Bayesian Neural Networks with semi-empirical signal processing for PEM Fuel Cell Fault Diagnosis

Proton exchange membrane fuel cell technology offers a sustainable low carbon alternative to fossil fuels, particularly for transport applications. Yet it still needs to be improved in terms of durability and cost to be competitive on the market. This work presents a method for determining its state of health using empirical Fourier decomposition and Bayesian neural networks, with a comparison with more conventional neural networks as classification tools. The proposed classifiers effectively distinguish between three operating conditions: normal, flooding, and drying. The classifiers demonstrated high performance, achieving accuracies ranging from 95% to 98%, with the Bayesian neural network exhibiting a slight advantage.