Graphical Modeling for Robust Prognostic of Proton Exchange Membrane Fuel Cell

The fuel cell (FC) is at present the alternative solution to the fossil fuels the most promising. It is however advisable to optimize it from a technical and financial point of view to see them appearing on the market for stationary and automobile applications. For that purpose, the improvement of the reliability and the availability of the FC system requires the implementation of algorithms capable not only of detecting and of identifying as soon as possible the failures but also of estimating in real time the state of health and forecasting of its remaining useful life. The methods of detection and isolation of faults (FDI) well developed in the literature have a major inconvenience owed mainly to the fact that they base themselves on the symptom of appearance of a failure. On the other hand, the prognostic consists in the estimation of the operating time before failure of a system and the risk of existence or later appearance of a failure. The methods of prognostics based on a physical model offer generally precise results once they do not requiring either learning or expertise of the operator. However, the problem for a FC system lies in the coupling of several phenomena (electrochemical, electric, thermodynamic), the uncertainty of the parameters of the model and the low instrumentation of the core of the fuel cell stack. In the thesis, we use uncertain models based on the Bond Graph tool well adapted to the multidisciplinary and multi-physical aspect of the FC. Concretely, the parameters uncertainties are integrated in the Bond Graph elements to generate a robust model (based on Bond Graphs LFT - Linear Fractional Transformation) of evolution of the powers (associated with the health of the FC). These dynamical models robust to parameters uncertainties are used for the detection of the beginning of the aging and the estimation of the degradation of the FC based on the causal and structural properties of the model. The generated model of degradation is used by an extended Kalman filter which allows the estimation of the state of health and the dynamics of the aging for any operating condition (of temperature, current and pressure). Furthermore, this algorithm allows to estimate the uncertainty which is used by the Bond Graph LFT model as well as by an Inverse First Order Reliability Method for the prediction of the remaining useful life and the inherent uncertainty. The global method was validated on various sets of data: - Constant load; - μ-CHP profile (Co-generation of Heat and Power); - Automotive profile. All the algorithms was integrated into a demonstrator developed under Matlab Guide. This one also allows the control by inversion of an EMR model (Energetic Macroscopic Representation) with time-varying parameters, robust to the aging (based on the state of health estimation). (author)