Empirical Evaluation of Machine Learning Models for Predictive Maintenance of Base Transceiver Stations in the South West Zone

Reactive maintenance of base transceiver stations results in unplanned outages that degrade network quality and generate emergency repair costs substantially above those of planned preventive maintenance. Machine learning-based predictive maintenance offers a data-driven approach to anticipating equipment failures before they occur, reducing outage duration and operational expenditure. This study empirically evaluates the performance of machine learning models for predictive maintenance of base transceiver stations using operational telemetry data from network sites in the South West geopolitical zone. A dataset of 36 months of equipment health telemetry from 180 base stations across Lagos, Ogun, Oyo, Osun, Ondo, and Ekiti States, obtained in collaboration with a major tower company, was used to train and validate random forest, gradient boosting, and long short-term memory neural network models for failure prediction. Features included power system parameters, rectifier efficiency metrics, battery health indicators, and environmental temperature logs. The gradient boosting model achieved the highest failure prediction accuracy at 89.4% with a false positive rate of 11.2% on the holdout test set. LSTM models demonstrated superior performance for multi-step ahead failure predictions of seven or more days, making them more useful for maintenance scheduling. Implementation simulation demonstrated that model-guided predictive maintenance scheduling could reduce unplanned outage duration by an estimated 41% and cut emergency maintenance costs by 33% relative to the current time-based preventive maintenance approach. Keywords: predictive maintenance, machine learning, base station, South West Nigeria, telemetry.