오토인코더 기반 전이학습을 통한 서스펜션 모듈 주파수 응답 예측
Prediction of Suspension Module Frequency Response through AutoEncoder-Based Transfer Learning
- 주제어 (키워드) Automotive suspension modules , NVH performance , Frequency response , Artificial Neural Network , Deep learning , Transfer learning , AutoEncoder , Latent space
- 발행기관 서강대학교 일반대학원
- 지도교수 이승엽
- 발행년도 2024
- 학위수여년월 2024. 2
- 학위명 석사
- 학과 및 전공 일반대학원 기계공학과
- 실제URI http://www.dcollection.net/handler/sogang/000000076996
- UCI I804:11029-000000076996
- 본문언어 한국어
- 저작권 서강대학교 논문은 저작권 보호를 받습니다.
초록
In the face of modern challenges in the automotive industry, especially in suspension module design impacting NVH(Noise, Vibration, Harshness) characteristics, the traditional FEM(Finite Element Method) analyses are often resource-intensive and time-consuming. This study introduces a novel neural network model, AEILSS(AutoEncoder-Inspired Latent Space Sharing), designed to enhance the frequency response prediction of automotive suspension modules, particularly in scenarios with limited datasets. AEILSS addresses this issue by leveraging a large dataset from an existing suspension module('D') to improve the frequency response prediction of a new module('K') with a smaller dataset. The research demonstrates the successful application of transfer learning and AutoEncoder concepts in integrating datasets from physically distinct suspension modules to augment prediction accuracy. AEILSS employs a pseudo-encoder structure to project various input dimensions into a unified latent space, facilitating the transfer of core weights between different modules. The model's effectiveness is validated through various experiments, including boundary and random sampling datasets, highlighting significant improvements in prediction accuracy. Key findings of this research include the necessity of finding the optimal layer combinations for transfer learning specific to each suspension module, and also the necessity of proportionate input latent vector dimensions to input data dimensions. Future research directions suggest that this study lays the foundational groundwork for exploring the generalization of frequency response predictions in suspension modules, utilizing existing datasets. This approach has the potential to significantly streamline the predictive modeling process in the automotive industry, offering a path towards more efficient and scalable solutions.
more목차
1. 서론 1
1.1. 연구 배경 1
1.2. 연구 동향 3
1.3. 연구 목표 6
2. 인공신경망 9
2.1. 인공신경망의 기초 구조 10
2.2. 인공신경망의 학습 과정 12
3. 연구 대상 데이터셋: 제약과 해결 방안 15
3.1. 데이터셋 구조 분석 15
3.2. 샘플링 부족 문제와 전이학습을 통한 접근 22
3.3. 데이터 차원 불일치 문제와 오토인코더를 활용한 접근 24
4. AEILSS: AutoEncoder-Inspired Latent Space Sharing 27
4.1. AEILSS의 신경망 구조 27
4.2. AEILSS의 학습 방법 29
4.3. AEILSS의 동작 원리 31
5. AEILSS 학습을 위한 설정 36
5.1. 데이터 전처리 36
5.2. AEILSS 하이퍼파라미터 설정 39
6. 실험 및 분석 40
6.1. 임의 샘플링 데이터셋을 통한 AEILSS 유효성 검증 40
6.2. 경계 데이터셋을 통한 AEILSS 유효성 검증 42
6.3. 입력 잠재벡터 차원 변경이 미치는 영향 검증 47
7. 결론 및 제언 51
8. Reference 54
