Physics-Informed Machine Learning for System Identification of an Autonomous Underwater Vehicle

Sammanfattning

Accurate dynamic modeling is crucial for the effective control and navigation of any Autonomous Underwater Vehicle (AUV). However, the identification of hydrodynamic parameters via first principles or expensive experimentation remains an open challenge. Machine learning techniques, however, offer an alternative to these via regressing the unknowns given sufficient data. Thus, this work investigates the use of Physics-Informed Machine Learning (PIML) methods, specifically a Physics- Informed Neural Network (PINN) model, for the purpose of system identification of the hydrodynamic damping matrix ⃗D of the Fossen model in an applied manner to Swedish Maritime and Robotics Center (SMaRC)s AUV Small & Affordable Maritime Robot (SAM). With data collected from a newly constructed water tank with integrated motion capture system the proposed models incorporate physical constraints and information to further extend standard Machine Learning (ML) methods. Three different models were developed, these being: a PINN model, a standard Neural Network (NN) model without physical constraints, and a physics-free or ”naive NN”. Each model has been compared against each other and the existing Fossen or ”white-box” dynamics model of SAM in the task of trajectory prediction, which has applications in state estimation, control and motion primitive generation for path planning. Results show the potential of physics-informed approaches in terms of prediction accuracy particularly when it comes to angular velocities. Though both the naive NN model and PINN model showed improvements in the heading of the vehicle, although only the naive NN was able to get accurate predictions for the translational speeds, but failing to predict the roll of the vehicle. Our results are however hindered by the inaccuracy of the existing white-box model used to embed the physics knowledge into the PINN model, whose empirical evaluation has also been part of this thesis.