Length and defect detection method for driving chain of escalator

Abstract

The invention relates to the technical field of special equipment detection, in particular to an escalator driving chain length and defect detection method. Three-dimensional point cloud data and a defect identification model of the chain are established by collecting image information of the driving chain at each angle, a smooth and continuous path is planned in the point cloud data by using an interpolation method, and then the path is segmented into line segments by key point cloud points for integral length calculation; a precise recognition model is obtained through a large amount of image training, verification and testing, so that the surface defects of the chain in the images are quickly and precisely captured and recognized. Compared with a traditional detection method, the detection cost and time can be saved, the detection efficiency can be improved, the labor intensity of workers can be reduced, the real-time state of the driving chain can be collected and fed back to detection personnel, and compared with original manual detection, the detection method is more convenient and direct, the accuracy of the detection result is higher, and the detection efficiency is improved. And the practical application value is higher.

Description

Technical Field

[0001] The present invention relates to the technical field of special equipment detection, in particular to a method for detecting the length and defects of an escalator drive chain. Background Art

[0002] Currently, there are three primary methods for measuring escalator drive chain length: manual, mechanical, and photoelectric sensor. When any of these methods detect abnormal chain expansion or contraction, they immediately shut down the main engine, providing protection. Manual measurement directly measures the chain length and is suitable for maintenance and troubleshooting. Mechanical measurement utilizes a mechanical structure to physically contact the chain and is suitable for static or low-speed environments. Photoelectric sensor measurement uses a non-contact photoelectric sensor to monitor the chain's motion and is suitable for dynamic, real-time monitoring.

[0003] Existing drive chain defect detection technologies include visual inspection, non-destructive testing (NDT), laser scanning, 3D imaging, and infrared testing. Visual inspection involves technicians directly observing the chain's exterior to identify visible defects. NDT uses ultrasonic, magnetic particle, and eddy current technologies to detect internal defects without damaging the object being inspected. Laser scanning utilizes high-precision laser scanners to obtain three-dimensional data of the chain's surface to assess wear and deformation. 3D imaging, combined with computer vision technology, generates a three-dimensional model of the chain for detailed analysis. Infrared testing uses thermal imaging to detect temperature anomalies and identify potential overheating areas.

[0004] However, in actual operation, when the main drive chain breaks, the microswitch of the safety monitoring device sometimes fails to respond. This may be due to excessive internal oil accumulation, excessive friction, or spring failure. To enhance the reliability of the drive chain safety protection system, researchers at home and abroad are focusing on the following aspects:

[0005] 1. Develop new chain materials with high strength and corrosion resistance to extend service life;

[0006] 2. Design a more sensitive and reliable chain break detection mechanism to ensure a rapid response when a break occurs;

[0007] 3. Improve lubricant formulation and technology to reduce wear and friction resistance;

[0008] 4. Introduce advanced sensing technology and data analysis algorithms to improve detection accuracy;

[0009] 5. Use programmable logic controller (PLC) to achieve automatic control and real-time monitoring;

[0010] 6. Use ultrasonic or infrasonic technology to detect changes in internal structures and provide early warning of potential problems.

[0011] While these methods improve the sensitivity and real-time response capabilities of protection systems, they mostly focus on post-event detection and lack comprehensive assessment of potential drive chain conditions and early warning capabilities. Currently, escalator safety inspections are still performed manually, a method that is not only inefficient and labor-intensive, but also prone to inaccurate measurements, poor repeatability, and overall low inspection efficiency.

[0012] In order to solve these problems, it is necessary to explore more efficient automated detection methods, provide more accurate and reliable detection results, and achieve early warning of potential safety hazards. Therefore, the present invention provides a method for detecting the length and defects of an escalator drive chain. Summary of the Invention

[0013] The technical problem to be solved by the present invention is to provide a method for detecting the length and defects of an escalator drive chain, thereby realizing accurate and automated measurement of the escalator drive chain and providing early warning of potential safety hazards.

