Method and system for detecting abnormal condition of accessory part of overhead line system

Abstract

The invention relates to the technical field of train maintenance, and discloses a method and a system for detecting abnormal conditions of accessories of a contact network. The method comprises the following steps: capturing images of the accessory parts of the overhead line system at a high speed to form a captured image set, and loading a region positioning algorithm model and a defect identification model; preprocessing the picture; after the region positioning algorithm model is adopted to extract picture feature information, regression classification is carried out to obtain possible positions of components and confidence values; carrying out defect identification on the component, reconstructing an original input image, and calculating a residual image; and carrying out gray processing and comparison on the residual image, and judging the existence condition of the defect. According to the invention, image snapshot instead of video acquisition frame-by-frame analysis is realized, the data volume is reduced, and meanwhile, the acquisition performance requirement on the camera is improved. According to the method, a complete data set is established on the basis of current and historical comparison, features of a training model are replaced, and generalization is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of train maintenance, and more particularly, to a method and system for detecting abnormal conditions of catenary accessory components. Background Art

[0002] The safety monitoring of urban rail transit infrastructure is an important part. The catenary suspension monitoring device is usually installed on a catenary inspection vehicle, a work vehicle or other special track vehicles, and dynamically monitors the catenary suspension system by non-contact means to achieve overall high-resolution imaging. The traditional layout and camera type have the following problems: In the traditional layout of the catenary accessory component detection system, video is continuously collected and analyzed frame by frame, which easily causes problems of large data volume and backlog, resulting in abnormal acquisition; Among the camera types, most are area array high-definition and cannot resist sunlight interference, showing a lack of adaptability in overall detection, especially for the monitoring requirements of outdoor operating vehicles; The traditional object detection algorithm uses methods such as comparing the current and historical data to identify foreign objects or abnormalities. This method has certain false alarms because the true validity of historical data and the complex diversity of environmental changes cannot be determined. Summary of the Invention

[0003] In view of this, the present invention provides a method and system for detecting abnormal conditions of catenary accessory components to solve the above problems.

[0004] To solve the above technical problems, the present invention provides a method for detecting abnormal conditions of catenary accessory components, including: Locate the catenary suspension support position, capture high-speed images of the front and back sides of the catenary accessory components, and classify and store the collected images according to the identification of the installation position to form a set of captured images; Initialize the default model, load the region location algorithm model and the defect recognition model; and retrieve pictures about abnormal accessory components and preprocess the pictures; After extracting the picture feature information by using the region location algorithm model, perform regression classification to obtain the possible positions and confidence values of the components; Filter the possible position areas of the components and input them into the defect recognition algorithm model. At the same time, input the original input image into the generator to obtain a reconstructed image, and then calculate the residual image formed by the residual; Calculate the average value and variance of the gray level of the residual image, compare the calculation results with the preset threshold, and judge the presence of defects.

[0005] As an optional method, retrieving pictures about abnormal accessory components and preprocessing the pictures includes: After converting the image to grayscale, enhance the target brightness through global histogram equalization, globally enhance the brightness through adaptive histogram equalization, and enhance the image contrast while suppressing noise through contrast-limited histogram equalization. Among them, contrast-limited histogram equalization includes: Divide the image into small blocks of a preset size, perform histogram equalization and cropping processing on each small block, adjust the grayscale value of the center point through a mapping function, and calculate the grayscale values in different regions to complete the enhancement processing.

[0006] As an alternative method, after extracting the picture feature information using a region localization algorithm model, perform regression classification to obtain the possible positions and confidence values of the components, including: Use a region localization algorithm model to extract the feature information of the catenary accessories from the vehicle-mounted detection grayscale image; Divide the grayscale image into a preset number of sub-blocks; Taking each sub-block as the center, determine the number of prediction boxes, and perform classification prediction on the objects within the prediction boxes; Optimize the prediction boxes, establish an automatic recognition network, and generate a tensor of a preset size after inputting the image; Use a loss function to compare the predicted object confidence with the actual value, and optimize the recognition accuracy of the network; Determine the positions and confidence levels of the accessory components through object detection and classification.

