AI Can Spot Methane Leaks That the Human Eye Cannot Detect
Researchers at the University of Tyumen (UTMN) have proposed using a combination of infrared cameras and artificial intelligence to detect gas leaks. The technology could serve as an additional intelligent layer in industrial safety systems.
Researchers at UTMN's School of Computer Science have developed a system that can detect leaks by analyzing data from infrared cameras. Its neural network identifies temperature anomalies: when gas escapes, the surrounding air cools, and the cameras capture the resulting temperature changes.
The neural network was trained on examples generated under laboratory conditions. More than 2,500 images of gas leaks were fed into the YOLOv11 Segmentation Medium and YOLOv8s models. The images covered a range of conditions, with varying light levels, camera angles and distances. The hybrid neural network architecture can both recognize objects and track changes over time. This helps filter out visual interference, including shadows and glare.
The researchers tested the trained system with 1,500 new images depicting different leak scenarios. The AI successfully detected gas leaks, including small ones.
Beyond detection, the AI also alerts an operator, assesses the severity of the situation and recommends response procedures, including the sequence for switching valves and notifying management when necessary. The system also features a web interface where users can upload video footage with timestamps marking detected leaks. Responsibility for decisions and subsequent actions, however, remains with human personnel, not the AI.
From a Laboratory Test Rig to Gas Fields and Power Plants
Testing under laboratory conditions, even when they closely replicate real-world scenarios, is only the first step. The system now needs to be tested at an industrial site. Gas fields, gas compressor stations and other facilities that use methane contain large amounts of equipment requiring continuous visual monitoring. These complex installations include pipelines, shutoff valves, pumps and connections. Comprehensive video surveillance is not feasible everywhere, so the technology developed by the Tyumen researchers could complement existing gas analyzers and safety systems.
Russian companies have already developed a range of infrared equipment, including devices designed to detect methane leaks. UTMN researchers will therefore also need to ensure that their system is compatible with equipment already available on the market.
Potential customers for the Tyumen-developed system include companies in the oil and gas and energy sectors. Most combined heat and power (CHP) plants, for example, use natural gas. The technology could also find applications in municipal utilities.
Visual Monitoring as a Foundation for AI Adoption
The idea of using cameras and algorithms to detect gas leaks remotely has already been put into practice. In 2021, researchers at R&D MSTU developed a system that uses video surveillance to monitor coke oven gas. Its neural network also detects gas emissions and alerts an operator. In 2022, teams from Messoyakhaneftegaz and Gazprom Neft tested laser, ultrasonic and infrared scanning equipment at the East Messoyakha field. In 2023, students at Ryazan State Radio Engineering University designed a hexacopter equipped with an infrared camera and machine vision technology to detect gas leaks from pipelines.
Leak Detectors Become More Effective When Paired With AI
Although industry still relies primarily on conventional leak detectors, which can be integrated with digital safety systems, combining video surveillance with AI offers considerable potential. Over the next two to three years, the researchers will need to develop a hardware and software prototype and test it at operating industrial facilities. If the system proves effective, potential investors could include both manufacturers of industrial safety equipment and major companies in the oil and gas and energy sectors. The Tyumen technology is not intended to replace existing systems entirely. Instead, it could add an intelligent layer that improves their reliability.
