검색 상세

해석과 실험의 전이학습을 통한 강성 특성 예측 연구

Prediction of Stiffness Properties Using Transfer Learning Between Analysis and Experimental Data

초록 (요약문)

The Finite Element Method (FEM) is a widely used tool for predicting and optimizing the properties of mechanical systems and structures. However, discrepancies between FEM simulations and experimental results are inevitable. Repeated experimental validation to address these discrepancies is often costly and time-consuming. To mitigate these challenges, Artificial Intelligence (AI) and Transfer Learning (TL) have garnered significant attention. TL enables the development of high-performance predictive models by leveraging knowledge learned in domain A and applying it to domain B, making it particularly effective in engineering fields where experimental data is scarce. This study proposes a transfer learning-based force-displacement prediction model for a Quasi-Zero Stiffness (QZS) structure, characterized by a stiffness profile with a zero slope region. The model is trained on 102 FEM-generated simulation datasets and 30 experimental datasets. By utilizing negative stiffness parameters, the proposed model predicts 25 data points on force-displacement graphs, analyzing the stiffness properties of the structure and the discrepancies between FEM simulations and experimental results. Additionally, seven different hidden layer architectures and five transfer learning scenarios were evaluated to identify the optimal neural network configuration and loss function. This approach significantly enhances predictive performance by integrating simulation data with limited experimental data, reducing the gap between FEM simulations and experimental results. The findings of this study hold potential for minimizing repeated experiments and improving the accuracy of experimental predictions across various engineering applications.

more

목차

1. 서론 1
1.1. 연구 배경 1
1.2. 연구 동향 3
1.3. 연구 목표 7
2. 이론적 배경 9
2.1. 준제로 강성 이론 및 응용 분야 9
2.2. 전이 학습 이론 및 적용사례 13
2.3. 딥러닝 신경망 학습 18
3. 연구방법 22
3.1. QZS 모델링 22
3.2. 소스/타겟 도메인 수집 및 전처리 24
3.3. 실제 실험 데이터 수집 및 전처리 36
3.4. 알고리즘 설계 41
3.5. 전이학습 모델 설계 43
3.6 하이퍼파라미터 설정 및 최적화 47
4. 실험 결과 및 분석 51
4.1. 예측 성능 비교 51
4.2. 전이학습 정량적 분석 52
4.3. 전이학습 정성적 분석 55
4.4. 실제 실험 전이학습 정량적 분석 61
4.5. 실제 실험 전이학습 정성적 분석 64
5. 전이 랜덤성 검증 결과 및 분석 70
5.1. 전이 랜덤성 문제 제기 70
5.2. 전이 랜덤성 검증 결과 71
6. 타 도메인 실험 검증 및 성능 비교 75
6.1. 선행 연구와의 비교 및 타 데이터 전처리 75
6.2. 타 도메인 데이터 전이 학습 결과 분석 78
7. 결론 83
8. Reference 86

more