Contextual anomaly detection method for multi-dimensional time series based on gated recurrent unit

Abstract

Most of the existing anomaly detection methods are only applicable to the anomaly detection in single-dimensional or simple patterns. These methods do not apply to multi-dimensional and complex data. Therefore, a method to solve the anomaly detection of complex time series was proposed, which involves cutting the time series into several sub-sequences and clustering similar sub-sequences to deviate the data from the context. Then, based on gated recurrent unit (GRU), and graph attention network (GAT), the method also includes training its unique prediction model on each category series, and modeling simultaneously at the time level and feature level to improve the credibility of time point prediction. Finally, the predicted value was compared with the actual value, and the threshold value was selected using the 3σ criterion to determine the outlier. The results show that the accuracy of the method on SMD and ASD data sets reaches 94.86% and 92.71% in the context of anomaly.

Keywords

• multi dimensions
• anomaly detection
• gated recurrent unit
• graph attention network
• context anomaly detection