Cross Fault Diagnosis Method Of Rotating Machinery Based On Unsupervised Domain Adaptation

In the process industries such as petrochemical,power,and metallurgy,rotating machinery is a widely used critical power equipment,and condition monitoring and fault diagnosis technology are the guarantee for safe,stable,efficient,and high-quality operation of this equipment.However,the information processing approaches must develop towards intelligent decision-making technology for industrial big data.Therefore,aiming at the poor application performance of traditional fault identification methods established under the assumption that the fault data have to satisfy independence and identical distribution,how to extract sensitive knowledge that effectively characterizes the operating status of mechanical equipment from the generating massive data under variable working conditions and construct an end-to-end intelligent cross-domain fault diagnosis theoretical model have become the key of developing advanced intelligent decision-making technology for rotating machinery.Consequently,this study takes rotating machinery as the research object,based on domain adaptation theory,utilizes convolutional neural networks as the technical means,and attempts to explore the unsupervised domain adaption method for addressing the cross-domain fault identification issues from four aspects,i.e.,domain distance measurement,adversarial domain,subdomain adaptation,and multi-source subdomain.The research progress and conclusions are drawn as follows:(1)Aiming at the issue that domain adaptation methods based on statistical quantities or distance measurements only consider the marginal probability distribution of data without taking conditional probability into consideration,which will result in insufficient adaptation of domain conditional probability,an unsupervised domain adaption cross-domain fault identification method named Joint Sliced Wasserstein Distance(JSWD)is proposed.By introducing Sliced Wasserstein Distance into the framework,the speed of network operation is accelerated.Furthermore,a strategy marked pseudo labels for target domain instances with cosine distance is designed.In an addition,the JSWD loss function is defined,and the JSWD model is established.Finally,the effectiveness of the proposed method is verified on four datasets,and the results suggest that the proposed pseudo labeling strategy for target domain samples through similarity of output probabilities from different training epochs can effectively improve the fault identification rate of the built model.In addition,noise sensitivity experiments demonstrate that the proposed method has better robustness and generalization performance.(2)Aiming at the problem that adversarial domain methods utilizing a single domain discriminator has limited performance in reducing domain discrepancy and do not consider whether features are transferable or not,an unsupervised domain adaption cross-domain fault identification approach named Joint Attention Adversarial Domain Adaptation(JAADA)is proposed.Firstly,the outputting features are divided into multiple feature blocks.In each feature block,both distance measurement and adversarial domain adaptation approaches are utilized to align the source domain and target domain feature distribution.Secondly,an attention mechanism is introduced,and a feature block weighting strategy about attention mechanisms is proposed.Additionally,the loss function is designed and unsupervised local,global,and joint attention adversarial domain adaptation models are established.The effectiveness and feasibility of this model are verified on four datasets.The experimental findings suggest that the proposed method c ould effectively identify transferable features and non-transferable features,enhance the domain adaptation matching ability of transferable features,and weaken the negative transfer caused by forced transfer of non-transferable features.(3)For the problem of existing researches have not considered data structure or domain shift between the same categories,resulting in insufficient adaptation of subdomain distribution discrepancy(i.e.,subdomain adaptation problem).Hence,an unsupervised structure subdomain adaption cross-domain fault identification method named Contrastive Cluster Center(CCC)is proposed to handle the forementioned issue.Instance structure features and pseudo labels for target domain samples are obtained by utilizing graph convolutional neural networks and softmax function,respectively.Additionally,a Dropout trick is designed to eliminate and refine d pseudo labels with incorrect annotations to ensure high confidence in target domain instance pseudo labels.Furthermore,the rationality of refining higher confidence sample pseudo labels is proved with domain adaptation theory.Moreover,the CCC loss function is defined,and the unsupervised structure subdomain adaption model is established.Finally,experiment on two datasets have conducted to validate the effectiveness of the proposed method,whose results reveal that the proposed method could effectively mine latent data structure attributes,establish a bridge between different categories of features in two domains,and effectively reduce domain shifts between different categories in two domains.(4)Some researches have only considered a single source domain without considering multi-source domain adaption adaptation problems.In view of this,the single source domain is extended to multiple source domains.Aiming at the e xisting multi-source domain adaptation research only adapts the distribution of each source domain to the target domain,but does not adapt the distribution of features between different source domains,and does not consider the distribution di screpancy between the same categories in different domains in multi-source domain adaptation,which will lead to insufficient adaptation issue in multi-source sub-domains.To address the forementioned matters,an unsupervised Multi-source Sub-domain adaption Contrastive Cluster Center(MS3C)cross-domain fault identification method is proposed.By embedding the CCC method in a single source domain into multiple source domains,a novel loss function is designed and a MS3 C model is established.To verify the effectiveness and superiority of the proposed method,two bearing datasets are selected.Simultaneously,the experimental findings demonstrate that the proposed method could bridge sub-domain shifts between the same categories in different domains,and is potential to help boost the accuracy and generalization ability of unsupervised multi-source sub-domain fault identification under complex and variable operating conditions.