Abnormality identification method for rubber sealing element of anti-side-rolling torsion bar connecting rod of subway

Abstract

The invention provides a subway anti-side-rolling torsion bar connecting rod rubber sealing element abnormity identification method, which comprises the steps of obtaining an image of a bogie anti-side-rolling torsion bar connecting rod rubber sealing element through a camera on a subway vehicle bottom inspection robot, and segmenting the rubber sealing element by using a semantic segmentation network based on a deep neural network. And the widths of visible parts on the left and right sides of the rubber part are obtained by image processing. Then, the width difference value of the visible parts on the left side and the right side is calculated and compared with a set threshold value so as to judge whether the anti-side-rolling torsion bar connecting rod rubber sealing piece is abnormal or not, an alarm is triggered when necessary, automatic detection and intelligent judgment are achieved, and digital and intelligent transformation of a train maintenance system is assisted.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent rail transit equipment, and in particular to a method for identifying abnormalities in a subway anti-roll torsion bar connecting rod rubber seal. Background Art

[0002] Urban rail transit is an important mode of transportation, and the smooth operation and ride comfort of subway trains have a significant impact on the passenger experience. Anti-roll torsion bars are usually connected to the car body or bogie through connecting rods. Their motion nodes use ball joint structures or rubber bushings to meet a certain degree of rotational freedom, and are injected with grease to reduce friction and extend service life. In order to prevent impurities such as dust and water vapor from invading the connection parts, or grease leakage causing lubrication failure, rubber seals are usually provided at the connecting rod connections. Rubber seals are exposed to high-frequency vibration, high temperature differences, and complex environments for a long time, and are prone to aging, cracking, and damage, resulting in seal failure. Once grease leaks or impurities enter, it will cause the ball joint to jam, rotation to fail, or even damage the mechanical structure. In severe cases, it may affect the normal operation of the anti-roll system and endanger the safety of train operation.

[0003] Therefore, during routine inspections after daily subway train stops, the rubber seals on the anti-roll torsion bar links must be checked for damage and oil leakage to ensure reliable system operation. Currently, this inspection relies primarily on manual visual inspection, a cumbersome process with inconsistent standards and the risk of missed detections and misjudgments. To improve inspection efficiency and accuracy, a reliable rubber seal anomaly detection solution is urgently needed that enables automatic detection and intelligent judgment, thereby facilitating the digital and intelligent transformation of train maintenance systems. Summary of the Invention

[0004] To address the challenges of the prior art, this paper provides a method for identifying abnormalities in the rubber seals of subway anti-roll torsion bar links. Using a camera mounted on a subway undercar inspection robot, the system captures images of the rubber seals in the bogie anti-roll torsion bar links. A semantic segmentation network based on a deep neural network is used to segment the rubber seals, and image processing is used to determine the widths of the visible areas on the left and right sides of the rubber seals. The difference in width between the visible areas is then calculated and compared with a set threshold to determine if the rubber seals are abnormal. An alarm is triggered when necessary, enabling automated detection and intelligent judgment, thereby facilitating the digital and intelligent transformation of train maintenance systems.

[0005] The present invention provides a method for identifying abnormalities in a subway anti-roll torsion bar connecting rod rubber seal, comprising the following steps:

[0006] Step 1) Before performing the inspection task, use the camera on the inspection robot to capture an image of the anti-roll torsion bar connecting rod rubber seal in a normal state as a template image, and select the minimum circumscribed rectangle of the rubber seal in the image;

[0007] Step 2) When performing the inspection task, use the camera on the inspection robot to capture an image of the rubber seal at the same position and angle as the task image;

[0008] Step 3) performing image registration on the rubber seal task image and the template image obtained in steps 1) and 2), mapping the annotation box in the template image to the task image, and obtaining the circumscribed rectangular box of the rubber seal in the task image, i.e., the area where the rubber seal is located on the task image;

[0009] Step 4) Extract the rubber seal area image from the task image in step 3) and feed it into the rubber seal semantic segmentation network. Using a semantic segmentation model based on a deep convolutional network, the backbone network extracts multi-scale features and outputs a pixel-level prediction mask of the same size as the original image. Foreground pixels with a value of 1 in the mask represent the rubber seal, and background pixels have a value of 0.

