RedoxFormer: A two-stage deep neural diagnosis framework for early-stage degradation in fuel cells via in-situ sensor measurements

Solid Oxide Fuel Cells (SOFCs) are promising clean energy conversion systems, yet their commercialization remains limited by vulnerability to unexpected degradation under real-world operating conditions. Early-stage diagnosis of degradation is therefore critical to enable predictive maintenance and prevent irreversible failures. Conventional diagnostics, such as Electrochemical Impedance Spectroscopy, provide valuable insights but are costly for real-time monitoring. This study presents an intelligent diagnostic framework for detecting the onset of nickel-redox degradation using readily available in-situ signals, such as power density. Ten identical anode-supported, lab-scale SOFCs were subjected to accelerated redox-cycling experiments to generate diverse run-to-failure datasets. The collected data were carefully labeled through an extensive knowledge-based analysis, providing a foundation for supervised fault detection. A novel two-stage algorithm is then proposed that integrates a Transformer-based neural forecasting model with a convolutional time-series classifier. This architecture aims to classify real-time streams of post-redox data into either healthy or faulty categories, enabling the detection of degradation signs far earlier than existing diagnostic methods. Benchmarking against state-of-the-art machine learning and deep learning classifiers demonstrates that the proposed framework not only achieves competitive performance across multiple metrics but also predicts failure events up to 81% earlier than existing models. By jointly considering diagnostic error performance and timeliness through the area under the Accuracy–Earliness Curve score, our model demonstrates a 40% improvement over the best-performing baselines. These results highlight the potential of the proposed framework to deliver accurate and timely detection of Ni-redox degradation, supporting predictive maintenance strategies that mitigate degradation effects and minimize SOFC downtime. • A large-scale SOFC degradation dataset (>5 million samples) from run-to-failure redox-cycling experiments is generated and made publicly available. • Degradation onset is labeled using a transparent, rule-based algorithm grounded in EIS/DRT peak evolution, ensuring reproducibility and physical consistency. • A novel two-stage RedoxFormer framework integrates Transformer-based long-horizon forecasting with time-series classification using only in-situ power density signals. • Interpretability and cross-degradation analyses confirm that RedoxFormer learns physically meaningful features and generalizes robustly across distinct SOFC degradation mechanisms. • RedoxFormer enables reliable early SOFC fault diagnosis, achieving up to 81% earlier detection and a 40% improvement in the Accuracy–Earliness score compared to state-of-the-art methods.