Research On Unsupervised Fault Diagnosis Method Of Rolling Bearing Under Varying Loads Based On Deep Transfer Metric Learning

Rolling bearings are the key parts of rotating machinery,and are widely used in industry and agriculture,national defense and other fields.Rolling bearings often work under severe conditions and are prone to failure.Once a bearing fails,it will probably lead to industrial shutdown and system collapse,resulting in huge economic losses,and even casualties.Therefore,how to accurately diagnose the bearing failure is of great significance.In actual production,the bearing load and rotating speed often change according to the needs,and the bearing faults under other loads cannot be accurately diagnosed by using a single load training model.However,it is often difficult to obtain labeled target domain data when training a varying loads fault diagnosis model according to various load data.At the same time,the existing methods fail to fully consider the optimizations of the inter-class and intra-class distances,resulting in low diagnosis accuracy.Therefore,an end-to-end unsupervised fault diagnosis method for rolling bearings under varying loads is proposed based on deep transfer metric learning.Aiming at the problem of poor performance of deep learning fault diagnosis model caused by less sample data,the sliding window is used to process the original vibration signal for data enhancement,and then in order to eliminate the dimensional influence,the data after data enhancement is normalized,so as to obtain the final data set for training and testing.In order to solve the problems of shallow artificial feature extraction,such as relying on expert experience,lack of single-domain feature information and redundancy of multi-domain feature information,a one-dimensional inception convolutional neural network with wide kernel at the first layer is proposed to extract the preprocessed data with adaptive deep features,which has advantages over shallow features without relying on expert experience.Because it is difficult for deep learning fault diagnosis model to diagnose faults under varying loads,an unsupervised deep transfer learning method based on improved deep adaptation network is proposed to solve the problem of rolling bearing fault diagnosis under varying loads when label data is scarce.At the same time,a deep metric learning method is proposed to solve the problem of low fault diagnosis accuracy caused by large intra-class dispersion and inter-class coupling.Center loss and Softmax cross entropy loss are added to the top layer of the network,and classification and discrimination performance are optimized to realize high-accuracy rolling bearing fault diagnosis under varying loads.Finally,the experimental results show that the proposed method can optimize the inter-class and intra-class distances,reduce the distribution difference between source domain and target domain,and achieve higher diagnostic accuracy under varying loads,and its performance is superior to rolling bearing fault diagnosis methods based on deep automatic encoder and multi-core semi-supervised transfer component analysis.