[0007] As an alternative method, filter the possible position areas of the components, including filtering the component areas using NMS, filtering out the background area results, and then filtering out the results that do not meet the quality requirements using image algorithms. The implementation method is as follows: Use non-maximum suppression to filter the detected component areas, removing the background areas and redundant detection boxes; Use image algorithms to further filter out the detection boxes that do not meet the quality requirements to ensure the accuracy and reliability of the detection results. Among them, Non-maximum suppression includes: Sort all the detection boxes by score, and select the detection box with the highest score; Calculate the intersection-over-union (IoU) of the detection box with the highest score and other detection boxes; Suppress the detection boxes with an IoU exceeding the set threshold; Repeat the above steps until only the detection boxes with no obvious overlap remain; only one high-confidence detection box is retained for each object.

[0008] As an alternative method, if a defect is detected, give the type and location information of the defect, and trace back to the above confidence level for coverage.

[0009] On the other hand, the present invention also provides a detection system for abnormal conditions of catenary accessory components, including: a high-definition image acquisition module, an on-vehicle control and analysis module, and a ground analysis center; The high-definition image acquisition module is installed on the roof of the inspection vehicle and is used for acquiring high-definition images of the catenary suspension support device; The on-vehicle control and analysis module is installed inside the inspection vehicle and is used for collecting and analyzing the high-definition image acquisition data; The ground analysis center is installed in the analysis room and is used for abnormal analysis of the image data and monitoring and alarming.

[0010] As an optional method, the high-definition image acquisition module is formed by constructing multiple roof cameras, and it includes: a panoramic high-definition camera, a high-definition camera, an out-of-tunnel pole number recognition camera, an in-tunnel pole number recognition camera, a clamp conductor suspension shooting high-definition camera, a continuous video monitoring camera; and, a radar speed measurement module and a GPS positioning module.

[0011] As an optional method, the on-vehicle control and analysis module is used for equipment power supply, camera triggering, data storage, and data analysis, and it includes: a cabinet, an acquisition computer, a control computer, a multi-functional control box, a power supply box, and an uninterruptible power supply.

[0012] As an optional method, the ground analysis center includes a ground analysis management platform and a defect recognition system, and the ground analysis management platform is erected using a B / S structure.

[0013] As an optional method, the image analysis of the on-vehicle control and analysis module includes: crack detection, foreign object detection, detection of loss and looseness of fastening ring nuts, and detection of protruding kit bolts.

[0014] The beneficial effects of the present invention are: The present invention realizes image capture instead of frame-by-frame analysis of video acquisition, reduces the amount of data, and at the same time improves the acquisition performance requirements for the camera. A complete data set is established based on the comparison of the current and historical data, and the features of the training model are replaced to improve the generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic flow chart of the method for detecting abnormal conditions of catenary accessory components provided by the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] In order to enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below in conjunction with the specific embodiments.

[0017] Please refer toFigure 1 , this embodiment provides a method for detecting abnormal conditions of catenary accessories, including: Locate the suspension support position of the catenary, capture high-speed front and back images of the catenary accessories, and classify and store the collected images according to the identification of the installation position to form a set of captured images; Initialize the default model, load the region localization algorithm model and the defect recognition model; retrieve pictures about abnormal accessories and preprocess the pictures; After extracting the picture feature information by using the region localization algorithm model, perform regression classification to obtain the possible positions and confidence values of the components; Filter the possible position areas of the components and input them into the defect recognition algorithm model. At the same time, input the original input image into the generator to obtain a reconstructed image, and then calculate the residual image formed by the residual; Calculate the average value and variance of the gray scale of the residual image, compare the calculation results with the preset threshold, and judge the existence of defects.

