Forecasting Reactant Ignition in Solid Oxide Fuel Cell Systems

Solid oxide fuel cell electrochemical stacks require high quality reformate for performance and durability. Insufficiently mixed reactants, carbon deposits, or improper chemical ratios thereof can result in reactant ignition during mixing prior to catalysis. Reactant ignition can warp and plug downstream components; therefore, it is desirable to predict and mitigate reactant ignition. Leading machine learning techniques were applied to the task of predicting ignition events in prototype (diesel-fueled) solid oxide fuel cells at a 30-second event horizon, using both current signal state and up to 30 seconds of signal history to make predictions. Based upon our analysis, first-order particle filtering using Fisher discriminant meta-reasoning provided the best cross-system performance when compared to other meta-reasoning methods (e.g., logistic regression, kernel support vector machine) as well as traditional vector quantization. In this paper, we demonstrate particle filter construction using data from eleven sensors, analyze predictive performance on real-world data, and discuss modifications to handle further system design changes.