Anomalijų atpažinimas industrijos gamybos procese naudojant giliuosius neuroninius tinklus

Ištrauka

Anomalijų atpažinimas industrijos gamybos procese naudojant giliuosius neuroninius tinklus

Contents
INTRODUCTION 6
PROCESS OF CONDUCTING THE RESEARCH 9
1.ANOMALY DETECTION 10
1.1.Supervised, unsupervised and semisupervised training 10
1.2.Patch Distribution Modeling Framework PaDiM 10
1.3.Semantic Pyramid Anomaly Detection 11
1.4.PatchCore 12
1.5.StudentTeacher Anomaly Detection 13
1.6.Asymmetric StudentTeacher 14
1.6.1.Normalizing flow 15
1.6.2.Teacher 15
1.6.3.Student 16
1.7.EfficientAD 16
1.7.1.Autoencoder 17
1.8.Conclusions of literature review 18
2.EXPERIMENT 19
2.1.MvTec AD dataset 19
2.2.MvTec LOCO AD dataset 20
2.3.Metrics 21
2.4.Implementation of EfficientAD model 23
2.5.Implementation of Asymmetric StudentTeacher model 25
2.6.Process of integrating AST into EfficientAD 26
2.6.1.Integration of teacher, student and autoencoder 26
2.6.2 Wrapper class for all components
26
2.6.3 Teacher configuration
27
2.6.4.Student configuration 27
2.6.5 Autoencoder configuration
27
2.6.6.Dataset 28
2.6.7.Early stopping 28
2.6.8 Teacher training loop
28
2.6.9.Studentautoencoder training loop 29
2.6.10 Training penalty
29
2.6.11 Percentile of output differences
29
2.6.12 Evaluation of modified model
29
2.7.Results of Experiment 30
2.7.1.Results on MvTec AD dataset 30
2.7.2.Results on MvTec LOCO AD dataset 38
2.7.3.Comparison with other models 44
RESULTS AND CONCLUSIONS 49
REFERENCES 52
APPENDIXES 55
Appendix 1 55

Introduction
As our society relies on industrial manufacturing more every day,
it is oſten a case that the manufacturing machines make mistakes,
and it is important to spot those mistakes as soon in the production
process as possible to reduce the cost of fixing and preventing them
from happening again [LXW+24]. One of the fundamental principles of
Lean manufacturing is called ”Jidoka”, which enables machines and
operators to detect an abnormal condition and immediately stop the
work [Sol20]. This is one of the main principles that helped the Toyota
company achieve great success and customer satisfaction by
delivering the best quality goods and reducing defects in the
manufacturing process. According to the article [S SW+21], deep
learning nowadays has become more relevant, and one must
understand its capabilities and utilize them in a way that could benefit
society. Anomaly detection is one of those technologies that could help
detect errors in various imagery representations. In this case, the ability
to automatically spot faulty products in the manufacturing process could
improve the quality of the products and lead to a fully automated
manufacturing process.
When training and testing deep neural networks for anomaly
detection tasks, one of the primary goals is to address the class
imbalance problem [JK19]. This is a problem when training data does
not consist of the same amount of different classes. In the anomaly
detection context, it is usually a case that faulty products consist of
only a minor part of the whole batch. To solve this problem,
unsupervised, supervised, and semisupervised methods [Lab20] of
training the models need to be reviewed to understand which of these
approaches would be best suited for the anomaly detection task in
industrial manufacturing.
There is a significant amount of deep neural networks that have
already been implemented to perform the anomaly detection task in
industrial manufacturing [BFS+19]. These models can be separated
into two categories: regular deep learning models and networks that
utilize other neural networks as backbones. The regular anomaly
detection models, such as autoencoders [F ad20] or Generative
Adversarial Networks (GANs) [DYW21], work by trying to recreate the
input data and spotting differences from the usual data patterns they
have learned. These models are built to understand standard data
and consider significant differences as anomalies. On the other hand,
models that use neural networks to pull out important features from
images start by using Convolutional Neural Networks (CNNs). The
important information they...