An Interpretable and Reliable Remaining Useful Life Prediction Approach Across Different Machines With Tensor Domain-Adversarial Regression Adaptation

Abstract

This article tries to address the concerns about remaining useful life (RUL) prediction across machines: 1) what data from source domain contributes more to transfer prediction? and 2) is the information transfer reliable enough? This article proposes a novel fault mode-oriented deep tensor domain-adversarial regression adaptation approach to achieve interpretable RUL transfer prediction across machines. First, by integrating fault mechanism and degradation characteristics, a new fault mode-oriented significance indicator (FSI) is constructed based on tensor representation to evaluate the importance of degradation data from source domain. Second, a multisubdomains adversarial regression adaptation network, in which each subsource domain corresponds to a fault mode, is constructed to purposefully transfer the degradation knowledge from source domain. The domain discriminator for each subsource domain is adaptively weighted by FSIs that are updated in each round of adversarial training. An alternating optimization algorithm is then designed to find the optimal knowledge representation and transfer effect. Moreover, an upper bound of prediction error is derived for the proposed approach, which offers a theoretical guarantee for cross-machine prognostic task. Experimental results on three benchmark datasets empirically validate the proposed approach under fixed and varying working conditions, and can reveal fault modes' significance for more trustworthy prediction.