Adaptive manifold partial domain adaptation for fault transfer diagnosis of rotating machinery

Abstract

When the source domain has more fault types than the target domain, the traditional domain adaptation (DA) performance will degrade. Therefore, exploring partial domain adaptation to improve the accuracy of cross-domain fault diagnosis is significant. This work develops a novel adaptive manifold partial domain adaptation (AMPDA) approach for implementing the partial transfer fault diagnosis by decreasing the distribution discrepancy and adaptively learning global geometrical structures. Specifically, AMPDA first proposes the extraction method of the global geometrical structure by an affinity matrix. Then, the two domains are aligned by decreasing the maximum mean discrepancy (MMD) between two classes. AMPDA applies the manifold regularization to the objective function of partial domain adaptation, adaptively changing the weight of the geometrical structure according to the feature mapping within the process aligning two domains. AMPDA can adaptively learn the global geometrical structure to reduce the man-made interference in the K-nearest neighbor and restrain negative transfer learning caused by forcing to keep the structure. AMPDA is successfully applied to partial transfer fault diagnosis of rolling bearings and planetary gearboxes with unlabeled target domain samples. The experimental results demonstrate that the AMPDA performs better than typical DA methods based on machine learning.