An Interpretable Neuro-Dynamic Scheme With Feature-Temporal Attention for Remaining Useful Life Estimation

With the wide application of deep learning in condition monitored system prognostics, its inadequate interpretability has always been questioned. This article proposes interpretable remaining useful life (RUL) estimation routine for RUL prediction, which consists of an augmenter network based on ordinary differential equations and an estimator network utilizing feature-temporal attention. The augmenter is implemented to suppress additive noise in original data and infer unobservable health-related variables with embedded formulas. Subsequently, the uncertainty-aware estimator is executed to predict RUL with quantile regressive module, while detecting dominant features and temporal dependencies via attention. Extensive evaluations are carried out on the N-CMAPSS aeroengine dataset. Compared with the baseline approaches, our method obtains 14% improvement for NASA's score and 7% for root-mean-square error. Moreover, the interpretability of our framework is further analyzed, in which the proposed method is able to imply physical constraints, detect abnormal subsystems and identify critical stages of flight with insufficient prior knowledge.