[0018] The catenary accessories are often directly exposed to sunlight. When the rail vehicle is running at high speed, due to weather changes, the overall gray value and contrast of the image are further reduced. Or there is a certain area of the accessory in the image that is fused with the background, resulting in a decrease in the gray value. For example, when the equipment is in an area with insufficient light or in a shadow area, it will increase the difficulty of extracting the target features. Based on this, the preprocessing method is to reduce the above effects. In the early stage of preprocessing, the image needs to be converted into a single-channel gray image, and the output of the on-vehicle catenary detection system is already a gray image, so no conversion is required. In the preprocessing algorithm: global histogram equalization makes the target the brightest; adaptive histogram equalization brightens the whole image and noise appears; contrast-limited histogram equalization has a better noise suppression effect. The specific method of contrast-limited histogram equalization is to divide the picture containing the accessory into blocks, usually divide the image into 8×8 sub-blocks, then perform a mapping function (histogram equalization and histogram clipping algorithm) on each sub-block, and then make the gray value of the center point c of the sub-block obtained by the mapping function, and finally calculate the gray value in regions. As an optional method, in this embodiment, retrieving pictures about abnormal accessories and preprocessing the pictures includes: After converting the image into a gray image, improve the target brightness through global histogram equalization, enhance the overall brightness through adaptive histogram equalization, and enhance the image contrast while suppressing noise through contrast-limited histogram equalization; among them, contrast-limited histogram equalization includes: dividing the image into 8×8 small blocks, performing histogram equalization and clipping processing on each small block, adjusting the gray value of the center point through the mapping function, and calculating the gray value in regions to complete the enhancement processing.

[0019] As an alternative method, after extracting the image feature information using the region localization algorithm model, the positions where the components may appear and the confidence values are obtained through regression classification, including: The vehicle-mounted detection grayscale image containing catenary accessories is divided into grids (equally divided into S×S sub-blocks). Taking each sub-block as the center, B prediction boxes are determined, and the objects within the prediction boxes are classified and predicted. At the same time, the IoU calculation is performed with the actual target box, and the optimal prediction box is selected through the non-maximum suppression method. An automatic recognition network is established and object detection is performed based on catenary accessories. The network converts the input image into a tensor of S×S×[B(4 + 1)+C], where S is equal to 7, B is equal to 2, C is the number of target categories equal to 20, and B(4 + 1) represents the 5-dimensional vector of the four coordinates of the border and whether there is a target within the box. After training, the predicted value and the actual value of the confidence of the detected object in the loss function are used to calibrate the confidence, that is, the recognition accuracy.

[0020] Filter the position areas where the components may appear, including filtering the obtained component areas using NMS, filtering out the background area results, and then filtering out the results that do not meet the quality requirements using image algorithms. The implementation method is as follows: Use non-maximum suppression to filter the detected component areas, removing the background areas and redundant detection boxes; Use image algorithms to further filter out the detection boxes that do not meet the quality requirements to ensure the accuracy and reliability of the detection results; among them, non-maximum suppression includes: Sort all the detection boxes by score and select the detection box with the highest score; Calculate the intersection over union (IoU) of the detection box with the highest score and other detection boxes; Suppress the detection boxes whose intersection over union exceeds the set threshold; Repeat the above steps until only the detection boxes with no obvious overlap remain; only one detection box with a high confidence level is retained for each target.

[0021] Specifically, there are many detection boxes of catenary accessories actually completed through training. These boxes all detect the same target, but ultimately only one detection box is needed for each target. NMS selects the detection box with the highest score (assumed to be C), and then calculates the corresponding IOU value between C and the remaining boxes. When the IOU value exceeds the set threshold (commonly set to 0.5, often set to 0.7 in object detection), that is, suppress the boxes that exceed the threshold. The method of suppression is to set the score of the detection box to 0. After such a round, continue to find the detection box with the highest score among the remaining detection boxes, and then suppress the boxes whose IOU with it exceeds the threshold until finally there will be almost no overlapping boxes left. In this way, basically only one detection box remains for each target.

[0022] As an alternative, if a defect is detected, the type and location information of the defect are given, and the above confidence level is traced back for coverage.

[0023] Based on the above solution, this embodiment considers increasing the diversity of the sample data set and the generalization of the model, and uses a multi-scale method to train the model and learn defect features. The data set incorporates more samples due to shooting angle rotation, color jitter, etc., aiming to improve the generalization of the algorithm. At the same time, the defect library is expanded by continuously supplementing new data collected on-site.

