Title Page
ABSTRACT
초록
Contents
NOMENCLATURE 13
CHAPTER 1. INTRODUCTION 14
1.1. Problem Definition 17
CHAPTER 2. BACKGROUND 19
2.1. Contrastive Autoencoder 19
2.2. Self-attention 22
2.3. Temporal Convolutional Network 24
CHAPTER 3. PROPOSED METHOD 27
3.1. Phase I 28
3.2. Phase II 32
CHAPTER 4. EXPERIMENTAL RESULT 34
4.1. Dataset 34
4.2. MMTS Forecast 36
CHAPTER 5. ANALYSIS 40
5.1. Ablation Study 40
5.2. Visual Analysis 41
CHAPTER 6. CONCLUSION AND FUTURE WORK 46
BIBLIOGRAPHY 47
Algorithm 1. Calculating the total loss 32
Table 1. Forecasting performance of comparison methods and the proposed method 38
Table 2. Variable MSE scores and average rank on CNC dataset 39
Table 3. Ablation results on CNC dataset 40
Figure 1. Architecture of Input Projection Layer 21
Figure 2. Figure of CNN Convolution Kernel 22
Figure 3. Self-attention in the proposed method, 1x1 convolution is applied to the input to create query, key, and value. The final output is added to the input, similar to the residual connection. 23
Figure 4. Basic structural explanation of TCN 26
Figure 5. Overall architecture of the proposed method 27
Figure 6. Architectural structure of SACAE encoder 28
Figure 7. Process of creating contextual representations with random cropping. From random cropping, the proposed method creates two augmented context representations Zc...[이미지참조] 29
Figure 8. Non-linear correlation pair grid of six variables in the dataset and an example of explicit non-linear correlation of variables x1_outputvoltage and y1_commandvelocity. 35
Figure 9. Fitted image of fourth order polynomial regression of variables x1_outputvoltage and y1_commandvelocity (Left) illustrates a clear non-linear correlation between two... 36
Figure 10. A prediction slice of the proposed method (top row) and (5) Fully Connected Layer (bottom row) on test set of CNC dataset, variables, X1_actualvelocity,... 41
Figure 11. Heatmap visualization of representation time series created from Phase I of experimented datasets. 42
Figure 12. Visualization of PMM original dataset with error identification (red), maintenance time points (blue) and predicted 20 time points (orange). 43
Figure 13. Original Cause Analysis and Zoom-in Cause Analysis Map of PMM dataset. Zoom-in Cause Analysis Map shows 85 to 205 time points of the Original Cause Analysis... 45