Overview of the virtualizing scheme of physical sensors in nuclear power plants. Credit: Computer-Aided Civil and Infrastructure Engineering (2025). DOI: 10.1111/mice.70051
Electrical equipment housed within auxiliary buildings of nuclear power plants—such as control panels, transformers, and emergency generators—are particularly vulnerable to vibrations. Notably, during the 2016 Gyeongju earthquake, while concrete structures remained intact, power facilities had to be shut down for safety inspections. Now, a new technology allows for quick identification of equipment requiring maintenance without the need for extensive manual inspections.
Researchers from the Department of Civil, Urban, Earth, and Environmental Engineering at UNIST, led by Professor Young-Joo Lee, in collaboration with Dr. Jaebeom Lee from the Division of Physical Metrology at the Korea Research Institute of Standards and Science, developed an advanced artificial intelligence (AI) model capable of estimating seismic responses at 139 detailed points within auxiliary buildings of nuclear power plants.
This AI model can analyze seismic data captured by a single high-quality sensor and, within 0.07 seconds, predict the acceleration response at 139 locations inside the building. These acceleration responses indicate how strongly and quickly equipment is shaken during an earthquake, helping prioritize which areas and devices require urgent inspection.
While installing hundreds of sensors would be necessary for direct measurement, this AI acts as a virtual sensor, accurately estimating responses across multiple points without physical sensors. This significantly reduces maintenance costs and minimizes disruptions to plant operations.
The research team designed the AI with six residual convolutional blocks, enabling it to learn a wide range of vibration patterns—from slow ground motions to rapid tremors. As a result, the model can precisely estimate both large-scale movements of the entire structure and amplified vibrations near critical equipment. The findings are published in the journal Computer-Aided Civil and Infrastructure Engineering.
In tests without noise, the model's prediction errors were as low as 0.44–0.59%. Even under artificially induced noise at 10dB, it maintained a low error rate of around 4%. Validation against real earthquake records (NGA-West 2 dataset) confirmed the model's reliability, even under seismic conditions set by safety standards used in Korea and the United States for nuclear facilities.
Professor Young-Joo Lee commented, "This technology drastically reduces inspection time, operational downtime, and maintenance costs for nuclear plants. Particularly in radiation-controlled zones where sensor installation and upkeep are highly restricted and expensive, this solution offers a fundamental improvement."
