Balanced Adaptation Regularization Based Transfer Learning for Unsupervised Cross-Domain Fault Diagnosis

Abstract

In fault diagnosis field, inconsistent distribution between training and testing data, resulted from variable working conditions of rotating machinery, inevitably leads to degradation of diagnostic performance. To address this issue, this study proposes a novel fault diagnosis method based on enhanced multi-scale sample entropies and balanced adaptation regularization based transfer learning. Specifically, different statistics-based multi-scale sample entropies are used to improve feature discriminability for different fault patterns under each working condition and enhance similarity of fault information between different working conditions. Then, based on these hand-crafted features, an improved transfer learning algorithm, referred to as balanced adaptation regularization based transfer learning, simultaneously exploring balanced distribution adaptation and balanced label propagation, is utilized to learn an adaptive classifier to perform cross-domain fault diagnosis. Finally, two public rolling bearing datasets verify that the proposed method can achieve an accurate diagnosis and outperform several existing transfer learning methods.