[0014] To solve the above technical problems, the present invention provides a method for detecting the length and defects of an escalator drive chain, comprising the following steps:

[0015] S101. Arrange at least two cameras around the escalator drive chain to obtain clear image information of the drive chain at different viewing angles;

[0016] S102. Capture the center of the axis pin point in the image information by image recognition technology and use it as a positioning reference to perform position matching on the image information to obtain image information of the same position of the drive chain at different viewing angles;

[0017] S103. Establishing a three-dimensional point cloud data and defect recognition model of the drive chain based on the image information from different perspectives;

[0018] S104. Determine the starting point of the drive chain in the three-dimensional point cloud data, and plan a smooth and continuous path along the drive chain between the point cloud points corresponding to the known axle pin center points using natural cubic spline interpolation;

[0019] S105. Discrete the path into a series of short line segments and then accumulate the lengths of the line segments to obtain a length measurement result of the drive chain;

[0020] S106. Import the real-time collected drive chain image information into the defect recognition model to perform real-time monitoring of the chain surface state and defect recognition.

[0021] In some implementations, in S101 , before collecting image information of the drive chain, each camera is calibrated using the Zhang Zhengyou calibration method.

[0022] In some implementations, in S102 , a multi-scale Hough transform is used to locate the center of the pivot pin.

[0023] In some embodiments, the specific method for locating the center of the pivot pin point includes the following steps:

[0024] S201. Process the original image information by grayscale and Gaussian filtering to reduce noise interference and smooth the image;

[0025] S202. Extract edge information from the image using the Canny edge detection algorithm;

[0026] S203. Perform gradient-based directional histogram analysis and / or least squares circle fitting in the pivot point area to achieve preliminary sub-pixel positioning;

[0027] S204. Construct a multi-scale image pyramid covering multiple resolution levels to accommodate pivot points of different sizes, and determine the minimum and maximum scales for Hough transform based on the expected diameter range of the pivot points in actual application scenarios;

[0028] S205. Apply the standard Hough Circle Transform (HTC) at each scale level, adjust the accumulator grid size and voting threshold, and obtain candidate circles.

[0029] S206. Perform non-maximum suppression (NMS) on all candidate circles and retain the most representative detection results;

[0030] S207. Reversely screen and verify the test results through geometric constraints and / or known dimensions of the drive chain, and then output the shaft pin point position information.

[0031] In some embodiments, in S103 , before establishing the three-dimensional point cloud data and defect recognition model of the drive chain, the collected image information is subjected to image preprocessing; the image preprocessing includes image filtering, image enhancement, dilation, and erosion.

[0032] In some embodiments, in S105 , the point cloud points corresponding to the pivot points divide the continuous path into short line segments, and the line segments are imported into MATLAB to establish a curve equation, and then the actual length corresponding to each line segment is obtained by integration.

[0033] In some embodiments, in S106 , the at least 1,400 captured images are divided into a training set, a validation set, and a test set in a ratio of 8:1:1 to train the defect recognition model.

[0034] In some embodiments, all image information in the training set is subjected to random angle rotation and / or horizontal and / or vertical mirroring, and a color jittering operation is added for rust defects to increase data diversity.

[0035] In some embodiments, the LabelImg image annotation tool is used to add a bounding box to the target object in each image and label its category to ensure that the annotation information is accurate and can correctly reflect the location and type of the target;

[0036] According to the characteristics of the processed dataset, the YOLOv8 configuration file is adjusted in real time for model training to ensure that the model gradually converges and achieves the expected results.

[0037] In some embodiments, a deformable convolution unit is integrated into the backbone network architecture of YOLOv8n, and the conventional convolution layer of the Bottleneck layer in the C2f module is replaced with a DCNv3 convolution layer.

[0038] After adopting such a design, the present invention has at least the following advantages:

[0039] First, a smooth and continuous path is generated between the point cloud points corresponding to the axis pin center through interpolation. This optimized path is then broken into a series of line segments at the point cloud points corresponding to the axis pin center. The point cloud coordinates of these line segments are imported into MATLAB to establish the curve equation. The actual length of each line segment is then integrated to obtain the length of the entire drive chain. In this process, the interpolation method achieves refined path modeling, and the accumulation of short line segments based on this method can fully guarantee the measurement accuracy of the chain when it is bent.

