Data-based Wave Filtering Using Neural Networks with Application to Dynamic Positioning

Abstract

Ships are constantly affected by environmental disturbances. Oscillating wave forces are one of these disturbances. In dynamic positioning applications, these wave disturbances lead to an oscillating motion of the ship about its mean position. Even though the overall position is not affected by the oscillating disturbance, the actuators of the ship will try to compensate for it. This behavior is unwanted since it leads to unnecessary wear of the actuators and high fuel consumption. A solution is to apply wave filtering techniques, which eliminate the oscillating motion induced by waves from the measured position. This enables control of the low-frequency motion and reduces oscillation of the actuators.

This study evaluates if it is possible to learn a wave filtering algorithm to predict the low-frequency motion without knowledge about system parameters. For this approach, available measurement and control data is utilized to extract the information needed. A predictor based on a recurrent neural network is proposed. The recurrent neural network is trained on a sliding window to learn the model adaptively. The required training data is created from past measured position data. Creating the training data is done by estimating the dominating frequency of the wave disturbance and applying noncausal filtering. The recurrent neural network learns to predict the low-frequency motion using measured positions, control inputs of the ship, and past predicted low-frequency positions as inputs to the neural network. After training, the low-frequency position is predicted by the predictor until the neural network is trained again. The training is performed regularly to enable convergence towards a wave-filtering model and to adapt to changing disturbances.

The performance is evaluated through simulations with increasing complexity. The available data can contains enough information to learn wave filtering. Simulations on a system with one degree of freedom indicate that a wave-filtering model can be learned if the system is excited. Increasing the complexity to three degrees of freedom shows that the predictor can be extended to predict the lowfrequency motion of more complex systems. However, the predictor struggles to learn rotational dependencies. Initial results from applying the predictor on data from a ship simulator are promising but need further work.