Unsupervised Method Based on Adversarial Domain Adaptation for Bearing Fault Diagnosis

Abstract

This paper contributes to improving a bottleneck residual block-based feature extractor as a set of layers for transforming raw data into features for classification. This structure is utilized to avoid the issues of the deep learning network, such as overfitting problems and low computational efficiency caused by redundant computation, high dimensionality, and gradient vanishing. With this structure, a domain adversarial neural network (DANN), a domain adversarial unsupervised model, and a maximum classifier discrepancy (MCD), a domain adaptation model, have been applied to conduct a binary classification of fault diagnosis data. In addition, a pseudo-label is applied to MCD for comparison with the original one. In comparison, several popular models are selected for transferability estimation and analysis. The experimental results have shown that DANN and MCD with this improved feature extractor have achieved high classification accuracy, with 96.84% and 100%, respectively. Meanwhile, after using the pseudo-label semi-supervised learning, the average classification accuracy of the MCD model increased by 15%, increasing to 94.19%.

Keywords

• bearing fault diagnosis
• adversarial domain adaptation
• unsupervised learning
• transfer learning
• transferability estimation