Cellular Network Antenna Tilt Anomaly Detection Using Federated Unsupervised Learning

An important issue for cellular network operators is how to maintain radio coverage so that any issues can be addressed before they impact user services. This is particularly important in dense small cell network deployment scenarios such as vehicular networks. Antenna electrical tilt is a key factor in this, as unintended deviations from the planned value can adversely affect coverage and service reliability. We propose a novel method to detect antenna tilt anomalies using existing data sources without the need for additional hardware to be deployed in the radio access network. Our approach goes beyond previous techniques by using federated unsupervised learning based on polar coordinates, together with a geometrical transformation to normalise data across multiple sites. By using this approach to combine scarce training data from multiple cells, we can achieve detection accuracy in excess of 95% in a way that minimises training data size as well as computing power and memory usage.