Physics-guided bidirectional synergistic fusion network for stern bearing fault diagnosis in ship propulsion systems

Stern bearing signals in ship propulsion shafting systems exhibit nonstationary modulation and distribution variations under harsh marine conditions. However, existing methods still struggle to construct robust physical priors and effectively integrate physical and data-driven features. A Physics-Guided Bidirectional Synergistic Fusion Network (PGSF-Net) is proposed for robust fault diagnosis of stern bearings in ship propulsion systems. A collaborative framework with physical-prior and data-driven branches is first constructed to alleviate feature misalignment under variable-speed conditions, providing an angular-domain representation basis and improving feature consistency by integrating modal-dominant physical priors with computed order tracking. Then, a bidirectional synergistic fusion mechanism is introduced to alleviate interference from strong marine background noise, which helps robust cross-branch feature learning and improves the representation of weak fault features. A hybrid local–global feature encoder is further designed to handle shifts in feature distribution caused by varying marine operating conditions, which can capture both multi-scale local transient impulses and long-range order dependencies by integrating multi-scale attention residual blocks with a position-adaptive global Transformer encoder. Finally, validated on the public KAIST and a private dataset, the proposed PGSF-Net achieves superior diagnostic performance. Our work shows robustness under varying speed conditions and unexpected operational scenarios, making it suitable for real-world maritime applications.