Fault Diagnosis Of Rotating Machinery Based On Domain Adaptation And Domain Generalization

With the development of China’s industry and the use of demand,rotating machinery(such as rolling bearings,gearboxes,etc.)as mechanical equipment in the use of the process of the most prone to damage devices,its safety and reliability are of great concern,and at the same time,with the increase in the degree of precision of machinery Rotating machinery in complex working conditions,the occurrence of failure does not have deterministic conditions,the suddenness is high,fault data is generally difficult to obtain.The cross-domain learning technology based on Domain Adaptation and Domain Generalization theory can ideally solve the pain points such as lack of data volume and complex working conditions,and use the existing small amount of data for effective fault diagnosis of rotating machinery to the maximum extent and with the highest efficiency,and bring out the maximum effect of the model.Therefore,domain adaptive and domain generalization theories are introduced to study the fault diagnosis method of rotating machinery with variable working conditions in complex operation environment,aiming to improve the fault diagnosis level of rotating machinery across working conditions and solve the problems of lack of labeled data,difficulty of fault feature extraction and poor generalization ability of the model,the main work accomplished is as follows:In response to issues such as lack of labeled data and difficulty in generalization of variable condition models,an improved adversarial domain adaptive model is proposed,which not only models the category labels and domain labels of fault data,but also focuses on data structure information to measure inter domain differences.The model models category labels and domain labels through classifiers and domain discriminators,and embeds data structure information into instance graph nodes through graph neural networks.Gaussian Wasserstein distance is used to measure the differences between instance graphs in different domains.Verify the effectiveness of the model by conducting experimental analysis of fault diagnosis of rotating machinery under multiple operating conditions.Furthermore,in response to the issues of invisible target domain data,potential failure of traditional domain adaptive networks,lack of task related datasets,and ineffective task independent datasets in traditional domain generalization networks,a confusion relationship transfer framework is proposed.Based on this,a fault diagnosis method that is compatible with domain adaptation and domain generalization is constructed.The model is divided into two stages: irrelevant domain adaptation and related domain generalization.In the irrelevant domain adaptation stage,an adversarial domain adaptive network with multiple domain discriminators is constructed to enhance domain confusion and ensure the effectiveness of the confusion relationship transfer framework,resulting in a universal feature extractor.At the same time,design and improve the maximum mean difference loss to enable the feature extractor to have more general domain confusion ability.In the phase of correlation domain generalization,the domain generalization diagnosis network is constructed,and the residual fusion classifier is used to prevent the fault classifier from overfitting and obtain higher diagnostic accuracy.Finally,experiments based on gearbox and bearing datasets verified the effectiveness and superiority of the proposed model.The research content expands the application scope of transfer learning,and provides a feasible method and valuable reference for rotating machinery fault diagnosis.