Machine Learning for Supervision in Computer Networks

The widespread adoption of telecommunications devices has significantly increased the complexity of network infrastructures, pushing providers to install a large number of devices to ensure communication over long distances. While having fully resilient networks would be ideal, the reality is that the network infrastructure requires maintenance and intervention from the team in charge. The high volume of alarms generated by the infrastructure of service providers poses a significant challenge, requiring the development of automated methodologies to support maintenance operations. This thesis explores the application of Machine Learning and Deep Learning techniques for the supervision and management of telecommunications networks. We focused our analyses on a specific and impactful failure type that potentially affects multiple nodes and cause severe downtime to end users. The aim of this project is to develop a machine learning algorithm capable of automatically identifying in real-time such fault type and enable timely intervention by the maintenance team, limiting service outage as much as possible. In particular, first, alarms are grouped together by means of a clustering algorithm. Then, each cluster of alarm is analyzed by the aforementioned machine learning model to identify whether or not it was caused by the previously cited failure. In this context, a cluster represents a set of alarms identified through spatial and temporal analysis, where each group is associated with a common root cause. Developed in collaboration with Fibercop S.p.A., the methodology is based on an autoencoder model on data provided by the maintenance team. The compressed representation of each cluster, generated by the encoder, is used to train a binary classification model to determine whether the cluster was caused by the severe fault under analysis. In addition, this thesis investigates the application of clustering and outlier detection techniques to refine clusters of alarms. The results indicate that the developed model successfully identifies clusters associated with the specific failure under analysis. However, it also tends to include unrelated clusters, suggesting potential areas for further refinement and development.