Transformer for gear fault diagnosis enhancing robustness through physics-informed multi-task learning

Abstract The effective diagnosis of gear faults significantly reduces downtime and cost and enhances the reliability of rotating machines such as wind turbines. In recent years, deep learning models have been increasingly applied in gear diagnostics at the cost of high expertise. Concurrently, it is important for deep learning models to extract features in accordance with physical phenomena to achieve a robust and reliable diagnostic performance. In this study, we constructed a multi-task diagnostic transformer model that focuses on the physical phenomena of gear fault vibrations. The tasks of estimating rotational shaft speed and gear mesh frequency (GMF) were simultaneously performed along with the diagnostic task. The proposed model was verified using acceleration data from an in-house gear test rig and five public datasets, under various operating conditions and gear geometries. The results show that multi-task learning (MTL) improves the diagnostic performance of gear faults in untrained conditions. Furthermore, we visualized the significant points of the input acceleration waveforms for the diagnostic results using explainable AI (XAI). The results showed that the proposed method allows periodic fault vibrations, in accordance with the physical phenomena of gears, to contribute more robustly to diagnosis.