A domestic research team has developed artificial intelligence (AI) technology capable of predicting seismic damage to nuclear power plants using a single sensor. This technology enables real-time estimation of the vibration status of key facilities within nuclear power plants without the need for hundreds of existing sensors, significantly improving inspection efficiency and safety.

Professor Young-Joo Lee's research team from the Department of Earth, Environmental, and Urban Construction Engineering at UNIST (Ulsan National Institute of Science and Technology) and Dr. Jae-Beom Lee's research team from the Non-Destructive Measurement Group at the Physical Measurement Center of the Korea Research Institute of Standards and Science (KRISS) announced on September 30 that they have jointly developed an AI model that predicts seismic acceleration responses at 139 points within a nuclear power plant auxiliary building using a single sensor.

This model takes a single seismograph input and produces vibration responses at 139 locations in just 0.07 seconds. Acceleration response is an indicator of the intensity of shaking that seismic waves cause to equipment, allowing for quick identification of equipment requiring priority inspection.

While previous systems required the installation of hundreds of sensors, this AI model can act as a virtual sensor, dramatically reducing maintenance costs. This is particularly effective in environments where sensor installation is limited, such as radiation control zones.

The research team designed a model using a residual block-based one-dimensional convolutional neural network (CNN). It was composed of six blocks to enable learning of a wide range of vibration patterns, from low to high frequencies, of seismic waves. Its performance was verified using actual earthquake records (NGA-West 2).

Prediction accuracy is also excellent. In noise-free conditions, the error rate was only 0.44-0.59%, and even in a 10dB noise environment, the error range remained low, around 4%. The model even learned noise removal on its own, demonstrating stable performance even in real-world driving environments.

Computational efficiency has also been proven. A model comprised of approximately 2 million parameters can predict 30 seconds of seismic response in just 0.07 seconds on an NVIDIA RTX 4090 GPU, making it suitable for real-time monitoring.

This technology has significant economic and safety benefits, reducing downtime due to nuclear power plant inspections and mitigating the risk of radiation exposure for workers. The research team emphasized that it is "a next-generation nuclear power plant safety monitoring technology that can overcome the limitations of building physical sensor networks."