[0024] On the other hand, this embodiment provides a detection system for abnormal conditions of catenary accessories. The system mainly consists of three parts, namely, an on-vehicle control and analysis module, a high-definition image acquisition module, and a ground analysis center. The high-definition image acquisition module is installed on the roof of the vehicle and is used to collect high-definition images of the catenary suspension support device; the on-vehicle control and analysis module is installed in the integrated inspection vehicle and is used to collect and archive the high-definition image acquisition data; the ground analysis center is installed in the analysis room and is used to perform abnormal intelligent analysis on the image data and alarm for abnormal data.

[0025] The on-vehicle control and analysis module is mainly responsible for the power supply of the equipment, the triggering of the camera, the storage of data, and the analysis of data. The hardware composition of the in-vehicle equipment consists of a cabinet, an acquisition computer, a control computer, a multi-functional control box, a power supply box, and an uninterruptible power supply, etc.

[0026] Roof Monitoring Image Acquisition Module The roof cameras mainly include, in terms of composition: panoramic high-definition cameras, high-definition cameras (optional models such as anti-glare and anti-reflection), out-of-tunnel pole number recognition cameras, in-tunnel pole number recognition cameras, high-definition cameras for shooting wire clamps, conductor suspension strings, continuous video monitoring cameras, radar speed measurement modules, GPS positioning modules, etc. Among them, the exposure time and sensitivity parameters of the high-definition cameras are optimized so that they can work under various lighting conditions, including direct sunlight, and can dynamically adjust the exposure time and sensitivity according to the ambient light, so as to better adapt to different lighting conditions and reduce the possibility of overexposure or underexposure. At the same time, special optical filters or coatings (special lens materials, anti-ultraviolet and anti-reflection coatings) are equipped to reduce the reflection or overexposure caused by sunlight. These designs can effectively reduce the reflection on the lens surface and improve the image quality.

[0027] The software system of the ground analysis center includes a ground analysis management platform and a defect recognition system. The ground analysis management system adopts a B / S structure, and users can directly access the 4C ground analysis center system through a browser on any host that can connect to the server, without the need to install other systems specially.

[0028] Based on the above, for the detection requirements of catenary accessories, the main functions of the system in this embodiment include: Accurately locate the position of the catenary boom mounting pillar (or suspension pillar); Take images of both the front and back of the support device, including images of the contact suspension (hanger, clamp, etc.) and images of the suspension pillar base area in the tunnel; Adopt a high-speed capture trigger mode, different from the traditional video frame capture, which can minimize the amount of collected data; Support the automatic identification function of the pillar number, which can identify each pillar number and store the identified pole number in the database to form a data base DB file; The image analysis objects include: crack detection, foreign object detection, detection of loss and looseness of fastening ring nuts, detection of protruding kit bolts, etc.; In this way, this embodiment changes from traditional video shooting and frame-by-frame analysis to image capture, reducing the size of data collection, improving the processing speed and time; the application upgrade of high-definition cameras improves the adaptability of the catenary accessory detection system, that is, it meets the outdoor daytime detection requirements and effectively improves the utilization rate of the time outside the skylight point.

[0029] The above is only the preferred embodiment of the present invention. It should be noted that the above preferred embodiment should not be regarded as a limitation of the present invention. The protection scope of the present invention should be subject to the scope defined by the claims. For those of ordinary skill in the art, without departing from the spirit and scope of the present invention, several improvements and refinements can still be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A method for detecting abnormal conditions of contact network accessory components, characterized in that: include: Locate the overhead contact network suspension support position, capture the front and back images of the overhead contact network accessories at high speed, and classify the collected images into the database according to the identification of the installation position to form a captured image set; Initialize the default model, load the regional positioning algorithm model and the defect recognition model; retrieve pictures of abnormal attached parts, and pre-process the pictures; After extracting the image feature information using the regional positioning algorithm model, regression classification is used to obtain the possible location and confidence value of the component; The possible location areas of the parts are filtered and input into the defect recognition algorithm model. At the same time, the original input image is input into the generator to obtain the reconstructed image, and then the residual image formed by the residual is calculated; The mean value and variance of the grayscale of the residual image are calculated, and the calculation results are compared with the preset threshold to determine the existence of defects.