[0010] Step 5) The rubber seal segmentation mask obtained in step 4) is divided into n regions according to the shooting direction, and the width of the mask is calculated for each region, as follows:

[0011] 5.1) Using the vertical axis of the segmentation mask's bounding rectangle in the image as the reference axis, divide the interior of the bounding rectangle into n equally spaced straight lines, and record the intersection of each line with the segmentation mask outline;

[0012] 5.2) For each line, calculate the horizontal distance between it and the leftmost and rightmost pixels of the two segmentation mask contours, and use it as the width w of the segmentation mask in the horizontal direction. i , the width calculation formula is:

[0013]

[0014] in and Respectively represent the minimum and maximum abscissa values ​​of the intersection of the i-th straight line and the j-th mask contour;

[0015] 5.3) If there are no foreground pixels in a region, i.e., it is completely occluded or the segmentation fails, then the width is recorded as 0 and is discarded or ignored in subsequent processing. Otherwise, 2n width values ​​are output, reflecting the width distribution of the left and right sides of the rubber seal at different vertical positions;

[0016] Step 6) After removing the widths corresponding to the top and bottom regions of the rubber seal segmentation mask, the remaining widths of the visible portions on the left and right sides are averaged; the difference between the mean values ​​of the left and right widths is then determined. If the difference is greater than a certain threshold, it indicates that the rubber seal has deviated to one side. Specifically,

[0017] 6.1) Eliminate the width values ​​corresponding to the top and bottom boundary areas on the left and right sides;

[0018] 6.2) Group the remaining 2n-4 width values ​​symmetrically and calculate the n-2 width values ​​w2, w3, ..., w on the left side respectively. n-1 and the right n-2 width values ​​w n+2 ,w n+3 ,...,w 2n-1 The average width is:

[0019]

[0020] 6.3) Calculate the difference between the left and right average widths: If the difference value Δw exceeds the set threshold value T, it means that the visible part of the rubber seal has obvious left-right asymmetry, and it is inferred that there is an abnormal trend of leaning to one side; otherwise, it is judged to be in a normal state.

[0021] Further improvement, the image registration method in step 3) is specifically as follows:

[0022] 3.1) Use feature extraction and matching algorithms to extract image feature points from the template image and the task image respectively, and obtain preliminary corresponding point pairs between the two images through descriptor matching;

[0023] 3.2) After removing the mismatched points, the least squares method is used to estimate the affine transformation moment between the two images;

[0024] 3.3) The minimum bounding rectangle manually annotated on the template image in step 1) is mapped to the task image using an affine matrix to obtain the location area of ​​the rubber seal on the task image. This area serves as the input of the subsequent semantic segmentation network.

[0025] Further improvements are made to the deep convolutional network-based semantic segmentation model described in step 4), where manually annotated segmentation masks are used as supervisory signals during the network training phase and optimized using a cross-entropy loss function. The deep convolutional network-based semantic segmentation model described in step 4) incorporates a multi-scale context fusion module and an edge attention mechanism to maintain good performance in real-world situations such as local occlusion, uneven lighting, and morphological changes. In the prediction phase, a binary segmentation mask is directly output, providing basic data support for subsequent rubber seal offset detection.

[0026] The beneficial effects of the present invention are as follows: using a camera on a subway car inspection robot to capture images of the bogie anti-roll torsion bar link rubber seal, segmenting the rubber seal using a semantic segmentation network based on a deep neural network, and using image processing to determine the widths of the visible areas on the left and right sides of the rubber seal. The difference in width between the visible areas on the left and right sides is then calculated and compared with a set threshold to determine whether the anti-roll torsion bar link rubber seal is abnormal. If necessary, an alarm is triggered, achieving automatic detection and intelligent judgment, and facilitating the digital and intelligent transformation of train maintenance systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0028] Figure 1 Flowchart for the execution of the process of the present invention;

[0029] Figure 2 Schematic diagram of the anti-roll torsion bar link rubber seal;

[0030] Figure 3 Schematic diagram of the segmentation mask for the anti-roll torsion bar link rubber seal;

[0031] Figure 4 Example diagram for calculating mask width for a rubber seal. DETAILED DESCRIPTION

[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0033] The present invention aims to provide a deep learning-based system for identifying anomalies in the rubber seals of anti-roll torsion bar links. This system uses a camera mounted on a subway car inspection robot to capture images of the rubber seals in the bogie's anti-roll torsion bar links. It then segments the rubber seals using a semantic segmentation network based on a deep neural network. Image processing is then used to determine the widths of the visible left and right sections of the rubber seals. The system then calculates the difference in width between the two visible sections and compares it with a set threshold to determine if the rubber seals are abnormal and trigger an alarm if necessary.

