Mixer-transformer: Adaptive anomaly detection with multivariate time series

Abstract

Anomaly detection is crucial for maintaining the stability and security of systems. However, anomaly detection systems often generate numerous false positives or irrelevant alerts, which obscure genuine security threats. To both reduce false positives in time series detection and accurately identify the source of anomalies, leveraging artificial intelligence techniques has emerged as a promising solution. These techniques can analyze strong temporal correlations and dynamic variations across different data frames. Existing detection methods face two primary challenges leading to false positives or negatives: (i) detecting anomalies in multivariate time series requires accounting for both temporal dependencies and complex interactions between variables; and (ii) traditional fixed-threshold approaches often struggle to adapt to dynamic environments. To address these issues, this paper proposes an anomaly detection method based on the Mixer-Transformer architecture. By combining the Mixer model with the Anomaly Transformer, the proposed method effectively captures global dependencies by alternately modeling interactions along both the channel and time dimensions, thereby enhancing its ability to extract complex spatiotemporal features. Additionally, an adaptive threshold update mechanism is employed to dynamically adjust the anomaly detection criteria in response to data fluctuations. The F1 scores on three real-world datasets — SMAP, MSL, and PSM — are 97.49%, 95.18%, and 98.20%, respectively. These results demonstrate that the proposed method outperforms existing technologies in reducing false positives and enhancing the detection accuracy of multivariate time series anomaly detection.