Anomaly Detection of LTE Cells using KNN Algorithms: The Case of Addis Ababa

No Thumbnail Available



Journal Title

Journal ISSN

Volume Title


Addis Ababa University


For mobile operators, delivering quality service to their customers is very important as service quality significantly affects their business. To achieve a consistent quality service delivery, they need to continuously monitor and analyze their network performance and timely address any obtained performance drops. Performance drops in mobile networks can be observed in key performance indicators (KPIs) of spatially distributed cells with different magnitude and at different time periods. As it is practically difficult to address all performance drops simultaneously, it is preferable to make prioritized corrective actions starting from cells with critical drops, also called anomaly cells. To detect anomaly cells, different automated methodologies have been proposed and analyzed. Yet, ethio telecom still applies manual and subjective anomaly detection method where measured KPI values are manually compared with fixed thresholds to determine if the measured values are within defined required ranges or not. Cells with KPI values out of the required range are analyzed for identifying performance drop root causes and taking corrective actions. The manual and subjective anomaly detection method is prone to detection errors and is maintenance time, manpower and then cost inefficient. These challenges of manual and subjective detection can be improved by applying advanced automatic methods based on machine learning algorithms. In this thesis work, KNN based anomaly detection algorithms such as KNN classification, local outlier factor (LOF) and connectivity outlier factor (COF) anomaly detection models is implemented, and their comparative evaluation are made for Addis Ababa LTE cells. The comparison is made based on type of output, complexity and their true positive rate (TPR) for time series and cell level detections. Unlike KNN classification, LOF and COF do not need heavy data set training and are able to provide anomaly scores instead of two-class labels. Experimentation results show that COF provides slightly better performance than the other models with negligible performance difference. For instance, the performance of COF with respect to TPR for RRC setup success rate in the experimentation is 97.91% for cell level detection and 88% for time series detection.



Anomaly detection, Connectivity outlier factor, Local outlier factor, KNN classification, Machine Learning, KPI, Alarm, LTE