Vegayan case study reports early warnings for broadband faults

Models trained on past faults helped prioritize problems in a network with 10 million customer endpoints, but key details needed to assess the reported accuracy were not provided.

Models trained on past faults helped prioritize problems in a network with 10 million customer endpoints, but key details needed to assess the reported accuracy were not provided.
Vegayan Systems presented a broadband fault-prediction system that, it said, gave operators three to four days of warning for potential problems at network aggregation nodes. APNIC Blog described Dr Girish Saraph's presentation at APNIC 62 in a September 30 report.
The system also groups alarms into events, assigns severity levels and ranks incidents by their impact on customers. The aim is to help network operations staff decide which problems need attention first, while identifying some faults before customers are affected. The performance figures came from Vegayan's presentation, rather than an independently verified evaluation.
The example network contained 10 million customer endpoints. Its infrastructure included optical network terminals (ONTs), which connect customers to the fibre network, and gigabit passive optical networks (GPONs).
Saraph described about 200 active issues each day and an average of four to six hours from diagnosis to repair. He said the workload exceeded the capacity of available network operations centre staff. Missing a major outage could also trigger substantial penalties under service-level agreements—the contracts that set an operator's service commitments.
According to the presentation, the platform uses several years of outage and fault records to train specialist machine-learning models: systems that learn patterns from data. Each model focuses on a particular type of problem. The models then process live network data to produce fault predictions and recommendations about which issues to prioritize.
| Network element | Share identified | Reported prediction measure |
|---|---|---|
| Customer access ports | 7.5% had recurring faults | 74% hit rate when predicting reviewed ports |
| GPON aggregation nodes | 5.8% classified as high risk | 65% predictive accuracy |
For GPON aggregation nodes, which combine customer connections, Vegayan reported the three-to-four-day prediction window. The presentation also said operators could better distinguish fibre cuts from hardware problems such as overheating or unstable voltage. That distinction could help staff decide what kind of repair to investigate.
The report did not give evaluation sample sizes, define “hit rate” or “predictive accuracy,” or report false-positive rates—how often the system incorrectly flagged a fault. Without those details, the two percentages cannot be treated as equivalent measures or used to establish that the system reduced outages.
Saraph cautioned against assuming that one machine-learning algorithm would be best everywhere. He recommended that operators test several approaches and choose the model that performs best in their own network.

Geofeeds let address holders publish approximate locations for blocks of internet addresses. SI-assisted checks can flag ambiguous place names and invalid codes, but operators remain responsible for accuracy.
3 min

The change covers all Copilot experiences. Users should update workflows and integrations, while enterprise administrators may need to enable replacement models.
1 min

Every Electric’s batteries charge overnight and power air conditioners during demand peaks. The loan requires a refundable $30 deposit per battery.
3 min