Industrial Anomaly Detection with Normalizing Flows

Produktbeschreibung

This thesis addresses deep learning-based methods for automatic anomaly detection in an industrial context. It involves image- or sensor-based detection of defects in the production process that can affect the quality of products. Automating this task provides a reliable and cost-effective alternative to humans, who perform this task manually by sighting. Since this setup has special requirements such as detecting previously unknown defects that traditional approaches cannot fulfill, this paper presents anomaly detection methods that learn without any examples of anomalies and include only normal data in the training process. Most of our proposed methods address the problem from a statistical perspective. Based on a deep-learning-based density estimation of the normal data, it is assumed that anomalies are considered unlikely according to the modeled distribution. The density estimation is performed by socalled Normalizing Flows, which, in contrast to conventional neural networks, can model a formally valid probability distribution due to their bijective mapping. Moreover, due to their flexibility, Normalizing Flows allow modeling of more complex distributions in contrast to traditional methods, which usually use strong simplifications about the distribution.

C O N T E N T S
1 Introduction 1
1.1 Anomaly Detection in the Industrial Context . . . . . . . . 3
1.2 Addressed Applications . . . . . . . . . . . . . . . . . . . . 4
1.3 Structure of the Thesis . . . . . . . . . . . . . . . . . . . . . 5
1.4 List of Publications . . . . . . . . . . . . . . . . . . . . . . . 7
1.4.1 Anomaly Detection . . . . . . . . . . . . . . . . . . 7
1.4.2 Other Publications . . . . . . . . . . . . . . . . . . . 10
2 Fundamentals 13
2.1 Anomaly Detection . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.1 Semi-Supervised Anomaly Detection . . . . . . . . 13
2.1.2 Taxonomy . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.2 Traditional Approaches . . . . . . . . . . . . . . . . 17
2.2.3 Generative Models . . . . . . . . . . . . . . . . . . . 19
2.2.4 Student-Teacher Networks . . . . . . . . . . . . . . 20
2.2.5 Density Estimation . . . . . . . . . . . . . . . . . . . 21
2.2.6 Synthetic Data . . . . . . . . . . . . . . . . . . . . . 22
2.2.7 Methods for Multivariate Machine Data . . . . . . 22
3 Datasets and Metrics 24
3.1 MVTec AD . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2 MVTec 3D-AD . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.3 Magnetic Tile Defects . . . . . . . . . . . . . . . . . . . . . . 28
3.4 Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4 DifferNet 31
4.1 Normalizing Flows . . . . . . . . . . . . . . . . . . . . . . . 33
4.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . 33
4.1.2 Related Work . . . . . . . . . . . . . . . . . . . . . . 35
4.1.3 Real-NVP . . . . . . . . . . . . . . . . . . . . . . . . 36
4.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.2.1 Architecture . . . . . . . . . . . . . . . . . . . . . . . 38
4.2.2 Training . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.2.3 Scoring Function . . . . . . . . . . . . . . . . . . . . 40
4.2.4 Localization . . . . . . . . . . . . . . . . . . . . . . . 40
4.3 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.3.1 Implementation Details . . . . . . . . . . . . . . . . 42
4.3.2 Detection . . . . . . . . . . . . . . . . . . . . . . . . 43
4.3.3 Localization . . . . . . . . . . . . . . . . . . . . . . . 44
4.3.4 Ablation Studies . . . . . . . . . . . . . . . . . . . . 45
4.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.5 Open Problems . . . . . . . . . . . . . . . . . . . . . . . . . 48
5 Cross-Scale-Flow 50
5.1 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.1.1 Cross-Scale Flow . . . . . . . . . . . . . . . . . . . . 52
5.1.2 Learning Objective . . . . . . . . . . . . . . . . . . . 54
5.1.3 Localization . . . . . . . . . . . . . . . . . . . . . . . 55
5.2 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.2.1 Implementation Details . . . . . . . . . . . . . . . . 55
5.2.2 Detection . . . . . . . . . . . . . . . . . . . . . . . . 57
5.2.3 Localization . . . . . . . . . . . . . . . . . . . . . . . 58
5.2.4 Ablation Studies . . . . . . . . . . . . . . . . . . . . 59
5.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.4 Open Problems . . . . . . . . . . . . . . . . . . . . . . . . . 61
6 Asymmetric Student-Teacher Networks 63
6.1 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6.1.1 Teacher . . . . . . . . . . . . . . . . . . . . . . . . . 66
6.1.2 Student . . . . . . . . . . . . . . . . . . . . . . . . . 67
6.1.3 3D Preprocessing . . . . . . . . . . . . . . . . . . . . 68
6.2 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
6.2.1 Implementation Details . . . . . . . . . . . . . . . . 70
6.2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . 71
6.2.3 Ablation Studies . . . . . . . . . . . . . . . . . . . . 75
6.3 Open Problems . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.4 Comparison with DifferNet and CS-Flow . . . . . . . . . . 79
6.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
7 Voraus-AD dataset 82
7.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
7.2 Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
7.3 Robot and Signals . . . . . . . . . . . . . . . . . . . . . . . . 86
7.4 Pick-And-Place Operation . . . . . . . . . . . . . . . . . . . 88
7.5 Anomalies . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
7.5.1 Process Errors . . . . . . . . . . . . . . . . . . . . . 88
7.5.2 Gripping Errors . . . . . . . . . . . . . . . . . . . . 91
7.5.3 Robot axis wear . . . . . . . . . . . . . . . . . . . . 91
7.6 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
7.6.1 Normal Data . . . . . . . . . . . . . . . . . . . . . . 91
7.6.2 Anomalies . . . . . . . . . . . . . . . . . . . . . . . . 92
7.6.3 Comparison to AURSAD . . . . . . . . . . . . . . . 94
7.6.4 Evaluation Protocol . . . . . . . . . . . . . . . . . . 95
7.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
VIcontents
8 MTS-Flow 96
8.1 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
8.1.1 Architecture . . . . . . . . . . . . . . . . . . . . . . . 97
8.1.2 Training . . . . . . . . . . . . . . . . . . . . . . . . . 98
8.1.3 Temporal Analysis . . . . . . . . . . . . . . . . . . . 99
8.2 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
8.2.1 Implementation Details . . . . . . . . . . . . . . . . 100
8.2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . 102
8.2.3 Ablation Studies . . . . . . . . . . . . . . . . . . . . 105
8.2.4 Temporal Analysis . . . . . . . . . . . . . . . . . . . 107
8.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
9 Conclusion 110
Bibliography 115

Keywords: Automatische Bildverarbeitung, Anomaliedetektion, Defekterkennung, Qualitätssicherung, Dichteschätzung, Normalizing Flows, Deep Learning, Automatische Bildverarbeitung, Anomaliedetektion, Defekterkennung, Qualitätssicherung, Dichteschätzung, Normalizing Flows, Deep Learning