합성곱 신경망과 복셀화를 활용한 선박 저항 성능 예측 = CNN-based ship resistance prediction using voxelization

다국어 초록 (Multilingual Abstract)

The resistance of a ship can be analyzed using computational fluid dynamics (CFD) or model tests. To explore a number of design candidates, for instance in the early design stages, it is too expensive to use the CFD and model tests because of their relatively long analysis time. Ship designers tend to often use statistical methods that are simple and need a short analysis time. The statistical methods provide such advantages, but are often relatively inaccurate due to their simplicity. To deal with the problem, we present a method for predicting ship resistance that is based on convolutional neural networks (CNNs). This converts input hulls into 3D voxels, which are the suitable data structure to use the CNNs. The CNNs extract only important features from the input hulls and this often allows for better convergence in the training of artificial neural networks (ANNs). In a case study, the proposed method was applied to developing ANNs for ship resistance prediction. It was compared with a parametric method which is also an ANNs, but the input of the ANNs used hull parameters such as the length overall, block coefficient, etc. The results of the case study show that the voxelized input improves the resistance prediction accuracy compared with the parametric input in developing ANNs for ship resistance prediction.

목차 (Table of Contents)

• 1. 서론 1
• 1.1 연구 배경 1
• 1.2 선행 연구 1
• 1.3 연구 목표 및 개요 3
• 2. 딥러닝 모델 5
• 2.1 딥러닝 개요 5
• 2.2 합성곱 신경망 8
• 2.2 임베딩 10
• 2.3 완전연결층 12
• 3. 데이터 구성 13
• 3.1 데이터 개요 13
• 3.2 선형 데이터의 복셀화 13
• 3.3 부가 정보 데이터 16
• 3.4 저항 데이터 17
• 4. 비교 시험 20
• 4.1 입·출력 데이터 수집 및 결정 21
• 4.1.1 데이터 수집 21
• 4.1.2 딥러닝 모델을 위한 데이터 결정 25
• 4.2 딥러닝 모델 결정 30
• 4.3 방법론 비교 34
• 5. 결론 38
• 5.1 요약 38
• 5.2 향후 과제 39