Dynamic deep graph convolution with enhanced transformer networks for time series anomaly detection in IoT

Anomaly detection of multi-time series data during the working process of Internet of Things systems that utilize sensors is one of the key aspects to prevent accidents in industrial information systems. The key challenge is to discover generalized normal patterns by capturing spatio-temporal correlations in multi-sensor data. However, most of the existing studies face the following challenges: (1) Complex topologies and nonlinear connectivity among sensors lack effective characterization methods. (2) Sophisticated correlations among time series need to be mined deeply. Therefore, we propose a novel dynamic deep graph convolution with enhanced transformer networks (DDGCT) for time series anomaly detection. We first construct a dynamic deep graph convolutional network to automatically learn the complex spatial dependencies of sensor data, which introduces \(l_0\) $l_0$ norm with Hard Concrete distribution to further guide the optimization of graph structure in graph learning. Meanwhile, we devise a new transformer model to deeply mine temporal dependencies from time-series data by designing a new positional encoding coupled with patch design as well as channel independence constraint. Then, DDGCT fuses and optimizes the captured temporal and deep spatial features using attention networks. Finally, anomaly scores are efficiently computed by prediction methods with threshold-based approaches to detect anomalies. Extensive experiments on real datasets show that DDGCT outperforms several state-of-the-art methods.