Research On Surface Defect Detection Method Of Train Bearings Based On Deep Neural Network Under Small Sample Conditions

Bearing is one of the core parts of train running gear,and its health status directly affects the safety level of train running.However,the poor working environment is easy to cause the train bearing scratch,peel,dent and other faults.At present,the maintenance unit will return to the factory to dismantle the train bearings for maintenance according to the train mileage and time.Maintenance personnel mainly detect surface defects of bearing components through artificial vision and experience.The machine vision detection method based on deep learning has the advantages of fast speed,low cost and intelligence.However,bearing faults are characterized by strong randomness and low probability,which leads to some difficulties in accurately detecting train bearing faults under the condition of small sample images.Based on this,this thesis builds a deep neural network model to carry out in-depth research on train bearing fault detection methods under the condition of small sample images.The main research contents are as follows:(1)The basic structure of train bearings is explained,the common failure forms of train bearings are introduced and their occurrence principles are analyzed;This thesis describes the bearing repair process in actual production,points out the problems of time consuming,low efficiency and strong dependence on the wisdom of experts and workers’ experience.The deep neural network is used to detect the surface defects of faulty bearings,which has the application prospect of adapting to industrial production.The convolutional neural network and image recognition theory are studied to lay the technical foundation for the subsequent construction of image classification model.(2)On the basis of theoretical research,collecting fault bearing images and conducting identification experiments on them,statistical analysis of the type and number of fault images,select the most common scratch,peel and dent three types of faults for research;By preprocessing the original image and enhancing the data,the small sample image data set of train bearing fault was constructed.The traditional convolutional neural network combined with twin network framework was used to construct an image recognition framework to realize image recognition and detection of faulty bearings.(3)Traditional twin networks have the problem of limited feature extraction ability,so a Multi-resolution Siamese neural network(MrSNN)model based on feature extraction enhancement is proposed.A Multi-resolution convolution fusion block(MrCFB),which combines two forms of general convolution and expansive convolution of different sizes,is constructed to extract detailed features and contour features of fault images.Double attention mechanism is introduced to adjust feature weights from different dimensions.Finally,the proposed method is verified by the above data set,which confirms the effectiveness of the proposed method.(4)The twin network model realizes image recognition and detection,but in practical application,the process of sample construction and analysis of detection results is complicated.The prototype network framework can improve the classification efficiency of train fault bearing images.By using Mahalanobis distance as the measurement function,the experimental results show that this method can extract image features better than the traditional feature extraction method.