Development of a Small-Scale Artificial Intelligence (A.I.) module to Generate Primary Computer Aided Drawing for the Initial Design Phases of the Machinery in the Shipbuilding Industry

Abstract

A design fixation problem happens when the quality of a naval architect’s designs degrades because of extensive working pressure. The research idea is inspired by the real-life stories of many naval architects. The research intends to find a solution that will provide the basic designs of marine vessels for the designers working in maritime industries. An artificially intelligent (AI) non-human naval designer – the main theme of this research, was therefore selected. During the entire research, two questions are searched for answers – the ‘necessity?’ and the ‘impacts?’ of the design optimization. The first one is found and the other is partially set with indicators that might provide answers soon when the project can be implemented on a larger scale. Building a fully operational AI-originated naval designing tool is a large workshop, so a small part of the project has been implemented throughout the current research, only limited to three components: Engine room, main engine, and generator. Basic designs of a ship start with the 2D general arrangement plan – where mainly the topology of the different design elements is reorganized as per the IACS codes and the shipowner’s opinions. The 2D topology optimization for machinery items requires datasets from the on-service ships. Gathering technical information for dataset preparation, feeding datasets into the machine learning platforms (‘Artificial Neural Networks [ANN]’ and ‘Generative Adversarial Networks [GAN]’), and developing codes and scripts in programming platforms (MATLAB and Google Colab) are the main working phases for topology optimization. Result validation with third-party software is one of the key points for the quality assurance (Q.A.) of the designs predicted by the machine. Therefore, the ‘Conditional Tabular Generative Adversarial Network (CTGAN)’ has been used for Q.A. Result findings are analyzed in terms of regression-based statistical methods, confusion matrices, and arithmetical methods. A smaller dataset of 152 real marine vessels, a medium augmented dataset of 750 virtual ships, and a larger augmented dataset of 6996 virtual ships have been used for the training of the novel AI model. Augmented datasets are developed with the published parametric relationships among ship’s particulars. Synthetic ship data generated by the novel AI model contains 150 samples for GAN and 10000 samples for CTGAN module per training session consisting of 500 epochs. There is no synthetic ship limitation for the ANN module. In the first and unsupervised training session, the rate of success of synthetic ship generation is nearly 0% for all three modules, though the convergence of the engine and generator inside the engine room shows a partial fulfillment of the basic ship design criteria. After the careful debugging procedures are applied, in the second and supervised training session, the rate of success has increased to 15% for the ANN, 0.06% for the GAN, and 0.15% for the CTGAN module. These success rates are calculated using the arithmetical method. Besides, the confusion matrices have also shown the primary version of the AI modules is in the presumed path of development. Limitations for enlarged research might have effects on the obtained results. A few numbers of previous research works, smaller datasets, design feature loss in CAD-to-Text data conversion, unsolved hidden bugs in the codes, and duration of human involvement in supervised machine learning are the main limitations of the current research. The unavailability of the survey data from the naval designers, clients, and shipbuilders for a similar product from the previous market history is one of the most important reasons for not obtaining the specific answer to the second research question. However, 85% of shipyard owners think of AI as the pioneering factor in shipdesigning in the future, and in addition, AI might reduce 25% designing time as well, according to a statistics report mentioned in the last chapter. Nevertheless, the ANN module unveiled a positive opportunity to carry on with the current research; probably on a larger scale with a view to the ship’s 3D topology optimization, and so a tentative future insight has been recommended by the author at the end of the research.