Study On Abnormal Detection Method Of Straddle Monorail Track Beam Based On Convolutional Neural Network

Straddle monorail is being used widely because of its flexibility.The track beam of straddle monorail possesses the actions of loading bearing and presenting,so abnormality on it is a serious threat to the safety of train.The current manual inspection methods have low efficient and accuracy.So it’s very urgent to develop an automatic method detecting the abnormality on track beam.To address this problem,based on images collected on the spot,this paper uses convolutional neural network to automatically detect abnormality on monorail track beam.The abnormality can be divided into cracks on beam surface and abnormal fasteners in finger plate area.We researched two different types of abnormality separately to reach better detection performance.Based on previous methods and the collected images of track beam,convolutional neural network was used to classify image patches for crack detection.Splitting into patches makes detecting cracks on original whole image very difficult,so we used binary classification of crack and no-crack for image patches.The final results of crack detection were obtained by combining patches which was classified to crack.Considering efficiency,MobileNet was chosen to be fine-tuned in our self-made monorail track beam dataset.Crack images on similar scenes were added to solve the problem of lack of crack data.We finally reached an accuracy of 99.085% on test set.To improve the efficient,pruning technique was used to prune the convolutional kernels.As a huge amount of redundant kernels have been removed,we improve the speed of model by 34.73% while the accuracy remains unchanged.For abnormal fastens detection on finger plate,we found no previous work can be used in monorail,so a method based on cascade convolutional neural network was presented in this paper to quickly detect abnormal fastens on finger plate.First,an objective-detection network locates all the fastens in the image.Then,we use proposed fastens key point regression network to predict the location of key point and get the features’ information.After comparing the key point’s information of same fasten at different time,we finally implement abnormality detection on fastens.The proposed method has been evaluated on finger-plate fastens dataset collected in Chongqing Rail Transit Line No.3.We reached an accuracy of 96.5% and a recall of 99%.The result shows that proposed method can be used in practical application.