[0034] The technical solution adopted by the present invention is specifically implemented according to the following steps: Figure 1 As shown:

[0035] Step 1: Before the inspection, use the robot's camera to capture a normal image of the anti-roll torsion bar link rubber seal as a template. The camera should be facing the rubber seal. A minimum bounding rectangle (LR) of the rubber seal should be manually selected in the image.

[0036] Step 2: When performing the inspection task, use the camera on the inspection robot to take an image of the rubber seal at the same position and angle as the task image. The camera should face the rubber seal when shooting.

[0037] Step 3: Perform image registration on the rubber seal task image and template image obtained in steps 1 and 2, map the annotation box in the template image to the task image, and obtain the circumscribed rectangular box of the rubber seal in the task image, that is, the area where the rubber seal is located on the task image.

[0038] Step 4: Extract the rubber seal area image from the task image in step 3 and feed it into the rubber seal semantic segmentation network to obtain the rubber seal segmentation mask.

[0039] Step 5: The rubber seal segmentation mask obtained in step 4 is divided into n regions according to the shooting direction, and the width of the mask is calculated for each region.

[0040] Step 6: After removing the widths corresponding to the top and bottom regions of the rubber seal segmentation mask, calculate the average of the remaining widths of the visible areas on the left and right sides. Then, determine the difference between the averages of the left and right widths. If the difference is greater than a certain threshold, it indicates that the rubber seal is deviating to one side.

[0041] A specific embodiment of the present invention is as follows:

[0042] Step 1: Template data collection:

[0043] Before performing daily inspection tasks, subway inspection robots need to collect images of rubber seals in a normal installation state as reference template images for subsequent inspection comparisons. The present invention selects images with relatively standard shooting angles and stable rubber seals as template images. Rubber seals are installed at the bottom of subway vehicles and occupy a small proportion of the entire image, making direct image recognition difficult. In order to improve the accuracy of subsequent image recognition, it is necessary to manually mark the minimum circumscribed rectangular area of ​​the rubber seal in the template image as the standard position frame of the rubber seal for subsequent image registration and local detection.

[0044] Step 2: Task graph collection:

[0045] During an inspection, the undercar inspection robot captures an image of the vehicle's rubber seal at the same location and angle. This image serves as the task map. The template map only needs to be collected once or updated as needed, eliminating the need for repeated collection with each inspection. The task map, however, is captured in real time during each inspection and serves as the basis for determining the correct task for that round of inspections.

[0046] Step 3: Image registration and target box mapping:

[0047] The key to determining whether the anti-roll torsion bar link rubber seal has an offset anomaly is to measure and compare the widths of the visible areas at both ends, and this measurement relies on the accurate segmentation mask of the seal in the image. Due to the complex environment under the vehicle and the seal being partially obscured by the structure, direct mask segmentation from the entire image is susceptible to background interference. Therefore, manually labeling prior information can effectively reduce the difficulty of the task. In order to accurately extract the position of the rubber seal area in the task image, the present invention uses an image registration method to obtain the mapping relationship between the task image and the template image.

[0048] First, the feature extraction and matching algorithm is used to extract the image feature points in the template image and the task image respectively, and the initial corresponding point pairs between the two images are obtained by descriptor matching. After removing the mismatched points, the least squares method is used to estimate the affine transformation moment M between the two images. affine The manually annotated minimum bounding rectangle on the template image in step 1 is mapped to the task image using an affine matrix to obtain the location of the rubber seal on the task image. This area serves as the input to the subsequent semantic segmentation network.

