Railway track induction plate vehicle-mounted dynamic detection method

Abstract

The application discloses a kind of track line induction plate vehicle dynamic detection methods, comprising the following steps: 2D image acquisition module gathers the image data of induction plate, support and bolt area, obtains original induction plate image;3D point cloud acquisition module gathers three-dimensional point cloud data of induction plate, left and right rail area;Collecting encoder pulse signal and deploying radio frequency tag signal on line, realize with line mileage synchronization, simultaneously output trigger pulse to induction plate data acquisition unit and interval collection data;Intelligent detection analysis is carried out to the collected induction plate image data, and the defect information including induction plate crack, support crack, bolt loss, bolt loosening is obtained;The collected induction plate point cloud data is handled, and the area, height, longitudinal displacement and wrong tooth information of each induction plate are obtained.The application can solve the technical problems that the detection efficiency of existing detection method is not high, the detection precision is insufficient and the detection content is not comprehensive.

Description

Technical Field

[0001] This application relates to the field of railway engineering inspection technology, and in particular to an on-board dynamic detection method for track line sensing plates based on computer vision technology. Background Technology

[0002] Currently, many urban rail transit lines in my country use linear motor induction plate systems. During operation, the timely detection and maintenance of induction plate anomalies plays a crucial role in ensuring train safety. Currently, the surface condition and geometric parameters of these induction plates are assessed manually, which is inefficient and consumes significant manpower and maintenance windows. Therefore, devices with induction plate detection capabilities have been continuously developed. Invention applications CN111123282A and CN101513883A utilize two point laser ranging sensors or ultrasonic ranging sensors to measure the height of the induction plate; while utility model patent CN215706365U proposes an induction plate foreign object detection device based on 3D image recognition. However, these devices are all mounted on handcarts and can only measure the height of the induction plate. They do not consider eliminating errors caused by vibrations during trolley movement, resulting in low detection accuracy and efficiency.

[0003] The following documents are relevant to this application in the prior art: Reference 1 is a Chinese invention application filed by Beijing Jiaotong University on March 12, 2018, and published on November 2, 2021, with publication number CN108573213A. This application discloses an automatic detection system and method for track fastener defects. The system includes: a first positioning module; a second positioning module; a segmentation module; a processing module; and a detection module. The system and method provided in this application can automatically locate and determine the condition of fasteners under different conditions, solving the problems of missed detections and the lack of objective accuracy of detection results that are difficult to guarantee with traditional manual methods. It also provides a new method and approach for the design of automated detection equipment for abnormal fastener conditions. This application can accurately and effectively identify abnormally damaged fasteners in track lines, significantly improving detection efficiency and providing a good foundation for meeting the requirements of safe and efficient online detection of rail transit lines. The system can achieve online detection, has a high detection speed, can adapt to the detection needs of different time periods under sufficient light, and has high reliability and accuracy.

[0004] Reference 2 is a Chinese invention application filed by Suzhou Licheng Zhiheng Electronics Technology Co., Ltd. on August 8, 2022, and published on September 16, 2022, with publication number CN115063416A. This application discloses a method and system for detecting the state of rail fasteners. The method includes acquiring an image of the rail fastener to be identified and inputting it into a first rail environment segmentation model and a second rail environment segmentation model to obtain the center point coordinates and contour point coordinates of the nut, rail, and fastener; obtaining a local image of the rail fastener to be identified based on the center point and contour point coordinates of the fastener, and inputting it into the fastener state detection model to obtain an initial predicted state; obtaining state auxiliary judgment information based on the center point and contour point coordinates of the nut, rail, and fastener, and further obtaining the state category of the rail fastener image to be identified. The detection method provided by this application obtains the position and shape of the fastener through the center point and contour point coordinates of the nut, rail, and fastener, obtains state auxiliary judgment information, and combines it with the initial predicted state of the fastener to achieve high-precision detection of the rail fastener state.

[0005] Reference 3 is a Chinese invention application filed by Southwest Jiaotong University on March 14, 2019, and published on June 28, 2019, with publication number CN109948690A. This application discloses a high-speed rail scene perception method based on deep learning and structural information, including the following steps: Step 1: Acquire track images, divide them into training and test sets, and annotate the images in the training set to form a dataset; Step 2: Construct an SSD network model and construct a loss function; Step 3: Use the dataset formed in Step 1 to iteratively train the network obtained in Step 2 to obtain a training model; Step 4: Input the video to be detected frame by frame into the training model obtained in Step 3, extract features, obtain the position and category information of fasteners and shoulders, and distinguish between turnouts and ordinary tracks based on the position and category information of fasteners and shoulders; Step 5: Cluster the position information of the positioning results in Step 4 to complete the perception of rails and sleepers. This application can detect and semantically segment track components in turnout areas, with high detection accuracy and fast detection speed.

[0006] Since there are many types of defects in the sensing plate, the above-mentioned references 1 to 3 do not cover the detection of abnormalities such as the appearance, displacement, misalignment, and missing or loose bolts of the sensing plate. They cannot achieve systematic detection coverage of the sensing plate status and also lack corresponding complete intelligent detection algorithms. Summary of the Invention

[0007] In view of this, the purpose of this application is to provide a vehicle-mounted dynamic detection method for track line induction panels, so as to solve the technical problems of low detection efficiency, insufficient detection accuracy and incomplete detection content of existing detection methods.

[0008] To achieve the aforementioned objectives, this application specifically provides a technical implementation scheme for a vehicle-mounted dynamic detection method for track line induction panels, comprising the following steps: S1) The 2D image acquisition module acquires image data I of the sensor plate, bracket and bolt area to obtain the original sensor plate image; S2) The 3D point cloud acquisition module acquires three-dimensional point cloud data P(x, y, z, i) of the sensing plate and the left and right rail areas, where x is the X-axis coordinate, y is the Y-axis coordinate, z is the Z-axis coordinate, and i is the reflection intensity of the point cloud. S3) Acquire encoder pulse signals and RFID tag signals deployed on the line to achieve synchronization with the line mileage, and simultaneously control the 2D image acquisition module and 3D point cloud acquisition module to acquire data at equal intervals; S4) Sensor plate surface defect detection: Intelligent detection and analysis of the collected sensor plate image data to obtain defect information including sensor plate cracks, bracket cracks, missing bolts, and loose bolts; Geometric parameter detection of sensor plates: The collected point cloud data of sensor plates is processed to obtain the area, height, longitudinal displacement and misalignment information of each sensor plate.

[0009] Furthermore, the 3D point cloud acquisition module includes three sets of camera components, characterized in that step S2) includes: S21) First, after acquiring point cloud data, perform filtering preprocessing and search for point clouds belonging to the same target between two frames of point clouds from adjacent camera components at the same mileage. S22) Then, the rotation matrix R and translation matrix T between the two frames of point clouds are calculated through point cloud feature registration.

[0010] Furthermore, the point cloud feature registration in step S22) includes the following process: S221) Perform preliminary coarse registration on the two point cloud data points PA and PB to obtain an initial rotation matrix R. ’ and translation matrix T ’ PB is obtained by performing the corresponding transformation on PB. ’ =PB*R ’ +T ’ ; Where PA is the point cloud data of the left camera component, and PB is the point cloud data of the camera component adjacent to the left camera component; S222) Sample the point cloud PA to obtain a set of sampling points; S223) in point cloud PB ’ Find the point closest to the sampling point to obtain a set of matching points; S224) Calculate the updated rotation matrix R based on the matching points. ’ and translation matrix T ’For PB ’ PB is obtained by performing the corresponding transformation. ’ ’ =PB ’ *R ’ +T ’ ; S225) Repeat steps S222) to S224), use the mean square error as the objective function to calculate the registration error. When the registration error is less than the set threshold T or the number of iterations is equal to the maximum number of iterations, the registration ends. Save the obtained rotation matrix R and translation matrix T to achieve point cloud feature registration.

