Machine learning-driven predictive maintenance models for hydrogen fuel cell systems in smart transportation networks

Hydrogen Fuel Cells (HFCs) are a central technology for advancing decarbonized mobility in Smart Transportation Networks (STN), yet their durability is limited by advanced electrochemical and thermal degradation. Anticipating such errors requires Predictive Maintenance Models (PMM) capable of extracting health indicators and predicting model behavior under dynamic operating conditions. This study develops Machine Learning (ML)-driven models for Fault Detection (FD), Remaining Useful Life (RUL) prediction, and prognostic reliability test in the Proton Exchange Membrane Fuel Cell (PEMFC) model. A 24-cell PEMFC stack dataset comprising 1500 h of operation under automotive load cycling was employed to analyze Supervised Learning (SL), Deep Temporal Networks (DTN), and a physics-guided hybrid residual model. Model training used cross-entropy and Mean Squared Error (MSE) objectives with causality-preserving temporal partitioning. Results proved that Deep Learning (DL) methods outperformed traditional classifiers, with the hybrid residual LSTM achieving 97.3% classification accuracy, 65.1 h RUL prediction RMSE, and early prognostic stabilization 85 h before error. Robustness analyses verified resilience against sensor noise, and computational profiling confirmed real-time feasibility with implication latency below 50 ms. These results establish that integrating physics-guided constraints into data-driven models yields accurate, deployable predictive maintenance for HFC, thereby enhancing safety, efficiency, and availability in STN.

Перевод пока недоступен