A hybrid fault diagnosis framework for rudder propellers: Physics-guided learning from data and simulation

Rudder propellers, as fully azimuthing thrusters, integrate propulsion and steering functions, playing a vital role in marine vessel performance and safety. However, their large size and complex installation limit the acquisition of fault data through physical experiments, posing challenges for intelligent fault diagnosis. To address this, a hybrid fault diagnosis framework is proposed, driven by both measured data and simulation models. A high-fidelity multibody dynamics model is developed with boundary conditions calibrated via CFD, enabling the generation of fault-labeled synthetic vibration signals. A dual-branch generative adversarial network (GAN), incorporating both time-domain and time–frequency features, is used to adapt simulated signals to real-world measurement conditions. The generated samples achieve high fidelity, with frequency-domain cosine similarity exceeding 0.88. A physics-informed neural network (PINN) with a matching dual-branch structure is then constructed for fault classification. Experimental results on a rudder propeller test platform show that the proposed method achieves up to 100 % classification accuracy. Compared to existing diagnostic approaches, this framework offers superior accuracy and robustness by effectively integrating physical modeling, adversarial domain adaptation, and physics-guided learning. • The main highlights of this work are as follows: • Built an f-CFD coupled simulator to generate high-fidelity vibration signals for gear/bearing faults across conditions. • Proposed SDDWGAN dual-branch model (waveform + wavelet packet) for sim-to-measurement adaptation; cosine similarity >0.88. • Developed a dual-branch PINN with characteristic-frequency constrains; up to 100% accuracy and noise robustness.