Investigation on the Capacity of Deep Learning to Handle Uncertainties in Remaining Useful Life Prediction

Remaining useful life (RUL) prediction is subjected to multiple uncertainty sources, such as measurement errors, operating conditions, and model representation capability. The quantification of the prediction uncertainty is important for assisting decision-making. In literature, stochastic processes have proven their efficiency in handling uncertainties in prognostics by providing RUL distribution. However, they have limitations in their adaptability to capture the dynamic behaviors of complex systems. To address this issue, it is recommended to employ deep learning (DL) methods that usually generate point-wise RUL predictions instead of RUL distribution. Therefore, the objective of this work is to investigate the capacity of DL methods to manipulate uncertainty in RUL predictions. Particularly, the probabilistic deep learning (PDL) framework is used to predict the RUL distribution instead of a point-wise RUL value. The obtained results by PDL are compared with the analytic solutions of the stochastic processes to highlight the uncertainty management capacity of PDL.

Keywords: Stochastic processes, Prognostics and health management, Deep learning, RUL prediction, Uncertainty management.

Download PDF