Discriminative manifold domain adaptation for cross-domain fault diagnosis of rotating machineries

Abstract

Domain adaptation aims to solve the issue of distribution shift. Manifold learning uses Riemann metrics as nonlinear distance to overcome the limitation of Euclidean metric in domain adaptation. However, it ignores the feature distortion and relationship change caused by the feature mapping framework. Class overlap leads to degradation of classifier performance. Thus, this study proposes a new discriminative manifold domain adaptation (DMDA) framework to solve these problems. DMDA can align manifold structures and domain distributions through discriminative manifold alignment (DMA) and joint distribution alignment, respectively. The DMA method is developed via a new pseudo label voting mechanism, the local fisher discriminant analysis, and the geodesic flow kernel. It can achieve the manifold structure alignment and learn the discriminative features. Moreover, a structural risk minimization classifier is constructed to achieve Laplacian regularization and align the joint distributions of the source domain and target domain. It also implements the fault classification task. The experimental results on bearing and planetary gearbox datasets prove that DMDA has better fault transfer diagnosis performance than the typical domain adaptation methods based on machine learning.