Deep imbalanced and semi-supervised regression scheme for aero-engine remaining useful life prediction with few run-to-failure data

Remaining useful life (RUL) prediction is the core of predictive maintenance for mechanical systems. Deep learning methods have been widely used in this field, leveraging the abundance of run-to-failure data. However, aero-engines are normally maintained before failure occurs due to the periodic maintenance strategy. Few run-to-failure data and large volume of degradation data prior to failure is collected in the real situation, which results in problems of data imbalance and unlabeled data. Hence, conventional deep learning methods are hard to work or have poor performance for RUL prediction of aero-engines in the real situation. In this work, a deep imbalanced and semi-supervised regression (DISSR) scheme is proposed for RUL prediction with few run-to-failure data. In the scheme, a data augmentation module based on multi-sensor time series forecasting is firstly proposed for complementing those operational data, and then an assigning strategy is proposed to obtain pseudo-labels. Secondly, a deep graph encoder-decoder architecture is proposed for adaptive feature extraction. Finally, a predictor is used for RUL prediction. The C-MAPSS dataset of aircraft engine is used to test the proposed method in this work. The experiment results show that the proposed method has higher accuracy than the conventional methods using few run-to-failure data.