CN-122023401-A - Intelligent detection system and device for tread damage of heavy-duty train wheel set

CN122023401ACN 122023401 ACN122023401 ACN 122023401ACN-122023401-A

Abstract

The invention discloses an intelligent detection system and device for tread damage of a wheel set of a heavy-duty train, and relates to the technical field of train detection. The system comprises four modules of image acquisition, pretreatment, tread defect detection and intelligent terminal display interface, and the device comprises an industrial personal computer and a wheel set tread defect detection device. The image acquisition module performs fixed-point shooting through symmetrically arranged array vision subsystems, pre-extracts tread areas through ROI cutting, and the preprocessing module performs anti-reflection operation and combines SIFT algorithm and weighted fusion to realize image splicing. The tread defect detection is carried out, a UGate-MSNet model is constructed to divide defect areas, feature extraction is enhanced through a GCM (global motion model) and an MSFM (MSFM) module, defects are precisely positioned through a DV-YOLO model, and the multi-scale feature fusion capability is improved through a double-branch structure and an XX-FPN module. The intelligent terminal realizes the visualization and alarm of the detection result, the device does not need to stop or slow down the train, and the dynamic and high-precision detection requirements of the safety operation and maintenance of the heavy-duty train are met.

Inventors

LIANG XIAO
WANG ZHIWEI
LUO PEIJIAN
LIU PENGFEI
WANG XUEWEI
SHEN YONGJUN
ZHANG CHAOGANG
YUAN LAI

Assignees

石家庄铁道大学

Dates

Publication Date: 20260512
Application Date: 20260410

Claims (8)

