Research On Fastener Status Detection Using Deep Learning

Railway fastener is an important component for ensuring the safe of trains.In the recent years,the maintenance of track infrastructure including fasteners has been put forward higher requirements since the high speed and heavy load of railway transportation.With the development of computer vision technology,many companies have proposed their own automatic fastener detection methods which are based on machine vision.Compared with the traditional manual inspection method,these automatic detection methods are faster and more efficient.However,these methods are still far from perfect.As an important part of artificial intelligence,deep learning technology has made enormous achievements in various fields.We want to build an intelligent,efficient and more general inspection system for disease fastener by deep learning technology.There is no public and available dataset in fastener detection field.In this paper,we present a method for fastener location based on semi-automatic labeling,and propose a semi-automatic labeling method,which can assist employee to complete labeling work of fastener data effectively.Here we collected and produced a dataset of 4,000 fastener data,including 4 categories,and completed the labeling work completely manually.This dataset could be used in the fastener location algorithm for learning and performance evaluation.We aim to detect the fasteners with diseases,in three kinds of situations,i.e.the breakage of elastic strips,deformation and displacement.Referring to Faster R-CNN method,we propose a fastener status detection method based on deep learning with multi-scale fusion features.In our method,weighted loss is also applied to improve the small sample's detection effect of disease fastener.In addition,the method costs less time in detection by optimizing the Anchor generation strategy.Because the number of track images containing disease fasteners is too small,we synthesized a batch of disease fastener data by simulating real samples.And on this dataset,we trained and tested the proposed method.In order to verify our method's generalization performance,we take the data of ballastless track from Hengyang to Changsha as the test set.The experimental results show that our proposed method is effective.