2. A method for detecting abnormal conditions of contact network accessory components according to claim 1, characterized in that: The retrieving pictures about abnormal accessory parts and preprocessing the pictures comprises: After converting the image into a grayscale image, the target brightness is increased by global histogram equalization, the overall brightness is increased by adaptive histogram equalization, and the image contrast is enhanced while suppressing noise by limited contrast histogram equalization; wherein the limited contrast histogram equalization includes: The image is divided into small blocks of preset size, and histogram equalization and cropping are performed on each small block. After adjusting the gray value of the center point through the mapping function, the gray value is calculated by region to complete the enhancement process.

3. A method for detecting abnormal conditions of contact network accessory components according to claim 1, characterized in that: After extracting the image feature information using the regional positioning algorithm model, regression classification is used to obtain the possible location and confidence value of the component, including: The regional positioning algorithm model is used to extract the characteristic information of the auxiliary parts of the contact network from the vehicle-mounted detection grayscale image; Divide the grayscale image into a preset number of sub-blocks; Taking each sub-block as the center, determine the number of prediction boxes and classify and predict the objects in the prediction boxes; Optimize the prediction box, build an automatic recognition network, and generate a tensor of preset size after inputting the image; Use the loss function to compare the predicted target confidence with the actual value to optimize the recognition accuracy of the network; Determine the location and confidence of attached parts through object detection and classification.

4. A method for detecting abnormal conditions of contact network accessory components according to claim 3, characterized in that: The filtering of the location area where the component may appear includes filtering the component area with NMS, filtering out the background area results and filtering out the results that do not meet the quality requirements with an image algorithm, which is implemented by the following method: Use non-maximum suppression to filter the detected component area and remove the background area and redundant detection frames; The image algorithm is used to further filter out the detection frames whose quality does not meet the requirements to ensure the accuracy and reliability of the detection results; wherein the non-maximum suppression includes: Sort all detection frames by score and select the detection frame with the highest score; Calculate the intersection and union ratio of the detection box with the highest score and other detection boxes; Suppress the detection boxes whose intersection-over-union ratio exceeds the set threshold; Repeat the above steps until only detection boxes with no obvious overlap remain; only one high-confidence detection box is retained for each object.

5. A method for detecting abnormal conditions of contact network accessory components according to claim 1, characterized in that: If it is judged that there is a defect, the type and location information of the defect is given, and the above confidence level is traced back to cover it.

6. A system for detecting abnormal conditions of auxiliary components of a contact network, characterized in that: include: High-definition image acquisition module, vehicle-mounted control and analysis module, and ground analysis center; The high-definition image acquisition module is installed on the roof of the inspection vehicle and is used to acquire high-definition images of the overhead contact network suspension support device; The vehicle-mounted control and analysis module is installed in the detection vehicle and is used to collect and analyze high-definition image acquisition data; The ground analysis center is installed in the analysis room and is used for performing abnormal analysis on the image data and monitoring alarms.

7. A contact network accessory component abnormality detection system according to claim 6, characterized in that: The high-definition image acquisition module is formed by a plurality of roof cameras, and includes: Panoramic HD camera, HD camera, tunnel pole number recognition camera, tunnel pole number recognition camera, wire clamp conductor hanging string shooting HD camera, continuous video monitoring camera; and, Radar speed measurement module and GPS positioning module.

8. A contact network accessory component abnormality detection system according to claim 6, characterized in that: The vehicle-mounted control and analysis module is used for equipment power supply, camera triggering, data storage and data analysis, and includes: Cabinet, acquisition computer, control computer, multi-function control box, power box and uninterruptible power supply.

9. A system for detecting abnormal conditions of auxiliary components of a contact network according to claim 6, characterized in that: The ground analysis center includes a ground analysis management platform and a defect identification system, and the ground analysis management platform is constructed using a B / S structure.

10. A system for detecting abnormal conditions of auxiliary components of a contact network according to claim 6, characterized in that: The image analysis of the vehicle-mounted control analysis module includes: crack detection, foreign matter detection, fastening ring nut loss and loosening detection, and kit bolt protrusion detection.

Images