Few-shot learning for estimating gear wear severity towards digital twinning

Digital twins hold the potential to revolutionize gear health monitoring by leveraging dynamic modeling and AI advancements, providing real-time insights and enabling proactive maintenance. Wear is one of the most critical fault mechanisms in gears, yet vibration-based gear diagnosis has mainly progressed for localized tooth faults due to the complexities associated with distributed wear. This study proposes a novel strategy for gear wear monitoring aligning with digital twins’ foundations, by integrating experiments, dynamic modeling, physical preprocessing, and machine learning. First, we propose an unsupervised learning-based sensitive health indicator for early wear detection, grounded in advanced physical preprocessing. Then, we introduce a novel few-shot learning approach for estimating wear severity, addressing the scarcity of labeled faulty measured data in real-world scenarios. We train a severity regressor using an extensive simulated training dataset, augmented and adapted to the inspected gear through transfer function estimation. The few-shot learning framework relies on measured training data, which includes healthy samples and a few faulty samples from a one label. The transfer of knowledge between the simulated and measured training datasets enables the severity regressor to make predictions on unseen measured data. The proposed strategy is demonstrated through controlled-degradation tests in gear wear, showcasing its superior performance in both fault detection and severity estimation, comparable to that of a fully supervised model, as if labeled measured data were available. Above all, this study highlights how a synergy between physics-based and AI-based approaches enhances performance while addressing practical challenges in wear diagnosis, paving the way for the implementation of digital twins in future endeavors.