A dynamic sparse transformer for fault detection of ship propulsion shaft system under noisy environments

Fault detection of ship propulsion shaft systems (SPSS) under noisy environments (FDUNE) is challenging because vibration signals are non-stationary and contaminated by noise. To address this challenge, we propose DSFormer, a dynamic sparse transformer integrating cross-shaped window self-attention (CSWSA) and a dynamic sparse (DynaSparse) block. CSWSA restricts attention to orthogonal stripe subspaces to prevent global noise diffusion. DynaSparse uses dual-path latent tokens and composite sparsity regularization to compact salient diagnostic features and suppress noise-dominated regions. On an SPSS dataset, DSFormer achieves 100% accuracy and F1 score under noise-free conditions. Under Gaussian noise, it maintains 94.37% accuracy at -4 dB and 99.95% at 10 dB; under impulsive noise, it reaches 99.95% accuracy and 99.94% F1 score, degrading by only 0.05% and 0.06%. In ten cross-SNR domain adaptation scenarios, it averages 99.18% accuracy and reaches 98.20% in the hardest 0→-4 dB case, exceeding the best baseline by 3.25 points. It has 11.8 M parameters, 2.9 G FLOPs, and 2.2 ms per-sample inference latency. These results show DSFormer provides accurate, robust, and efficient FDUNE for SPSS and other rotating machinery. Further validations on bearing and gearbox benchmarks confirm its versatility, establishing DSFormer as a robust and efficient solution for noisy‑environment fault diagnosis in marine propulsion systems.