Modeling Voice Traffic Patterns for Anomaly Detection and Prediction in Cellular Networks Based on CDR Data

The available telecommunication services nowadays make connecting users easier. In return, vast streams of data are generated every day. However, mining useful information in data requires relevant techniques and procedures. In this article, a novel study is proposed consisting of a multi-algorithm approach to understand, detect and predict anomalies in cellular networks. The study holds 37 million Call Detail Records data from a cellular network with deployment of the latest network generations including 5G. The research is divided into two phases. In the first phase, we outline the voice traffic profile, where utilizing certain algorithms are meant to target specific attributes and scenarios in the data to understand the typical voice traffic patterns. Gaussian Mixture model is used to define the regular groups of call duration and Mean Shift clustering algorithm is employed to detect the peak hours on a daily basis. In the second phase, we deseasonalize the data for higher accuracy followed by the distribution function to comprehend the patterns in the data. We introduce three algorithms to detect and predict anomalies in the cellular network. The performance evaluation shows that DBSCAN and Isolation Forest algorithms provide the highest accuracy with 98% compared to the Z-score algorithm.