LT-PEMFC remaining useful life prediction for predictive maintenance
Résumé
The Low-Temperature Proton Exchange Membrane Fuel Cell (LT-PEMFC) is one of the most promising systems for decarbonizing key sectors such as the automotive or marine industries. To make this type of system competitive in terms of performance and cost, the discipline of Prognostic and Health Management (PHM) has been applied to LT-PEMFC. This innovative approach makes it possible to optimize maintenance (moving from preventive to predictive maintenance), forecast health status and ultimately extend system lifetime, thanks to sophisticated algorithms enabling real-time monitoring. In ref [1], an ESN was chosen to predict the degradation of a LT-PEMFC over time. The authors report very good results, outperforming particle filtering (PF) in terms of accuracy and computation time. In ref [2], an ESN was also used to predict the voltage degradation of an LT-PEMFC cell. Again, prediction results were good, however, measuring cell voltages is not always accessible in a commercial system. In order to be compatible with a commercial system, in this paper, the voltage of the PEMFC stack will be used as an indicator of Remaining Useful Life (RUL). A data-driven method will be developed on an open source dataset from our lab [3]. This dataset is based on 500W LT-PEMFC, operating under nominal load with current ripples of +-5%, for a total duration of 1,055h. The data-driven approach was used to predict the RUL of an LT-PEMFC, using an Echo State Network (ESN) and a Long-Short Term Memory (LSTM). The aim is to predict the RUL at least 100h in advance to carry out the maintenance intervention before the LT-PEMFC can no longer perform its mission.