Research On Imbalanced Sample Learning For Fastener Defect Detection

In order to ensure the safe operation of the railway,it is urgent to use machine learning,image processing and other technologies to carry out reliable,real-time and efficient inspection of track infrastructures such as the rail,fastener,etc.And it is also an important part of intelligent high-speed rail in the future.The incidence of fastener defect is very low on railways that are in service,and it takes a long time to collect track images containing defective fasteners.This makes the number of normal and abnormal fasteners(defective fasteners are also called abnormal fasteners)seriously imbalanced,which could seriously affect the performance of machine learning algorithm.Therefore,this paper research on the imbalanced sample learning for fastener defect detection.Based on the deep object detection model Faster R-CNN,this paper improves the model from the two perspectives of improving the classification and representation ability of the model.And it incorporates prior information such as geometry structure of tracks into the model,in order to achieve high precision and recall rate for the detection of defective fasteners at the same time.The main work and innovations of the paper are as follows:(1)A method of fastener defect detection based on data synthesis is proposed.This method mainly improves the classification ability of the model from the two aspects of data and loss function.At the data level,two methods of data synthesis,FIT and FIB,are proposed to expand the number of abnormal fasteners in training data.At the loss function level,the weighted center loss function can solve the problems of imbalanced sample,small inter-class difference and large intra-class difference at the same time.The method also uses a lightweight backbone network to improve the detection speed of the model.And a threshold pruning algorithm is designed to filter the background area misidentified as a fastener to improve the precision of the model.The simplified Ro I head can reduce the complexity of the model.Finally,the overall performance of the model and the influence of each part on the performance of the model are verified by comparison and ablation experiments respectively.(2)A method of fastener defect detection based on multi-attention mechanism is presented.This method mainly designs the feature attention mechanism(FAT)and the position attention mechanism(PAT)to improve the model representation ability.Because RPNs and Ro I heads share features but have different tasks,feature attention mechanisms are designed to enable the two modules to automatically learn to take different attention to shared features.Based on the analysis of the experimental results of object detection,this paper suggests that the features of different positions have different importance in identifying the categories of objects in the area.Then,the position attention mechanism is designed to enable the model to automatically learn to take different attention to the features of different locations.A multivariate Io U filtering algorithm is proposed to solve the problem that the model misidentifies the background area as a fastener.The algorithm uses X-Io U,Y-Io U,H-Io U,W-Io U to depict the geometric relationship between the bounding boxes predicted by the model,and then filters out the misidentified background areas based on the prior information of transverse parallel,vertical alignment and uniform size of the fasteners.Finally,the overall performance of the model and the influence of each part on the performance of the model are verified by comparison and ablation experiments.