[0040] Second, incorporating deformable convolutional units into the YOLOv8n backbone network architecture enables more in-depth and detailed extraction of image features, which will facilitate feature fusion in subsequent stages and improve the prediction accuracy of the detection head. In order to introduce the deformable convolution module without significantly modifying the original YOLOv8n architecture and simplifying the entire development process, we chose to adaptively improve the C2f module of YOLOv8. In the C2f module, the conventional convolutional layer in the Bottleneck is directly replaced with the DCNv3 convolutional layer. This allows us to retain the functionality of the Bottleneck layer while leveraging the advantages of DCNv3, especially when dealing with tasks such as target detection, with better responsiveness to targets of various scales and shapes.

[0041] Third, all images in the training set are randomly rotated (for example, within the range of ±30 degrees) and / or mirrored horizontally and / or vertically, and color jittering is added for rust defects to increase data diversity and improve the generalization ability of the model, ensuring the accuracy of recognition of drive chain surface defects. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The above is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0043] Figure 1 Schematic diagram of the drive chain length and defect detection method of the present invention;

[0044] Figure 2 Schematic diagram of the steps of the center positioning method of the axle pin point of the present invention. DETAILED DESCRIPTION

[0045] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0046] A method for detecting the length and defects of an escalator drive chain: By establishing three-dimensional point cloud data and a defect recognition model of the drive chain, the drive chain length can be accurately measured and surface defects can be accurately determined. Figure 1 As shown, the specific length measurement and surface defect identification steps are as follows:

[0047] S101. Arrange at least two cameras around the escalator drive chain to obtain clear image information of the drive chain at different viewing angles;

[0048] S102. Capture the center of the axis pin point in the image information by image recognition technology and use it as a positioning reference to perform position matching on the image information to obtain image information of the same position of the drive chain at different viewing angles;

[0049] S103. Establishing a three-dimensional point cloud data and defect recognition model of the drive chain based on the image information from different perspectives;

[0050] S104. Determine the starting point of the drive chain in the three-dimensional point cloud data, and plan a smooth and continuous path along the drive chain between the point cloud points corresponding to the known axle pin center points using natural cubic spline interpolation;

[0051] S105. Discrete the path into a series of short line segments and then accumulate the lengths of the line segments to obtain a length measurement result of the drive chain;

[0052] S106. Import the real-time collected drive chain image information into the defect recognition model to perform real-time monitoring of the chain surface state and defect recognition.

[0053] In S101, before collecting image information of the drive chain, each camera is calibrated using the Zhang Zhengyou calibration method;

[0054] Multiple cameras capture images of a precision optical calibration plate. Image recognition technology is then used to extract checkerboard feature points from the images. Based on the correspondence between the feature points in each image set, the least squares method is used to estimate the rotation and translation vectors between the two cameras, thereby obtaining the camera calibration parameters. These calibration parameters include the camera's intrinsic and extrinsic parameters, distortion coefficients, translation matrix, and rotation matrix.

[0055] In S102, multi-scale Hough transform is used to locate the center of the pivot pin.

[0056] The specific method of center positioning of the pivot pin point is as follows:

[0057] Process the original image information through grayscale and Gaussian filtering to reduce noise interference and smooth the image;

[0058] Use the Canny edge detection algorithm to extract edge information from the image and provide clear input for subsequent processing;

[0059] To improve the initial positioning accuracy, gradient-based directional histogram analysis and / or least squares circle fitting are performed in the pivot point area to achieve preliminary positioning at the sub-pixel level, ensuring accurate locking of the target position during subsequent Hough transform.

[0060] Construct a multi-scale image pyramid covering multiple resolution levels to accommodate pivot points of different sizes, and determine the minimum and maximum scales for Hough transform based on the expected diameter range of pivot points in actual application scenarios;

[0061] At each scale level, a standard Hough Circle Transform is applied, adjusting the accumulator grid size and voting threshold to ensure that small pivot points are effectively detected while not missing larger objects, thereby obtaining candidate circles.

[0062] In order to prevent multiple detections of the same pivot point, non-maximum suppression (NMS) is performed on all candidate circles, and the most representative detection results are retained;

[0063] By back-screening and validating the detection results using geometric constraints and / or known drive chain dimensions, the output of the pin point location information ensures that the final output of the pin point center position is both accurate and reliable. This approach not only improves the accuracy of pin point detection but also enhances the system's adaptability to complex backgrounds and changing environments.

[0064] In S103 , before establishing the three-dimensional point cloud data and defect recognition model of the drive chain, the collected image information is subjected to image preprocessing; the image preprocessing includes image filtering, image enhancement, dilation and erosion.

