SpectroSwinFusion: Multi-Scale Time-Frequency Transformer with Adaptive Spatiotemporal Fusion for fault diagnosis

Fault diagnosis remains a critical research focus for industrial equipment. However, the performance of existing fault diagnosis methods faces the dual challenges of data imbalance and large parameter volumes. To address this, this paper proposes a novel method named the Multi-Scale Time-Frequency Transformer with Adaptive Spatiotemporal Fusion (SpectroSwinFusion). Firstly,the new signal preprocessing module is designed to intelligently convert time-series signals into structured 2D data by integrating Continuous Wavelet Transform (CWT), dynamically parameterised Short-Time Fourier Transform (STFT), and a novel Hilbert Envelope 2D Transform. Secondly, the Cyclic-Shift Window Transformer Block (CSwin-TB) achieves global modelling of local features through spatially aligned residual connections and cyclic-shift operations, while simultaneously reducing computational costs via channel-compressed attention and bias-free positional encoding. Finally, the Spatiotemporal Adaptive Fusion module captures spatial topological dependencies through pseudo-time-series reconstruction and Bidirectional LSTM, followed by feature calibration using lightweight attentional gating. The proposed method was evaluated on two bearing datasets under three imbalance ratios, consistently achieving recognition rates exceeding 98.67%. Ablation studies and comparisons with state-of-the-art methods demonstrate its superior recognition accuracy with significantly reduced parameters.