[0011] Furthermore, photoelectric encoders are installed in the wheels of the track inspection train, and radio frequency tag readers are installed under the track inspection train. Step S3) includes the following processes: S31) During the operation of the track inspection train, the processing board collects the equally spaced pulse waveforms emitted by the photoelectric encoder, and uses the pulse waveforms after frequency doubling and frequency division as the input source for sampling triggering of the 2D image acquisition module and the 3D point cloud acquisition module, and calculates the preliminary mileage accordingly. S32) Scan the RFID tag signal at the set sampling frequency. If the RFID tag signal is detected, look up the real mileage information represented by the tag in the tag ID and actual mileage comparison table, and then map this real mileage information onto each pulse signal to achieve synchronization with the line mileage.

[0012] Furthermore, in step S3), the current line mileage is corrected according to the following formula: Let the actual mileages of two adjacent RFID tags be M1 and M2, respectively. n The corresponding trigger pulse counts are N1 and N. n Then in N1 and N n The actual distance between M i =M1+(N i -N1)*(M n -M1) / (N n -N1), i=1,2,3……n.

[0013] Furthermore, the intelligent detection analysis in step S4) combines the target detection network and the classification network for judgment; the defect information includes the defect category, the mileage, the coordinates of the rectangle in the image (x, y, w, h), the actual physical size, and the defect severity level, where x and y are the x and y coordinates of the top left vertex, and w and h are the width and height of the rectangle, respectively.

[0014] Furthermore, the detection of apparent defects in the induction plate in step S4) includes the following process: S41) Perform intelligent detection and analysis on the collected sensor plate image data to obtain defect information including sensor plate cracks, bracket cracks, missing bolts, and loose bolts; S42) First, the target detection network model detects cracks in the sensing plate, cracks in the bracket, missing bolts, and normal bolts. Then, the sub-image of normal bolts is sent to the classification network model to determine whether they are loose.

[0015] Furthermore, the target detection network model annotates the original sensor plate image with target detection boxes, dividing it into a training set and a validation set. The label categories for the target detection box annotations include sensor plate cracks, support cracks, missing bolts, and normal bolts, represented by 0, 1, 2, and 3 respectively. The entire dataset is divided into 80% of the samples as the training set and 20% of the samples as the validation set.

[0016] Furthermore, in step S42), an object detection network model is constructed based on YOLOv10. The feature extraction part uses CSPDarknet as the backbone network and path aggregation network as the neck network. Large convolutional kernels and partition attention are used to enhance the model's context learning ability. Spatial-channel decoupled downsampling and rank-guided modules are used to improve the overall efficiency of the model.

[0017] Furthermore, the input size of the object detection network model is 1280, and the output detection box of the object detection network model includes the object category, the object box, and the confidence score. The training set is loaded into the object detection network model for iterative training, with a maximum iteration count of 400 epochs. After 10 epochs of training, the Maosic enhancement operation is disabled, and only conventional image enhancement operations, including horizontal flipping, vertical flipping, and chroma transformation, are retained. A validation set test is performed after each epoch, and the average accuracy (mAP) of the validation set is recorded. Simultaneously, the model parameter with the highest average accuracy (mAP) during the entire iterative training process is saved to the local disk as best.pt, and the model parameters obtained from the latest iteration are saved as last.pt. Each image I to be detected is input into the network model loaded with the best.pt parameter for forward inference, resulting in several object detection boxes. Object categories 0, 1, 2, and 3 correspond to induction plate crack, bracket crack, missing bolt, and normal bolt, respectively. The ROI region is extracted from image I using the object detection box of category 3 to obtain the bolt sub-image S. i , i=1,2,3……n, where n is the number of bounding boxes for target category 3, S iThey are respectively sent into the classification network model to determine whether the bolt is loose. The classification network model divides the induction plate bolt sub-image data set into two categories: normal bolts and loose bolts, and the corresponding labels are 0 and 1 respectively. 80% of the normal bolt and loose bolt images are used as the classification training set, and the remaining 20% are used as the classification verification set. The classification network model loads the induction plate bolt sub-image data set for iterative training. The maximum number of iterations is set to 300, the learning rate is set to 0.001, and image enhancement operations such as random rotation, horizontal flipping, vertical flipping, and chromaticity transformation are adopted. The verification set is tested once every 1 epoch iteration, the classification accuracy ACC information is recorded, and the model parameters with the maximum accuracy ACC during the entire iterative training process are saved as the best.param file, and the model parameters obtained from the latest round of iterative training are saved as the last.param file. Bolt sub-image S i After being input into the classification model loaded with best.param and inferred, the confidence conf0 belonging to category 0 and the confidence conf1 belonging to category 1 are obtained. If conf0 < conf1, this bolt is loose; otherwise, it is a normal bolt.

[0018] Furthermore, in the point cloud processing in step S4), according to the distribution characteristics of the induction plate, the front, back, left, and right edge positions of each induction plate are extracted through edge detection, so as to segment the point cloud of each induction plate. The height information of each position of the induction plate is further obtained from the coordinate z of the point cloud. The longitudinal displacement information of the induction plate is obtained by calculating the offset of the gap between two adjacent induction plates exceeding the standard value range. The misalignment information of the induction plate includes the lateral difference, height, and angle deviation between adjacent induction plates. The vertices of the induction plate are obtained through the boundary intersection points of the induction plate, and the lateral and height deviations of the induction plate are obtained by comparing the coordinate x deviation and coordinate z deviation between the corresponding vertices of adjacent induction plates. Then, the angle deviation is solved by fitting the normal vector of the induction plate plane calculated.

[0019] Furthermore, the detection of the geometric parameters of the induction plate in step S4) includes the following process: S411) First, the induction plate is segmented by block, and the upper, lower, left, and right edge positions of the induction plate are extracted through gradient edge detection, so as to segment the point cloud Psb of each induction plate i , i = 1, 2, 3... n; S412) Perform bilateral filtering on the point cloud Psb i to eliminate noise and obtain Psb i ’ , and the z coordinate of each point in the induction plate point cloud Psb i ’ is the height value of each position of the induction plate. Extract Psb i ’The maximum value Max and the minimum value Min of the z coordinate; S413) If the maximum value Max > the allowable ultra-high value, it is determined that the induction plate is abnormally ultra-high. Conversely, it is judged whether the minimum value Min < the allowable ultra-low value holds. If it holds, it is determined that the induction plate is abnormally low. Otherwise, the height of the induction plate is normal; S414) The longitudinal displacement information of the induction plate is obtained by calculating the offset of the gap between two adjacent induction plates exceeding the standard value range. d1 is the longitudinal displacement between the first induction plate Psb1 ’ and the second induction plate Psb2 ’ d2 is the longitudinal displacement between the second induction plate Psb2 ’ and the third induction plate Psb3 ’ ; S415) If the allowable minimum displacement Offset Min < d1 < the allowable maximum displacement Offset Max , it is judged that the longitudinal displacement of the first induction plate is normal. Otherwise, it is judged that the longitudinal displacement of the first induction plate is abnormal.

[0020] Furthermore, the misalignment information of the induction plate includes the lateral, height, and angle deviations between adjacent induction plates. The detection of the geometric parameters of the induction plate in step S4) includes the following processes: S421) The vertices of the induction plate are obtained through the boundary intersection points of the induction plate. The vertices of the first induction plate are A1, B1, C1, D1, and the vertices of the second induction plate are A2, B2, C2, D2, and so on. Each vertex includes x, y, z coordinates; S422) c1 is the lateral deviation between the first induction plate and the second induction plate, which is obtained by taking the absolute value of the difference between the X-axis coordinate D1(x) of D1 and the X-axis coordinate C2(x) of C2, that is, |D1(x) - C2(x)|. c2 is the lateral deviation between the second induction plate and the third induction plate. If the lateral deviation is greater than the lateral deviation threshold Tx, it is considered misaligned. Otherwise, it is determined that the lateral deviation is normal; The height deviation is obtained by taking the absolute value of the difference between the z coordinates of the corresponding vertices of adjacent induction plates. The height deviation between the first induction plate and the second induction plate is |D1(z) - C2(z)| and |B1(z) - A2(z)|. If the height deviation is greater than the height deviation threshold Tz, it is determined that there is misalignment. Otherwise, it is determined that the height deviation is normal.

