System Identiﬁcation of Ship Models using Deep Learning[go to overview]
Objective of this thesis is the system identiﬁcation of ships for multistep prediction, i.e. simulation, with deep learning methods. First-principles modeling of ships is a challenging and expensive task, as it requires complex numerical computations, model tests, sea trials, and expert knowledge in marine engineering. The collection of sensor data during the routine operation of ships enables system identiﬁcation methods for deriving mathematical models of the vessel dynamics.
While traditionally system identiﬁcation requires data collected during maneuvering tests to accurately identify the system with linear models, deep learning is able to extract patterns from large and noisy datasets and thereby effectively model ship dynamics from the aforementioned collection efforts. In this thesis, a ship model including an open-loop control system to simulate human inputs and environmental disturbances in the form of wind and wind-induced waves is derived from ﬁrst-principles for simulation. The simulation allows the generation of sensor data for varying sea states and maneuvers. Linear time-invariant models (LTI) and recurrent neural networks (RNN) are evaluated on the multistep prediction problem to analyze the beneﬁt of deep neural networks in the system identiﬁcation task. Unlike related work, a focus is set on identifying a system affected by signiﬁcant environmental disturbances. A novel loss function, mean squared gradient error (MSGE), is proposed and shown to outperform the mean-squared error (MSE) w.r.t. predicting roll motion of the vessel. In addition to common evaluation metrics, domain-speciﬁc criteria, e.g. maneuvering characteristics and frequency analysis, are employed to validate the physical plausibility of the identiﬁed model. Additionally, it is shown that RNNs can identify physically plausible models and outperform LTI models w.r.t. to prediction accurac
09.05.19 - 10:15