<p>Gearbox fault diagnosis is critical for ensuring the reliability and safety of industrial rotating machinery operating under variable speed and load conditions. Traditional vibration-based diagnostic methods rely on handcrafted features and often suffer from reduced robustness when operating conditions change. This study proposes a unified hybrid framework that integrates physics-informed signal processing with a Cross-Scale Time–Frequency Transformer (CS-TFT) for robust gearbox fault diagnosis. Experiments are conducted on the PHM Society 2009 gearbox dataset, which comprises 560 multi-sensor vibration runs collected across five shaft speeds (30–50&#xa0;Hz), two load levels, and multiple gear, bearing, and shaft fault types with multi-label annotations. Raw vibration signals are processed using tachometer-guided segmentation, time-synchronous averaging, and multi-resolution time–frequency representations, including the continuous wavelet transform, spectral kurtosis, and variational mode decomposition. The proposed CS-TFT is pre-trained using contrastive self-supervised learning and further enhanced with adversarial domain adaptation to improve generalization across unseen operating conditions. Performance is evaluated under a strict cross-speed testing protocol, where training and testing speeds are disjoint. Compared with classical machine learning models (Random Forest, SVM, XGBoost) and deep learning baselines (1D-CNN and CNN–LSTM), the proposed framework achieves a 40–60% reduction in Hamming loss and consistently higher macro- and micro-F1 scores on unseen speeds. Visualization and attention analyses further demonstrate that the model focuses on physically meaningful fault-related frequency components. These results indicate that the proposed hybrid transformer-based approach provides a robust and generalizable solution for practical gearbox condition monitoring.</p>