[0021] Furthermore, step S423) includes: The point cloud of the first sensor panel is defined by vertices A1, B1, C1, and D1. Then, the plane equation Ax + By + Cz + D = 0 of the sensor panel surface is calculated iteratively using a plane fitting algorithm based on random sampling consistency. Its normal vector is n1(A, B, C). The normal vector n2(A, B, C) of the second sensor panel is then calculated. Finally, the angle deviation is solved by the dot product of the fitted sensor panel plane normal vectors. , if the angle deviation If the angle deviation threshold is met, it is considered a misaligned tooth; otherwise, it is considered normal.

[0022] Furthermore, the method includes: S5) After the detection result is obtained through step S4), the detection result information and the corresponding raw data are transmitted to the ground user terminal, and the operators are notified to formulate an operation plan.

[0023] By implementing the technical solution of the on-board dynamic detection method for track induction panels provided in this application, the following beneficial effects are achieved: (1) The on-board dynamic detection method for induction plates of rail transit in this application automatically processes the three-dimensional point cloud data of the induction plate and analyzes the geometric parameter information of the induction plate, such as abnormal height, longitudinal displacement, and misaligned teeth, to provide a reliable basis for the operation and maintenance of the induction plate. It realizes the on-board dynamic detection of induction plates of rail transit, greatly improves the detection efficiency and accuracy, and comprehensively covers the detection of the appearance and geometric parameters of the induction plate. (2) The on-board dynamic detection method for induction plates of track lines in this application adopts target detection network and classification network to intelligently analyze surface cracks, missing bolts, and loosening defects of induction plates, realizing end-to-end detection of surface defects of induction plates, and further improving detection efficiency and accuracy; (3) The on-board dynamic detection method for track induction plate of this application uses a 2D camera component to collect the surface texture features of the induction plate body and the support, and a 3D camera component to collect the rail outline and three-dimensional point cloud information of the induction plate. It can be mounted on a train to collect and process induction plate data at high speed, realize automatic and efficient detection of surface defects and geometric parameters of the induction plate, and guide relevant management and operation personnel to maintain the induction plate. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other embodiments can be obtained based on these drawings without creative effort.

[0025] Figure 1This is a structural block diagram of a specific embodiment of the on-board dynamic detection system for track line induction plates, on which the method of this application is based; Figure 2 This is a schematic diagram of the 3D point cloud module distribution of a specific embodiment of the vehicle-mounted dynamic detection method for track line induction panels in this application. Figure 3 This is a flowchart illustrating the joint calibration of a 3D camera assembly in a specific embodiment of the on-board dynamic detection method for track line induction panels according to this application. Figure 4 This is a block diagram of the structural composition of the mileage positioning and synchronization unit in a specific embodiment of the on-board dynamic detection system for track line induction plates on which the method of this application is based; Figure 5 This is a flowchart illustrating the detection of surface defects of the induction plate in a specific embodiment of the on-board dynamic detection method for induction plates of track lines in this application. Figure 6 This is a schematic diagram of geometric parameter points for a specific embodiment of the vehicle-mounted dynamic detection method for track line induction plates in this application.

[0026] In the diagram: 1-Induction board data acquisition unit, 2-Obstacle positioning and synchronization unit, 3-Intelligent analysis and processing unit, 4-Communication unit, 5-Power supply unit, 6-Storage unit, 7-Induction board, 8-Bracket, 9-Bolt, 10-Rail, 11-2D image acquisition module, 12-3D point cloud acquisition module, 121-Left 3D camera assembly, 122-Middle 3D camera assembly, 123-Right 3D camera assembly, 21-Photoelectric encoder, 22-Processing board, 23-RFID tag reader. Detailed Implementation

[0027] For the sake of clarity and reference, the technical terms, abbreviations, or acronyms used below will be recorded as follows: ICP: Iterative Closest Point algorithm; YOLOv10: An object detection network; CSPDarknet: A feature extraction module; PAN: Path Aggregation Network; ResNet-18: A lightweight classification network model; 2D: 2-dimensional; 3D: 3-dimensional; FPGA: Field Programmable Gate Array; CPU: Central Processing Unit; GPU: Graphics Processing Unit; JPG: An image compression format.

[0028] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] As attached Figure 1 To be continued Figure 6 As shown, a specific embodiment of the vehicle-mounted dynamic detection method for track line induction plates of this application is given. The application will be further described below with reference to the accompanying drawings and specific embodiments.

[0030] To achieve full-element acquisition of various defect features of the induction board and realize automatic and efficient detection of surface defects and geometric anomalies of the induction board, this application provides a highly efficient and robust on-board dynamic detection method for track induction boards, which can adapt to the complex and ever-changing railway line environment and maintain high detection efficiency while ensuring detection effect.

[0031] Example 1 An embodiment of the on-board dynamic detection method for track line induction panels according to this application specifically includes the following steps: S1) The 2D image acquisition module 11 acquires image data I of the area of ​​the sensor plate 7, bracket 8 and bolt 9 to obtain the original sensor plate image; S2) 3D point cloud acquisition module 12 acquires three-dimensional point cloud data P(x, y, z, i) of the area of ​​the sensing plate and the left and right rails 10, where x is the X-axis coordinate, y is the Y-axis coordinate, z is the Z-axis coordinate, and i is the reflection intensity of the point cloud. S3) Acquire encoder pulse signals and RFID tag signals deployed on the line to achieve synchronization with the line mileage, and simultaneously control the 2D image acquisition module 11 and the 3D point cloud acquisition module 12 to acquire data at equal intervals; S4) Sensor plate surface defect detection: Intelligent detection and analysis of the collected sensor plate image data to obtain defect information including sensor plate cracks, bracket cracks, missing bolts, and loose bolts; Sensor plate geometric parameter detection: The collected sensor plate point cloud data is processed to obtain the region, height, longitudinal displacement and misalignment information of each sensor plate 7; Steps S1) to S3) can be executed in any order.

[0032] The 3D point cloud acquisition module 12 further includes three sets of camera components, and step S2) further includes: S21) First, after acquiring point cloud data, perform filtering preprocessing and search for point clouds belonging to the same target between two frames of point clouds from adjacent camera components at the same mileage. S22) Then, the rotation matrix R and translation matrix T between the two frames of point clouds are calculated through point cloud feature registration.

[0033] As attached Figure 3 As shown, the point cloud feature registration in step S22) further includes the following process: S221) Perform preliminary coarse registration on the two point cloud data points PA and PB to obtain an initial rotation matrix R. ’ and translation matrix T ’ PB is obtained by performing the corresponding transformation on PB. ’ =PB*R ’ +T ’ ; Where PA is the point cloud data of the left camera component, and PB is the point cloud data of the camera component adjacent to the left camera component; S222) Sample the point cloud PA to obtain a set of sampling points; S223) in point cloud PB ’ Find the point closest to the sampling point to obtain a set of matching points; S224) Calculate the updated rotation matrix R based on the matching points. ’ and translation matrix T ’ For PB ’ PB is obtained by performing the corresponding transformation. ’ ’ =PB ’ *R ’ +T ’ ; S225) Repeat steps S222) to S224), use the mean square error as the objective function to calculate the registration error. When the registration error is less than the set threshold T or the number of iterations is equal to the maximum number of iterations, the registration ends. Save the obtained rotation matrix R and translation matrix T to achieve point cloud feature registration.

[0034] The track inspection train has photoelectric encoders 21 installed in its wheels and radio frequency tag readers 23 installed under its undercarriage. Step S3) further includes the following processes: S31) During the operation of the track inspection train, the processing board 22 collects the equally spaced pulse waveforms emitted by the photoelectric encoder 21, and uses the pulse waveforms after frequency doubling and frequency division as the input source for sampling triggering of the 2D image acquisition module 11 and the 3D point cloud acquisition module 12, and calculates the preliminary mileage accordingly. S32) Scan the RFID tag signal at the set sampling frequency. If the RFID tag signal is detected, look up the real mileage information represented by the tag in the tag ID and actual mileage comparison table, and then map this real mileage information onto each pulse signal to achieve synchronization with the line mileage.

