Abstract:
Wireless cellular networking in the world goes through a tremendous structural change where many advances in technology find an opportunity to present themselves for assistance. 5G cellular network, the most recent generation wireless network cur rently undergoing implementation, welcomes artificial intelligence with the novel net work data analytics function (NWDAF). NWDAF is a data analytics mechanism where other components of 5G can request information from in order to utilize their oper ations. In this thesis, the structure and protocols of NWDAF are described. A 5G network data set is generated by using the fields obtained from the technical specifi cation documents provided by 3rd Generation Partnership Project (3GPP). To bring the generated data set closer to reality, randomly created anomalies are added. Sev eral machine learning (ML) algorithms are trained to study two aspects of NWDAF, namely network load prediction and anomaly detection. Linear regression (LR), re current neural network (RNN) and long-short term memory (LSTM) algorithms are implemented and trained using the generated data set and a data set obtained from a real enterprise network for network load prediction [1, 2]. Mean absolute error and mean absolute percentage error performance metrics indicate that RNN and LSTM outperform LR in both generated and real life data sets. LSTM is the best perform ing algorithm for the real life data set. Logistic regression and a tree-based classifier, XGBoost are implemented for anomaly detection, and trained using the generated data set to maximize the area under receiver operating characteristics curve. The re sults indicate that tree-based classifier XGBoost outperforms logistic regression. These predictions are expected to assist 5G service-based architecture through NWDAF to increase its performance.
