Remaining useful life prediction under variable operating conditions via multisource adversarial domain adaptation networks

Abstract

Domain adaptation (DA) methods have been extensively applied for intelligent remaining useful life prediction (RUL) tasks, addressing domain shift issues under variable operating conditions. Almost all existing methods focus on exploiting the single-source unsupervised domain adaptation (SUDA) approach which only learns prognostic knowledge from a single domain. Nevertheless, SUDA methods have significant limitations due to the disregard for the valuable information contained in historical data from numerous sources. Compared to conventional SUDA approaches, multisource unsupervised domain adaptation (MUDA) approaches can acquire invariant domain representation from a variety of sources and offer higher generalization in transfer tasks. Therefore, a novel MUDA framework multisource adversarial domain adaptation network (MADAN) is proposed for regression tasks. MADAN can efficiently organize and utilize multiple source domain data through a two-stage adversarial domain alignment, of which adversarial domain distribution alignment and inter-domain regression alignment modules are designed to minimize domain discrepancy and regressor discrepancy between each pair of source and target domains. Experimental results indicate that MADAN achieves accurate cross-domain predictions by using multiple source domains, and a comparison demonstrates that it performs better than the state-of-the-art approaches.