[0065] Image filtering uses a mean filter to smooth the image, reduce random noise, and improve the effectiveness of edge detection and other feature extraction methods. The bilateral filter can remove noise while maintaining edge information, facilitating better maintenance of chain contour information.

[0066] Image enhancement increases the contrast of the image through histogram equalization or adaptive histogram equalization (CLAHE), making the drive chain stand out more clearly from the background. If the original image is too dark or too bright, the brightness can be adjusted through linear transformation to ensure that edge contour details are not lost. Applying the Laplacian operator or other sharpening filters can strengthen the boundaries of the drive chain in the image.

[0067] Dilation and erosion operations, erosion followed by dilation, can be used to eliminate small objects without significantly changing the size of larger objects. Dilation followed by erosion is used to fill small holes or disconnected gaps inside objects, facilitating subsequent image processing.

[0068] In S105 , the point cloud points corresponding to the pivot points divide the continuous path into short line segments, and the line segments are imported into MATLAB to establish a curve equation, and then the actual length corresponding to each line segment is obtained by integration.

[0069] The starting point of the chain is determined based on the three-dimensional point cloud obtained in step S103, and natural cubic spline interpolation technology is used to generate a smooth and continuous path between the point cloud points corresponding to the known axle pin center points. This interpolation method can ensure a smooth path while being as close as possible to the original point cloud data. The optimized path is discretized into a series of short line segments with the point cloud points corresponding to the axle pin center points as nodes. These line segments are imported into MATLAB to establish the curve equation, and the actual length corresponding to each line segment is then obtained by integration. The total path length is calculated by accumulating the lengths of these line segments. This method not only takes into account the actual bending of the chain, but also uses interpolation to achieve refined modeling of the path, ultimately providing high-precision length measurement results.

[0070] In S106, at least 1400 pieces of collected image information are divided into a training set, a validation set, and a test set in a ratio of 8:1:1 to train the defect recognition model.

[0071] All image information in the training set is subjected to random angle rotation and / or horizontal and / or vertical mirroring processing, and a color jittering operation is added for rust defects to increase data diversity.

[0072] Use the LabelImg image annotation tool to add bounding boxes to the target objects in each image and label their categories, ensuring that the annotation information is accurate and correctly reflects the location and type of the target;

[0073] According to the characteristics of the processed dataset, the YOLOv8 configuration file is adjusted in real time for model training to ensure that the model gradually converges and achieves the expected results.

[0074] The deformable convolution unit is integrated into the backbone network architecture of YOLOv8n, and the conventional convolution layer of the Bottleneck layer in the C2f module is replaced with the DCNv3 convolution layer.

[0075] Integrating deformable convolutional units into the YOLOv8n backbone network architecture enables more in-depth and detailed image feature extraction, which facilitates subsequent feature fusion and improves prediction accuracy in the detection head. To introduce the deformable convolutional module without significantly modifying the original YOLOv8n architecture and simplifying the overall development process, the C2f module of YOLOv8 was adapted and modified. In the C2f module, the Bottleneck layer is a key structure used to compress and expand feature dimensions while maintaining efficient computation. DCNv3 is a convolutional technology that enhances the model's adaptability to changes in object position and size. It can replace or supplement traditional convolutional layers. Simply replacing the conventional convolutional layers in the Bottleneck layer with DCNv3 convolutional layers preserves the functionality of the Bottleneck layer while leveraging the advantages of DCNv3. This, in particular, improves responsiveness to objects of various scales and shapes in tasks such as object detection. This improves model runtime and training efficiency without sacrificing performance, and enhances the ability to capture and recognize object features in object detection tasks.

[0076] At the same time, in order to ensure that the original image of the drive chain is clear, a shadowless light source can be selectively set up in the inspection area to ensure that the surface of the inspected object is evenly illuminated, reducing visual misjudgments caused by shadows, making minor defects easier to detect, thereby improving the accuracy and reliability of detection. The brightness level can also be changed through the light source controller. The digital light source controller uses a PWM digital dimming mode. It has two light control channels, the maximum output power of a single channel is 65W, and each channel has 255 levels of brightness adjustment. The digital light source controller has two ways to adjust the brightness: manual adjustment and host computer adjustment. In order to facilitate the inspection personnel to adjust the appropriate brightness according to the drive chain background, the host computer control is adopted and connected to the computer via the RS422 communication interface. The brightness level of each channel can be set, and the current brightness level of each channel can be read.

