University of Oulu

Machine learning based anomaly detection in release testing of 5g mobile networks

Saved in:
Author: Khairi, Moustafa1
Organizations: 1University of Oulu, Faculty of Information Technology and Electrical Engineering, Department of Computer Science and Engineering, Computer Science and Engineering
Format: ebook
Version: published version
Access: open
Online Access: PDF Full Text (PDF, 2.3 MB)
Pages: 65
Persistent link: http://urn.fi/URN:NBN:fi:oulu-202204131561
Language: English
Published: Oulu : M. Khairi, 2022
Publish Date: 2022-04-19
Thesis type: Master's thesis (tech)
Tutor: Siirtola, Pekka
Reviewer: Suutala, Jaakko
Siirtola, Pekka
Description:

Abstract

The need of high-quality phone and internet connections, high-speed streaming ability and reliable traffic with no interruptions has increased because of the advancements the wireless communication world witnessed since the start of 5G (fifth generation) networks. The amount of data generated, not just every day but also, every second made most of the traditional approaches or statistical methods used previously for data manipulation and modeling inefficient and unscalable. Machine learning (ML) and especially, the deep learning (DL)-based models achieve the state-of-art results because of their ability to recognize complex patterns that even human experts are not able to recognize. Machine learning-based anomaly detection is one of the current hot topics in both research and industry because of its practical applications in almost all domains. Anomaly detection is mainly used for two purposes. The first purpose is to understand why this anomalous behavior happens and as a result, try to prevent it from happening by solving the root cause of the problem. The other purpose is to, as well, understand why this anomalous behavior happens and try to be ready for dealing with this behavior as it would be predictable behavior in that case, such as the increased traffic through the weekends or some specific hours of the day.

In this work, we apply anomaly detection on a univariate time series target, the block error rate (BLER). We experiment with different statistical approaches, classic supervised machine learning models, unsupervised machine learning models, and deep learning models and benchmark the final results. The main goal is to select the best model that achieves the balance of the best performance and less resources and apply it in a multivariate time series context where we are able to test the relationship between the different time series features and their influence on each other. Through the final phase, the model selected will be used, integrated, and deployed as part of an automatic system that detects and flags anomalies in real-time. The simple proposed deep learning model outperforms the other models in terms of the accuracy related metrics. We also emphasize the acceptable performance of the statistical approach that enters the competition of the best model due to its low training time and required computational resources.

see all

Subjects:
Copyright information: © Moustafa Khairi, 2022. This publication is copyrighted. You may download, display and print it for your own personal use. Commercial use is prohibited.