Decoding spatiotemporal signatures of membrane drying in PEM fuel cell via distributed sensing and attentive deep learning

Accurate and timely diagnosis of membrane drying in proton exchange membrane (PEM) fuel cells remains challenging, primarily due to the limited spatial resolution of conventional, externally measured lumped sensor data. To address this gap, this paper proposes an attentive deep learning framework that leverages distributed sensing to directly capture and decode spatiotemporal patterns and internal current density heterogeneity associated with membrane drying. This paper pioneers the use of internal current density distribution for diagnosis, directly linking its spatiotemporal evolution to gradient membrane drying conditions. The developed CNN-LSTM-SEAM framework effectively integrates convolutional neural network (CNN), long short-term memory network (LSTM), and squeeze-and-excitation attention mechanism (SEAM) to jointly model spatial distributions and temporal dynamic. The attention-weighted reverse decoding mechanism innovatively provides interpretable insights into characteristic contributions. An attention-weighted reverse decoding mechanism is designed to provide interpretable insights by quantifying the contribution of spatiotemporal features to the diagnosis. Extensive validation confirms diagnostic accuracy above 99.30% and identifies the inlet/outlet as critical zones. This work presents a novel, interpretable approach that enhances both the precision and mechanistic understanding of membrane drying fault diagnosis in PEM fuel cells.