Claims

1. A method for detecting the length and defects of an escalator drive chain, characterized by: The following steps are involved: S101. Arrange at least two cameras around the escalator drive chain to obtain clear image information of the drive chain at different viewing angles; S102. Capture the center of the axis pin point in the image information by image recognition technology and use it as a positioning reference to perform position matching on the image information to obtain image information of the same position of the drive chain at different viewing angles; S103. Establishing a three-dimensional point cloud data and defect recognition model of the drive chain based on the image information from different perspectives; S104. Determine the starting point of the drive chain in the three-dimensional point cloud data, and plan a smooth and continuous path along the drive chain between the point cloud points corresponding to the known axle pin center points using natural cubic spline interpolation; S105. Discrete the path into a series of short line segments and then accumulate the lengths of the line segments to obtain a length measurement result of the drive chain; S106. Import the real-time collected drive chain image information into the defect recognition model to perform real-time monitoring of the chain surface state and defect recognition.

2. The method for detecting the length and defects of an escalator drive chain according to claim 1, wherein: In S101 , before collecting image information of the drive chain, each camera is calibrated using the Zhang Zhengyou calibration method.

3. The method for detecting the length and defects of an escalator drive chain according to claim 1, wherein: In S102, multi-scale Hough transform is used to locate the center of the pivot pin.

4. The method for detecting the length and defects of an escalator drive chain according to claim 3, wherein: The specific method for center positioning of the pivot pin point comprises the following steps: S201. Process the original image information by grayscale and Gaussian filtering to reduce noise interference and smooth the image; S202. Extract edge information from the image using the Canny edge detection algorithm; S203. Perform gradient-based directional histogram analysis and / or least squares circle fitting in the pivot point area to achieve preliminary sub-pixel positioning; S204. Construct a multi-scale image pyramid covering multiple resolution levels to accommodate pivot points of different sizes, and determine the minimum and maximum scales for Hough transform based on the expected diameter range of the pivot points in actual application scenarios; S205. Apply the standard Hough Circle Transform (HTC) at each scale level, adjust the accumulator grid size and voting threshold, and obtain candidate circles. S206. Perform non-maximum suppression (NMS) on all candidate circles and retain the most representative detection results; S207. Reversely screen and verify the test results through geometric constraints and / or known dimensions of the drive chain, and then output the shaft pin point position information.

5. The method for detecting the length and defects of an escalator drive chain according to claim 1, wherein: In S103 , before establishing the three-dimensional point cloud data and defect recognition model of the drive chain, the collected image information is subjected to image preprocessing; the image preprocessing includes image filtering, image enhancement, dilation and erosion.

6. The method for detecting the length and defects of an escalator drive chain according to claim 1, wherein: In S105 , the point cloud points corresponding to the pivot points divide the continuous path into short line segments, and the line segments are imported into MATLAB to establish a curve equation, and then the actual length corresponding to each line segment is obtained by integration.

7. The method for detecting the length and defects of an escalator drive chain according to claim 1, wherein: In S106, at least 1400 pieces of collected image information are divided into a training set, a validation set, and a test set in a ratio of 8:1:1 to train the defect recognition model.

8. The method for detecting the length and defects of an escalator drive chain according to claim 7, wherein: All image information in the training set is subjected to random angle rotation and / or horizontal and / or vertical mirroring processing, and a color jittering operation is added for rust defects to increase data diversity.

9. The method for detecting the length and defects of an escalator drive chain according to claim 8, wherein: Use the LabelImg image annotation tool to add bounding boxes to the target objects in each image and label their categories, ensuring that the annotation information is accurate and correctly reflects the location and type of the target; According to the characteristics of the processed dataset, the YOLOv8 configuration file is adjusted in real time for model training to ensure that the model gradually converges and achieves the expected results.

10. The method for detecting the length and defects of an escalator drive chain according to claim 9, characterized in that: The deformable convolution unit is integrated into the backbone network architecture of YOLOv8n, and the conventional convolution layer of the Bottleneck layer in the C2f module is replaced with the DCNv3 convolution layer.

Images