Research On Damaged Tread Detection Of Train Wheel Set Under Machine Vision

The damaged tread of train wheel set will impact upon the rail running safety.Since the current manual damaged tread detection of train wheel is low efficiency,high labor intensity,it is very impotant for damaged tread detection of train wheel to achive automation and intelligentialize.In the thesis,the tread damage detection of train wheel set was studied.By using digital image processing and optoelectronic measurement,a system of tread damage detection was developed under machine vision for dynamic on-line detection of train wheel tread.For a system of tread damage detection,there are three aspects were studied on preprocessing of the wheel image,damage detection of the wheel tread and judgment of damage degree.The main work includes as follows:1)Preprocessing of the wheel image.It includes train wheel tread extraction and enhancement of train wheel tread image.Aiming at the problem of train wheel tread extraction,this paper presents a wheel rim fitting method based on Random Sample Consensus to realize the extraction of train wheel tread based on the characteristics of train wheel tread damage detection system fixed-point imaging in machine vision system.By using the gradient information of train wheel rims,the wheel rim points are obtained,and RANSAC is used to fit the wheel rims of the train to achieve the train wheel tread extraction.In order to solve the problem of nonuniform illumination in the image,the homomorphic filtering technique based on illumination-reflection model and the Retinex enhancement technique based on multi-scale are studied.By using the multi-scale Retinex enhancement technology,a significant processing effect was achieved.2)Damage detection of the wheel tread.This paper proposes a train wheel tread damage detection method based on texture clustering and region growing.By using the Gray Level Co-occurrence Matrix,Variance,Contrast and Homogeneity is used to describe the tread area texture characteristics as feature vector.K-Means++ clustering algorithm is used to identify the suspicious regions with damage.Damage profile points are detected by using the damage region contour gradient information.Select the intersection point of the contour point as the seed point,and the region growing technique based on the 3 sigma principle is proposed to extract the damage region.Based on morphological methods of directional structural elements,the similar damage region of tread is merged to realize train wheel tread damage detection.3)Judgment of damage degree.Aiming at the problem of tamper detection in tread damage detection,By contrasting the generating gray scale,generating interval and generating direction of the gray level co-occurrence matrix to the contrastive experiment which describes the damage characteristic,the appropriate gray level co-occurrence matrix parameters are selected to generate contrast and correlation and homogeneous as the texture feature vector.Back-propagation(BP)neural network and Support Vector Machine(SVM)are introduced to judge the tread damage.The experimental results show that the generalization of damage judgment of SVM is better than that of BP neural network when the number of training samples is small.The damage detection system based on the above-mentioned technology has been tested online at Transportation Technical Center Incorporation(TTCI),and the accuracy of damage judgment has been reached 94.375%,which lays a good foundation for the production of on-line inspection system of train wheel tread.