Multiple Source Domains Knowledge Transfer For Intelligent Fault Diagnosis Of Rotating Machinery

Rotating machinery is the most common transmission unit in mechanical equipment,whose operational reliability significantly influent the production efficiency and the safety of life and property.Recently,deep learning based intelligent fault diagnosis technique becomes a promising tool for ensuring the reliability and security of rotating machinery by its distinguished feature extraction and nonlinear fitting capability.However,the effectiveness of deep learning based fault diagnosis methods heavily rely on the quantity and quality of labeled training samples,which are required to be independent and identically distributed with future testing samples.In practical application,the occurrence and discovery of faults are scarce and random,which can hardly meet the requirement of deep learning based fault diagnosis methods.Thus,aiming at the scarcity of labeled training samples in deep learning based fault diagnosis methods,based on deep transfer learning,this dissertation studies the multiple source domains knowledge transfer for intelligent fault diagnosis of rotating machinery.Four kinds of fault diagnosis knowledge transfer scenarios are constructed from two perspetives:the availability of target domain data,and the relations of label space.The main research contens are as follows:1)Based on dynamic distribution adaptation,Dynamic Distribution Adaptation based Transfer Network(DDATN)is proposed to improve the performance of fault diagnosis model in unsupervised knowledge transfer fault diagnosis scenary.Cosidering the impacts of marginal distribution,conditional distribution and the dynamic fluctiuation of their contributions on inter-domain distribution discrepancy,a novel dynamic distribution discrepancy metric termed Instance-weighted Dynamic Maximum Mean Discrepancy(IDMMD)is proposed.DDATN utilizes IDMMD and convolutional neural network(CNN)for unsuperivised cross-domain fault diangosis.The bearing fault diagnosis experiment demonstrated the advantages of dynamic distribution adaptation over marginal distribution adaptation and joint distribution adaptation,and the superiority of IDMMD and DDATN.2)Aiming at the problem of semi-supervised domain(working condition)generalization fault diagnosis,which target domains datasets are unavailable and just one labeled and one unlabeled source domains datasets are available,Deep Semi-supervised Domain Generalization Network(DSDGN)is proposed.DSDGN utilized WGAN-GP based adversarial training to train feature extractor that can extract domain-invariant fault features,simultanesly,pseudo-label learning is utilized to improve the generalization performance of fault diagnosis model.The rotating machinery fault diagnosis experiments demonstrated that DSDGN can effectively generalize fault diagnosis model to fault diagnosis task under unseen working conditions.3)In practical application,it is hard to collcted complete fault dataset under one working condition,and the target domain label space is generally unknown in unsupervised cross-domain fault diagnosis,which can lead to heterogeneous label space.To solve above problem,Collaborative Adaptation Network(Co AN)is proposed,which introduces multiple source domains universal knowledge transfer fault diagnosis framework.Co AN utilizes uncertainty and similarity metrics to evaluate the probalibity that samples belong to shared classes and assign training weights accordingly.By assign larger training weights to shared classes samples,CoAN can eliminate the effect of private classes samples and learn general fault diagnosis knowledge from multiple source domains whose label spaces are heterogeneous.In addition,the classifier adversarial strategy is utilized for unknown fault recognition.The bearing fault diagnosis experiments verified the effectiveness of Co AN in multiple source domains universal knowledge transfer fault diagnosis task.4)Heterogeneous label space phenomenon is common in rotating machinery fault diagnosis,aiming at the problem of domain(working condition)generalization fault diagnosis which has multiple heterogeneous label space source domains,Joint Generalization Fault Diagnosis Network(JGFDN)is proposed.JGFDN combines maximum mean discrepancy(MMD)based conditional distribution adaptation and episodic training to learn general fault diagnosis knowledge from multiple heterogeneous label space source domains and generalizes it to fault diagnosis task in unseen working conditions.Bearing fault diangnosis experiments demonstrated the effectiveness and superiority of JGFDN.