[0035] In step S3), the current line mileage is further corrected according to the following formula: Let the actual mileages of two adjacent RFID tags be M1 and M2 respectively. n The corresponding trigger pulse counts are N1 and N. n Then in N1 and N n The actual distance between M i =M1+(N i -N1)*(M n -M1) / (N n -N1), i=1,2,3……n.

[0036] In step S4), the intelligent detection and analysis combines the target detection network and the classification network for judgment. The defect information further includes the defect category, mileage, the coordinates of the rectangle in the image (x, y, w, h), the actual physical size, and the defect severity level. x and y are the x and y coordinates of the top left vertex, respectively, and w and h are the width and height of the rectangle, respectively.

[0037] As attached Figure 5 As shown, step S4) of the sensor plate surface defect detection further includes the following process: S41) Perform intelligent detection and analysis on the collected sensor plate image data to obtain defect information including sensor plate cracks, bracket cracks, missing bolts, and loose bolts; S42) First, the target detection network model detects cracks in the sensing plate, cracks in the bracket, missing bolts, and normal bolts. Then, the sub-image of normal bolts is sent to the classification network model to determine whether they are loose.

[0038] The object detection network model annotates the original sensor plate images with object detection boxes, dividing them into training and validation sets. The label categories for the object detection box annotations include sensor plate cracks, support cracks, missing bolts, and normal bolts, represented by 0, 1, 2, and 3 respectively. The entire dataset is divided into 80% of the samples as the training set and 20% of the samples as the validation set.

[0039] In step S42), an object detection network model is constructed based on YOLOv10. The feature extraction part uses CSPDarknet as the backbone network and path aggregation network as the neck network. Large convolutional kernels and partition attention are used to enhance the model's context learning ability. Spatial-channel decoupled downsampling and rank-guided modules are used to improve the overall efficiency of the model.

[0040] The input size of the object detection network model is 1280. The output detection boxes of the object detection network model include the object category, the object box, and the confidence level. Load the training set and send it into the object detection network model for iterative training. Set the maximum number of iterations epoch = 400. After training for 10 epochs, turn off the Maosic augmentation operation, and only retain the conventional image augmentation operations including horizontal flipping, vertical flipping, and chromaticity transformation. Conduct a test on the validation set after each 1 epoch iteration, and record the mean average precision (mAP) information of the validation set. At the same time, save the model parameters with the highest mAP during the entire iterative training process to the local disk, and name the file best.pt. Save the model parameters obtained from the latest round of iterative training as the last.pt file. Input each image I to be detected into the network model loaded with the parameters best.pt for forward inference, and obtain several object detection boxes. The object categories 0, 1, 2, and 3 correspond to the crack of the induction plate, the crack of the bracket, the missing bolt, and the normal bolt respectively. Use the object detection box with the object category 3 to intercept the ROI region in the image I to obtain the bolt sub-image S i , i = 1, 2, 3... n, where n is the number of object detection boxes with the object category 3, and S i are respectively sent into the classification network model to determine whether the bolt is loose. The classification network model divides the induction plate bolt sub-image dataset into two categories: normal bolts and loose bolts, and the corresponding labels are 0 and 1 respectively. Use 80% of the normal bolt and loose bolt images as the classification training set, and the remaining 20% as the classification validation set. The classification network model loads the induction plate bolt sub-image dataset for iterative training, sets the maximum number of iterations to 300, sets the learning rate to 0.001, and adopts image augmentation operations such as random rotation, horizontal flipping, vertical flipping, and chromaticity transformation. Conduct a test on the validation set after each 1 epoch iteration, record the classification accuracy (ACC) information, and save the model parameters with the maximum ACC during the entire iterative training process as the best.param file. Save the model parameters obtained from the latest round of iterative training as the last.param file. The bolt sub-image S i is input into the classification model loaded with best.param, and after inference, obtain the confidence level conf0 belonging to the category 0 and the confidence level conf1 belonging to the category 1. If conf0 < conf1, then this bolt is loose; otherwise, it is a normal bolt.

[0041] In the point cloud processing in step S4), according to the distribution characteristics of the induction plates 7, the front, rear, left, and right edge positions of each induction plate 7 are extracted through edge detection, so as to segment the point cloud of each induction plate 7. The height information of each position of the induction plate 7 is further obtained from the z coordinate of the point cloud. The longitudinal displacement information of the induction plate 7 is calculated by calculating the offset of the gap between two adjacent induction plates 7 exceeding the standard value range. The induction plate misalignment information includes the lateral difference, height, and angle deviation between adjacent induction plates 7. The vertices of the induction plate 7 are obtained through the boundary intersection points of the induction plate 7. The lateral and height deviations of the induction plate 7 are obtained by comparing the coordinate x deviation and coordinate z deviation between the corresponding vertices of adjacent induction plates 7. Then, the angle deviation is solved through the induction plate plane normal vector obtained by fitting calculation.

[0042] The detection of the geometric parameters of the induction plate in step S4) further includes the following process: S411) First, the induction plates are segmented by block, and the upper, lower, left, and right edge positions of the induction plate 7 are extracted through gradient edge detection, so as to segment the point cloud Psb of each induction plate 7 i , i = 1, 2, 3... n; S412) Perform bilateral filtering on the point cloud Psb i to eliminate noise and obtain Psb i ’ , and the z coordinate of each point in the induction plate point cloud Psb i ’ is the height value of each position of the induction plate. Extract the maximum value Max and minimum value Min of the z coordinates in Psb i ’ ; S413) If the maximum value Max > the allowed ultra-high value, it is determined that the induction plate is abnormally ultra-high. On the contrary, it is judged whether the minimum value Min < the allowed ultra-low value holds. If it holds, it is determined that the induction plate is abnormally too low. Otherwise, the height of the induction plate is normal; S414) The longitudinal displacement information of the induction plate 7 is calculated by calculating the offset of the gap between two adjacent induction plates 7 exceeding the standard value range. d1 is the longitudinal displacement between the first induction plate Psb1 ’ and the second induction plate Psb2 ’ , and d2 is the longitudinal displacement between the second induction plate Psb2 ’ and the third induction plate Psb3 ’ ; S415) If the allowed minimum displacement Offset Min < d1 < the allowed maximum displacement Offset Max , it is determined that the longitudinal displacement of the first induction plate is normal. Otherwise, it is determined that the longitudinal displacement of the first induction plate is abnormal.

[0043] The information on misaligned sensor plates includes the lateral, height, and angular deviations between adjacent sensor plates 7. Step S4) further includes the following process for detecting the geometric parameters of the sensor plates: S421) The vertices of the sensing plates are obtained through the boundary intersection of the sensing plates 7. The vertices of the first sensing plate are A1, B1, C1, and D1, the vertices of the second sensing plate are A2, B2, C2, and D2, and so on. Each vertex includes x, y, and z coordinates. S422) c1 is the lateral deviation between the first and second sensing plates, which is obtained by taking the absolute value of the difference between the X-axis coordinate D1(x) of D1 and the X-axis coordinate C2(x) of C2, i.e. |D1(x)-C2(x)|. c2 is the lateral deviation between the second and third sensing plates. If the lateral deviation is greater than the lateral deviation threshold Tx, it is considered to be misaligned; otherwise, it is judged to be normal. S423) The height deviation is obtained by taking the absolute value of the difference between the coordinates z of the corresponding vertices of the adjacent sensing plates 7. The height deviation between the first sensing plate and the second sensing plate is |D1(z)-C2(z)| and |B1(z)-A2(z)|. If the height deviation is greater than the height deviation threshold Tz, it is judged as misaligned tooth; otherwise, it is judged as normal height deviation.

