Discriminative feature-based adaptive distribution alignment (DFADA) for rotating machine fault diagnosis under variable working conditions

Abstract

Recent years, cross-domain fault diagnosis of rotating machinery has been a hot topic, and various kinds of methods taking advantage of transfer learning are proposed correspondingly. Despite their success, they mainly focus on marginal distribution alignments, which ignore weighing between marginal and conditional distributions in network training. However, this kind of weighting can boost diagnosis network performance further and make it more robust. Hence, a novel transfer learning method called discriminative feature-based adaptive distribution alignment (DFADA) is proposed, which can extract discriminative features and conduct a two-stage adaptive distribution alignment on L2 ball. In DFADA, maximum mean discrepancy (MMD) and graph Laplacian regularization are fused to extract discriminative and task-specific features. Meanwhile, for comprehensive and adaptive distribution alignments, the distributions of datasets are pre-matched via MMD and further matched in feature classifier via dynamic distribution alignment (DDA), which can not only reduce both marginal and conditional distribution discrepancies but also weigh their importance adaptively. Finally, a DFADA-based fault diagnosis method for rotating machinery with volatile working conditions is constructed correspondingly. The validity of the proposed method is also confirmed by extensive experiments and comparisons with some state of the arts on 18 transfer learning cases.