Speed and load changes in rail transit bearings cause pronounced vibration-distribution shifts, while fault labels under new conditions are difficult to obtain, limiting cross-condition adaptation. This paper proposes a time-frequency contrastive prototype alignment network. Each vibration window forms two views from its time-domain segment and spectral representation. Supervised contrastive learning enhances source-feature discrimination, while source-label-only class prototypes and entropy confidence align target features. Domain adversarial learning and time-frequency consistency further constrain cross-domain discrepancy and target cross-view representations. Evaluated on six CWRU and SEU transfer tasks under transductive checkpoint selection using target-validation labels, the network achieves 97.06% equal-weight macro-average accuracy; removing both constraints lowers Avg. by 3.09 percentage points. These results show that class-prototype-based, time-frequency-consistent cross-domain representation learning improves recognition across multiple speed-load combinations and offers a practical approach to intelligent rail transit bearing monitoring.