Generalised Fault Diagnostics of Polymer Electrolyte Fuel Cells Using Machine Learning

The ongoing threat of global warming necessitates a shift towards clean energy sources to meet rising demands while reducing carbon emissions. Polymer electrolyte fuel cells (PEFCs) represent a promising technology for both mobile and stationary applications but their poor operational lifetimes and frequent faults are barriers to their commercial ubiquity. Thus, detecting and addressing faults rapidly is crucial to extend PEFC lifetimes and enhance their viability for alternative electricity generation. White-box and black-box models are widely used to diagnose PEFC faults; this work introduces a novel black-box approach using multifrequency Walsh function perturbation signals to diagnose polymer electrolyte membrane PEFC faults. This method improves signal-to-noise ratios in the electrical response of the fuel cell, increasing measurement accuracies without causing cell damage from excessive perturbation amplitudes. Using the voltage response of the PEFC as the diagnostic variable, dense neural networks (DNNs), 1-dimensional convolutional neural networks (1D-CNNs), and support vector machines (SVMs) were investigated for fault classification. Initial testing revealed that all models could accurately detect normal, drying, and starvation conditions in an individual PEFC, with the 1D-CNN and SVMs achieving 100% diagnostic accuracy. When tested on data from a different PEFC, the models exhibited poor generalisation abilities; nevertheless, combining data from multiple PEFCs significantly improved diagnostic accuracy, with the 1D-CNN displaying superior generalisation performance, particularly when trained with only a small portion of new data. The network’s convolutional architecture facilitates effective parameter sharing and local connectivity, enhancing computational efficiency and reducing errors. Thus, it emerges as the most suitable model for the diagnostic framework, capable of managing varying datasets from different PEFCs while maintaining high accuracies.