Fault Diagnosis For Rolling Bearing Based On Deep Transfer Learning

In the evolution of Industry 4.0,the unmanned and intelligent technologies are crucial for the future industrial process.To be the key component of industrial production,the automatic detecting of intelligent fault diagnosis for various equipment plays an important role in industrial production process.Traditional fault diagnosis is based on statistical characteristics of vibration signals,combined with expert knowledge,to identify the failures.It is difficult to apply on large-scale cases,though the preferable performance could be acquired in some specific scenarios.Deep learning is capable of learning large amounts of data.However,enough labeled samples is required to complete model training.Therefore,transfer learning is introduced into deep neural network to study intelligent fault diagnosis methods.The learning of source domain related to target task can facilitate the target classification task.The main contributions are summarized as follows.First,the research background and significance of rolling bearing fault diagnosis are analyzed.Then,the state of the art intelligent fault diagnosis methods are reviewed.Due to the high cost of human annotation and the large distribution discrepancy,deep transfer learning is used for fault diagnosis of rolling bearings.Based on this,the concepts and definition about transfer learning,and Convolutional Neural Network(CNN)are described as well.Second,one of the main contributions is to study the marginal and conditional distribution adaptation and weighting of the previous distribution adaptation.A deep transfer learning method,named as Domain Adaptive based on Feature Distribution Matching(DAFDM),is proposed,where the minimizing of marginal and conditional distribution discrepancy is done on the multilayer feature extracted by CNN for extracting the domain-unbiased feature.Experiments on CWRU and BV datasets show that the proposed method can effectively improve the fault diagnosis performance.Third,to further extract the unbiased feature,self-supervised learning is used to extract the substantial feature for both source and target domains.Then,a new method of Domain Adaptive based on Self-Supervised Learning(DASSL)is proposed.After two-domain data is transformed by different time-frequencies methods,the selfsupervised learning network is trained on the data by various time-frequency transformation methods to recognize the corresponding transformation methods,where the unbiased feature is learned.After that,clustering is done on the feature of target domain by exploiting the distribution structure of target domain.The pseudo-labels are updated according to the "strong clustering" rule.The confidence of pseudo-labels is further designed to improve the usability of labels.Experiments on CWRU and BV datasets demonstrate the superiority of the proposed method.Finally,the main contributions and innovations of the thesis are summarized.And the unresolved problem and future work are described.The paper contains 24 figures,10 tables,and 82 references.