Title Page
ABSTRACT
요약문
Contents
1. INTRODUCTION 16
2. BACKGROUND THEORY 20
2.1. Surrogate Model 20
2.2. Structure-Borne Noise Analysis 21
2.3. Graph Neural Network (GNN) 22
2.4. Graph Convolutional Network (GCN) 24
3. PROBLEM DEFINITION 26
4. PROPOSED METHOD 28
4.1. Overview of Proposed Method 28
4.2. Predicting Natural Frequency of Panel and Cavity using GCN─Step 1 30
4.3. Predicting Average Response of Sub-system using Inference Function -Step 2 32
4.4. Predicting Interior Sound Pressure of Panel-Cavity System using CNN─Step 3 34
5. EXPERIMENTATION AND RESULTS 36
5.1. Dataset - Panel 36
5.2. Dataset - Cavity 38
5.3. Training and Results - Step 1 41
5.4. Training and Results - Step 2 43
5.4.1. Training and Results 43
5.4.2. Effect of Inference function 47
5.5. Training and Results - Step 3 48
6. APPLICATION - DIFFERENT BOUNDARY CONDITON OF PANEL 51
6.1. Boundary condition setup and result 51
7. APPLICATION - INVERSE PROBLEM 54
7.1. Differential evolution algorithm 54
7.2. Panel natural frequency maximization 56
8. DISCUSSION 58
REFERENCES 62
Appendix A. Deep convolutional neural network with new training method and transfer learning for structural fault classification of vehicle instrument panel structure 65
A. ABSTRACT 65
A.1. INTRODUCTION 66
A.2. PROPOSED METHOD 69
A.3. APPLICATION OF STRUCTURE FAULT CLASSIFICATION: INSTRUMENT PANEL STRUCTURE'S BOUNDARY CONDITIONS 73
A.4. APPLICATION TO GEAR FAULT DIAGNOSIS BASED ON SOUND SIGNALS 80
A.5. APPLICATION OF STRUCTURE FAULT CLASSIFICATION: INSTRUMENT PANEL STRUCTURE WITH A WELD MISSING AND CRACKS 83
A.6. CONCLUSIONS 94
A. REFERENCES 95
Table 1. Panel natural frequency of flat panel, maximum frequency of training set and from Differential evolution method 57
Table 2. Proposed method, predictions, and the achieved r-square values at each step. 59
Figure 1. Mesh-based structure-borne noise model which consisted of the panel and cavity. 27
Figure 2. Proposed method for prediction of structure-borne noise using graph-convolutional-network, neural-network and convolutional-neural-network-based surrogate model. 29
Figure 3. Graph convolutional network for panel natural frequency prediction. Numbers in each layer is the number of outputs. 31
Figure 4. Graph convolutional network for cavity natural frequency prediction. Numbers in each layer is the number of outputs. 31
Figure 5. Inference function as an input for average response prediction. 32
Figure 6. Neural network for predicting average response of panels and cavity. Numbers in each layer is the number of outputs. 34
Figure 7. Convolutional neural network for predicting panel-cavity coupled structure-borne noise. Numbers in each layer are the convolution filter sizes and the number of outputs. 35
Figure 8. Panel samples which morphed with 4 handles in a panel center area for panel dataset. 36
Figure 9. Flat/morphed panels natural frequency and mode shapes 1st to 5th mode: (a) panel size 0.4 m x 0.6 m and 0.5 m x 0.6 m; (b) panel size 0.6 m x 0.6 m and 0.7 m x 0.6 m. 37
Figure 10. Cavity dataset with cutting the volume from 4 sizes of hexahedron: (a) cavity size 1.2 m x 1.0 m x 0.8 m, 1.3 m x 0.8 m x 0.8 m; (b) cavity size 1.5 m x 1.0 m x 0.9 m, 1.6m x... 39
Figure 11. Cavity shape with cutting the volume from hexahedron and mode shapes up to 5th mode: (a) cavity size 1.2m x 1.0m x 0.8m and 1.3m x 0.8m x 0.8m; (b) cavity size 1.5m x 1.0m... 40
Figure 12. Panel natural frequency prediction using proposed method─Step 1: (a) accuracy of validation data; (b) accuracy of test data. Blue─target; red─predicted. 42
Figure 13. Cavity natural frequency prediction using proposed method─Step 1: (a) accuracy of validation data; (b) accuracy of test data. Blue─target; red─predicted. 42
Figure 14. Panel average response prediction─Step 2. Blue─target; red─predicted. 45
Figure 15. Cavity average response prediction─Step2. Blue─target; red─predicted. 46
Figure 16. GCN-NN network for predicting panel frequency response up to 500Hz in 1/12 octave band 47
Figure 17. Panel vibration frequency response predicted by GCN-NN network 48
Figure 18. Panel-cavity coupled interior noise prediction─Step 3. Blue─target; red─predicted. 50
Figure 19. Panel natural frequency prediction using proposed method for simply supported boundary condition ─Step 1: (a) accuracy of validation data; (b) accuracy of test data. Blue─... 51
Figure 20. Panel average response prediction of simply supported boundary condition─Step 2. Blue─target; red─predicted. 52
Figure 21. Panel-cavity coupled interior noise prediction of simply supported boundary condition─Step 3. Blue─target; red─predicted. 53
Figure 22. Panel natural frequency distribution: Training data and maximized panel frequency using DE 57