EST transformer: enhanced spatiotemporal representation learning for time series anomaly detection

Anomaly detection in time series data is of great importance across various domains, which aims to identify outliers in a sequence of time series data generated from various sensors or system operations that deviate significantly from the norm. Systems are often not composed of a single variable, and the detection of anomalies in multidimensional time series has attracted more attention from researchers and engineering practitioners. Multidimensional time series not only exhibit strong temporal characteristics but also have the inter-correlations between dimensions are extremely significant. But existing reconstruction-based models for multivariate time series anomaly detection often struggle to differentiate between normal and anomalous sequences during the reconstruction process. To address these issues, in this paper, we propose an novel method, the Enhanced Spatiotemporal Transformer (EST Transformer), which improves the conventional self-attention mechanism by incorporating state information to strengthen the model’s sensitivity to contextual information. Our proposed method amplifies the differences between the normal and anomalous sequence segments based on the state discrepancies within the context. Additionally, the proposed method employs variable self-attention to model dimensional correlations, thereby enhancing the model’s sequence reconstruction capabilities. The effectiveness of the proposed method has been verified on four publicly available datasets. And the experimental evaluations demonstrate that the proposed EST Transformer outperforms the various established baseline models across multiple benchmark datasets in terms of anomaly detection accuracy.