Machine-Learning-Assisted Failure Prediction in Microwave Networks based on Equipment Alarms

Modern microwave networks must cope with strict Quality of Services (QoS) requirements, such as low latency, high bandwidth and high availability. As network failures can affect service availability, failure management is crucial for service maintenance and, recently, application of Machine Learning (ML) for automated failure management is becoming pervasive. In particular, ML promises to deliver predictive maintenance capabilities, where failure occurrence is anticipated thanks to ML prediction capabilities. In this study we developed two workflows, based on a modular ML implementation, capable of short- and long-horizon failure predictions, while taking into consideration computational complexity constraint. As input data, we used real alarms coming from deployed equipment of a nation-wide microwave network. Our ML-based failure-prediction system learns from human experience through labelled data, performs alarms forecasting, detects future failure occurrence and identifies failure root causes. In our numerical results, we compare the prediction performance of different ML models in terms of various standard ML performance metrics. Overall accuracy over 95% is achieved in all prediction scenario simulated within an hour, suggesting that microwave network operators can gain actual operational benefits by deploying this framework in realworld infrastructures.