Title Page
Contents
ABSTRACT 10
Ⅰ. Introduction 11
1.1. Research Background and Significance 11
1.2. Literature Review 13
1.3. Main Research Objectives 15
Ⅱ. Air Compressors of Railway Vehicles 17
Ⅲ. Deep Learning Theory and Anomaly Detection Technology 21
3.1. Deep Learning Theory 21
3.2. Long Short-Term Memory 23
3.3. Bidirectional Long Short-Term Memory 27
3.4. Autoencoder 29
3.5. Attention Mechanism 32
Ⅳ. Time Series Data Theory and Pre-processing 34
4.1. Time Series Data Theory 34
1. Point anomaly 35
2. Interval anomaly 35
3. Sequence anomaly 35
4.2. Air Compressor Data Acquisition 36
4.3. Data Pre-processing 39
1. Standardization 40
2. Normalization 40
3. Data regularization 40
4.4. Data Windowing 41
4.5. Data Threshold 44
Ⅴ. Anomaly Detection Model Construction and Learning 46
5.1. Experimental Environment and Setup 46
5.2. LSTM-AE Model Structural Design 47
5.3. BiLSTM-AE Model Structural Design 49
5.4. BiLSTM-AE with Attention Model Structural Design 50
Ⅵ. Anomaly Detection of an Air Compressor of a Railway Vehicle 53
6.1. Anomaly Data Setting 53
6.2. Abnormal Diagnostic Evaluation Index 54
6.3. Anomaly Detection Experiment Process 55
6.4. Anomaly Detection Results and Analysis 57
Ⅶ. Conclusion 64
References 65
국문초록 68
Table 3-1. Ordinary autoencoder training process 32
Table 5-1. Experimental environment of Anomaly Detection Model 46
Table 5-2. Anomaly detection model reconstruction error 52
Table 6-1. Abnormal scenarios and manipulated values 53
Table 6-2. Relationship between actual situation and predicted result 54
Table 6-3. Anomaly detection process 57
Table 6-4. Anomaly detection model threshold value 58
Table 6-5. Anomaly detection different algorithm score results 58
Figure 2-1. The air end unit of the screw air compressor (Courtesy of Ingersoll Rand) 17
Figure 2-2. Electric motor-driven screw air compressor for railway vehicles (Seoul Subway Line 8). (a) Exterior (b) Actuator CMSB (Courtesy of Eugene Machinery) 19
Figure 2-3. Screw air compressor drawings and the installation positions of the main sensors 20
Figure 3-1. Schematic diagram of a LSTM unit 25
Figure 3-2. Schematic diagram of BILSTM unit structure 28
Figure 3-3. Schematic diagram of autoencoder 29
Figure 4-1. Rail vehicle air compressors sensor photo at the lower end of Seoul Subway Line 4 (Gwacheon Ansan Line) 36
Figure 4-2. Temperature sensor 37
Figure 4-3. U,V Current sensor 37
Figure 4-4. Main reservoir (MR) pressure sensor 38
Figure 4-5. Raw data appearance representation 39
Figure 4-6. Rail Vehicle Air Compressors real-time monitoring platform 41
Figure 4-7. The flow chart of data pre-processing process 43
Figure 4-8. Regularize preprocessed part of the data 44
Figure 4-9. Normal distribution curve 45
Figure 5-1. The structure of the designed LSTM-AE model 48
Figure 5-2. LSTM-AE model learning results 49
Figure 5-3. BiLSTM-AE model learning results 50
Figure 5-4. The structure of the BiLSTM-AE with attention model 51
Figure 5-5. BiLSTM-AE with attention model learning results 52
Figure 6-1. Model training process for detection of an air compressor 56
Figure 6-2. Main reservoir pressor anomaly detection time results. (a)LSTM-AE Model(b)BiLSTM-AE Model(c)BiLSTM-AE with Attention Model 60
Figure 6-3. Air compressor pressor anomaly detection time results. (a)LSTM-AE Model(b)BiLSTM-AE Model(c)BiLSTM-AE with Attention Model 61
Figure 6-4. Oil temperature anomaly detection time results. (a)LSTM-AE Model(b)BiLSTM-AE Model(c)BiLSTM-AE with Attention Model 63