Raine Viitala

This work presents a physics-informed learning method for torsional vibration estimation. A discrete-time torsional vibration model is used to constrain the learning process. The proposed estimator learns an output-to-input inverse mapping from measured shaft rotating speed and torque along with motor torque to propeller excitation, after which the full drivetrain state is reconstructed by forward simulation. The reconstructed state is filtered with a discrete-time Kalman filter to consider minimal sensing requirements. The method was evaluated with a supervised training objective and an indirect training objective, where examples of the target quantity were unavailable even within the training samples. Experiments are conducted on a laboratory thruster test bench under steady-state, impulse, and ice excitations. The results suggest that physics-informed learning and filtering can regularize structurally ill-posed drivetrain estimation problems without changing the problem formulation. The supervised formulation achieved accurate estimates with one realistically available torque transducer measurement point. The reconstruction-based formulation also produced usable estimates in selected configurations, but showed stronger dependence on sensor availability.

This work presents a physics-informed learning method for torsional vibration estimation. A discrete-time torsional vibration model is used to constrain the learning process. The proposed estimator learns an output-to-input inverse mapping from measured shaft rotating speed and torque along with motor torque to propeller excitation, after which the full drivetrain state is reconstructed by forward simulation. The reconstructed state is filtered with a discrete-time Kalman filter to consider minimal sensing requirements. The method was evaluated with a supervised training objective and an indirect training objective, where examples of the target quantity were unavailable even within the training samples. Experiments are conducted on a laboratory thruster test bench under steady-state, impulse, and ice excitations. The results suggest that physics-informed learning and filtering can regularize structurally ill-posed drivetrain estimation problems without changing the problem formulation. The supervised formulation achieved accurate estimates with one realistically available torque transducer measurement point. The reconstruction-based formulation also produced usable estimates in selected configurations, but showed stronger dependence on sensor availability.

In this paper, a torque reconstruction method for electric motor driven powertrains is presented. Accurate knowledge of shaft torque in electric powertrains is essential for performance optimization, torsional vibration diagnostics, and optimal control. Direct measurement of shaft torque typically requires sensors at various points in the system, increasing costs in manufacturing and maintenance. Reducing the number of required sensors would be a valuable improvement, which makes various torque estimation methods an attractive alternative to physical sensors. This paper presents an estimation framework that leverages data on motor torque and shaft rotation velocity combined with a convex optimization algorithm to provide torque information for the entire powertrain. The method applies piecewise trend filtering to suppress measurement noise and reconstruct the powertrain's unknown inputs and states. A simulation study is performed on a powertrain to create an initial analysis of the optimization performance. These results indicate a potential method of shaft torque estimation as a robust and cost-effective solution for powertrain monitoring, control, and fault detection in shaftline systems.