[0049] Through the image registration method, even if there is a slight difference between the task image and the template image in shooting angle or viewing angle, the corresponding area of ​​the rubber seal in the task image can still be accurately obtained through geometric mapping, thereby improving the robustness and adaptability of the system.

[0050] Step 4: Get the rubber seal segmentation mask:

[0051] To improve segmentation accuracy and robustness, the present invention appropriately expands the minimum bounding rectangle of the rubber seal obtained in step 3 and then crops the local image region containing the rubber seal from the task image as the input image for the neural network. The image segmentation of the rubber seal uses a semantic segmentation model based on a deep convolutional network. The backbone network is responsible for extracting multi-scale features and outputting a pixel-level prediction mask of the same size as the original image. In the mask, foreground pixels represent the rubber seal (value 1), while background pixels have a value of 0.

[0052] The network training phase uses manually annotated segmentation masks as supervisory signals and uses a cross-entropy loss function for optimization. To further enhance the ability to recognize boundaries, the network introduces a multi-scale context fusion module and an edge attention mechanism, enabling it to maintain good performance in real-world situations such as partial occlusion, uneven lighting, and morphological changes. The prediction phase directly outputs a binary segmentation mask, providing basic data support for subsequent rubber seal offset detection. Image examples and processing results are shown in the figure below. Figure 2 and Figure 3 As shown in the figure, the segmentation result can effectively fit the outer contour of the rubber seal and has practical engineering value.

[0053] Step 5: Mask division and width calculation, such as Figure 4 As shown:

[0054] The rubber seal segmentation mask obtained in step 4 is a binary image, with the foreground region representing the visible portion of the rubber seal. In practice, the seal itself is obscured by the vehicle structure, and only the symmetry of its two ends can be observed. Therefore, this step aims to assess the visible width of the rubber seal on both sides of the horizontal image to support subsequent offset determination.

[0055] First, using the vertical axis of the segmentation mask's bounding rectangle in the image as the reference axis, divide the bounding rectangle into 10 equally spaced straight lines. Record the intersection of each line with the segmentation mask outline. For each line, calculate the horizontal distance between it and the leftmost and rightmost pixels of the two segmentation mask outlines. This is the width w_i of the segmentation mask in that horizontal direction. The width calculation formula is:

[0056]

[0057] in and where represents the minimum and maximum abscissa values ​​of the intersection of the i-th line and the j-th mask outline. If a region contains no foreground pixels (i.e., completely occluded or segmentation failed), the width is recorded as 0 and is subsequently discarded or ignored. This process will output 20 width values, representing the width distribution of the left and right sides of the rubber seal at different vertical positions.

[0058] Step 6: Determine the difference in left and right visible widths:

[0059] Since the top and bottom of the rubber seal are often affected by strong occlusion or incomplete segmentation, their width values ​​have large errors. To improve the robustness of offset judgment, this step first removes the width values ​​corresponding to the top and bottom boundary areas on the left and right sides. The remaining 16 width values ​​are symmetrically grouped, and the 8 width values ​​on the left side, w2, w3, ..., w9 (the visible part on the left side of the rubber seal) and the 8 width values ​​on the right side, w12 ,w 13 ,...,w 19 The average width of the visible part on the right side of the rubber seal is recorded as:

[0060]

[0061] Then calculate the difference between the left and right average widths: If the difference Δw exceeds the set threshold T, it indicates that the visible portion of the rubber seal has significant left-right asymmetry, suggesting an abnormal tendency toward one side. Otherwise, it is considered normal. This judgment criterion takes into account both structural stability and practical application tolerances, and an appropriate threshold can be set through experimentation to suit different vehicle models.

