal, and the data of the next passing train is continued to be processed.

[0087] As Figure 5 shown, it is the detection result after processing the above steps. It can be seen from the figure that there are dense peeling defects throughout the entire circumference of this wheel. After being reviewed and verified by the owner, it has been included in the wheel turning maintenance plan.

[0088] The present invention has been disclosed above with preferred embodiments, but they are not used to limit the present invention. Anyone familiar with this art can make various changes or modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention should be defined by the protection scope of the claims of this application.

Claims

1. A method for detecting the dense peeling of the entire circle of train wheel tread, It is characterized in that The method comprises: (1) Tread image acquisition: multiple linear array camera modules are used to collect images of the wheels of a moving train; (2) Tread image correction and stitching: locate the tread image to obtain the boundary points on both sides of the tread; perform curve fitting on the boundary points to obtain two boundary lines of the tread; normalize and interpolate the tread within the boundary line according to the row direction to obtain a regular tread image after correction and flattening; tread images collected by different linear array camera modules are fused and stitched after correction and flattening to obtain a 360-degree panoramic image of the tread; (3) Defect analysis and sample processing: The 360-degree panoramic image of the tread obtained in the above steps (1) and (2) is used as a model sample, including an image of the tread with a peeling defect and an image of a normal tread; (4) Model training and generation: Using the large number of model samples obtained in step (3), the improved RT-DETR deep learning method is used to train and verify the detection model; (5) Tread image detection and review: The collected image is processed through steps (1) and (2) and then input into the detection model for peeling defect detection. If an alarm is generated in the detection result, manual review is required.

2. A method for detecting intensive peeling of a train wheel tread as claimed in claim 1, It is characterized in that The improved RT-DETR deep learning method: S4 and S5 are fused separately to meet the needs of smaller tread peeling detection, where S4 is a feature map extracted by the RT-DETR backbone network according to 16 steps, and S5 is a feature map extracted by the RT-DETR backbone network according to 32 steps.

3. A method for detecting intensive peeling of a train wheel tread as claimed in claim 2, It is characterized in that The fusion processing of S4 and S5 is performed respectively: 1) Perform feature dimensionality reduction processing on S4 and S5 respectively: Q4=K4=V4=Flatten(S4) and Q5=K5=V5=Flatten(S5), where Flatten represents feature dimensionality reduction processing; Q4, K4, V4 and Q5, K5, V5 are processing results respectively; 2) Complete the information interaction within the scale: F4 = Reshape (AIFI (Q4, K4, V4)) and F5 = Reshape (AIFI (Q5, K5, V5)), where AIFI is the information interaction network within the scale, Reshape is the feature size change, and F4 and F5 are the results after size transformation respectively; 3) Merge the fusion results: F = Concat (F4, F5), where Concat is the feature merging processing module and F is the merging processing result; 4) The combined processing result F as well as S3 and S4 are input into the cross-scale fusion network to fuse the features of different scales, that is, O=CCFF({S3, S4, F}), where CCFF is the cross-scale fusion network and O is the fusion processing result.

4. A method for detecting intensive peeling of a train wheel tread as claimed in any one of claims 1 to 3, It is characterized in that Defect analysis and sample processing in step (3): If dense peeling defects appear throughout the entire circumference of the wheel tread, record the number of peeling defects as N. N contour textures similar to the peeling shape will appear in the panoramic image. If the distance Dis between the centers of two adjacent peelings in the image is less than the distance threshold Dthresh between the centers of two peeling defects in the image, and the overall length is greater than the length threshold Lthresh of the peeling defect, then it can be marked as one defect sample. If Dis is not less than Dthresh and the lengths of the two peelings are greater than Lthresh, then they are marked as two defect samples. According to this rule, in the panoramic tread image with N dense peelings, at most N defect samples can be marked.

5. A method for detecting dense peeling throughout the entire circumference of a train wheel tread according to any one of claims 1 - 3, characterized in that, In step (4), model training and generation: For the large number of model samples obtained in step (3), first perform image contrast stretching to obtain sample data with more obvious tread peeling defect features, then perform normalization processing on the sample data, and then use it as the training set and / or validation set for model training.

6. A method for detecting dense peeling throughout the entire circumference of a train wheel tread according to any one of claims 1 - 3, characterized in that, In step (5), the alarm level and defect size are calculated as follows: First, pre - calibrate the horizontal resolution and vertical resolution of the image. Secondly, calculate the actual physical lengths of the tread defects in the horizontal and vertical directions by detecting the pixel sizes of the tread defects in the image. Finally, determine the alarm level based on the defect length and the number of detected defects.

7. A method for detecting dense peeling throughout the entire circumference of a train wheel tread according to claim 6, characterized in that, In step (5), the alarm level and defect size are calculated as follows: 1) Pre - calibrate the horizontal resolution and vertical resolution of the image, denoted as Xdpi mm / pixel and Ydpi mm / pixel respectively; 2) Denote the pixel sizes of the detected tread defects in the horizontal and vertical directions as Xp pixels and Yp pixels respectively; 3) Calculate the actual physical quantities of the tread defects in the horizontal and vertical directions, denoted as Xm mm and Ym mm respectively, that is, Xm = Xp * Xdpi, Ym = Yp * Ydpi; 4) The hierarchical alarm is set according to the following steps, that is (a): Determine the sizes of Xm and Ym, and take the maximum value between the two, denoted as L = max(Xm, Ym); (b): Determine the number of detected tread defects, denoted as Num; (c): Set the hierarchical alarm limit levels as r1, r2, r3 respectively, with the unit of mm, If r1 ≤ L < r2 and Num < 3, then the system finally displays a level - three alarm; If r2 ≤ L ≤ r3 and 3 ≤ Num ≤ 6, then the system finally displays a level - two alarm; If L > r3 and Num > 6, then the system finally displays a level - one alarm.

8. A method for detecting dense peeling throughout the entire circumference of a train wheel tread according to any one of claims 1 - 3, characterized in that, Before the tread image correction and splicing in step (2), the train wheel image needs to be segmented to obtain a tread image.

9. A method for detecting dense peeling of a train wheel tread as claimed in any one of claims 1 to 3, It is characterized in that The interpolation process includes nearest neighbor interpolation, bilinear interpolation or cubic interpolation.

10. A method for detecting dense peeling of a train wheel tread as claimed in any one of claims 1 to 3, It is characterized in that A plurality of linear array camera modules are arranged on both sides of the track, and the linear array camera modules include linear array cameras and laser transmitters.

Images