Title Page
Abstract
국문 초록
Preface
Contents
Chapter 1. Introduction 14
Chapter 2. Related Work 17
Chapter 3. Proposed Method 19
3.1. Problem Statement 19
3.2. Assumptions 20
3.3. Model Training 21
3.3.1. Variance Prediction 22
3.3.2. AE Block 23
3.3.3. Self-Attention Block 23
3.3.4. LSTM Block 24
3.3.5. Residual Stacking 25
3.3.6. Loss 26
3.4. Evaluation 27
Chapter 4. Experiments 28
4.1. Public Datasets 28
4.2. Experimental Setup 29
4.2.1. Data Preparation 29
4.2.2. Architecture 29
4.2.3. Evaluation Metrics 30
4.2.4. Baselines 31
4.3. Results 33
4.3.1. Benchmarking Performance 33
4.3.2. Analysis of Wtrain[이미지참조] 33
4.3.3. Visualization on Low-dimensional Representation 34
4.3.4. Analysis of Forecasting Performance 34
4.3.5. Analysis of Number of Stacks 35
4.3.6. Analysis of Ratio of Reconstruct Part and Forecast Part 36
Chapter 5. Conclusion 39
Reference 41
Table 4.1. Dataset summary. (%) is the percentage of anomalous data points in the dataset. 28
Table 4.2. Anomaly detection results for different datasets 32
Table 4.3. Forecasting performance aggregated over each dataset 36
Figure 3.1. Stacked blocks and the overall model structure. A stack consists of an AE block, a self-attention block, and a LSTM block. The black lines represent the compu-... 21
Figure 3.2. Visualization of the three blocks: (a) AE block, (b) self-attention block, and (c) LSTM block. The figure illustrates the outputs of each block, including the forecast... 22
Figure 4.1. Visualization of the weights between variance prediction errors and input sequences of the reconstruct errors for the training data in the SMAP dataset. The figure... 34
Figure 4.2. Data representations from SMAP dataset: (a) the original data representation and (b) the low-dimensional representation learned by our model. The red points in the... 35
Figure 4.3. The prediction reuslts (Horizon = 13) from SMAP dataset. The blue line represents the actual time series plot, while the orange line depicts the forecasted time... 36
Figure 4.4. The impact of number of stacks on F1-Score performance from SMAP dataset. The results indicate that the optimal performance is attained when three stacks... 37
Figure 4.5. Comparative analysis of the effects of ratio variation on the reconstruct and forecast parts. Graph (a) plots the influence of different ratios and input sequence lengths... 38