A fault diagnosis method using frequency-aware adaptive domain alignment for bearings

Abstract

In practical applications of rolling bearings, variations in the measurement data distribution caused by diverse operating conditions result in complicated domain adaptation tasks and significantly impair the effectiveness and generalizability of conventional models. Therefore, this study proposes a frequency-domain-aware adaptive domain alignment network (FADAN). This framework initially transforms the measured vibration signal into the frequency domain and subsequently extracts multiscale spectral features via a parallel multiresolution branch optimised using a frequency-domain attention mechanism. To enhance the generalisation capability, a class-aware domain classifier is employed along with an adaptive gradient-inversion layer, enabling precise and dynamic domain alignment. A comprehensive evaluation across the three datasets demonstrates that FADAN consistently outperforms the domain adaptation methods in the control group in terms of both diagnostic accuracy and stability. These results validate the effectiveness and robustness of the proposed framework and establish it as a generalizable solution for bearing fault diagnosis under complex real-world operating conditions.