Superb AI, a Vision Intelligence company, said on Sept. 1 it jointly developed an AI model with the National Fire Research Institute to detect electric vehicle fires early in underground parking garages.

The company said the model focuses on capturing faint smoke in the early stage of battery thermal runaway. Research results showed detection time was shortened by 65 percent and the false alarm rate fell by 80 percent compared with existing fire detection methods.

Unlike existing systems trained on publicly available data, the new model is specially designed to recognise faint smoke in underground parking garages and from the underside of electric vehicles, including smoke from a long distance.

Superb AI first processed 100,000 images of fire, smoke and vehicle objects to build a training dataset. It combined data from actual fire incidents with experimental data from controlled environments, and supplemented rare cases that do not occur often with synthetic data created using generative AI.

During dataset construction, the company managed quality by checking the distribution and patterns of fire and smoke data using the Superb platform's Scatter View, a feature that spreads data across a coordinate plane by similarity to visually confirm distributions and outliers.

The model is implemented as a real-time monitoring system that runs on edge devices. It analyses video on site without using the cloud, can be deployed on existing CCTV infrastructure, and issues immediate alerts with on-screen warnings and audible alarms when a fire is detected while keeping a log history. Users can also check detection status immediately without an external server by accessing a local web server from a PC on the same network.

The two sides are conducting the final round of research and aim to optimise the model so it can run in real time even on small edge devices, completing it as a field-deployable version. Once the research is completed, they plan to derive quantitative testing standards such as recognition within a certain number of seconds at a certain distance, which is expected to be used to set performance standards for future electric vehicle fire detection systems.

Kim Hyun-soo (김현수), Superb AI's CEO, said, "This research fills field-specific characteristics that cannot be learned from public data with a specialised dataset and synthetic data, and is a case of running the model in real time on edge devices." He added, "We will continue to lead AI adoption from data construction to model development and field verification."