Adapted Anomaly Detection for RAN Performance

Abstract

Complexity of mobile radio access networks is increasing rapidly. At the same time, customers’ expectations for the service quality are keeping the pressure on operators. It requires huge human efforts to maintain a sustainable and high- performance radio access network. To meet the highest quality and service level requirements, more and more network management and maintenance functions are required to be passed to automated solutions. There are many already - like SON, MLB, automatic antenna tilting which helps to scale operations. Those are dedicated to improving specific functions, but they do not pay much attention to service KPIs, which are a direct reflection of customers’ experience. Most often, changes in network performance KPIs indicate some interruption or deterioration of service quality, which needs to be restored as soon as possible. Automated solutions based on machine learning process enables engineers to notice those network issues much faster. There are many algorithms that have been created for this purpose, but most of them are specialized and tuned to detect an anomaly in the special datasets, like fraud, abnormal CPU usage, earthquake, cyber-attack. Consequently, it is either not applicable or not performing well while identifying anomalies on telecommunication network performance data. The research on this paper proofs anomaly detection accuracy problem, lack of functions for telecommunication data in existing R libraries. The only R “changepoint” package has a functionality to detect not a single local anomaly, but also anomalous time series, which is the most relevant in RAN (radio access network) performance monitoring. Default function meanvar() of method – PELT, penalty - BIC returns up to 30 false positive anomalies in retainability and integrity KPI and up to 10 in mobility KPI group on one month time series data. To extract true positive anomalies additional algorithm is applied with a cost of false negative detection increase up to 10% in a single time series data and occurs not more than in 5% of used datasets. These results are compared against other popular R libraries, like “Anomalize”, “AnomalyDetection”, “bcp”, “changepoint” and proofed to be the most accurate on radio network performance data. This adapted anomaly detection algorithm can be treated as a scalable generic algorithm to any kind of high-dimensionality data as it does not require additional tuning for different KPI group or technology. It is the solution for automated network performance monitoring that can take off a lot of workload from daily network engineer routine and improve the quality while tackling network performance issues.