This thesis addresses the estimation of the Remaining Useful Life (RUL) of a proton exchange membrane fuel cell by combining physical modeling with artificial intelligence. The objective is to develop reliable and robust predictive maintenance methods. A nonlinear degradation model has been developed to represent the evolution of the fuel cell’s performance over time. The proposed approach is hybrid: it integrates a robust Kalman filter with recurrent neural networks to estimate states whose dynamics are not explicitly known. This combination leverages both physical knowledge and the iterative structure of the model, while exploiting the predictive power of sequential learning methods. Particular attention is given to numerical stability issues and computational costs typically associated with classical methods. Different hybridization strategies, loose and tight, have been investigated in order to balance accuracy, robustness, and efficiency. The use of robust functions and semi-supervised learning compensates for the lack of direct access to internal states. Experimental results are obtained from a dataset generated through accelerated aging tests. The contribution offers a trade-off between RUL prediction accuracy and computational cost. Comparisons with alternative approaches, purely model-based, purely data-driven, or loosely coupled, confirm the superiority of the proposed method. In particular, it improves prediction accuracy while mitigating training instabilities.
