Variations in axlebox bearing geometry, rotational speed, and railway operating conditions shift fault characteristic frequencies, thereby reducing the cross-domain generalization capability of diagnostic models. To address this problem, this paper proposes a method that integrates envelope-spectrum-based spectrum normalization (SPNE) with a Fourier neural operator (FNO). Based on bearing geometry and the candidate fault type, SPNE maps the envelope spectrum to a unified dimensionless coordinate, aligning the fundamental and harmonic peaks of the same fault type across domains. The global receptive field of the FNO is then used to extract features from the normalized spectrum and identify faults. The method is evaluated on the PU and CWRU datasets, field railway measurements, and axlebox bearing test-rig data. Under strong noise (−5 dB), the proposed method achieves accuracies of 97.82% and 98.83% on the PU and CWRU datasets, respectively. It achieves 99.88% accuracy on the field railway measurements. These results indicate that the method maintains high diagnostic accuracy across different noise levels and datasets.

axlebox bearing,field railway measurements,spectrum normalization,Fourier neural operator,cross-domain fault diagnosis