A physics-informed deep clustering framework with single-cell voltage features for early fault identification in automotive PEM fuel cell

In automotive proton exchange membrane fuel cells (PEMFCs), dynamic loads can induce inconsistencies within single cells and localized faults that are challenging to detect using stack-voltage-based diagnostic criteria. To facilitate earlier and more accurate fault identification, a key single-cell feature identification and determination method aimed at early anomaly capture and localization is first developed that quantifies temporal voltage discrepancies and selects the most fault-indicative single cells for diagnostic feature construction. Building on these features, a deep clustering method integrating DenseNet with physics-informed consensus clustering (DPICC) is proposed. Under a semi-supervised framework, a triple physics-informed integration mechanism constrains clustering center initialization, clustering process, and feasibility validation, jointly optimizing representations and cluster structure to improve accuracy and stability with limited labels while relying solely on single-cell voltage features. Experimental results demonstrate that fault identification can be advanced by up to approximately 35 s versus stack-voltage-based methods. Compared with Support Vector Machine (SVM), DPICC improves accuracy by 6.03%, and generalizes better across operating conditions. Ablations confirm that the triple physics-informed mechanism enhances clustering stability and interpretability, yielding an overall gain of approximately 16.6%. The proposed method provides a feasible pathway for early fault identification and reliable diagnosis of automotive PEMFCs.