Unsupervised Anomaly Detection and Localization for Multivariate Time Series and Their Applications in Structural Health Monitoring

Abstract

This thesis advances the field of anomaly detection in multivariate time series by addressing key challenges in anomaly detection, localization, and severity assessment. Through the development of EdgeConvFormer and U-GraphFormer, this research offers robust, interpretable, and efficient solutions applicable across various domains, with a particular focus on SHM. Extensive evaluations across diverse multivariate time series datasets and real-world scenarios demonstrate the potential of these models to enhance the monitoring and maintenance of critical systems, ensuring their safety and longevity.