Publication

Hyperautomation Artificial Intelligence

노이즈에 강건한 베어링 고장진단 모델 개발을 위한 규제 기법

본문

Conference
대한기계학회 CAE 및 응용역학부문 2025년 춘계학술대회
Author
김민재, 김성종, 이승윤, 이지원, 윤헌준, 윤병동
Date
2025-04-17
Presentation Type
구두
Abstract

Bearing failures can lead to severe accidents, making real-time monitoring and early fault diagnosis essential. Although deep learning-based methods achieve high accuracy, they often struggle with generalization in noisy environments. To improve robustness, we propose a multi-task learning (MTL)-based regularization method that considers task correlation and balances task difficulty. This approach enables effective shared representation learning and robust constraints without negative transfer or instability. Our method ensures balanced task difficulty and better regularization ability by integrating signal processing techniques, domain knowledge, Deep Operator Networks (DeepONet), and Many-to-One mapping. Experiments conducted on the CWRU dataset demonstrate that our model maintains a fault diagnosis accuracy of over 80% under -6 dB noise conditions, significantly outperforming the baseline model. These results confirm the effectiveness of the proposed method for noise-robust fault diagnosis.