Reliable cross-domain lifetime prediction for fuel cells: A time-frequency fusion transfer learning architecture

Current transfer learning-based methods for predicting the degradation of proton exchange membrane fuel cells (PEMFCs) commonly face insufficient elimination of inter-domain distribution shifts and low knowledge transfer efficiency. Traditional approaches typically extract domain-invariant features by filtering multi-dimensional operational parameters. This process is not only complex but may also introduce interference weakly correlated with voltage degradation, thereby compromising the stability and reliability of cross-domain predictions. To address these challenges, this paper proposes a time-frequency fusion transfer learning (TF-TL) architecture. Unlike existing methods that filter domain-invariant features from multi-dimensional operational parameters, the proposed approach directly aligns domains by leveraging the intrinsic frequency-domain characteristics of voltage signals. The integrated Frequency Domain Adaptation (FDA) technique employs Fast Fourier Transform (FFT) to extract intrinsic low-frequency components from voltage signals. It then aligns features between source and target domains using an enhanced Maximum Mean Discrepancy (MMD) metric. This voltage-centric, frequency-domain alignment avoids dependence on complex auxiliary parameters, thus improving the interpretability and stability of transfer learning. Experimental results show that TF-TL significantly improves generalization across devices and operating conditions, with FDA effectively reducing inter-domain distribution divergence. In multiple transfer tasks, the framework achieves highly accurate and reliable remaining useful life estimation with a relative error below 1.9%. The proposed architecture provides a theoretical foundation for deploying rapid and reliable online prognostic systems for PEMFCs under complex operating conditions.