Remaining useful life estimates of a PEM fuel cell stack by including characterization-induced disturbances in a particle filter model.

Proton Exchange Membrane Fuel Cells (PEMFC) are available for a wide variety of applications such as transportation, micro-cogeneration or powering of portable devices. However, even if this technology becomes close to competitiveness, it still suffers from too short life duration to pretend to a large scale deployment. In a perspective of a longer lifetime, prognostics aims at tracking and anticipating degradation and failure, and thereby enables deciding mitigation actions to increase life duration. Yet, the complexity of degradation phenomena in PEMFC can make prognostic implementation really tough. Indeed, a PEMFC implies multiphysics and multiscale phenomena making the construction of a physics-based aging model very complex. Moreover, prognostics should also take into account external events influencing the aging. Among them, characterization techniques such as electrochemical impedance spectroscopies and polarization curves introduce disturbances in the stack behavior, and a voltage recovery is observed at the end of characterizations process. It means that irreversible degradation and reversible decrease of performances have to be considered. This work proposes to tackle this problem by setting a prognostics system that includes disturbances' effects. We propose a hybrid prognostics approach by combining the use of empirical models and available data. In an evolving system like a fuel stack, a particle filtering framework seems to be really appropriate for life prediction as it offers the possibility to compute models with time varying parameters and to update them all along the prognostics process. Moreover, it offers a great adaptability to include characterization effects and allows giving prediction with a quantified uncertainty. The logic of the work is the following. First, it is shown that simple empirical models only taking into account the aging are very limited in terms of prognostics performances. Then, some features describing the impact of characterization on the stack behavior and aging are extracted and a more complete prognostics model is built. Finally, the new prognostic framework is used to perform remaining useful life estimation and the whole proposition is illustrated with a long term experiment data set in constant current solicitation and stable operating conditions.