[0062] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, the above is only a preferred embodiment of the present invention. Since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited to this. Any technical personnel familiar with this technical field is within the technical scope disclosed by the present invention. For ordinary technical personnel in this technical field, changes or replacements that can be easily thought of should be covered within the protection scope of the present invention without departing from the principle of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for identifying abnormalities in rubber seals of subway anti-roll torsion bar connecting rods, characterized in that The following steps are involved: Step 1) Before performing the inspection task, use the camera on the inspection robot to capture an image of the anti-roll torsion bar connecting rod rubber seal in a normal state as a template image, and select the minimum circumscribed rectangle of the rubber seal in the image; Step 2) When performing the inspection task, use the camera on the inspection robot to capture an image of the rubber seal at the same position and angle as the task image; Step 3) performing image registration on the rubber seal task image and the template image obtained in steps 1) and 2), mapping the annotation box in the template image to the task image, and obtaining the circumscribed rectangular box of the rubber seal in the task image, i.e., the area where the rubber seal is located on the task image; Step 4) Extract the rubber seal area image from the task image in step 3) and feed it into the rubber seal semantic segmentation network. Using a semantic segmentation model based on a deep convolutional network, the backbone network extracts multi-scale features and outputs a pixel-level prediction mask of the same size as the original image. Foreground pixels with a value of 1 in the mask represent the rubber seal, and background pixels have a value of 0. Step 5) The rubber seal segmentation mask obtained in step 4) is divided into n regions according to the shooting direction, and the width of the mask is calculated for each region, as follows: 5.1) Using the vertical axis of the segmentation mask's bounding rectangle in the image as the reference axis, divide the interior of the bounding rectangle into n equally spaced straight lines, and record the intersection of each line with the segmentation mask outline; 5.2) For each line, calculate the horizontal distance between it and the leftmost and rightmost pixels of the two segmentation mask contours, and use it as the width w of the segmentation mask in the horizontal direction. i , the width calculation formula is: in and Respectively represent the minimum and maximum abscissa values ​​of the intersection of the i-th straight line and the j-th mask contour; 5.3) If there are no foreground pixels in a region, i.e., it is completely occluded or the segmentation fails, then the width is recorded as 0 and is discarded or ignored in subsequent processing. Otherwise, 2n width values ​​are output, reflecting the width distribution of the left and right sides of the rubber seal at different vertical positions; Step 6) After removing the widths corresponding to the top and bottom regions of the rubber seal segmentation mask, the remaining widths of the visible portions on the left and right sides are averaged; the difference between the mean values ​​of the left and right widths is then determined. If the difference is greater than a certain threshold, it indicates that the rubber seal has deviated to one side. Specifically, 6.1) Eliminate the width values ​​corresponding to the top and bottom boundary areas on the left and right sides; 6.2) Group the remaining 2n-4 width values ​​symmetrically and calculate the n-2 width values ​​w2, w3, ..., w on the left side respectively. n-1 and the right n-2 width values ​​w n+2 ,w n+3 ,...,w 2n-1 The average width is: 6.3) Calculate the difference between the left and right average widths: If the difference value Δw exceeds the set threshold value T, it means that the visible part of the rubber seal has obvious left-right asymmetry, and it is inferred that there is an abnormal trend of leaning to one side; otherwise, it is judged to be in a normal state.

2. The method for identifying abnormalities in subway anti-roll torsion bar connecting rod rubber seals according to claim 1, characterized in that: Step 3) The image registration method is specifically as follows: 3.1) Use feature extraction and matching algorithms to extract image feature points from the template image and the task image respectively, and obtain preliminary corresponding point pairs between the two images through descriptor matching; 3.2) After removing the mismatched points, the least squares method is used to estimate the affine transformation moment between the two images; 3.3) The minimum bounding rectangle manually annotated on the template image in step 1) is mapped to the task image using an affine matrix to obtain the location area of ​​the rubber seal on the task image. This area serves as the input of the subsequent semantic segmentation network.

3. The method for identifying abnormalities in subway anti-roll torsion bar connecting rod rubber seals according to claim 1, characterized in that: In step 4), the semantic segmentation model based on the deep convolutional network uses manually annotated segmentation masks as supervisory signals during the network training phase and is optimized using a cross-entropy loss function.

4. The method for identifying abnormalities in subway anti-roll torsion bar connecting rod rubber seals according to claim 1 or 3, characterized in that: In step 4), the semantic segmentation model based on the deep convolutional network is introduced with a multi-scale context fusion module and an edge attention mechanism, so that it can still maintain good performance when dealing with actual situations such as local occlusion, uneven lighting, and morphological changes. In the prediction stage, a binary segmentation mask is directly output to provide basic data support for subsequent rubber seal offset detection.

Images