1. The intelligent detection system for the tread damage of the heavy-duty train wheel set is characterized by comprising an image acquisition module, an image preprocessing module, a tread defect detection module and an intelligent terminal display interface; the image acquisition module acquires wheel set images, transmits the wheel set images to the image preprocessing module, the image preprocessing module inputs the preprocessed images into the tread defect detection module to obtain defect information, and the intelligent terminal display interface acquires the preprocessed tread images and the preprocessed defect information and performs visual display Constructing UGate-MSNet models and tread defect region data sets in the tread defect detection modules to divide tread defect regions; And further constructing a DV-YOLO model and a tread defect detection data set to accurately detect and position the defects.
2. The intelligent detection system for tread damage of heavy-duty train wheel set according to claim 1, wherein the image acquisition module shoots tread images through a camera at fixed points, acquires train wheel set tread images with consistent positions and pixel sizes of the images of the train, pre-extracts tread areas through an ROI image automatic clipping algorithm, and inputs the tread images into the image preprocessing module; the image preprocessing module performs anti-reflection processing and image stitching on the tread image; The anti-reflection treatment comprises the following steps: Firstly, converting a picture from an RGB color space to an HSV color space, then using a Retinex multi-scale Gaussian filter to obtain an illumination component, then using a two-dimensional Gamma function to carry out brightness correction on a V component of the HSV space of an original picture, and finally converting the HSV color space back to the RGB color space to obtain a picture after reflection removal; The image stitching is used for stitching the same wheel tread of different parts of the plurality of cameras in the array vision subsystem to obtain a complete wheel tread image; The method comprises the following specific steps: (1) Extracting image feature points by using a SIFT algorithm; (2) Performing feature matching search by adopting K-dtree and BBF algorithm, and performing primary screening according to the distance ratio of the nearest neighbor to the next nearest neighbor; (3) Estimating geometric parameter transformation between images to be spliced by using a RANSAC algorithm and splicing the images; (4) And (5) realizing image fusion by adopting a weighted average weight method so as to eliminate splicing marks.
3. The intelligent detection system for tread damage of heavy-duty train wheel set according to claim 1, wherein the UGate-MSNet model is based on a U-Net model and is divided into two parts of a decoder and an encoder, wherein an original image is firstly subjected to Patch division of 4*4 size by a PatchEmbedding module, then a GCM module initially extracts tread defect region characteristic representation, further rich tread defect characteristic representation is extracted layer by MSFM, and Patch downsampling with low information loss is performed between scale characteristics by PATCHMERGE layers; Further, a loss function is constructed for the model, and the specific steps are as follows: Based on the cross entropy loss, the formula is as follows, ; Wherein N represents the number of pixels, C represents the number of categories, Representing the probability that the i-th pixel is predicted to be of class c; Indicating whether the i-th pixel is of class c; Extending the sub-set function into a piecewise linear convex function based on the Lov sz expansion of the sub-set function, and taking the piecewise linear convex function as a convex proxy of Jaccard loss, so that the original non-convex IoU loss can be convexly optimized on a continuous domain; The formula for the lovassz expansion of the set function is as follows: ; Smooth continuation is performed through Lovasz expansion, and a constructed Lovasz loss function formula is as follows: ; Wherein father is a set function, i is the ith pixel point, For the constructed error vector, the formula is as follows: ; whether the i-th pixel is the value of the true label of class C, Predicting the ith pixel point as the c-th probability; the formula of (2) is as follows: ; Wherein pi is the descending order of m elements; The mixed loss function combining the cross entropy loss and the Lovasz loss is designed at the segmentation head, and the formula is as follows: ; wherein k is a dispensing factor.
4. The intelligent detection system for the tread damage of the heavy-duty train wheel set according to claim 3, wherein the GCM module performs layer normalization on input features first and then divides the input features into two branches, each branch uses a Linear layer to map the features to a high-dimensional space, the first high-dimensional branch uses parallel deep convolution to extract multi-scale features, the kernel sizes are respectively designed to be 7,5 and 3, feature graphs of the parallel branches are added, nonlinear activation and layer normalization of SiLU are performed once, jump connection is added for the features, then a result of the branch is obtained, meanwhile, the other high-dimensional branch is activated only by using a SiLU activation function, element-by-element multiplication operation is performed on the high-dimensional features of the two branches, then the Linear layer is used for mapping back to an original dimension, and final output is obtained after the input features are added.
5. The intelligent detection system for tread damage of heavy-duty train wheel set according to claim 3, wherein the MSFM module firstly extracts multi-scale characteristic information of input characteristics through three GCM modules, wherein the multi-scale characteristic information comprises GCM-S, GCM-M and GCM-L, the three GCM modules are provided with different convolution kernel sizes, the convolution kernel size of GCM-S is set to be [9,7,5], the convolution kernel size of GCM-M is set to be [15,13,11], the convolution kernel size of GCM-L is set to be [21,19,17], and then the three extracted scale characteristics are input to the improved SE attention SE-LIP module for characteristic reselection; The SE-LIP module calculates the space attention and the inverse space attention of the input feature respectively, the space attention calculation firstly uses the global average pooling of the channel dimension to aggregate the input feature graph into a single channel, and generates a pixel level attention diagram after 15×15 convolution and Sigmoid function After the inverse spatial attention is calculated through the same processing as the spatial attention, an inverse attention map is generated through a 1-operation And then, carrying out global normalization on the forward and reverse attention weights to obtain global importance weights thereof respectively And The expression is: ; Wherein Ω represents the global domain taken; Further, the input features are subjected to weighted average pooling by using global importance weights, and the obtained forward and reverse pooling features are respectively shown as follows: ; ; Wherein the method comprises the steps of For the forward pooling feature, Is a reverse pooling feature, and finally, the output feature is obtained by the sum of the forward pooling features, namely F out = Here, point-by-point addition is shown, and finally for the resulting spatial compression characteristic F out , it is input to a MLP with SiLU activation function to compress and excite the channel information, and then a Sigmoid function is passed to obtain the final channel attention, which is multiplied by the input characteristic to obtain the output of the final SE-LIP module.
6. The intelligent detection system for tread damage of heavy-duty train wheel set according to claim 1, wherein the construction of the DV-YOLO model and tread defect detection data set accurately detect and position defects; the DV-YOLO model takes YOLOv as a basic detection framework and comprises a trunk, a feature extraction branch and an XX-FPN module; The main part carries out feature extraction and dimension compression on the image through a downsampling module, a mixed attention feature fusion module is added between the downsampling modules, a double convolution module in the feature extraction branch dynamically generates unshared attention weights for each receptive field sliding window, the importance of key areas in the feature image is emphasized, feature information is extracted, the feature information is fused with the feature information extracted by the main part through the mixed attention feature fusion module, and finally the fused information is spliced with the branches and the XX-FPN module.
7. The intelligent detection system for the tread damage of the heavy-duty train wheel set is characterized in that the XX-FPN module comprises an up-sampling module, a feature fusion node and a C3 module, wherein the XX-FPN module encodes picture features, receives multidimensional features which are output by the mixed attention feature fusion module in a trunk part and comprise a lowest-dimensional feature, a middle-dimensional feature and a highest-dimensional feature, and the obtained features are mutually spliced in multiple paths and multiple layers through up-sampling the high-dimensional features and down-sampling the low-dimensional features to obtain the multi-scale features.
8. An intelligent detection device for tread damage of a heavy-duty train wheel set, which is used for realizing the detection system according to any one of claims 1-7, The device comprises an industrial personal computer and a wheel set tread defect detection device; the wheel set tread defect detection device acquires wheel set image data and inputs the data into the industrial personal computer; The wheel set tread defect detection device comprises 4 groups of array vision subsystems which are symmetrically arranged on two sides of a rail, wherein the array vision subsystems comprise a shooting device and a control device; the shooting device comprises a camera (1), a compensation light source (2) and a triggering device (3); The two groups of array vision subsystems which are positioned on the same side of the rail and are respectively used for shooting the front wheels and the rear wheels on the same bogie of the train are symmetrically arranged, wherein the distance between the first group of array vision subsystems and the second group of array vision subsystems on the same side is determined by the length and the pitching angle of the bogie; the array vision subsystems comprise a plurality of groups of shooting devices which are uniformly arranged at intervals along the rail direction, and the shooting devices in the first group of array vision subsystems are sequentially provided with a triggering device (3), a camera (1) and a compensation light source (2) from the near end to the far end relative to the second group of array vision subsystems; the triggering device (3), the camera (1) and the compensating light source (2) in the shooting device are respectively connected with the control device; The array vision subsystem comprises a control device, the shooting device is connected with an industrial personal computer, and the triggering device (3) adopts a PNP three-wire proximity sensor.

