Machine Learning–Based Detection of Network Failures in Core Network Cellular Data

The increasing complexity of mobile networks and the large volume of traffic data generated by multiple network elements make it difficult for Mobile Network Operators (MNOs) to accurately identify operational irregularities and reliably distinguish true network failures from normal network traffic. This study analyses mobile network traffic volume across four major geographical regions of South Africa—Johannesburg, Cape Town, Pretoria, and Durban—using multivariate time series data obtained from the core network of a MNO. The research proposes a systematic methodology for data filtering, including the identification and removal of time intervals characterised by unusually low traffic volumes, as well as the ap-plication of statistical anomaly detection techniques to remove abnormal traffic fluctuations. Both offline and online learning approaches are investigated for network traffic volume prediction and network failure classification. These predictive modelling are leveraged to reconstruct normal operational traffic patterns. For offline traffic volume prediction, three models are evaluated: Linear Regression, Random Forest Regressor, and Feed-forward Neural Networks. The Online prediction is performed using a Linear Regression model. Experimental results for the offline approach indicate that the Random Forest Regressor consistently outperforms the other models across the datasets for all four geographical regions. Furthermore, offline and online classification models are applied to distinguish between network failures and normal traffic conditions. The results demonstrate that the offline classification model achieves substantially better performance than its online counterpart.

Grantor dc:publisher

Stellenbosch : Stellenbosch University

Year dc:date.issued

2026

Author dc:creator

• Fita, Gospel Antonio

Advisors dc:contributor.advisor

• Wolhuter, Riaan
• Niesler, Thomas
• Du Toit, Jaco

Language dc:language.iso

en

Repository record dc:identifier.uri

https://scholar.sun.ac.za/handle/10019.1/135986

OAI identifier oai:identifier

oai:scholar.sun.ac.za:10019.1/135986