AD2T: Multivariate Time-Series Anomaly Detection With Association Discrepancy Dual-Decoder Transformer

Abstract

Multivariate time-series (MTS) anomaly detection is of great importance in both condition monitoring and malfunction identification in multisensor systems. Current MTS anomaly detection approaches are typically based on reconstruction, prediction, or association discrepancy learning algorithms. These methods detect anomalies by learning hidden representations of entire sequences, modeling dependencies at a single time-step level, or calculating an association-based metric inherently distinguishable between regular and deviant points. However, most existing methods typically fail to leverage all three types of models simultaneously to enhance overall performance as well as often disregard the correlations between different sensors. To address the issues above, this article proposes a novel deep learning-based unsupervised MTS anomaly detection algorithm called association discrepancy dual-decoder transformer (AD2T). AD2T employs a dual-decoder architecture to accommodate reconstruction, prediction, and association discrepancy learning tasks, thereby effectively utilizing information across these tasks to better characterize MTS data. We further develop a min-max training strategy to jointly optimize all the aforementioned tasks. Additionally, we propose a compound embedding module based on dilated causal convolution to simultaneously capture correlations in both temporal and sensor dimensions. Extensive empirical studies on five multisensor system datasets from the aerospace, server, and water treatment domains have demonstrated the superiority of our method, achieving an average improvement of 1.96% in the $F1$ -score compared to state-of-the-art (SOTA) methods.