Mechanical system cross-domain diagnostics based on signal fusion and cycle consistency domain adaptation network

In actual industrial situations, the operating conditions of mechanical systems usually change over time and the inconsistency of data distributions leads to poor generalization ability of intelligent diagnosis models. Therefore, this research proposes an intelligent cross-domain diagnostic method based on signal fusion and cycle consistency domain adaptation network (SF-CCDAN) for mechanical systems under time-varying conditions. Firstly, a multi-source signal fusion strategy using the weighted information entropy is designed to enhance the richness of fault information. Subsequently, a condition-immune feature extractor is constructed with residual networks and sparse regularization techniques, which aims to extract the time-varying condition-immune fault features. Simultaneously, a cycle consistency domain adaptation network is designed for the inter-transformation of source and target domain features, aiming at enhancing the domain-invariant features. Finally, the cross-domain diagnosis of mechanical systems is realized through the collaborative training of the feature extractor with label classifier, domain discriminator, and cycle consistency network. Experimental validations of the proposed SF-CCDAN method have been carried out on two machinery datasets considering various time-varying conditions, and diagnostic accuracies of 98.38% and 98.31% are yielded respectively, which adequately demonstrate the superior domain adaptation capability and transfer diagnostic performance of the proposed SF-CCDAN method over several mainstream comparison methods.