Multisensor Time Series Forecasting-Based Augmentation Framework for Remaining Useful Life Prediction of Aero Engines

Remaining useful life (RUL) prediction for aero engines is critical for predictive maintenance. Deep learning methods have been widely used for RUL prediction. In real scenarios, the late-stage degradation data is a minority of historical data, resulting in the problem of data imbalance. Hence, the conventional deep learning-based methods have poor accuracy for RUL prediction of aero engines. To solve this problem, a multi-sensor forecasting-based augmentation (MSFA) framework is proposed to enhance RUL prediction. In this framework, a multi-scale and multi-component transformer (MSMCT) model is proposed to forecast multi-sensor time series for completing censored data. A wasserstein distance-based selector is used to augment the training dataset with forecasted data. Finally, a dynamic graph spatio-temporal network (DGSTN) is trained on the augmented dataset for accurate RUL prediction. Experimental validation on the C-MAPSS dataset demonstrates the framework’s superior performance in improving prediction accuracy under imbalanced data scenarios.