Advanced Anomaly Detection in Mobile Networks: A Hybrid Approach Based on Statistical and Machine Learning Techniques

Abstract

Network traffic analysis (NTA) is a technique used by network administrators to monitor network activity, ensure availability, and detect unusual patterns to identify potential anomalies. However, traditional traffic monitoring systems often struggle to detect these anomalies accurately because they rely on rigid models and a limited pool of data. Additionally, anomaly detection is particularly challenging, as anomalies exhibit patterns that differ from most network activities, making their identification based on prior knowledge difficult. This underscores the necessity for an automated and unsupervised approach capable of detecting various types of anomalies despite these limitations. In this paper, we propose an unsupervised framework that combines three statistical machine learning (ML) methods for mobile network failure detection, such as the lower control limit, the cumulative sum algorithm (CUSUM), and the robust stat detector model. Compared to previous studies, models often struggle with large datasets and generate high rates of false positives. However, our approach has proven to be better suited to the demands of traffic monitoring in large telecommunications infrastructures. It offers improved data handling and significantly reduces the rate of false positives while achieving an impressive 98% detection rate for anomalies on telecommunications sites.