[0044] Step S423) includes: The point cloud of the first sensing plate is defined by vertices A1, B1, C1, and D1. Then, the plane equation Ax + By + Cz + D = 0 of the surface of sensing plate 7 is iteratively calculated using a plane fitting algorithm based on random sampling consistency. Its normal vector is n1(A, B, C). The normal vector n2(A, B, C) of the second sensing plate is then calculated. Finally, the angle deviation is obtained by solving the dot product problem using the fitted plane normal vectors of the sensing plates. , if the angle deviation If the angle deviation threshold is met, it is considered a misaligned tooth; otherwise, it is considered normal.

[0045] The method for on-board dynamic detection of track line induction panels further includes: S5) After obtaining the detection result through step S4), the detection result information and the corresponding raw data are transmitted to the ground user terminal, and the operators are notified to formulate a work plan.

[0046] Example 2 As attached Figure 1 As shown, an embodiment of the track line induction plate vehicle-mounted dynamic detection system 10 based on the method described in Embodiment 1 is installed on a track inspection train and specifically includes: The sensor board data acquisition unit 1 includes a 2D image acquisition module 11 and a 3D point cloud acquisition module 12; The 2D image acquisition module 11 acquires image data I of the area of ​​the sensor plate 7, the bracket 8 and the bolt 9 to obtain the original image of the sensor plate; The 3D point cloud acquisition module 12 acquires three-dimensional point cloud data P(x, y, z, i) of the sensor plate and the left and right rails 10 areas; where x is the X-axis coordinate, which refers to the positive direction of the X-axis from left to right along the direction of sleeper placement when facing the extension direction of the rail 10; y is the Y-axis coordinate, which is the positive direction of the Y-axis along the forward direction of the rail 10; z is the Z-axis coordinate, which is defined by the X and Y axes through the right-hand screw rule, indicating the height of the sensor plate 7; and i is the reflection intensity of the point cloud. Mileage positioning and synchronization unit 2 collects encoder pulse signals and RFID tag signals deployed on the line to achieve synchronization with the line mileage, and at the same time outputs trigger pulses to the sensor board data acquisition unit 1 to collect data at equal intervals; The intelligent analysis and processing unit 3 includes a sensor plate surface defect detection module and a sensor plate geometric parameter detection module. The sensor plate surface defect detection module sends the collected sensor plate image data to the intelligent detection model for analysis, obtaining defect information including sensor plate cracks, bracket cracks, missing bolts, and loose bolts. The sensor plate geometric parameter detection module sends the collected sensor plate point cloud data to the point cloud processing model, obtaining the region, height, longitudinal displacement, and misalignment information for each sensor plate 7.

[0047] The onboard dynamic detection method for track line induction panels also includes a communication unit 4, comprising a wired or wireless communication module. After the intelligent analysis and processing unit 3 calculates the detection results, it sends the detection result information and corresponding raw data to the ground user terminal through the communication unit 4, and notifies the operators to specify the work plan. The dynamic detection system also includes a storage unit 6, used to store the collected raw data in a designated location. The dynamic detection system also includes a power supply unit, providing power to the entire system, including AC / DC conversion and voltage regulation.

[0048] The 2D image acquisition module includes two 2D camera assemblies, each consisting of a high-speed linear scan camera, a high-resolution lens, a laser, and peripheral circuitry. These assemblies are responsible for acquiring image data I of the sensor plate (body) 7, the support bracket 8, and the bolt 9 area. The 3D point cloud acquisition module 12 further includes three camera assemblies, each comprising a high-speed 3D area scan camera, a high-resolution lens, a line laser source, and peripheral circuitry. The 2D image acquisition module can be tilted towards the central axis of the sensor plate by adjusting its installation angle to ensure clear visibility of the sensor plate support area. Each acquisition module utilizes a dedicated laser source for supplemental lighting, ensuring the system can operate normally under different lighting conditions.

[0049] like Figure 2As shown, the 3D point cloud acquisition modules 12 have a certain range of overlapping fields of view between each pair of adjacent (3D) camera components, which is used to calibrate the three 3D point cloud acquisition modules 12 to the same coordinate system. The calibration process is attached. Figure 3 As shown, the 3D point cloud acquisition module 12 first acquires point cloud data and then performs filtering preprocessing to search for point clouds belonging to the same target between two frames of point clouds from adjacent camera components at the same mileage. Then, it calculates the transformation matrix (including rotation matrix R and translation matrix T) between the two frames of point clouds using a point cloud feature registration algorithm (such as ICP).

[0050] The 3D point cloud acquisition module 12 performs preliminary coarse registration on the two point cloud data sets PA and PB to obtain an initial rotation matrix R. ’ and translation matrix T ’ PB is obtained by performing the corresponding transformation on PB. ’ =PB*R ’ +T ’ PA represents the point cloud data of the left camera component, and PB represents the point cloud data of the camera component adjacent to the left camera component. Point cloud PA is sampled to obtain a set of sampled points, which are then used in point cloud PB. ’ Find the point closest to the sampling point to obtain a set of matching points, and then calculate the updated rotation matrix R based on the matching points. ’ and translation matrix T ’ For PB ’ PB is obtained by performing the corresponding transformation. ’ ’ =PB ’ *R ’ +T ’ For PB ’ The process involves iterative updates, using mean square error loss (MSE) as the objective function to calculate the registration error. Registration ends when the registration error is less than a set threshold T or the number of iterations equals the maximum number of iterations. The resulting rotation matrix R and translation matrix T are then saved to achieve point cloud feature registration.

[0051] As attached Figure 4As shown, the mileage positioning and synchronization unit 2 further includes a photoelectric encoder 21, a processing board 22, and an RFID reader 23. The photoelectric encoder 21 is installed in the wheels of the track inspection train. During the operation of the track inspection train, the processing board 22 collects the equally spaced pulse waveforms emitted by the photoelectric encoder 21, and uses the pulse waveforms after frequency multiplication and division as the input source for sampling triggering of the 2D image acquisition module 11 and the 3D point cloud acquisition module 12. The sampling interval of one trigger pulse cycle is set to 0.5 mm, then the distance after 1000 pulses is 0.5 m, and the preliminary mileage is calculated accordingly. Since the wheel diameter inevitably changes due to wear or other conditions during operation, the sampling interval is not strictly equal to 0.5 mm. Therefore, mileage calibration and synchronization are required through RFID tags deployed on the line. In this embodiment, the RFID reader 23 is installed under the track inspection train and scans the RFID tag signal at a set sampling frequency (e.g., 1000Hz). If the RFID tag signal is sensed, the reader queries the tag ID and actual mileage table to find the actual mileage information represented by the tag, and then maps this actual mileage information onto each pulse signal to achieve synchronization with the line mileage.

[0052] Mileage positioning and synchronization unit 2 further corrects the line mileage according to the following formula: Let the actual mileages of two adjacent RFID tags be M1 and M2 respectively. n The corresponding trigger pulse counts are N1 and N. n (The mileage calculated based on a constant pulse interval of 0.5mm is N1*0.5mm, N) n *0.5mm), then in N1 and N n The actual distance between M i =M1+(N i -N1)*(M n -M1) / (N n -N1), i=1,2,3……n.

[0053] To ensure storage efficiency, the raw data collected in this application is stored in storage unit 6, which can specifically be a solid-state drive. Due to the large amount of point cloud data, a 16-bit depth map is used to store the raw data. To further reduce the data volume, a data compression algorithm is used to compress and store the raw data. Since 3D data is sensitive to distance information, lossy compression methods can lead to data distortion and detection errors. Therefore, a lossless compression method is required. First, differential encoding is performed, followed by positive and negative number remapping, and a combination of run-length encoding and variable-length encoding to reduce the number of bits in the entire image. The 2D raw image data can be compressed using a common lossy compression algorithm. The system uses a JPG compression algorithm to compress 2000 images and store them in a single large file to ensure data storage and copying efficiency.

