WT-STCA-ViT: Window-transform synergistic spectro-temporal channel attention Vision Transformer for robust bearing fault diagnosis

Abstract Reliable bearing fault diagnosis under variable operating conditions and severe noise remains challenging because empirically selected time-frequency front ends may produce degraded representations, while generic attention mechanisms lack explicit sensitivity to fault-related spectral structures. To address these limitations, this study proposes a robust cascaded framework termed window-transform synergistic spectro-temporal channel attention Vision Transformer (WT-STCA-ViT), which integrates a window-enhanced Fourier-based synchrosqueezing transform (FSST) with a spectro-temporal channel attention Vision Transformer (STCA-ViT). Specifically, a parameterized $\\alpha {\\rm{ - exponential Kaiser - Blackman}}$ $(\\alpha {\\rm{ - EKB}})$ window family and a task-constrained parameter screening strategy are established to optimize spectral concentration and transient localization, mitigating energy spreading across noisy scenarios. The resulting high-resolution time-frequency maps are encoded using a pretrained Vision Transformer (ViT) backbone and refined via the proposed frequency-axis-structure-aware spectro-temporal channel attention (STCA) mechanism, which generates frequency-dependent channel weights to calibrate band-level features while suppressing noise-dominated and irrelevant feature responses. Furthermore, a lightweight post-Transformer encoder and LoRA adapters are coupled to achieve global token-level refinement and parameter-efficient optimization. Extensive benchmarks on the public CWRU dataset and a self-built cylindrical roller bearing dataset demonstrate competitive diagnostic accuracy, improved weak-fault recognition, and stable feature separability, with computational cost analysis further confirming a favorable trade-off among diagnostic robustness, trainable-parameter efficiency, and potential for low-latency online deployment.