Proton exchange membrane fuel cell (PEMFC)

as core energy conversion devices in the hydrogen economy

exhibit significant performance variations across different structural specifications and operating conditions. Although deep learning technology has become the mainstream approach for performance degradation prediction

its predictive robustness diminishes markedly under small-sample conditions. While existing transfer learning methods can alleviate data bottlenecks

they struggle to overcome knowledge transfer barriers caused by inter-domain discrepancies. To address these challenges

this paper proposes a cross-domain degradation prediction method based on frequency-domain adaptive transfer learning. This technique employs fast Fourier transform to extract global frequency- domain features from voltage signals. It then couples maximum mean discrepancy with kernel density estimation to collaboratively identify low-frequency voltage consensus components between source and target domains

fundamentally eliminating domain shifts in degradation features. Utilizing the reconstructed shared feature space

a time-frequency hybrid model is further employed

so as to achieve cross-domain transfer prediction. Experimental results demonstrate that the proposed method achieves accurate predictions with minimal training samples under cross-condition and cross-device scenarios. Compared with existing transfer learning techniques

it reduces the root mean square error by up to 34.1%. This work introduces the integration of time-frequency analysis into degradation transfer modeling

establishing a novel framework for PEMFC lifetime prediction that synergizes dynamic feature decoupling with knowledge transfer.