The past decade the packing approach was developed at the TU Delft. This tool automatically generates tens of thousands of ship designs in order to fully explore the design space. This is opposed to the traditional way of ship design, where just a limited number of designs can be
...
The past decade the packing approach was developed at the TU Delft. This tool automatically generates tens of thousands of ship designs in order to fully explore the design space. This is opposed to the traditional way of ship design, where just a limited number of designs can be elaborated. The result is a set of ship designs that can be used both to obtain initial sizing parameters of the ship, and to relate the impact of design decisions on performance characteristics.
It is questioned whether resulting sets of designs really contain tens of thousands of different ships, or just a couple of really different ships that have only minor variations. It is thus questioned whether such a set of ship designs can be divided into families. This is investigated in this thesis.
In order to attack this problem in a generic fashion, the ship designs were approached from a numerical perspective. It was found that in the field of data science, clustering algorithms exist which are devoted to find clustering structures in data. Therefore these techniques, such as PCA and k-means, were applied to data at hand. In a test case this method is applied to divide the set of designs from a mine countermeasures vessel into families.
First it was questioned whether families could be used in order to assess the survivability of machine systems of these ships. The resulting families matched the families regarding the position of the gun, which were already known upfront. Although these families are not sufficient to fully assess the survivability of these ships, this analysis showed that most probably no other structure is present, stimulating more straightforward definitions of families.
In a second test case the same MCMV was divided into families regarding the layout of the designs. It then appeared that the dataset at hand was built by ten distinct runs of the packing approach which were combined to this one total set. The method showed that it could be pointed out, just by looking at the designs, which design was generated in which run. This showed that the dataset might not be as diverse as it seems. Since it is the core idea for the packing approach to explore a big part of the design space, thus generating a diverse set of ship designs, this motivates new research to tackle this issue.