Abstract
The safe operation of ships requires the condition of propulsion components to be maintained. Digital twins are a promising alternative for intelligent monitoring of these complex systems. Digital twins require models which ensure that the digital representation is able to mimic the behavior of the physical system. Alternate modeling solutions must be found when intellectual property restrictions or lack of available information limit the usability of physics-based models. This paper considers such a case where a system model of the propulsion system requires a real-time capable model of the propeller hydrodynamic torque. The creation of a data-driven hydrodynamic torque model based on full-scale, operational measurements is discussed. The described method focuses on the significant challenges associated with data cleaning and preparation while also evaluating whether well-known machine learning methods are suited for this application. The methods use speed-over-ground, heading, course, rotational speed, and propeller pitch as inputs. The outputs of the models are the single quadrant propeller torque coefficient and the amplitude of harmonic torsional excitation. These outputs are then combined to create a holistic prediction of the torque. Results indicate that both a polynomial least-squares fit and a shallow neural network predict the mean and the amplitude of harmonic components of the torque well. This prediction can be used to isolate the hydrodynamic torque when more than one torque source is present or to simulate what-if scenarios in a digital twin environment.