Finding Anomalous eNodeBs
Faidi, Sameh (2018)
Metropolia Ammattikorkeakoulu
2018
All rights reserved
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2018082014570
https://urn.fi/URN:NBN:fi:amk-2018082014570
Tiivistelmä
A typical telecommunication operator could easily have over 10,000 eNodeBs. Huge amount of data and logs are collected from these network elements in daily basis. Existing tools are used to analyze the collected data and generate reports about the status and health of the network. The amount of reports and the information in each is overwhelming which make it virtually impossible for the maintenance team to find problems in their radio network. This lead to many cases where problems go undetected which might result in service degradation and eventually revenue loss.
The objective of this thesis was to use machine leaning, particularly anomaly detection, to rank the eNodeBs in the order of their probability of being anomalous using KPIs data. Maintenance teams can save time by focusing on the short list of top anomalous eNodeBs. By performing further investigation and analysis on the anomalous network elements, maintenance teams will be able to apply any required changes and fixes before problems escalate and cause service degradation and eventually loss of revenue.
Three different anomaly detection methods were applied to a selected sub set of the KPIs time series, HW related KPIs. The methods and their results were compared and evaluated based on their advantages and disadvantages. The result of this thesis shows that the distance based with custom references method is the most suitable for detecting anomalous eNodeBs as it requires the least number of hyper parameters and it does not seem to be sensitive to the choice of selected threshold.
The objective of this thesis was to use machine leaning, particularly anomaly detection, to rank the eNodeBs in the order of their probability of being anomalous using KPIs data. Maintenance teams can save time by focusing on the short list of top anomalous eNodeBs. By performing further investigation and analysis on the anomalous network elements, maintenance teams will be able to apply any required changes and fixes before problems escalate and cause service degradation and eventually loss of revenue.
Three different anomaly detection methods were applied to a selected sub set of the KPIs time series, HW related KPIs. The methods and their results were compared and evaluated based on their advantages and disadvantages. The result of this thesis shows that the distance based with custom references method is the most suitable for detecting anomalous eNodeBs as it requires the least number of hyper parameters and it does not seem to be sensitive to the choice of selected threshold.