Anomaly Detection in A Multivariate DataStream in a Highly Scalable and Fault Tolerant Architecture

Sammanfattning: The process of monitoring telecommunication systems performance by investigatingKey Performance Indicators (KPI) and Performance Measurements(PMs) is crucial for valuable sustainable solutions and requires analysts' interventionwith profound knowledge to help mitigate vulnerabilities and risks.This work focuses on PMs anomaly detection in order to automate the processof discovering unacceptable Radio Access Network (RAN) performance byleveraging K-meansjj algorithm and producing an anomaly scoring mechanism.It also oers a streaming, fault tolerant, scalable and loosely coupled architectureto process data on the y based on a normal behavior model. Theproposed architecture is used to test the anomaly scoring system where variousdata patterns are ingested. The tests focused on inspecting the anomaly score'sconsistency, variability and sensitivity. The results were highly impacted by thereal-time standardization process of data, and the scores were not entirely sensitiveto changes in constant features; however, the experiment yielded acceptableresults when the correlation between features was taken into account.