Time-Interval-Driven Sequence Construction with a Transformer-Based Autoencoder for Anomaly Detection in NPP DCS Controller Logs

This study proposes a Time-Interval-Driven Sequence Construction (TIDSC) framework combined with a Transformer-based Autoencoder (Transformer-AE) for anomaly detection in nuclear power plant (NPP) Distributed Control System (DCS) controller logs. The proposed framework enhances log sequence construction by segmenting log streams according to temporal intervals, thereby helping preserve temporally coherent behavioral sequences. Based on the constructed behavior-oriented sequences, the Transformer-AE with a positional-only query decoder is trained using normal operational sequences through reconstruction-based learning and identifies anomalies based on reconstruction errors. Experimental results demonstrate improved performance over fixed-window and sliding-window baselines under the considered experimental settings. These results further suggest that temporal structure-aware sequence construction can improve behavioral representation learning and enhance anomaly separability in reconstruction error space. Overall, the proposed framework provides a potential approach for anomaly detection in the investigated NPP DCS controller log dataset, highlighting the value of integrating temporal sequence modeling with reconstruction-based learning.