Interpretable data-augmented adversarial variational autoencoder with sequential attention for imbalanced fault diagnosis

Abstract

Data augmentation with imbalanced samples is tentative and tricky for rolling bearing fault diagnosis in practice. However, its complex hierarchical knowledge and opaque decision rules prevent users from trusting the outputs. In this paper, an interpretable data-augmented adversarial variational autoencoder with sequential attention (AVAE-SQA) is proposed for assisting imbalanced fault diagnosis. Imbalance fault diagnosis is supported by AVAE-SQA, which synthesizes data to supplement the imbalanced dataset for meeting the needs of intelligent diagnostic models. Firstly, an adversarial variational inference is defined, which is optimized by the reliable control limit (RCL) with dual bounds. The data distributions are estimated better due to dual bounds; that is, inter-class data augmentation is facilitated by RCL. Then, AVAE with an adversarial process is designed to ensure the feasibility and controllability of RCL while improving its accuracy and generalization. Thirdly, SQA is constructed for global dependencies without overfocusing. Furthermore, SQA illustrates the decision rules of AVAE, which coincide with the failure mechanism of rolling bearings. Various cases are adopted to evaluate the effectiveness of the proposed method. Experiments confirm that AVAE-SQA is preferred over other prevailing approaches in fault diagnosis with imbalanced samples and is potentially promising for engineering applications.