Quantifying Multicollinearity in Ship Manoeuvring Modeling by Variance Inflation Factor

Parameter drift is a tricky problem in system identification, and collinearity in the identified model is considered to be its cause in the field of statistical learning. To diminish the parameter drift and identify the model parameters more accurately, a better understanding of the characteristics of collinearity is necessary. System identification is one of the effective modeling methods in the study of ship manoeuvrability. This paper aims at quantifying and analyzing the collinearity in the modeling of ship manoeuvring motion by using Variance Inflation Factor (VIF). By utilizing the multiple datasets including those from zigzag and turning tests and the combinations of manoeuvres, as well as the data processed by D-optimizing design or difference method, the VIF is applied to quantify the severity of collinearity of different model structures. The results show that the selected manoeuvring models have high collinearity under the data of single standard manoeuvre. The pre-processed test data and an appropriate model structure can alleviate the collinearity, hence to diminish the parameter drift. However, the collinearity of the selected models cannot be eliminated. Some suggestions are given for selecting more appropriate training data and mathematical model structures of ship manoeuvring to accurately estimate the model parameters.