ISTAD: Scaled Transformer with Interactive Learning for Multivariate Time Series Anomaly Detection

Anomaly detection of multivariate time series is of great significance in establishing an efficient anomaly detection system. However, the current research does not take into account the characteristics of different time scales, namely, short-term mode and long-term patterns. This paper proposes a method based on Interactive Scaled-Transformer Anomaly Detection (ISTAD) to solve the above problems. The main idea is to cross the original data sequence to obtain two sub-sequences, and then learn feature sharing in the interactive learning structure. At the same time, the scaled transformer module is embedded to extract the long-term correlation of the time sequence interactively. In addition, the Gated Recurrent Unit (GRU) is introduced to capture the short-term correlation of the time sequence. In terms of model design, a joint optimization strategy based on a forecasting-based model and a reconstruction-based model is adopted, and the fusion of single timestamp prediction and reconstruction of the entire time sequence represent information sequences. We conducted experimental verification on the widely used benchmark datasets, and the results showed that our method was better than other methods and reached the most advanced performance level.