Description

Intelligent detection system and device for tread damage of heavy-duty train wheel set Technical Field The invention relates to the technical field of train detection, in particular to an intelligent detection system and device for tread damage of a wheel set of a heavy-duty train. Background Wheels are used as key components of the railway vehicle, and the health condition of the wheels is directly related to the safety and stability of train operation. Under the complex interaction of wheel and rail, the tread of the wheel is easy to generate various damages such as flat, peeling, pit abrasion, polygonal abrasion and the like, and the defects can cause severe vibration and noise and even threaten the driving safety. In order to cope with the challenge, the wheel detection technology is gradually developed from early manual static measurement to a dynamic monitoring mode based on a sensor, and aims to realize real-time and accurate perception of wheel damage, so that technical support is provided for safe operation and intelligent operation and maintenance of a train. The problems still exist in the prior art that dynamic and on-line detection cannot be realized, part of ultrasonic flaw detection can be performed only by disassembling wheels after a thunder car is stopped or is greatly slowed down, normal transportation order is seriously disturbed, detection efficiency is extremely low, daily preventive maintenance cannot be realized, image imaging is difficult, illumination is received, vibration interference is large, defect dimensions are extremely tiny, apparent characteristics are weak, recognition difficulty is caused by physical characteristics of defects per se due to high form diversity, and accuracy is low. Disclosure of Invention The invention aims to solve the technical problem of providing an intelligent detection system and device for tread damage of a wheel set of a heavy-duty train, which have high precision and high efficiency. In order to solve the technical problems, the invention adopts the following technical scheme: the intelligent detection system for the tread damage of the heavy-duty train wheel set is characterized by comprising an image acquisition module, an image preprocessing module, a tread defect detection module and an intelligent terminal display interface; the image acquisition module acquires wheel set images, transmits the wheel set images to the image preprocessing module, the image preprocessing module inputs the preprocessed images into the tread defect detection module to obtain defect information, and the intelligent terminal display interface acquires the preprocessed tread images and the preprocessed defect information and performs visual display Constructing UGate-MSNet models and tread defect region data sets in the tread defect detection modules to divide tread defect regions; And further constructing a DV-YOLO model and a tread defect detection data set to accurately detect and position the defects. Preferably, the image acquisition module shoots the tread image through a camera fixed point, acquires the tread image of a train wheel set with consistent image position and pixel size of the train, pre-extracts the tread region through an ROI image automatic clipping algorithm, and inputs the tread image into the image preprocessing module; the image preprocessing module performs anti-reflection processing and image stitching on the tread image; The anti-reflection treatment comprises the following steps: Firstly, converting a picture from an RGB color space to an HSV color space, then using a Retinex multi-scale Gaussian filter to obtain an illumination component, then using a two-dimensional Gamma function to carry out brightness correction on a V component of the HSV space of an original picture, and finally converting the HSV color space back to the RGB color space to obtain a picture after reflection removal; The image stitching is used for stitching the same wheel tread of different parts of the plurality of cameras in the array vision subsystem to obtain a complete wheel tread image; The method comprises the following specific steps: (1) Extracting image feature points by using a SIFT algorithm; (2) Performing feature matching search by adopting K-dtree and BBF algorithm, and performing primary screening according to the distance ratio of the nearest neighbor to the next nearest neighbor; (3) Estimating geometric parameter transformation between images to be spliced by using a RANSAC algorithm and splicing the images; (4) And (5) realizing image fusion by adopting a weighted average weight method so as to eliminate splicing marks. The method comprises the steps of dividing an original image into two parts, namely a decoder and an encoder, by using a PatchEmbedding module, firstly carrying out Patch division of 4*4 size on the original image, then preliminarily extracting tread defect region characteristic representation by using a GCM module, further ext