AI-Assisted performance evaluation and health monitoring of PEM fuel cells: an experimental study of component-level aging, LSTM-based degradation forecasting and random forest fault diagnosis

Proton-exchange-membrane fuel cells (PEMFCs) are at the core of clean hydrogen energy systems, but their long-term operation is constrained by coupled ageing of the membrane-electrode assembly and operational faults such as flooding, membrane drying, fuel starvation, and catalyst-layer degradation. This paper presents an experimental and data-driven study of PEMFC stack behaviour during 1500 hours of continuous operation using two complementary artificial-intelligence models: a Long Short-Term Memory (LSTM) network for voltage-degradation forecasting and remaining-useful-life (RUL) estimation, and a Random Forest (RF) classifier for fault diagnosis. Polarization and power-density characteristics were measured at 0 h, 500 h, 1000 h and 1500 h, while six operating parameters—cell voltage, stack temperature, relative humidity, pressure drop, current ripple and charge-transfer resistance—were recorded for fault-signature extraction. The revised analysis explicitly relates the observed stack-level degradation to the ageing pathways of the carbon paper, membrane, catalyst and ionomer and discusses their cumulative effects. A component-screening database was organized by mapping each key component and ageing mechanism to measurable sensor signatures and AI-learning variables; direct component-level chemical or microscopic measurements were not part of the present experiment and are therefore treated as a limitation rather than as synthetic data. The LSTM model predicted the voltage-degradation trajectory with RMSE of 6.85 mV, MAE of 4.51 mV and R² of 0.970 and predicted RUL with R² of 0.960. The RF classifier achieved 97.3% accuracy, 97.4% macro-precision, 97.3% macro-recall and 97.3% F1-score on five operating-condition classes. Results demonstrate that combining component-aware degradation interpretation with sequence-learning prognostics and ensemble fault classification provides a practical health-monitoring framework for PEMFC systems.