Few-Shot Classification in Deep Learning based Anomaly Detection of Noisy Industrial Data

Publikationen: Thesis / Studienabschlussarbeiten und HabilitationsschriftenMasterarbeit

Standard

Few-Shot Classification in Deep Learning based Anomaly Detection of Noisy Industrial Data. / Freyler, Patricia.
2023.

Publikationen: Thesis / Studienabschlussarbeiten und HabilitationsschriftenMasterarbeit

Bibtex - Download

@mastersthesis{e3380965363e4c83b2a21ae4a2bfe30e,
title = "Few-Shot Classification in Deep Learning based Anomaly Detection of Noisy Industrial Data",
abstract = "Manufacturing processes can be improved by early detection of anomalies using Deep Learning methods. These methods require large volumes of data, however in manufacturing processes, there is usually only a small amount of information about anomalies, which leads to biased data. These small number of occurrences and a resulting small amount of data together with the variety of process anomalies in the manufacturing processes in modern production plants pose challenges for traditional Deep Learning methods. This reduces the potential performance of neural networks in the production environment due to the lack of transferability of the individual models among each other. The large volume of data needed to build neural networks places high demands on the quality and quantity of data labeling, which results in high costs. The field of Few-Shot Learning, which focuses on the design of high-performance neural networks with limited data sets, promises a potential remedy. The purpose of the thesis is to transfer knowledge from state-of-the-art computer vision methods to the new application domain of noisy industrial data and investigate an efficient labeling system using advanced Deep Learning Few-Shot classification methods for data streams collected during production. The main results of this work are the following: The Prototypical Network (PN) using the Euclidean Distance reached an F1-score of 93.92 % on the verification task when trained based on 70 good and 21 bad samples (dataset 1) and an F1-score of 80.01 % with 17 good and six bad samples (dataset 3). The Matching Network (MN) reached an F1-score of 87.34 % and 71.81 %. By implementing Cosine Distance as the final classification, PN achieves an F1-score of 95.21 % and MN an F1-score of 91.46 % with dataset 1 (Table 4.2). The DOT-product achieves an F1-performance of 93.51 % for the PN and 88.70 % for the MN. The number of shots for the support set, should be about 5 to 7 shots with an F1-score of 93.92 % and 94.82 %. Three shots are not sufficient for the support and query set with an F1-score of 92.92 %. Few-Shot Learning in quality control can significantly reduce the need for training data. Different distance/similarity methods improve the performance of the networks. These techniques provide good results for the data used in this work.",
keywords = "Machine Learning, Deep Learning, Anomaly Detection, Machine Learning, Deep Learning, Anomalieerkennung",
author = "Patricia Freyler",
note = "no embargo",
year = "2023",
doi = "10.34901/mul.pub.2023.33",
language = "English",
school = "Montanuniversitaet Leoben (000)",

}

RIS (suitable for import to EndNote) - Download

TY - THES

T1 - Few-Shot Classification in Deep Learning based Anomaly Detection of Noisy Industrial Data

AU - Freyler, Patricia

N1 - no embargo

PY - 2023

Y1 - 2023

N2 - Manufacturing processes can be improved by early detection of anomalies using Deep Learning methods. These methods require large volumes of data, however in manufacturing processes, there is usually only a small amount of information about anomalies, which leads to biased data. These small number of occurrences and a resulting small amount of data together with the variety of process anomalies in the manufacturing processes in modern production plants pose challenges for traditional Deep Learning methods. This reduces the potential performance of neural networks in the production environment due to the lack of transferability of the individual models among each other. The large volume of data needed to build neural networks places high demands on the quality and quantity of data labeling, which results in high costs. The field of Few-Shot Learning, which focuses on the design of high-performance neural networks with limited data sets, promises a potential remedy. The purpose of the thesis is to transfer knowledge from state-of-the-art computer vision methods to the new application domain of noisy industrial data and investigate an efficient labeling system using advanced Deep Learning Few-Shot classification methods for data streams collected during production. The main results of this work are the following: The Prototypical Network (PN) using the Euclidean Distance reached an F1-score of 93.92 % on the verification task when trained based on 70 good and 21 bad samples (dataset 1) and an F1-score of 80.01 % with 17 good and six bad samples (dataset 3). The Matching Network (MN) reached an F1-score of 87.34 % and 71.81 %. By implementing Cosine Distance as the final classification, PN achieves an F1-score of 95.21 % and MN an F1-score of 91.46 % with dataset 1 (Table 4.2). The DOT-product achieves an F1-performance of 93.51 % for the PN and 88.70 % for the MN. The number of shots for the support set, should be about 5 to 7 shots with an F1-score of 93.92 % and 94.82 %. Three shots are not sufficient for the support and query set with an F1-score of 92.92 %. Few-Shot Learning in quality control can significantly reduce the need for training data. Different distance/similarity methods improve the performance of the networks. These techniques provide good results for the data used in this work.

AB - Manufacturing processes can be improved by early detection of anomalies using Deep Learning methods. These methods require large volumes of data, however in manufacturing processes, there is usually only a small amount of information about anomalies, which leads to biased data. These small number of occurrences and a resulting small amount of data together with the variety of process anomalies in the manufacturing processes in modern production plants pose challenges for traditional Deep Learning methods. This reduces the potential performance of neural networks in the production environment due to the lack of transferability of the individual models among each other. The large volume of data needed to build neural networks places high demands on the quality and quantity of data labeling, which results in high costs. The field of Few-Shot Learning, which focuses on the design of high-performance neural networks with limited data sets, promises a potential remedy. The purpose of the thesis is to transfer knowledge from state-of-the-art computer vision methods to the new application domain of noisy industrial data and investigate an efficient labeling system using advanced Deep Learning Few-Shot classification methods for data streams collected during production. The main results of this work are the following: The Prototypical Network (PN) using the Euclidean Distance reached an F1-score of 93.92 % on the verification task when trained based on 70 good and 21 bad samples (dataset 1) and an F1-score of 80.01 % with 17 good and six bad samples (dataset 3). The Matching Network (MN) reached an F1-score of 87.34 % and 71.81 %. By implementing Cosine Distance as the final classification, PN achieves an F1-score of 95.21 % and MN an F1-score of 91.46 % with dataset 1 (Table 4.2). The DOT-product achieves an F1-performance of 93.51 % for the PN and 88.70 % for the MN. The number of shots for the support set, should be about 5 to 7 shots with an F1-score of 93.92 % and 94.82 %. Three shots are not sufficient for the support and query set with an F1-score of 92.92 %. Few-Shot Learning in quality control can significantly reduce the need for training data. Different distance/similarity methods improve the performance of the networks. These techniques provide good results for the data used in this work.

KW - Machine Learning

KW - Deep Learning

KW - Anomaly Detection

KW - Machine Learning

KW - Deep Learning

KW - Anomalieerkennung

U2 - 10.34901/mul.pub.2023.33

DO - 10.34901/mul.pub.2023.33

M3 - Master's Thesis

ER -