Application of Transformer-based Anomaly Detection in Dam Structural Strong Motion Monitoring Data

Abstract

In the context of structural monitoring systems, the persistent acquisition of high-fidelity data is crucial for ensuring the reliability of system outputs. However, anomalies are inevitably present due to the intricate operational status of sensing and acquisition devices. This study presents a neural network architecture predicated on the Transformer model, tailored for the detection and categorization of anomalies within dam structural health monitoring data. The temporal sequences of the monitoring data serve as direct inputs to the Transformer framework, with feature extraction facilitated through an elaborate six-layer encoder module. The input temporal sequences are eventually categorized into four discrete classes: normal, missing, drift, and burr. The experimental results substantiate the robust efficacy of the proposed approach across binary and multi-class classification paradigms. The deployment of this method on continuous monitoring data underscores its capacity to ensure the correctness of data, thereby bolstering the reliability of dam monitoring systems.