This work presents a probabilistic diagnostic framework for solid oxide electrolyser cell systems based on electrochemical impedance spectroscopy. Instead of assuming a fixed operating point, the framework explicitly accounts for variations in operating conditions. Spectral data are deconvoluted using an equivalent circuit model whose parameters are obtained via variational Bayes inference, resulting in probabilistic estimates. Parameters inferred under nominal (healthy) operating conditions are then described by a Gaussian process model and subsequently used to check for faults. Specifically, deviations from nominal behaviour are quantified by means of the Wasserstein distance, which measures the discrepancy between the predicted and experimentally obtained parameter distributions. These distance-based residuals form a set of features that enable fault detection and fault isolation using a support vector classifier. The complete framework is validated on data from a 6-cell solid oxide electrolyser short stack collected over a 90-day experimental campaign comprising more than 600 electrochemical impedance spectroscopy spectra. On the test set, the proposed approach achieved 97% accuracy for fault detection and 96% accuracy for multiclass fault isolation under variable operating conditions, demonstrating robust diagnostic performance in dynamically operated solid oxide electrolysis cell systems. • Fault diagnosis under varying conditions using impedance spectroscopy. • Probabilistic model captures operating point effects on system behaviour. • Distribution comparison enables robust detection of system deviations. • Achieves 97% fault detection and 96% fault isolation accuracy. • Validated on 90-day electrolysis stack experiment with 600+ spectra.
Žnidarič et al. (Wed,) studied this question.