Enhancing the Control System of a Reconfigurable Underwater Vehicle Through Dynamic Parameters Identification

The development of autonomous underwater vehicles (AUVs) has recently experienced exponential growth in both mechanical design and control systems. Traditional control strategies based lonely on Proportional-Integral-Derivative (PID) controllers are becoming less prevalent, making way for control laws that leverage mathematical models to describe the system dynamics more accurately. This paper presents a method for estimating the dynamic parameters of an AUV using specialized neural networks known as Physics-Informed Neural Networks (PINNs). Unlike conventional neural networks, PINNs incorporate the physical model of the system directly into the training process. The application of this approach has demonstrated significant efficiency in estimating the added mass and drag coefficients of the vehicle under study, achieving an average estimation error of 1.42% for added mass and 3.02% for drag coefficients. Finally, to highlight the practical benefits of these parameter estimations, a feedforward control system was implemented on the vehicle. The results illustrate the substantial improvements in the vehicle's dynamic behaviour, emphasizing the advantages of integrating such estimations into control strategies.