Abstract
The increasing threat of marine plastics and other debris to wildlife and ecosystems has underscored the need for efficient marine cleanup solutions. Traditional methods, often manual and risky, are inefficient over large areas. In response, water-cleaning Unmanned Surface Vehicles (USVs) like WasteShark have emerged as promising alternatives, capable of safely and effectively collecting floating trash in hazardous areas.
This work focuses on adapting Model Predictive Control (MPC) to the dynamic changes caused by trash accumulation within a USV's hull. We developed parameter estimation methods, including Moving Horizon Estimator (MHE) and Extended Kalman Filter (EKF), to track these hydrodynamic changes. Their performance was evaluated through extensive simulations and tested with real data from WasteShark operations in a pool. MHE demonstrated robustness to measurement noise but proved computationally expensive.
An MHE-based MPC framework was proposed where MHE continuously estimates hydrodynamic parameters online and updates the MPC's prediction model. Although MHE effectively tracked changes in linear drag, it required sufficient input excitation to accurately estimate added mass. However, MHE-based MPC did not enhance path-following accuracy. The path-following performance of MPC degraded with the addition of a small surface current, a finding validated by field experiments conducted at Bergse Voorplas lake in Rotterdam under moderate wind conditions. The experiments showed that currents induced by wind had a more dominant effect than trash accumulation inside the hull.
Additionally, model-free controller approach, which does not explicitly involve parameter estimation, was also proposed. This approach involved developing various controllers and comparing them against MPC through simulations. Compared to MPC, this approach demonstrated better path-following performance under external disturbances due to surface currents. This highlights that MPC requires an accurate dynamic model to perform well, and disturbances due to currents were not modeled within the MPC prediction framework.
