MLP-Based Condition Assessment Method for Marine Propulsion Shafting Driven by Measured and Digital Twin Data Fusion

Marine propulsion shafting operates under complex, variable conditions, where stern bearing wear and pedestal settlement redistribute loads, migrate vibration energy, and distort the shaft orbit, threatening safe operation. To address insufficient fault samples, difficulty in measuring degradation states directly, and limited interpretability of purely data-driven models, we propose a multilayer perceptron (MLP) condition assessment method driven by measured and digital twin data fusion. A digital twin framework comprising geometric, physical, rule-based, and behavioral models was constructed. A Timoshenko beam finite element model and modal truncation reduce computational cost while describing displacement-induced load and vibration responses. Static and dynamic tests are performed on a shafting test rig, and condition matching, outlier processing, and feature-level fusion produce a dual-source dataset. The MLP takes rotational speed, support forces of the four bearings, and normalized temporal position as input features; temperature scaling calibrates the class probabilities, while a separate severity score indicates the condition trend. The reduced model maintains first three natural-frequency errors below 10% and representative bearing-load errors below 5%. On an independent temporal hold-out test set, the MLP achieves 96.78% accuracy and 87.96% fault recall. Within the tested laboratory-scale, settlement-dominated operating range, these results demonstrate an effective and physically interpretable approach for condition assessment and risk early warning of marine propulsion shafting.