A Meta-Transfer-Driven Method for Predicting the Remaining Useful Life of Rolling Bearing with Few Shot Data

Abstract

As a key component of rotating machinery, the operation, maintenance, and health management of bearings are of great significance, yet challenges such as the low bearing remaining useful life (RUL) prediction accuracy and the poor generalization persist. To address the above issues, this paper proposes a meta-transfer-driven method for cross-domain RUL prediction of rolling bearings with limited data. Firstly, the features with strong monotonicity and trendability in the time domain, frequency domain and time-frequency domain of bearing are selected. Then a meta-transfer learning framework is built based on task adaptation, incorporating the affine transformation parameters in the inner loop to enable adaptive model updates. In addition, the multi-kernel maximum mean difference (MK-MMD) is employed to minimize the differences between the two different domains. Finally, two cases validate the superior prediction results and generalization performance of the presented method.

Keywords: Remaining useful life, Transfer learning, Few-shot learning, Meta learning