Unsupervised Multi-Head Attention Autoencoder for Multivariate Time-Series Anomaly Detection

Cyber-Physical Systems (CPSs) are driving innovation and advancement in various industry sectors such as manufacturing, logistics, and chemical processes. These systems are susceptible to security attacks, and indeed, serious hacking incidents are on the rise. Various anomaly detection methods have been attempted to mitigate such threats in CPSs, and further developments are essential. In this study, we propose a novel unsupervised deep learning-based anomaly detection approach for multivariate time-series data called Unsupervised Multi-head Attention Autoencoder (UMAA). This architecture can focus on the relationships between time-series data and effectively learn the temporal characteristics of the data. we evaluate the performance of UMAA on four real-world CPS datasets with state-of-the-art models. Our experimental results show that UMAA can achieve higher F1 scores compared to other models across all datasets. Specifically, UMAA improves the F1 score by 26% and 12 % over the best-performing baseline model on the WADI and SMAP datasets, respectively.