Anomaly Detection of Periodic Multivariate Time Series under High Acquisition Frequency Scene in IoT

Anomaly Detection of Multivariate Time Series is an intensive research topic in data mining, especially with the rise of Industry 4.0. However, few existing approaches are taken under high acquisition scene, and only a minority of them took periodicity of time series into consideration. In this paper, we propose a novel network Dual-window RNN-CNN to detect periodic time series anomalies of high acquisition frequency scene in IoT. We first apply Dual-window to segment time series according to the periodicity of data and solve the time alignment problem. Then we utilize Multi-head GRU to compress the data volume and extract temporal features sensor by sensor, which not only solves the problems caused by high acquisition but also adds more flexible transfer ability to our network. In order to improve the robustness of our network in different periodic scenes of IoT, three different kinds of GRU mode are put forward. Finally we use CNN-based Autoencoder to locate anomalies according to both temporal and spatial dependencies. It should also be note that Multi-head GRU broadens the receptive field of CNN-based Autoencoder. Two parts of experiment were carried to verify the validity of Dual-Window RNN-CNN. The first part is conducted on UCR/UEA benchmark to discuss the performance of Dual-Window RNN-CNN under different structures and hyper parameters, for datasets in UCR/UAE benchmark contain enough timestamps to monitor the high acquisition and periodicity in IoT. The second part is conducted on Yahoo Webscope benchmark and NAB to compare our network with other classic time series anomaly detection approaches. Experiment results confirm that our Dual-Window RNN-CNN outperforms other approaches in anomaly detection of periodic multivariate time series, demonstrating the advantages of our network in high acquisition scene.