[0054] The intelligent analysis and processing unit 3 consists of an industrial-grade computer and corresponding detection software. To improve detection efficiency, this application uses a CPU (Central Processing Unit) for data processing and a GPU (Graphics Processing Unit) for computational acceleration. The sensor plate appearance defect detection module sends the collected sensor plate image data to the intelligent detection model for analysis, obtaining defect information including sensor plate cracks, support cracks, missing bolts, and loose bolts. The intelligent detection model combines a target detection network model and a classification network model for judgment. (See attached...) Figure 5 As shown, the intelligent detection model first detects cracks in the induction plate, cracks in the bracket, missing bolts, and normal bolts through a target detection network model. Then, the sub-image of normal bolts is fed into a classification network model to determine whether they are loose. The defect information further includes the defect category, the mileage, the coordinates of the rectangle in the image (x-coordinate and y-coordinate of the top left vertex, width w, and height h of the rectangle), as well as the actual physical dimensions and the severity level of the defect.

[0055] The object detection network model annotates the original sensor plate images with object detection boxes, dividing them into training and validation sets. The label categories for the object detection box annotations include sensor plate cracks, support cracks, missing bolts, and normal bolts, represented by 0, 1, 2, and 3 respectively. The entire dataset is divided into 80% of the samples as the training set and 20% of the samples as the validation set.

[0056] The intelligent analysis and processing unit 3 constructs an object detection network model based on YOLOv10. The feature extraction part uses CSPDarknet as the backbone network and Path Aggregation Network (PAN) as the neck network. Large convolutional kernels and partition attention are used to enhance the model's context learning ability. Spatial-channel decoupled downsampling and rank-guided modules improve the overall efficiency of the model.

[0057] The input size of the object detection network model is 1280 (an integer multiple of 32). The output detection boxes of the object detection network model include the object category, the object box (the x and y coordinates x1, y1 of the upper left vertex and the x, y coordinates x2, y2 of the lower right vertex), and the confidence. The training set is loaded and sent into the object detection network model for iterative training. The maximum number of iterations is set to epoch = 400. After training for 10 epochs, the Maosic augmentation operation is turned off, and only the conventional image augmentation operations including horizontal flipping, vertical flipping, and chromaticity transformation are retained. After each epoch iteration (traversing all training samples), the validation set is tested once, and the average accuracy mAP (mean Average Precision) information of the validation set is recorded. At the same time, the model parameters with the highest average accuracy mAP during the entire iterative training process are saved to the local disk, and the file is named best.pt. The model parameters obtained from the latest round of iterative training are saved as the last.pt file. Each image I to be detected is input into the network model loaded with the parameters best.pt for forward inference, and several object detection boxes are obtained. The object categories 0, 1, 2, 3 correspond to the induction plate crack, the bracket crack, the bolt missing, and the normal bolt respectively. The ROI (Region Of Interesting) region is intercepted from the image I using the object detection box with the object category 3 to obtain the bolt sub-image S i , i = 1, 2, 3... n, where n is the number of object detection boxes with the object category 3, and S i are respectively sent into the classification network model to determine whether the bolt is loose. The classification network model divides the induction plate bolt sub-image dataset into two categories: normal bolts and loose bolts, and the corresponding labels are 0 and 1 respectively. 80% of the normal bolt and loose bolt images are used as the classification training set, and the remaining 20% are used as the classification validation set. The classification network model loads the induction plate bolt sub-image dataset for iterative training. The maximum number of iterations is set to 300, the learning rate is set to 0.001, and the image augmentation operations of random rotation (angle limited within 5°), horizontal flipping, vertical flipping, and chromaticity transformation are adopted. After each epoch iteration (traversing all training samples), the validation set is tested once, the classification accuracy ACC information is recorded, and the model parameters with the maximum accuracy ACC during the entire iterative training process are saved as the best.param file. The model parameters obtained from the latest round of iterative training are saved as the last.param file. The bolt sub-image S i is input into the classification model loaded with best.param. After inference, the confidence conf0 belonging to category 0 and the confidence conf1 belonging to category 1 are obtained. If conf0 < conf1, this bolt is loose; otherwise, it is a normal bolt.

[0058] According to the distribution characteristics of the induction plate 7, the point cloud processing model extracts the front, rear, left, and right edge positions of each induction plate 7 through edge detection, thereby segmenting the point cloud of each induction plate 7. The height information of each position of the induction plate 7 is further obtained from the coordinate z of the point cloud. The longitudinal displacement information of the induction plate 7 is calculated by calculating the offset of the gap between two adjacent induction plates 7 exceeding the standard value range. The misalignment information of the induction plates includes the lateral difference, height, and angle deviation between adjacent induction plates 7. The vertices of the induction plate 7 are obtained through the boundary intersection points of the induction plate 7. The lateral and height deviations of the induction plate 7 are obtained by comparing the coordinate x deviation and coordinate z deviation between the corresponding vertices of adjacent induction plates 7. Then, the angle deviation is solved by fitting the normal vector of the induction plate plane calculated.

[0059] As shown in the appendix Figure 6 As shown, the induction plate geometric parameter detection module first divides the induction plate into blocks. When further segmenting the induction plate (body) 7, the significant height difference between the induction plate (body) 7 and its surroundings is fully utilized. The upper, lower, left, and right edge positions of the induction plate 7 can be extracted through gradient edge detection, thereby segmenting the point cloud Psb of each induction plate 7 i , i = 1, 2, 3... n. For the point cloud Psb i Perform bilateral filtering to eliminate noise to obtain Psb i ’ , the z coordinate of each point in the induction plate point cloud Psb i ’ is the height value of each position of the induction plate. Extract the maximum value Max and minimum value Min of the z coordinate in Psb i ’ . If the maximum value Max > the allowable ultra-high value, it is determined that the induction plate is abnormally ultra-high. On the contrary, it is judged whether the minimum value Min < the allowable ultra-low value holds. If it holds, it is determined that the induction plate is abnormally too low. Otherwise, the height of the induction plate is normal. The longitudinal displacement information of the induction plate 7 is calculated by calculating the offset of the gap between two adjacent induction plates 7 exceeding the standard value range. d1 is the longitudinal displacement between the first induction plate Psb1 ’ and the second induction plate Psb2 ’ , d2 is the longitudinal displacement between the second induction plate Psb2 ’ and the third induction plate Psb3 ’ . If the allowable minimum displacement Offset Min < d1 < the allowable maximum displacement Offset Max , it is determined that the longitudinal displacement of the first induction plate is normal. Otherwise, it is determined that the longitudinal displacement of the first induction plate is abnormal.

[0060] The information on misaligned sensor plates includes the lateral, height, and angular deviations between adjacent sensor plates 7. The sensor plate geometric parameter detection module obtains the vertices of the sensor plates through the boundary intersections of the sensor plates 7. The vertices of the first sensor plate are A1, B1, C1, and D1, the vertices of the second sensor plate are A2, B2, C2, and D2, and so on. Each vertex includes x, y, and z coordinates. c1 is the lateral deviation between the first and second sensor plates, obtained by taking the absolute value of the difference between the x-axis coordinate D1(x) of D1 and the x-axis coordinate C2(x) of C2, i.e., |D1(x)-C2(x)|. c2 is the lateral deviation between the second and third sensor plates. If the lateral deviation is greater than the lateral deviation threshold Tx, it is considered a misalignment; otherwise, it is considered normal. The height deviation is obtained by taking the absolute value of the difference between the coordinates z of the corresponding vertices of the adjacent sensing plates 7. The height deviation between the first sensing plate and the second sensing plate is |D1(z)-C2(z)| and |B1(z)-A2(z)|. If the height deviation is greater than the height deviation threshold Tz, it is judged as misaligned tooth; otherwise, it is judged as normal height deviation.

[0061] The induction plate geometric parameter detection module defines the point cloud of the first induction plate using vertices A1, B1, C1, and D1. Then, it iteratively calculates the plane equation Ax + By + Cz + D = 0 for the surface of induction plate 7 using a plane fitting algorithm based on Random Sample Consensus (RANSAC), with its normal vector n1 (A, B, C). It then calculates the normal vector n2 (A, B, C) for the second induction plate. Finally, the angle deviation is obtained by solving the dot product of the fitted induction plate plane normal vector. , if the angle deviation If the angle deviation threshold is met, it is considered a misaligned tooth; otherwise, it is considered normal.

