A task-incremental semi-supervised meta learning method for few-shot RUL prediction

Abstract

Accurate remaining useful lifetime (RUL) prediction is essential for predictive maintenance. In recent years, with the advances in industrial artificial intelligence, many deep learning (DL)-based data-driven methods for RUL prediction have made significant progress. However, most of the existing deep learning models require an avalanche of expensive labeled data for model training. This severely limits their scalability to real-world industrial applications where the equipment is often maintained or replaced before failure, resulting in only a small number of failure historical units and a large number of suspension historical units. In other words, the RUL prediction becomes a few-shot issue with a small amount of labeled data and a large amount of unlabeled data. To address such issue, this article proposes a task-incremental semi-supervised meta-learning method for few-shot RUL prediction. We incorporate meta-learning to learn each unit-specific model as a task, which leverages unit-specific information to make more accurate prognostics. Different from prior works that leverage unlabeled data for self-supervised pretraining, the proposed method predicts the pseudo labels for suspension historical units. After calibration and selection, these pseudo-labeled units can augment the training task set to improve the model generalization capability. A theoretical generalization error bound theorem is proven that calibration and task increment can decrease the generalization error bound of the meta-learning model. Finally, a case study of aircraft engines shows that the proposed method is competitive with the state-of-the-art semi-supervised methods on few-shot RUL prediction.