Skip to content
All projects
2026

Root-Cause Analysis for a 5G Network Operations Centre

I owned root-cause analysis in an 8-person 5G Academy proof of concept: from alarms, performance readings and the network map, a learned model ranks every candidate cause and rule checks confirm one before it is claimed.

17 to 185

Nodes per simulated network

Problem
One fault in a mobile network sets off alarms on many elements at once, and operators need the element that actually failed. A wrong confident answer is worse than an honest "not sure", because operators may act on it.
My role
Team of 8, 5G Academy (Federico II, with Nokia, TIM and PagoPA). I owned the root-cause analysis stage; the simulator, network model, alarm stage, backend and security module are the team's.
Result
On a sealed simulated hold-out (read once on 20 Oct 2026, networks of 17 to 185 nodes), it named the true root cause first on 497 of 710 single-fault runs (70.0%), against 137 (19.3%) for the team's earlier ranker, and sent zero wrong causes with a recommended action.

Honest limits

  • Simulated runs only; the earlier ranker cannot name 3 fault classes and is read strictly, so "70% vs 19%" overstates the gain over a fair baseline.
  • 213 of the 710 single faults got no claimed cause, and with several faults at once only 499 of 933 events were found.
  • Faults that only the proposed simulator extension can inject were found 22 to 29% of the time. The code is in the team's private repository.

Overview

Final project of the 5G Academy postgraduate programme at Federico II: an AI-assisted network operations centre built by an 8-person team, where I owned the root-cause analysis component; the simulator, network model, alarm stage, backend and security module are the team's. Given correlated alarms, KPI time series and the network topology, a learned model ranks every candidate cause and rule checks confirm one before it is claimed; an unconfirmed or uncertain answer becomes "investigate" instead of a recommended action. On a sealed simulated hold-out (H-final, read once on 20 Oct 2026, networks of 17 to 185 nodes) it named the true root cause first on 497 of 710 single-fault runs, against 137 for the team's earlier ranker, and sent zero wrong causes with a recommended action. Simulated runs only. The code lives in the team's private repository, so there is no public source or demo.

Tech stack

PythonLightGBMNetworkXpandasPyYAMLJSON Schema

Private team code

This was built inside a team's private repository, so there is no public source or demo. Message me on LinkedIn for details.