Multi-Scale Sampling Based MLP Networks for Anomaly Detection in Multivariate Time Series

Haojie Li, Hongzuo Xu, Wei Peng, Chiran Shen, Xianwen Qiu

Abstract: Multivariate time series data are produced in various domains, including AIOps, space crafts, and healthcare. Identifying anomalies within them is significant to ensure the stability of target systems. Current anomaly detectors mainly focus on devising intricate network structures based on recurrent neural networks, Transformers, or graph neural networks to model the temporal and inter-variate dependencies of the input time series. Nevertheless, complex models often lead to computational burden in training and inferencing stage. Also, system operators have to understand the technical details of these complex models to fine-tune them or add new modules for different needs. This motivates us to consider an intriguing question: can we construct an anomaly detection model merely based on simple networks? In this paper, we propose FlightAD, a light but effective anomaly detection model using only multi-layer perceptron (MLP) networks. FlightAD applies MLP networks on top of a multi-scale sampling strategy, followed by an informationfusing mechanism to fuse the learned features. More specific, the multi-scale sampling strategy is used to extract abundant temporal patterns of time series, and the MLP blocks applied to it can simultaneously model the dependencies along the time axis and dependencies between different sampled sub-sequences (i.e., different variables). Finally, we wield an informationfusing module to fully integrate the learned multi-scale features. Extensive experiments on six real-world datasets show that our model outperforms seven state-of-the-art competitors on average by 2.1%-18.2% in F 1 score and 3.8%-38.9% in AUC-PR.