A Neural Network-Based Sensitivity Analysis Approach for Data-Driven Modeling of Ship Motion

Abstract

Researchers have been investigating data-driven modeling as a key way to achieve ship intelligence for years. This paper presents a novel data analysis approach to data-driven modeling of ship motion. We propose a global sensitivity analysis (GSA) approach combining artificial neural network (ANN) and sparse polynomial chaos expansion (SPCE) techniques to accommodate high-dimensional sensor data collected from ship motion. An ANN is constructed as a surrogate model to associate ship sensor data with a certain type of ship motion. To account for the computational efficiency of GSA, an SPCE is integrated into the GSA to decrease the need for Monte Carlo (MC) samples generated by the ANN. A probe variable is designed to couple with the MC samples, which plays a role in determining the degree of convergence of variable importance. A test on benchmark function demonstrates the efficiency and accuracy of the proposed approach. A case study of ship heading with and without environment effects is conducted. The experimental results show that the proposed approach can identify and rank the most sensitive factors of ship motion. The proposed approach highlights the application of GSA in data-driven modeling for ship intelligence.