A cross-domain few-shot remaining useful life estimation framework based on model-agnostic meta-learning with task embeddings

Remaining useful life (RUL) estimation aims to predict the time until system failure based on monitoring data, facilitating proactive maintenance actions. Precise RUL estimation can significantly enhance system reliability and safety. However, when new system failures emerge, predictive models trained on historical failures data often encounter difficulties in accurate estimation. The distribution shift between historical and new failures data, coupled with extremely few new failures data, results in cross-domain few-shot prognostic scenarios, posing a significant challenge to many deep-learning-based RUL estimation methods. In response to the challenge, this paper proposes a novel cross-domain few-shot RUL estimation framework based on model-agnostic meta-learning (MAML) with task embeddings. First, a segmentation strategy is adopted to construct more meta-tasks, which can capture more comprehensive degradation information for efficient meta knowledge extraction. Then, task embeddings that are independent of backbone network are designed to encode task-specific degradation knowledge into efficient low-dimensional vectors, which alleviates overfitting caused by limited labeled data, thus improving RUL estimation performance. Moreover, the encoded degradation knowledge is only injected into feature extractor, making representation change dominant for better cross-domain adaptability. Experimental results on turbofan engine and wind turbine gearbox datasets reveal the effectiveness and superiority of the proposed framework. Estimation results evaluated by RMSE and Score improve 9 % and 31 %, respectively.