Deep Learning for Anomaly Detection
Loading...
Authors
Author (Corporation)
Publication date
2021
Typ of student thesis
Master
Course of study
Collections
Type
11 - Student thesis
Editors
Editor (Corporation)
Supervisor
Parent work
Special issue
DOI of the original publication
Link
Series
Series number
Volume
Issue / Number
Pages / Duration
Patent number
Publisher / Publishing institution
Hochschule für Wirtschaft FHNW
Place of publication / Event location
Olten
Edition
Version
Programming language
Assignee
Practice partner / Client
Abstract
In the age of Big Data, data analysis becomes ever more important. To analyse the data, many researchers nowadays focus on artificial intelligence. Artificial intelligence does not rely on labour-intensive feature engineering like the traditional machine learning or statistical models. Therefore, the use of AI, such as neural networks, can save a lot of development time. Two widely used architectures of neural networks are the Convolutional Neural Networks and the Recurrent Neural Networks. A Convolutional Neural Network is generally used when a task is related to image recognition, whereas Recurrent Neural Networks are used for the prediction of time series. Recently an approach was proposed to analyse time series data with Convolutional Neural Networks. The strengths and weaknesses of this approach, however, are currently unknown and are further investigated in this paper. To examine the usefulness, the practically relevant use case of anomaly detection was chosen. Within the scope of this work, different approaches on anomaly detection, that employ convolutional or recurrent neural networks are investigated. Since the architectures should be compared regarding their performance, ways to evaluate the performances are assessed. After elaborating the methodology applied in this work, it is described how the hyper-parameters were determined to make the models comparable. As a main part of this work, three experiments on different datasets are conducted. The datasets used, vary in complexity and contain different types of anomalies. The first experiment was carried out on a synthetic dataset with synthetic anomalies. In the second experiment a real-world dataset with synthetic anomalies was used. Finally, an official benchmark dataset was employed in the third experiment. On the obtained results the two architectures are assessed according to their usefulness for anomaly detection. Further, it is classified how helpful deep learning is for forecasting time series and detecting anomalies. At last, the insights of this work are presented together with suggestions for future research.
Keywords
Subject (DDC)
Event
Exhibition start date
Exhibition end date
Conference start date
Conference end date
Date of the last check
ISBN
ISSN
Language
English
Created during FHNW affiliation
Yes
Strategic action fields FHNW
Publication status
Review
Open access category
License
Citation
Saner, K. (2021). Deep Learning for Anomaly Detection [Hochschule für Wirtschaft FHNW]. https://irf.fhnw.ch/handle/11654/48603