Unsupervised Anomaly Detection of Anchor Bolt Loosening in NPP Electrical Cabinets Using a 1D-CNN Autoencoder

Seismic Category I electrical cabinets performing safe shutdown functions in nuclear power plants (NPPs) must maintain anchorage integrity during earthquakes; however, the scarcity of actual failure data limits the application of conventional supervised learning. Therefore, this study proposes an unsupervised 1D-CNN autoencoder framework to detect anchor-bolt loosening using exclusively normal-condition data. For this purpose, a dataset comprising one normal condition and four loosening scenarios (40, 30, 20, and 10 N⋅m) was constructed through full-scale shaking-table tests conducted in accordance with NRC guidelines. Acceleration time-series data were transformed into spectrograms using STFT (Short-Time-Fourier-Transform), and the model was trained with SGD to minimize the reconstruction mean squared error (MSE). Anomaly scores were derived from the reconstruction MSE, and optimal thresholds were determined using Youden's J statistic. The proposed framework achieved an AUC of 0.831 and a recall exceeding 0.95 across all scenarios. These results demonstrate its practicality for structural health monitoring of NPP electrical cabinets, even in the absence of failure data.