Mobile network faults prediction using machine learning

In the last years, the world has faced a significant growth of network data traffic performed by mobile devices. While users rely always more on wireless cellular networks, the service's reliability expected by customers increases day by day. This study investigates the impact of using machine learning and data mining techniques for improving the availability of the base stations, the core of a mobile network. In particular, we design a framework for applying the predictive maintenance on the air conditioners of indoor base stations. In this research, we describe the entire process done to design the final model: study the context, retrieve, clean, transform and integrate the data and, lastly, predict the failures and evaluate the results. From a business perspective, the trade-off between hit-rate and false alarms is something crucial. It is a key feature avoiding to perform lots of useless maintenances for only preventing few breakdowns. For being effective, a maintenance schedule should not only prevent the failures but also having operating costs as low as possible. We tested five different classification algorithms (Naive Bayes, Decision Tree, XGBoost, Adaboost and Random Forest) against a baseline, comparing their performances over 47 weeks with a sliding window validation technique. The results were evaluated with several metrics: for each model, we scored the confusion matrix, the ROC curve and the F-measure. In particular, this latter was used to select the approach with the most balanced outcome between precision and recall, necessary condition for improving the baseline. Results indicate that the proposed framework, using XGBoost and Random Forest, can be effectively used for predicting the failures of mobile network's base stations.