A Machine Learning Framework to Model Extreme Events for Nonlinear Marine Dynamics

Extreme events such as large motions and excess loadings of marine systems can result in damage to the device or loss of life. Since the system is exposed to a random ocean environment, these extreme events need to be understood from a statistical perspective to design a safe system. However, analysis of extreme events is challenging because most marine systems operate in the nonlinear region, especially when extreme events occur, and observation of the extreme events is relatively rare for a proper design. Conducting high-fidelity simulations or experimental tests to observe such events is cost-prohibitive. In the current research, a novel framework is proposed to randomly generate test environments that lead to a large response of the system. With the generated environment, large responses that would take a very long time to achieve can be observed within a much shorter time window. The time-domain context around the extreme event provides the user with rich insights towards the improvement of the design. The proposed framework consists of two modules, which are named as Threshold Exceedance Generator (TEG) and Design Response Estimator (DRE). The framework is data-driven, and its application requires minimal knowledge about the system from the user. The DRE module can identify a nonlinear marine system based on collected data. The TEG module can generate ocean environments that lead to large system response based on the system identification by the DRE module. Machine learning methods, especially neural networks, are heavily used in the proposed framework. In the thesis, the extreme generation problem in the marine field is described and addressed from a machine-learning perspective. To validate the framework, marine examples including linear wave propagation, nonlinear wave propagation, nonlinear ship roll, tank sloshing, and a floating object in waves are explored. Examples from such a wide range show that the framework can be used for linear or nonlinear systems and Gaussian or non-Gaussian environments. The cost and the amount of data to apply the method are estimated and measured. The comparison between the results from the framework and Monte Carlo Simulation fully demonstrates the accuracy and feasibility of using the data-driven approach.PhDNaval Architecture & Marine EngineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/163268/1/wenzhe_1.pd