Dual-domain transformer with co-attention mechanism for time series anomaly detection and diagnosis

초록

Multivariate time series anomaly detection remains challenging as it requires capturing both temporal dependencies and inter-variable relationships. In this paper, we propose a dual-domain transformer encoder for anomaly detection and diagnosis (DTE-AD), which jointly models the time and frequency domains using a co-attention mechanism to enhance robustness in detection. The time-domain encoder captures temporal dependencies, while the frequency-domain encoder leverages the 2D discrete Fourier transform (2D-DFT) to model correlations among variables. By integrating these complementary representations through co-attention, DTE-AD achieves up to a 10% improvement in anomaly detection and a 13% improvement in variable-level interpretability compared with state-of-the-art baselines across six benchmark datasets. These results demonstrate that integrating complementary domain representations significantly improves the reliability of complex multivariate time series analysis. © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2026.

키워드