Foundation models promise reusable representations for industrial fault diagnosis, but heterogeneous machinery, sampling rates, and sensing configurations make cross-source signals difficult to align. Index-space preprocessing may assign the same token coordinate to different physical durations or frequencies. We observe that global fast Fourier transforms of equalduration windows share the same frequency resolution across native sampling rates, enabling overlapping bins to represent consistent physical frequencies without resampling. Based on this insight, we propose UniFreq, a physically consistent dualbranch foundation model. Its time branch encodes equal-duration patches, while its frequency branch models the normalized global spectrum on a common physical-frequency grid. Time– Frequency Alignment connects the paired views, Amplitude Scaling Contrastive learning reduces dependence on acquisitionspecific magnitude, and channel-independent encoding accommodates varying channel counts. We evaluate UniFreq on a 15-dataset benchmark covering bearing, gearbox, and motor systems, with ten unlabeled pretraining datasets and five disjoint evaluation datasets forming seven downstream tasks. UniFreqBase achieves the highest average accuracy under all three labelefficient fine-tuning settings, while the 431K-parameter UniFreqTiny also surpasses the strongest baseline. Ablations further validate the global FFT and both self-supervised objectives.

Fault diagnosis,foundation model,heterogeneous industrial signals,self-supervised learning