Spatial-Temporal Interaction Decoding Transformer for Unsupervised Multivariate Time Series Anomaly Detection

Abstract

Time series data consists of a temporal dimension and features associated with each timestamp. Anomaly detection in this context necessitates the consideration of both temporal and spatial features. However, existing work focuses on separately addressing temporal and spatial features, neglecting the interactive features between them. In this paper, we aim to leverage spatial-temporal interaction and propose a Spatial-Temporal inTeraction Decoding (STTD) model for time series anomaly detection. First, we employ the parallel transformer encoder to capture temporal dependencies at various scales and spatial dependencies among variables. Second, we propose a parallel transformer decoder with cross-attention to fuse spatial-temporal features. Moreover, we also utilize channel-attention to aggregate spatial features for better fusion. Experimental results on eight public datasets show that STTD outperforms state-of-the-art models, which shows the effectiveness of capturing spatial-temporal interaction.