A Time-Space Fusion Early Diagnostic Framework for PEM Fuel Cell Faults with Deep Learning

An accurate fault diagnostic method of proton exchange membrane (PEM) fuel cells can guide adjustments for control strategies and enable the system to operate under optimal conditions. However, the performance of existing PEM fuel cell fault diagnostic methods remains subject to uncertainty and limited verifiability due to the lack of real fault data, and the consideration of non-uniformity inside the stack is insufficient under faulty conditions. To address this issue, a time-space fusion early diagnostic framework for PEM fuel cell faults is proposed in this study, which is designed to provide early warnings at the onset of electrochemical mechanism deviation rather than the manifestation of performance degradation. Polarization tests are conducted to extract the related abnormalities data, and three temporal latent indicators are constructed for developing a deep learning (DL)-based method to predict the internal dynamic state of the stack. Besides, two spatial indicators are defined to characterize cell-level voltage non-uniformity as complementary information for early fault detection. By integrating the states of predicted temporal latent indicators and spatial indicators, a set of early diagnostic rules for membrane dry-out, flooding, and oxygen starvation is ultimately formulated. The results show that the proposed framework has higher credibility and mechanism consistency.