Unsupervised Bearing Fault Diagnosis via a Multi-Layer Subdomain Adaptation Network

Abstract

Bearings play a crucial role in the functioning of rotating machinery, making it essential to monitor their condition for maintaining system stability and dependability. In recent years, intelligent diagnostic techniques for bearing issues have made significant progress due to advancements in artificial intelligence. These methods rely heavily on data, requiring data collection and labeling to develop the learning model, which is often highly challenging and nearly infeasible in industrial settings. As a result, a domain adaptation-based transfer learning approach has been suggested. This approach aims to minimize the difference between the distribution of accessible data and the unlabeled real-world data, enabling the model trained on public data to function effectively with actual data. In this paper, we introduce a sophisticated subdomain adaptation technique for cross-machine bearing fault diagnosis using vibration, termed multi-layer subdomain adaptation. Verification experiments were conducted, and the findings indicate that the proposed approach offers relatively high accuracy up to 97.47% and excellent transferability. Comparative experiments revealed that the proposed method is a superior technique for bearing fault diagnosis and slightly outperforms other methods (3-5%) in both predictive and noise-ignore capabilities. Comprehensive validation experiments were conducted using the HUST dataset.