Machine learning based call drop healing in 5G

Abstract: Self-Organizing Network (SON) functions include self-configuration, dynamic optimization and self-healing of networks. In the era of 5G, mobile operators are increasingly exploring areas of SON through Machine Learning (ML) techniques. It is seen that 5G packet switched networks are often hit with radio link failures, an important Key Performance Indicator (KPI). Reasons for a dropped call ranges from a failed handover to coverage/capacity issues. In current networks, such issues are resolved by KPI analysis, but these metrics are not always service/user specific. The aim with this master’s thesis work is to investigate how well ML techniques can be applied to predict a call drop in real-time networks. In the thesis, ML techniques, namely neural networks and logistic regression were used to classify the link status. Initially, the parameters which characterize a link connection, e.g. the Reference Signal Received Power (RSRP), Block Error Rate (BLER) and similar parameters were investigated. This was followed by applying ML to the selected parameter(s) and classifying a bad link (with failure) from a good link (without failure), this was the first phase of the thesis. The next phase was forecasting a radio link failure before one occurs. This forms phase two of the thesis and the start of the self heal process where, counter measures could be taken to avoid a radio link failure. Counter measures for self-healing was not covered in the thesis. This thesis only focuses on phase one and two of predicting a radio link failure.