Riding with SK Telecom's AI patrol vehicle that watches over Korea's buried fiber network

On Tuesday afternoon, on a road near SK Telecom's Bundang office in Seongnam, Gyeonggi Province, a monitor inside a bus showed a construction site on one side of the road. An orange rectangle appeared over the site on the small camera feed, and a notification promptly read: "Construction site detected."

Construction work frequently nicks fiber-optic cables buried underground, a mishap that can knock out communications for tens of thousands of households in the surrounding area. That is precisely why detecting construction sites and issuing alerts matters.

The scene was SK Telecom's AI-powered telecommunications infrastructure solution, known as Topda, in action. Over the course of a roughly 3-kilometer drive, Topda analyzed 1,820 images captured by the onboard camera in real time.

SK Telecom has moved to commercialize Topda, a telecom infrastructure safety inspection solution it developed in-house.

The solution works by analyzing footage from AI cameras installed on SK Telecom service vehicles to detect construction sites and other anomalies. It screens for 11 categories of risk, including construction sites, heavy machinery and defects in line facilities.

Telecommunications infrastructure is scattered across roads nationwide — at construction sites, utility poles, cables and manholes. About 200,000 kilometers of fiber-optic cable, enough to circle the Earth five times, runs underground, while roughly 2.35 million utility poles line the roads above. Manual inspection of all of it has clear limits.

To address that, SK Telecom deployed Topda for wired-network inspections. The solution operates in sequence: cameras capture images and footage, relay the data to an AI platform server, and the server analyzes and classifies hazards.

Three AI models developed in-house — a vision AI, a context AI and a spatial AI — handle the analysis. The vision AI classifies 11 types of risk factors, such as the presence of excavators or construction sites, leaning utility poles and sagging cables. The context AI goes further, determining not just whether an excavator is present but whether it is actively digging or being transported on a truck. The spatial AI combines high-precision GPS data with camera footage to plot the location of each hazard on a map.

SK Telecom plans to sharpen Topda's capabilities through continuous data accumulation and machine learning. The ultimate goal is an autonomous network — one that connects field data with AI to detect and respond to anomalies on its own, without human judgment.

The first milestone is lifting the vision AI's detection accuracy from its current 88 percent to 95 percent or higher next year.

The company also plans to expand its fleet of equipped vehicles. SK Telecom ran a pilot with six vehicles from May through August, a roughly four-month trial. Next year it aims to scale up to about 150 vehicles, extending inspection coverage to 70 percent of the network. By 2028, the target rises to more than 90 percent.

More broadly, SK Telecom is applying AI transformation across its entire network operations. Two flagship systems illustrate the effort: A-One, a wireless network operations system, and Spider, an integrated control system for the core network.

A-One forecasts crowd sizes and traffic loads for large events such as concerts and fireworks festivals, cutting network preparation time from a minimum of five days to about 30 minutes. Spider uses AI to analyze the large volume of alarms generated in the core network and recommend the cause of faults and remediation steps. Since Spider's introduction, the actual failure rate has fallen 53 percent and average recovery time has dropped 24 percent.

"We are working to integrate databases across domains to analyze the correlation and impact of alarms," said Bok Jae-won, SK Telecom's head of network operations. "Going forward, we will continue to improve the safety and efficiency of network operations based on AI and field data, and accelerate the transition to an autonomous network."