Study on Marine Propeller Design Using Artificial Neural Networks and Multivariate Analysis

The designers of marine propellers usually encounter the problems of characteristics studies of propellers. These characteristics would include the relationships among design variables and performances, matching between propeller and ship hull, and choosing a compromise design solution from many alternatives. In this work, the study of these characteristics is presented based on artificial neural networks and multivariate analysis. First, the propeller design solutions are generated from design of experiment (DOE) approach using B-series charts for a given effective power-speed curve of ship, while taking geometry, cavitation constraints into consideration. The geometrical parameters and performance of propellers are included in one high-dimensional dataset. Then, the characteristics of this dataset are investigated by a famous artificial neural network, Self-Organizing Maps (SOM) and several multivariate methods. The relationships among design variable and performance are studied by correlation analysis and SOM graphs. The effects of geometrical parameters upon open water efficiency are found by main effect analysis. The clustering of solutions on SOM graphs are obtained through partitive clustering approach according to DB index. The significance of variables in these clusters is also studied. A parallel coordinate plot with cluster information also provides decision help to designers. The relationships among design variables are examined by hierarchical clustering then. In order to find the intrinsic structure of this high-dimensional dataset, a well-known dimensionality reduction skill, Isomap is utilized. The relationships among these variables are shown in 2D plots. Finally, a TOPSIS (Technique for Order Preference by Similarity to Ideal Solution) procedure and entropy weight method are employed for finding final compromise solution from this dataset. A numerical example is provided in this work to illustrate the analysis process. The results show that the SOM and multivariate approaches are effective and they can mine useful information from dataset for propeller designers.