Adaptive periodicity-based anomaly detection in LTE networks

Abstract

In contemporary cellular networks utilizing Long-Term Evolution (LTE), accurately identifying anomalies in multivariate time series is crucial for ensuring network security and reliability. Although existing anomaly detection methods effectively identify anomalous time points, they typically fail to capture anomalies across multiple temporal scales, such as hourly or daily patterns, which are significant for practical network management. To address this limitation, this dissertation proposes a novel anomaly detection framework specifically designed to analyze LTE traffic data collected over a 39-day period from an operational base station. The proposed method explicitly captures both short-term and long-term temporal characteristics by identifying inherent periodic patterns within LTE traffic data and reorganizing one-dimensional series into structured two-dimensional representations. Furthermore, an adaptive aggregation approach is employed to integrate periodic components at various scales, enabling anomaly detection across different time granularities. Experimental evaluations demonstrate the proposed method's superior performance and practical applicability in real-world LTE network scenarios.