Research On Intelligent Fault Diagnosis Based On Deep Transfer Learning For Bearings With Few Samples

Rolling bearings are key components in most of rotating machinery.When a defect exists in the bearing,its failure may cause the failure of the whole equipment,even shutdown and accidents.Therefore,it is necessary to carry out the prognostics and health management for these bearings.In recent years,deep learning greatly improves the learning ability and intelligent diagnosis level of industrial big data,but it still needs to solve the problems of diagnosis performance descending caused by few fault data,varying working conditions,and strong noise.First,most of deep neural networks relies on a large amount of complete data,and the scarcity of fault samples makes deep learning inadequate,which leads to low training accuracy and reduced diagnosis accuracy.Second,the generalization performance of the deep network is affected by data variation,varying working condition,noise intensity and other factors,which limits the engineering applications of the intelligent diagnosis model.To solve these problems,the research investigates the small-sample learning and transfer learning based on deep networks for intelligent fault diagnosis.The main contents and innovations in this research are summarized as follows:（1）The sensor data sets usually contain a large amount of data without labels and a small amount of data with labels.The calibration of data without labels is time-consuming and laborious,even impossible.To solve this problem,a semi-supervised fault diagnosis method based on generative adversarial network（GAN）is proposed for bearings with few labeled data.In this method,the adversarial learning and semi-supervised learning work together to enhance the recognition ability of the deep convolutional network by using few labeled data.Furthermore,an enhanced feature matching algorithm is proposed to improve the convergence speed of the network training by using depth features of middle layers.Experimental results indicate that the proposed method can effectively improve the accuracy of fault diagnosis when few labelled samples are available.（2）For the bearing in service,in its whole life time,most of its data are healthy and only few data are collected when it is faulty.In this study,an improve Siamese network（SN）with multi-scale convolution is proposed to solve this problem.First,the multi-scale convolutional neural network（CNN）is constructed to enhance its learning ability and avoid the effects of over-fitting and high computation caused by the increase of the network layer.Secondly,the SN is introduced to discriminate the distribution difference between healthy data and few faulty data,and then the original Euclidean distance is replaced by the maximum mean discrepancy to measure the similarity between samples,so that the trained network can distinguish different states of data.In the experiments,different numbers of faulty data are used for training.The results show that the proposed network can obtain high diagnosis accuracy when few faulty data can be obtained and solve the problem of insufficient learning for deep networks.（3）Due to bearing replacement,varying working condition and other factors,the distributions of training and test data sets are greatly different,and the generalization performance of the network obtained by big data learning and complex training process is poor.To solve this problem,a transfer learning method based on improved domain conditioned adaptation network（DCAN）is proposed in this study.First,using the DCAN,the attention mechanism and feature warping are introduced to build the domain condition channel module and improved feature wrapping module,the former of which is used to solve the difficulty of setting network weights caused by the distribution difference between the source and target domains and complete the individual study of feature channels,the latter of which uses nonlinear transformation to replace the original linear transformation and improve the fitting degree between source and target domains.As a result,the transfer learning from source domain to target domain is realized.The experimental verification and method comparison show that the proposed network can realize cross-working condition and cross-bearing transfer learning,and the diagnosis knowledge obtained from experimental bearings can accurately identify the faults of other experimental bearings or actual wheelset bearings.In this research,vibration data sets from experimental bearings and wheel bearings are used for transfer learning and fault diagnosis.The results indicate that the proposed methods have higher diagnosis accuracy and better generalization ability.Meanwhile,some strategies are adopted to reduce the computation cost and make these methods more suitable for applications.Therefore,this research not only improves the structure design and classification performance of deep transfer learning,but also provides effective methods and valuable reference for intelligent fault diagnosis of mechanical equipment.