Fuel cell fault diagnosis using Kalman filtering and extreme learning machine

Green technologies such as fuel cells are needed to reduce greenhouse gas emissions. However, fuel cells can experience faults such as hydrogen crossover, where hydrogen leaks through the membrane, resulting in oxygen starvation. Leak faults and the ensuing starvation can accelerate degradation, reducing fuel cell life. This thesis develops and refines techniques to estimate hydrogen leak faults and oxygen starvation to help mitigate fault progression. Specifically, this thesis develops fault detection techniques using traditional artificial neural networks (ANN) as well as extreme learning machines (ELM), a special subset of machine learning algorithms. The thesis also develops extended Kalman filters (EKFs) that are used in conjunction with ANN and ELM to mitigate the effect of the noise. The data for the training is generated by adding input and measurement noise to a previously developed pseudo-2D model of the fuel cell. It was found that the ELM trained much more quickly than the ANN, but that the accuracy of the ANN and ELM were similar. The EKF model of the fuel cell agreed with the pseudo-2D model for normal fuel cells, but not for oxygen-starved fuel cells. Hence, when using the EKF as a prefilter for the machine learning algorithms, the machine learning estimate for hydrogen crossover leakage and oxygen starvation improved for normal fuel cells, but not for starved fuel cells. The disagreement between the EKF model and the pseudo-2D model likely stems from the failure of the EKF to account for losses due to hydrogen pumping and hydrogen crossover leakage, resulting in a significant reduction in accuracy for the resulting machine learning data. The best way to deal with this is to either account for hydrogen pumping and hydrogen crossover in starved fuel cells or to not use the EKF as a prefilter for voltage in starved fuel cells or when classifying fuel cells as normal or starved or starved.