Tight Hybridization of Long Short-Term Memory Neural Networks and Robust Kalman Filter for Remaining Useful Life Estimation of Proton Exchange Membrane Fuel Cells

As part of industry 4.0, predictive maintenance aims to achieve the best compromise between risk prevention and repair costs. It leverages diagnostic and prognostic methods that can be either model-based or data-driven. More recently, a combination of both has been considered. This paper deals with the estimation of the remaining useful life of a proton exchange membrane fuel cell by predicting the evolution of its power over time. We propose to combine a Kalman Filter (KF) with a Long Short-Term Memory (LSTM) neural network. Contrary to recent works, a tight coupling is considered. The LSTM neural network is trained to learn only a part of the KF prediction model whereas the remaining dynamics are described by knowledge models. Moreover, the available data exhibit outliers due to interruptions during the measurement campaign. To mitigate their impact, a robust version of the KF and a robust cost function for the training of the LSTM are considered as a second contribution.