Generalized fault diagnostics of polymer electrolyte fuel cells using machine learning

Polymer electrolyte fuel cells (PEFCs) are promising for mobile and stationary applications, but short operational lifetimes and frequent faults limit their commercial viability. This work introduces a black-box diagnostic method using multifrequency Walsh function perturbation signals to detect water management and starvation faults. The approach improves signal-to-noise ratio and accuracy without harming the cell. Using voltage response as the diagnostic variable, dense neural networks (DNNs), 1D convolutional neural networks (1D-CNNs), and support vector machines (SVMs) were tested. All models accurately classified normal, drying, and starvation conditions in a single PEFC, with 1D-CNN and SVMs reaching 100% accuracy. However, model generalization to a different PEFC was poor. Including data from multiple PEFCs significantly improved performance, with the 1D-CNN showing superior generalization, even when trained on limited unseen data. This establishes the 1D-CNN as the most effective model for robust, scalable PEFC diagnostics across varied datasets.