Scientists Find Way to Speed ​​Up Fire Response in Energy Sector Tenfold

Scientists at Gubkin University (National Research University) have developed an artificial intelligence (AI)-based system capable of reducing fire response times at energy sector facilities tenfold. Researchers estimate that implementing the system could cut fire-related damage by 58%.

As part of the project, the scientists shifted the paradigm of AI application in industrial safety from mere "detection" to "meaningful analysis." Previously, neural networks were used only for binary classification (fire present vs. absent) or basic detection. The Gubkin University researchers developed an AI system that can assess the severity of an incident and rapidly suggest a course of action to the operator.

The neural network analyzes camera footage and assigns a hazard classification to the incident. The system was tested using images of fires that occurred at an actual energy sector facility; operator response time dropped from a couple of minutes to ten seconds.

The neural network is built on a vision-language model architecture capable of simultaneously analyzing an image and describing it.

"In other words, the model doesn't just say 'fire'; it answers questions like: 'What type of incident is this?', 'How dangerous is it?', and 'What exactly should the operator do?'" explains project participant and Gubkin University PhD student Andrey Evsikov.

Operator consoles are continuously flooded with signals from various sensors and video surveillance cameras. It can take several minutes before an incident is analyzed, and the data often requires verification. The new neural network-based system instantly filters out false positives and generates a detailed report on actual threats. This saves operators from wasting time manually reviewing low-priority false alarms, as they receive a ready-made, structured analysis of significant incidents. Consequently, response times are reduced to an average of ten seconds.

"The time it takes to detect a fire and for the fire brigade to arrive directly impacts successful evacuation, casualty reduction, property damage minimization, and fire containment within buildings. Studies from last year showed that if a fire is detected within 10.8 seconds and quickly extinguished, the risk to human life drops by 75.6%, and material damage is reduced by 58%," said Andrey Evsikov.

As part of the project, scientists found a way to eliminate so-called neural network "hallucinations"—a pressing issue in AI responses where the model generates false data due to insufficient information without flagging the questionable facts.

"We crafted a prompt that eliminated 'hallucinations.' To achieve this, we imposed a strict response framework on the model, mandating objectivity: it must rely solely on visible evidence and avoid inferring or fabricating additional information. Tests confirmed that the quality of the prompt has a greater impact on the outcome than the model's architecture itself," the researcher explained.

The university is implementing this "AI firefighter" project with the support of companies in the fuel and energy sector.

Researchers estimate that implementing the new system at a single facility will cost a maximum of 500,000 rubles. It does not require an internet connection, thereby eliminating risks associated with data leaks or unauthorized external access. The neural network can be integrated into existing security systems, where it will operate alongside traditional sensors. According to Andrey Evsikov, the system’s development potential extends beyond fire detection and fuel and energy sector facilities. The neural network will be capable of operating in warehouses, manufacturing plants, airports, seaports, transport hubs, shopping malls, and other high-occupancy venues with stringent fire safety requirements. A broader industrial analytics system is already being developed based on this technology; it will be able to monitor equipment usage, employee compliance with personal protective equipment (PPE) protocols, and other aspects that cannot be anticipated or programmed using traditional methods.