DAEJEON — Artificial intelligence will help nuclear power plants set inspection priorities after an earthquake, enabling faster safety checks.
A joint research team led by Lee Jae-beom, a senior researcher at the Korea Research Institute of Standards and Science (KRISS), and Lee Young-joo, a professor at the Ulsan National Institute of Science and Technology (UNIST), has developed a deep-learning technology that predicts shaking at multiple points inside a nuclear plant in real time using signals from just one seismometer, KRISS said on the 26th. The technology quantifies the likelihood of danger and allows inspection priorities to be assigned.
When a magnitude-5.8 earthquake struck Gyeongju in 2016, units 1 through 4 of the Wolsong nuclear plant were restarted only after roughly 80 days of performance testing and detailed inspections. Similarly, after a strong earthquake hit Kumamoto, Japan, in July this year, TSMC's Kumamoto Fab 1 temporarily halted operations and returned to normal in stages following equipment checks and adjustments.
Even when an earthquake causes no major damage, the process of confirming safety can prolong shutdowns, resulting in significant time and economic burdens.
The KRISS researchers developed a virtual sensing technology in which AI analyzes seismic wave signals measured by a single seismometer to infer, in real time, the vibration responses at 139 points inside a nuclear plant that have no sensors. The approach can quickly narrow down the areas most affected by an earthquake without installing sensors at every location, and it showed strong predictive performance even on records from actual earthquakes not used in training.
The team also accounted for uncertainty arising from factors such as structural characteristics, quantifying on a scale of 0 to 100% the likelihood that the vibration response at each location will exceed a predetermined danger threshold. Rather than simply classifying locations as "safe" or "dangerous," the system ranks them from the highest probability of danger, helping experts quickly determine inspection priorities.
The researchers also sharply improved the efficiency and versatility of the AI design. By developing a design formula that derives the optimal AI structure based on a structure's natural frequency, they reduced the trial and error of repeatedly designing and comparing multiple models.
The AI model designed this way achieves accuracy comparable to that of the latest deep-learning models more than 200 times larger by parameter count, while substantially lowering the computational burden and improving its usability even in constrained environments.
Because the AI model can be flexibly designed to match a structure's vibration characteristics, it is expected to be broadly applicable to major industrial facilities such as semiconductor plants and data centers in the future.
"In the field of safety, AI must not only make accurate predictions but also recognize the limits of its own judgment," said Lee Jae-beom of KRISS. "We will continue to develop trustworthy AI technology that flags uncertain situations and supports additional human judgment."
The findings were published in three papers, including one in Reliability Engineering & System Safety (IF: 13.7), an international journal ranked in the top 1.4% in the field of civil engineering.
The work was carried out jointly by the research teams of senior researcher Lee Jae-beom at KRISS and professor Lee Young-joo at UNIST, with graduate researcher Lee Jin-gu as first author, supported by programs including a basic research grant from the National Research Foundation of Korea and KRISS's core project. The findings were published in three papers, including one in Reliability Engineering & System Safety (IF: 13.7), an international journal ranked in the top 1.4% in the field of civil engineering.
