PREDICTING RESIDUARY RESISTANCE FOR A YACHT USING HYBRID LEARNING ALGORITHMS

Abstract

With the use of modern analysis tools, naval architect and ship design industry is required to adopt innovative and non-conventional approaches for more efficient results. However, the ship design remains a complex area of research and implementation primarily due to the unpredictable dynamics of oceanic waters. Therefore, new ships design must be approached in a systematic manner. In case there are no prototypes available for any defined vessel, then the designer must ‘get it right’ over the course of three distinct phases – conceptual, contractual and detail design. In this thesis one such parameter, known as Residuary Resistance per unit weight of displacement, which is directly related to ship’s propulsive power and speed calculations, is evaluated using the Hybrid Neural Network Algorithms. The models used for the same are ANFIS and ELANFIS. Earlier research on this subject has focused on conventional artificial neural networks wherein a Multi-layer Perceptron (MLP) regression was trained. The dataset used in the same is a Yacht Hydrodynamics dataset, which has been taken from UCI Machine Learning Repository. All parameters have been formulated from the ships’ characteristics and are non-dimensional. The results are validated with root mean square error (RMSE) and comparative analysis of both algorithms is presented. Results achieved by ANFIS were found to be encouraging compared to MLP. However, in case of ANFIS, the training time observed to be large and generalization performance was observed to be moderate. To overcome the same, ELANFIS was trained and both training time and generalisation performance were found to be improved.