Integrated deep stacked autoencoder–BiLSTM networks and online geometric electrochemical impedance spectroscopy data for real-time fault diagnosis and health assessment of PEMFCs

This paper introduces a novel end-to-end framework for multi-fault diagnosis and state-of-health (SoH) tracking in proton exchange membrane fuel cells (PEMFCs). The approach combines geometric analysis of electrochemical impedance spectroscopy (EIS) with a physics-informed deep stacked autoencoder-bidirectional long short-term memory (dSAE-BiLSTM) network. An extended 11-element equivalent circuit produces twelve interpretable descriptors, centers and radii of four depressed Nyquist semicircles, representing ohmic resistance, anode/cathode activation, and mass-transport processes. This compact feature set condenses complex spectral data into a twelve-dimensional vector, enabling on-board extraction from a single frequency sweep without manual thresholding. A dSAE denoises and compresses these geometric features into a latent manifold, preserving electrochemical interpretability. A BiLSTM processes temporal sequences of latent codes to classify five fault modes, membrane dehydration, cathode flooding, catalyst degradation, reactant blockage, and membrane structural damage. Physics-aligned multi-task loss functions enforce monotonicity and geometric consistency, anchoring outputs to mechanistic relations among circuit parameters. Optimized regression heads quantify fault severity and track SoH in real time. Experimental validation demonstrates a fault classification accuracy above 98.5% and a SoH estimation achieving an average accuracy exceeding 92%, with a peak reaching 97.3%, under dynamic load profiles. Importantly, the proposed method showcases exceptional resilience; even when subjected to 30% noise applied to data extracted from EIS curves, it maintains an accuracy above 86.1% by leveraging its integrated filtering and circle-fitting technique. Comparative and ablation studies confirm superior performance and robustness over model-based, deep-learning, and prior hybrid approaches.