Transfer learning for fault classification in heavy rotating machines

Health monitoring of rotating machines, particularly in the aviation industry, poses a significant challenge: developing machine learning models without direct access to labeled faulty data from the monitored machines. This challenge stems from the aviation industry's stringent safety standards, which minimize fault occurrences and, consequently, limit the availability of faulty data. In contrast to traditional fault classification methods, our approach addresses this critical gap by enabling fault detection and classification in scenarios of limited or no faulty data. We propose a novel domain adaptation methodology, employing transfer function estimation to enable Transfer across Different Sensors (TDS). In this study, we evaluated a case study involving gear wear faults, demonstrating our method’s ability to classify fault severity with an increase of at least 20% in accuracy, showcasing its adaptability across domains. This work lays the foundation for advancements in the development of transfer across different machines and digital twins, providing an innovative framework for predictive maintenance and health monitoring in safety-critical industries such as aviation.