A SVM-based Sensitivity Analysis Approach for Data-Driven Modeling of Ship Motion

Abstract

This paper presents a novel method that combines support vector machine (SVM) with sensitivity analysis (SA) to analyze sensor data for ship motion modeling. In order to investigate how each model input contributes to the model's output, the PAWN method which is based on cumulative distribution function (CDF) is used. Considering the limitation of the PAWN method that it cannot be applied to sensor data directly, a surrogate model using SVM is integrated into PAWN method as it has a better solution of non-linearity and high-dimension problem as well as excellent generalization capability. Implementation of whole systematic workflow is elaborated in detail and related experiments are made such as comparison of four distance metric methods, benchmark test and SA of vessel data. The results show the proposed method can be used for analyzing vessel data to seek prime parameters that affect ship motion.

A sensitivity quantification approach to significance analysis of thrusters in dynamic positioning operations

“…One advantage of PAWN is that it hugely reduces computational cost because there is no need to compute unknown parameters for the approximation of empirical CDF. Another In this study, based on our previous experience [16], the PAWN method is adopted to conduct an SA of thrusters. In addition, we make some modifications and improvements according to features of DP data.…”

“…In this study, positioning capability refers to how well the DP vessel is positioned, instead of the extremity of the environmental conditions the vessel can counteract, as underlined by [11]. Based on prior studies and our SA method [16], positioning capability is quantified by time-series ship parameters such as ship position, heading, and consumed power. Some aforementioned statistics of time-domain data to analyze the DP capability of offshore vessels were accepted and adopted.…”

“…For PAWN, is set to 10 based on the samples of data as well 444 as experience as described in other papers [16,31].…”

“…374(BP), regularized extreme learning machine (RELM), and 375 SVM, are introduced into training models[16,33,34]. Next, 376 PAWN combined with these three models is used to compute 377 sensitivity indices of three parameters of Ishigami function.378Finally, SA results are compared with a benchmark to iden-379 tify the optimal ML method for analyzing the significance of…”