A Physics-Informed Training Approach for Data-Driven Method in Remaining Useful Life Estimation

Abstract

Data-driven models are currently used extensively for remaining useful life (RUL) estimation of equipment with multisensor signals. But the low controllability is one of their common limitations. This study proposes a systematic method to predict RUL with multisensor data under dynamic operating conditions and failure modes. The proposed method integrates a physics-informed loss function with data-driven methods to achieve the targets of safe and controllable predicting. A delayed prediction penalty mechanism-based loss function is introduced into the deep learning model training. Finally, the proposed method is validated on the Commercial Modular Aero-Propulsion (C-MAPSS) dataset. Comparisons with other advanced forecasting methods show that the predictions are more safety while ensuring high-fitting accuracy. The controllability and flexibility of the deep learning model are improved in practice.