Rolling bearing fault diagnosis method based on fusion of STFT-statistical features and AL-SOA optimized bagging tree.

To address the non-stationarity of rolling bearing vibration signals and the interference under complex working conditions, this paper proposes a fault diagnosis method based on the fusion of short-time Fourier transform (STFT) and time-domain statistical features, optimized by an adaptive Lévy-seagull optimization algorithm (AL-SOA) for Bagging Tree (denoted as STSF-AL-SOA-BT). To simultaneously capture both the frequency-domain impulsive components and the time-domain waveform structure while controlling the input dimensionality, a lightweight dual-channel attention fusion module (AFF) is designed. The module adaptively weights and concatenates a 256-dimensional STFT spectrum with six time-domain statistical features, followed by dimensionality reduction via PCA to mitigate redundancy and overfitting. For optimizing the key structural parameters of the Bagging Tree, an AL-SOA strategy is developed, integrating Lévy flight and linear inertia weighting while maintaining an external elite archive to preserve the Pareto optimal set. This approach achieves a multi-objective balance among accuracy, model complexity, and training efficiency. The proposed method is systematically validated on the CWRU, SEU, and self-built SUT test rigs, achieving average accuracies of 98.88%, 98.50%, and 97.53%, respectively. Ablation studies demonstrate that both the attention mechanism and AL-SOA make significant contributions; the introduction of attention improves accuracy by 0.6-1.2% (p < 0.01), while the Pareto front analysis confirms an effective trade-off between accuracy and complexity. Overall, the proposed method balances discriminative capability, interpretability, and engineering deployability, providing a feasible solution for lightweight bearing fault diagnosis under variable operating conditions and strong noise interference.