[0062] After the intelligent analysis and processing unit 3 calculates the detection results, it needs to send them to the ground user terminal (i.e., the ground management platform) via the communication unit 4, and notify the operators to formulate a work plan. The communication unit 4 further transmits the detection result information and corresponding raw data to the ground user terminal via 4G / 5G wireless communication according to HTTP / HTTPS, TCP / IP, and UDP protocols. The power supply for the entire system is provided by the power supply unit 5. First, the power supply unit 5 connects to the AC 220V power supply from the vehicle-mounted terminal, and then converts the AC / DC power supply to the DC 110V and DC 24V voltage outputs required by the system through the AC / DC conversion module, which are then used by the various units of the system.

[0063] In the description of this application, it should be noted that when an element is referred to as being "fixed to" or "set on" another element, it can be directly set on the other element or indirectly set on the other element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or indirectly connected to the other element.

[0064] It should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0065] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" or "several" means two or more, unless otherwise explicitly specified.

[0066] It should be noted that the structures, proportions, sizes, etc., shown in the accompanying drawings of this specification are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the conditions under which this application can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and purposes that this application can produce, should still fall within the scope of the technical content disclosed in this application.

[0067] By implementing the technical solution of the on-board dynamic detection method for track induction panels described in the specific embodiments of this application, the following technical effects can be achieved: (1) The on-board dynamic detection method for the induction board of the rail transit described in the specific embodiment of this application automatically processes the three-dimensional point cloud data of the induction board and analyzes the geometric parameter information of the induction board such as abnormal height, longitudinal displacement, and misaligned teeth, providing a reliable basis for the operation and maintenance of the induction board, realizing the on-board dynamic detection of the rail transit induction board, greatly improving the detection efficiency and accuracy, and comprehensively covering the detection of the appearance and geometric parameters of the induction board. (2) The on-board dynamic detection method for track line induction plate described in the specific embodiments of this application uses target detection network and classification network to intelligently analyze surface cracks, missing bolts, and loosening defects of the induction plate, realizing end-to-end detection of surface defects of the induction plate, and further improving detection efficiency and accuracy; (3) The on-board dynamic detection method for track induction plate described in the specific embodiments of this application uses a 2D camera assembly to collect the surface texture features of the induction plate body and the support, and a 3D camera assembly to collect the rail outline and the three-dimensional point cloud information of the induction plate. It can be mounted on a train to collect and process induction plate data at high speed, realize the automatic and efficient detection of surface defects and geometric parameters of the induction plate, and guide relevant management and operation personnel to maintain the induction plate.

[0068] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0069] The above description is merely a preferred embodiment of this application and is not intended to limit this application in any way. Although this application has been disclosed above with reference to preferred embodiments, it is not intended to limit this application. Any person skilled in the art can make many possible variations and modifications to the technical solutions of this application using the methods and techniques disclosed above, or modify them into equivalent embodiments with equivalent changes, without departing from the spirit and technical essence of this application. Therefore, any simple modifications, equivalent substitutions, equivalent changes, and modifications made to the above embodiments based on the technical essence of this application without departing from the content of the technical solutions of this application shall still fall within the protection scope of the technical solutions of this application.

Claims

1. A method for onboard dynamic detection of track line induction panels, characterized in that, Includes the following steps: S1) The 2D image acquisition module (11) acquires image data I of the sensor plate (7), bracket (8) and bolt (9) area to obtain the original sensor plate image; S2) 3D point cloud acquisition module (12) acquires three-dimensional point cloud data P(x, y, z, i) of the area of ​​the induction plate and the left and right rails (10), where x is the X-axis coordinate, y is the Y-axis coordinate, z is the Z-axis coordinate, and i is the reflection intensity of the point cloud; S3) Collect encoder pulse signals and RFID tag signals deployed on the line to achieve synchronization with the line mileage, and at the same time control the 2D image acquisition module (11) and 3D point cloud acquisition module (12) to collect data at equal intervals; S4) Sensor plate surface defect detection: Intelligent detection and analysis of the collected sensor plate image data to obtain defect information including sensor plate cracks, bracket cracks, missing bolts, and loose bolts; Geometric parameter detection of sensor plate: The collected point cloud data of sensor plate is processed to obtain the area, height, longitudinal displacement and misalignment information of each sensor plate (7); The point cloud processing in step S4) is based on the distribution characteristics of the sensing plate (7). The front, back, left and right edge positions of each sensing plate (7) are extracted by edge detection, thereby segmenting the point cloud of each sensing plate (7). The height information of each position of the sensing plate (7) is further obtained from the coordinate z of the point cloud. The longitudinal displacement information of the sensing plate (7) is obtained by calculating the offset of the gap between two adjacent sensing plates (7) that exceeds the standard value range. The misalignment information of the sensing plate includes the lateral difference, height and angle deviation between adjacent sensing plates (7). The vertex of the sensing plate (7) is obtained through the boundary intersection of the sensing plates (7). The lateral and height deviations of the sensing plate (7) are obtained by comparing the coordinate x deviation and coordinate z deviation between the corresponding vertices of adjacent sensing plates (7). Then, the angle deviation is solved by fitting the normal vector of the sensing plate plane.

2. The on-board dynamic detection method for track line induction board according to claim 1, wherein the 3D point cloud acquisition module (12) comprises three sets of camera components, characterized in that, Step S2) further includes: S21) First, after acquiring point cloud data, perform filtering preprocessing and search for point clouds belonging to the same target between two frames of point clouds from adjacent camera components at the same mileage. S22) Then, the rotation matrix R and translation matrix T between the two frames of point clouds are calculated through point cloud feature registration.

3. The on-board dynamic detection method for track line induction panels according to claim 2, characterized in that, The point cloud feature registration in step S22) is further... Includes the following processes: S221) Perform preliminary coarse registration on the two point cloud data points PA and PB to obtain an initial rotation matrix R. ’ and translation matrix T ’ PB is obtained by performing the corresponding transformation on PB. ’ =PB*R ’ +T ’ ; Where PA is the point cloud data of the left camera component, and PB is the point cloud data of the camera component adjacent to the left camera component; S222) Sample the point cloud PA to obtain a set of sampling points; S223) in point cloud PB ’ Find the point closest to the sampling point to obtain a set of matching points; S224) Calculate the updated rotation matrix R based on the matching points. ’ and translation matrix T ’ For PB ’ PB is obtained by performing the corresponding transformation. ’ ’ =PB ’ *R ’ +T ’ ; S225) Repeat steps S222) to S224), use the mean square error as the objective function to calculate the registration error. When the registration error is less than the set threshold T or the number of iterations is equal to the maximum number of iterations, the registration ends. Save the obtained rotation matrix R and translation matrix T to achieve point cloud feature registration.

4. The on-board dynamic detection method for track line induction plates according to any one of claims 1 to 3, wherein photoelectric encoders (21) are installed in the wheels of the track inspection train, and radio frequency tag readers (23) are installed under the track inspection train, characterized in that, Step S3) further Includes the following processes: S31) During the operation of the track inspection train, the processing board (22) collects the equally spaced pulse waveforms emitted by the photoelectric encoder (21), and uses the pulse waveforms after frequency doubling and frequency division as the input source for sampling triggering of the 2D image acquisition module (11) and the 3D point cloud acquisition module (12), and calculates the preliminary mileage accordingly. S32) Scan the RFID tag signal at the set sampling frequency. If the RFID tag signal is detected, look up the real mileage information represented by the tag in the tag ID and actual mileage comparison table, and then map this real mileage information onto each pulse signal to achieve synchronization with the line mileage.

5. The on-board dynamic detection method for track line induction panels according to claim 4, characterized in that, In step S3), the current line mileage is further corrected according to the following formula: Let the actual mileages of two adjacent RFID tags be M1 and M2, respectively. n The corresponding trigger pulse counts are N1 and N. n Then in N1 and N n The actual distance between M i =M1+(N i -N1)*(M n -M1) / (N n -N1), i=1,2,3……n.

