Research On Rail Surface Defect Detection System Based On Machine Vision

With the rapid development of China’s rail transit construction,the mileage of rail lines has steadily ranked first in the world.With the long-term service of rail lines,a variety of rail defects and damage have gradually increased,making the maintenance workload of rail components sharply increased.As the core of the rail components,the rail head is often slightly defective due to its direct contact with the wheel,which affects the driving stability.If it cannot be detected in time,and it may develop into severe defects,endangering the driving safety.Therefore,in recent years,rail defect detection technology has become the key research direction of the industry,and the research on intelligent detection technology and algorithm of rail defects has been fully developed.In order to realize intelligent identification of rail surface defects and provide efficient track inspection and maintenance means,high quality automatic detection methods and technical systems are needed.Based on machine vision technology,the intelligent detection system of rail surface defects is studied in this paper.The system includes two parts: the hardware system of image acquisition and the algorithm of rail surface defect recognition.To identify defects on the surface of rail,the acquisition of a stable and clear track image is necessary.This paper designs a modular track image acquisition system based on linear camera and laser light source.The system integrates the light source and the camera,simplifies the module composition,and only needs to provide the power supply and data interface to realize the orbit image acquisition function.In terms of algorithm,a rail edge prediction model is proposed based on onedimensional convolutional neural network,which is used to extract the rail region from the original track image and simplify the subsequent defect identification steps.Compared with the traditional rail region extraction algorithm,the proposed method has a higher recognition accuracy for a variety of complex scene track images.In the test set,the average absolute error between the predicted boundary and the real boundary is only0.01,and the processing speed of the algorithm can reach 34 fps on the CPU alone.Secondly,based on this algorithm,constructs the rail surface defect dataset in which only contain rail surface region.The joint,scaling,burn and crack which appear on the rail surface has carried on the annotation.Based on statistical analysis,the data sets the label of the scale and characteristics of all kinds of defects are analyzed,and to provide data support for the research of defect recognition algorithm.Then,for the sake of accurate localization and classification of rail surface defects,based on the rail surface defect dataset,benchmark experiments are carried out on 8 classical deep learning models of object detection.Based on the characteristics of the rail defect images,we put forward the texture feature pyramid augmentation method and weighted fusion pooling method to optimize the classical deep learning models of object detection,it is used to enhance the learning ability of the feature extractor to the texture features in the rail image.After using the optimized method,the detection performance of the benchmark model has been improved,and the average detection accuracy of the four types of defects has reached85.1%.Finally,based on the digital image processing technology,the Weber differential excitation method was improved to achieve the segmentation of rail surface defect texture,and the obtained defect location and classification information was used to complete the pixel level classification and recognition of rail surface defects.