Graph Transformer Network Incorporating Sparse Representation for Multivariate Time Series Anomaly Detection

• DOI: 10.3390/electronics13112032
• Publication Date: 2024-05-23T13:04:25Z
• Authors: Qian Yang; Jiaming Zhang; Junjie Zhang; Cailing Sun; Shanyi Xie; Shangdong Liu; Yimu Ji
• Keywords: 0202 electrical engineering, electronic engineering, information engineering; 02 engineering and technology

Cyber–physical systems (CPSs) serve as the pivotal core of Internet Things (IoT) infrastructures, such smart grids and intelligent transportation, deploying interconnected sensing devices to monitor operating status. With increasing decentralization, surge in sensor expands potential vulnerability cyber attacks. It is imperative conduct anomaly detection research on multivariate time series data that these sensors produce bolster security distributed CPSs. However, high dimensionality, absence labels real-world datasets, intricate non-linear relationships among present considerable challenges formulating effective algorithms. Recent deep-learning methods have achieved progress field detection. Yet, many either rely statistical models struggle capture or use conventional deep learning like CNN LSTM, which do not explicitly learn inter-variable correlations. In this study, we propose a novel unsupervised method integrates Sparse Autoencoder with Graph Transformer network (SGTrans). SGTrans leverages for dimensionality reduction reconstruction high-dimensional series, thus extracting meaningful hidden representations. Then, are mapped into graph structure. We introduce multi-head attention mechanism from structure learning, constructing forecasting module. This module performs attentive information propagation between long-distance nodes complex temporal dependencies them enhance prediction future behaviors. Extensive experiments evaluations three publicly available datasets demonstrate effectiveness our approach.