6. The on-board dynamic detection method for track line induction plates according to claim 1, 2, 3 or 5, characterized in that: The intelligent detection and analysis in step S4) combines the target detection network and the classification network to make a judgment; the defect information includes the defect category, the mileage, the coordinates of the rectangle in the image (x, y, w, h), the actual physical size, and the defect severity level, where x and y are the x and y coordinates of the top left vertex, and w and h are the width and height of the rectangle, respectively.

7. The on-board dynamic detection method for track line induction panels according to claim 6, characterized in that, The detection of apparent defects in the induction plate in step S4) further includes the following process: S41) Perform intelligent detection and analysis on the collected sensor plate image data to obtain defect information including sensor plate cracks, bracket cracks, missing bolts, and loose bolts; S42) First, the target detection network model detects cracks in the sensing plate, cracks in the bracket, missing bolts, and normal bolts. Then, the sub-image of normal bolts is sent to the classification network model to determine whether they are loose.

8. The on-board dynamic detection method for track line induction panels according to claim 7, characterized in that: The target detection network model annotates the original sensor plate image with target detection boxes and divides it into training set and validation set. The label categories of the target detection box annotation include sensor plate crack, support crack, missing bolt and normal bolt, which are represented by 0, 1, 2 and 3 respectively. The entire dataset is divided into 80% of the samples as training set and 20% of the samples as validation set.

9. The on-board dynamic detection method for track line induction plates according to claim 7 or 8, characterized in that: In step S42), an object detection network model is constructed based on YOLOv10. The feature extraction part uses CSPDarknet as the backbone network and path aggregation network as the neck network. Large convolutional kernels and partition attention are used to enhance the model's context learning ability. Spatial-channel decoupled downsampling and rank-guided modules are used to improve the overall efficiency of the model.

10. The on-board dynamic detection method for track line induction panels according to claim 9, characterized in that: The input size of the target detection network model is 1280. The output detection box of the target detection network model includes the target category, the target box, and the confidence score. The training set is loaded into the target detection network model for iterative training. The maximum number of iterations is set to 400 epochs. After training for 10 epochs, the Maosic enhancement operation is turned off, and only the conventional image enhancement operations, including horizontal flipping, vertical flipping, and color transformation, are retained. After each epoch, a validation set test is performed, and the average accuracy (mAP) of the validation set is recorded. At the same time, the model parameters with the highest average accuracy (mAP) during the entire iterative training process are saved to the local disk and named best.pt. The model parameters obtained from the latest round of iterative training are saved as last.pt. Each image to be detected, I, is input into the network model loaded with parameters best.pt for forward inference to obtain several target detection boxes. The target categories 0, 1, 2, and 3 correspond to the sensor plate crack, the bracket crack, the missing bolt, and the normal bolt, respectively. Intercept the ROI region in image I using the object detection box with the target category of 3 to obtain the bolt sub-image S i , i = 1, 2, 3... n, where n is the number of object detection boxes with the target category of 3, and S i are respectively fed into the classification network model to determine whether the bolt is loose; the classification network model divides the induction plate bolt sub-image dataset into two categories: normal bolts and loose bolts, with the corresponding labels being 0 and 1 respectively. 80% of the normal bolt and loose bolt images are used as the classification training set, and the remaining 20% is used as the classification validation set; the classification network model loads the induction plate bolt sub-image dataset for iterative training. The maximum number of iterations is set to 300, the learning rate is set to 0.001, and image enhancement operations such as random rotation, horizontal flipping, vertical flipping, and chromaticity transformation are adopted. The validation set is tested once every 1 epoch iteration, and the classification accuracy ACC information is recorded. The model parameters with the maximum accuracy ACC during the entire iterative training process are saved as the best.param file, and the model parameters obtained from the latest round of iterative training are saved as the last.param file; the bolt sub-image S i is input into the classification model loaded with best.param for inference to obtain the confidence conf0 belonging to category 0 and the confidence conf1 belonging to category 1. If conf0 < conf1, this bolt is loose; otherwise, it is a normal bolt.

11. The on-board dynamic detection method for track line induction panels according to any one of claims 1, 2, 3, 5, 7, 8 or 10, characterized in that, The detection of the geometric parameters of the induction plate in step S4) is further... Includes the following processes: S411) First, the sensor plate is divided into blocks. The top, bottom, left and right edge positions of the sensor plate (7) are extracted by gradient edge detection, thereby segmenting the point cloud Psb of each sensor plate (7). i , i = 1, 2, 3...n; S412) for point cloud Psb i Psb is obtained by performing bilateral filtering to eliminate noise. i ’ Psb point cloud of sensor panel i ’ The z-coordinate of each point is the height value of the sensor plate at each position. Psb is then extracted. i ’ The maximum and minimum values ​​of the z-coordinate in the middle. S413) If the maximum value Max > the allowed high value, then the sensor panel is determined to be abnormally high. Otherwise, it is determined whether the minimum value Min < the allowed low value. If it is true, the sensor panel is determined to be abnormally low. Otherwise, the sensor panel height is normal. S414) The longitudinal displacement information of the sensing plate (7) is obtained by calculating the offset of the gap between two adjacent sensing plates (7) that exceeds the standard value range, where d1 is the first sensing plate Psb1 ’ With the second sensor board Psb2 ’ The longitudinal displacement between them, d2 is the second sensing plate Psb2 ’ With the third sensor board Psb3 ’ Longitudinal displacement between them; S415) If the minimum allowable displacement Offset is satisfied Min <d1 < the maximum allowable displacement Offset Max , it is determined that the longitudinal displacement of the first induction plate is normal; otherwise, it is determined that the longitudinal displacement of the first induction plate is abnormal.

12. The on-board dynamic detection method for track line induction panels according to claim 11, characterized in that, The misalignment information of the sensing plate includes the lateral, height and angular deviations between adjacent sensing plates (7), and the sensing plate geometric parameter detection in step S4) further includes the following process: S421) The vertex of the sensing plate is obtained through the boundary intersection of the sensing plate (7). The vertices of the first sensing plate are A1, B1, C1, and D1, the vertices of the second sensing plate are A2, B2, C2, and D2, and so on. Each vertex includes x, y, and z coordinates. S422) c1 is the lateral deviation between the first and second sensing plates, which is obtained by taking the absolute value of the difference between the X-axis coordinate D1(x) of D1 and the X-axis coordinate C2(x) of C2, i.e. |D1(x)-C2(x)|. c2 is the lateral deviation between the second and third sensing plates. If the lateral deviation is greater than the lateral deviation threshold Tx, it is considered to be misaligned; otherwise, it is judged to be normal. S423) The height deviation is obtained by taking the absolute value after subtracting the coordinates z of the corresponding vertices of the adjacent sensing plates (7). The height deviation between the first sensing plate and the second sensing plate is |D1(z)-C2(z)| and |B1(z)-A2(z)|. If the height deviation is greater than the height deviation threshold Tz, it is judged as misaligned tooth; otherwise, it is judged as normal height deviation.

13. The on-board dynamic detection method for track line induction panels according to claim 12, characterized in that, Step S423) further includes: The point cloud of the first sensing plate is defined by vertices A1, B1, C1, and D1. Then, the plane equation Ax + By + Cz + D = 0 of the surface of the sensing plate (7) is calculated iteratively using a plane fitting algorithm based on random sampling consistency. Its normal vector is n1 (A, B, C). The normal vector n2 (A, B, C) of the second sensing plate is also calculated. The angle deviation is then solved by the dot product of the plane normal vector of the sensing plate calculated by fitting. , if the angle deviation If the angle deviation threshold is met, it is considered a misaligned tooth; otherwise, it is considered normal.

14. The on-board dynamic detection method for track line induction panels according to claims 1, 2, 3, 5, 7, 8, 10, 12 or 13, characterized in that, The method further includes: S5) After the detection result is obtained through step S4), the detection result information and the corresponding raw data are transmitted to the ground user terminal, and the operators are notified to formulate an operation plan.

Images