Partial Transfer Learning using Domain Knowledge Filter for Fault Diagnosis of Bearing
본문
- Conference
- ASME 2025 International Design Engineering Technical Conferences & Computers and Information in Engineering Conference
- Date
- 2025-08-18
- Presentation Type
- Oral
Abstract
Rolling element bearings are essential components of rotating machinery, and their failure may result in substantial element bearings are essential components of rotating machinery, and their failure may result in substantial economic losses and pose serious safety risks. Accurate fault diagnosis of rolling bearings is essential for maintaining the stability and performance of rotating machinery. However, in industrial environments, changes in operating conditions and limited availability of labeled data often degrade the performance of deep learning models. Moreover, conventional domain adaptation techniques assume that both source and target domains share the same label space-a condition rarely met in real world applications. This mismatch can lead to negative transfer, where irrelevant source classes interfere with learning in the target domain. To address these limitations, we propose a partial transfer learning framework that introduces two key modules: a domain knowledge filter and a gradient alignment module. The domain knowledge filter extracts physically meaningful features by generating short-term and long-term envelop signals from raw vibration data, allowing the model to better capture fault-specific patterns. Meanwhile, the gradient alignment module ensures effective distribution matching by selectively aligning shared class features between domains, thereby avoiding negative transfer and stabilizing training. Experimental results on publicly bearing datasets confirm the robustness and accuracy of the proposed method under various partial and imbalanced label scenarios, demonstrating its practicality for real-world industrial applications.