Neural Network Vehicle Modelling for Underwater Vehicles

Underwater assets such as wind turbine foundations and other infrastructure require regular inspections. These inspections are currently performed by extensively trained divers or pilots of remotely operated underwater vehicles. Compact unmanned underwater vehicles could perform many of these tasks, and myriad others, with reduced risk and cost. However, navigation and position estimation is difficult under water, especially due to the absence of satellite navigation signals. Therefore, the fusion of information from sensors into a single position estimate is a common approach. These approaches often suffer from drift in the position estimate, which can accumulate during sub-surface operation. Furthermore, critical sensors can be unavailable in certain segments of the operation. In addition, information from the vehicle’s actuators is often disregarded, while these signals could improve position estimation. In this work, the focus is therefore on the inclusion of actuator signals for modelling vehicle behaviour. In particular, the use of neural networks for modelling the vehicle dynamics is explored in a machine learning approach, using real-world data. It is shown that these neural network models are able to capture the relevant dynamics in these data. The feedforward neural network severely limited the effects of position drift when used as plant model in a Kalman filter together with IMU measurements. The presented approach is therefore a promising method for mitigating the error due to unavailability of critical sensors in a real-world setting