Fault Diagnosis of a Gearbox using Physics-information-based Artificial Intelligence

Research on vibration-based gearbox fault diagnosis using artificial intelligence (AI) has been actively conducted in recent years. However, a problem exists whereby the performance of AI models that have been optimized to solve certain problems such as classification based on the pattern analysis of training data can be significantly degraded when the characteristics of the measured vibration signals are varied by the various changes in the operating environments. In this study, a physics-based kernel that can extract fault-related features based on physical information on the defect characteristics is designed. Unlike typical AI model kernels, the proposed physics-based kernel was fixed to prevent it from being learned, allowing physical information to be reliably extracted. This allows the development of a robust gearbox fault diagnosis in various domains, including non-stationary operating conditions that occur in actual industrial settings. The proposed idea was the validated on a gearbox testbed operated under various operating conditions. Consequently, the proposed method outperforms the conventional method, while retaining its physical robustness and capability for accurate gearbox fault diagnosis.