Virtual sensing of seismic floor responses for rapid prioritization of critical equipment inspection in nuclear power plants

Abstract

Critical equipment in nuclear power plant auxiliary buildings such as control cabinets, panels, transformers, and diesel generators often malfunction before structural damage occurs, demanding rapid post-earthquake inspection prioritization. However, direct walkdown inspection or dense sensor networks are impractical due to restricted accessibility in radiological zones and the high costs associated with maintenance. To address this, we propose a residual convolutional network-based virtual sensing framework that supports urgent inspection prioritization by predicting acceleration at 139 locations from a single high-quality seismometer. The model employs six residual blocks with progressively downsized kernels to capture multi-scale features, while skip connections prevent vanishing gradients. Trained on artificial earthquakes with 10 dB noise and validated against unseen Next Generation Attenuation-West 2 ground motions matched to Nuclear Regulatory Commission Regulatory Guide 1.60 and Korean uniform-hazard spectra, the model achieves a maximum mean absolute percentage error of 0.44%-0.59% for noise-free case and <= 4.23% at 10 dB, demonstrating robust generalization. The resulting rapid, noise-tolerant virtual sensor network enables actionable equipment-level decision making in nuclear facilities at a fraction of conventional monitoring cost.