DBAD: A Dual-Branch Time Series Anomaly Detection Method Based on Transformer and MLP
Yang Sun, Ning Zhang, Chun Zhang, M. Wang, Chenhao Shi · 2025
This paper proposes a novel anomaly detection method, DBAD, which is a dual-branch time series anomaly detection method based on transformer and MLP. DBAD employs channel-independent prediction and channel-mixing reconstruction methods to better capture the internal dependencies of data. The dual-branch structure of DBAD enhances the stability of anomaly detection, while the anomaly point fusion strategy improves the model's generalization ability. Experimental results on multiple datasets demonstrate that DBAD outperforms several state-of-the-art methods in terms of anomaly detection accuracy and stability.