ML Optimization of Cell-Range Overshooting Detection in Real LTE Networks

The dynamic evolution of mobile access networks, marked by increasing size and complexity, necessitates innovative approaches to automatic anomaly detection and resolution. Traditional manual methods proves insufficient in handling the complexi- ties of modern networks. Consequently, network optimizers must shift focus towards developing algorithmic solutions that enable automation. This thesis emerges from the AI/ML Optimization Program 2024-2026, a collabo- rative research initiative between Telef ́onica’s Radio Access Network Optimization teams and the Universidad Polit ́ecnica de Madrid (UPM). The primary objective is to develop an automated system for managing and monitoring access networks, leveraging data analysis and Machine Learning (ML) techniques to optimize network performance. This project concentrates in studying the behavior of commonly used network Key Performance Indicators (KPIs) under various typical issues, referred to as use cases, with the goal of automatically suggesting appropriate adjustments. Specifically, the thesis focuses on the development of a detection system for cell-range overshoot phenomenon in LTE networks using Nokia telecommunication equipment. Cell- range overshoot occurs when the coverage area of a cell extends beyond its intended boundaries, leading to inefficient network resource usage, degraded signal quality, and disrupted handover procedures. This research, after reviewing basic LTE access procedures and optimization tech- niques, introduces the cell-range case study. To this end, a detailed analysis of commonly used Key Performance Indicators (KPIs) is addressed. Specifically, for this study a carefully designed database of real KPIs collected from the actual Telef ́onica’s Nokia LTE deployment in Spain has been created. This database also includes true labels of cell-range anomalies manually detected by current Telefo ́nica Optimization Teams. Several Machine Learning models have been designed and evaluated to test their capabilities to automatically detect cell-range overshooting. Key findings include the development of primary classification models capable of detecting problematic cells. To this end, tree ensemble models (namely Random Forest and eXtreme Gradient Boosting) are chosen both for their performance and their ability to express the feature importance analysis on which the algorithms build their classification criterion. The models are constructed and trained performing K- fold cross-validation techniques over the KPIs database. These models provide around 70% accuracy in detecting cell-range anomalies and they have been deployed in a real testing probe by Telefo ́nica. This research study highlights the challenging task of automatically detecting specific cellular anomalies, which are notably similar to one another and closely linked to the intrinsic characteristics of access networks. Future research should focus on incorporating additional information specific to cell- range overshooting to enhance detection mechanisms for this particular use case. In summary, this thesis contributes to the ongoing efforts to automate and optimize Telef ́onica’s mobile network management, showcasing the potential of ML techniques to improve network performance and efficiency. The detection solutions developed in this thesis are anticipated to extend seamlessly to Telefo ́nica’s entire network infrastructure, including future technologies like 5G.