Probabilistic Adaptive Graph Transformer for Uncertainty-Aware Multivariate Time Series Anomaly Detection

Multivariate time series anomaly detection (MTAD) is important for ensuring reliable operation and improving service quality in industrial systems. Forecasting-based methods have been a primary approach for MTAD, detecting anomalies by assuming that anomalous data points produce higher forecasting errors than normal ones. However, existing forecasting-based MTAD methods often have limited anomaly detection performance because they generally overlook forecast uncertainty, leading to overconfident detections. In this study, we propose a probabilistic forecasting-based MTAD method that evaluates forecasting errors by comparing observed values with their predictive distributions, reflecting forecast uncertainty, and uses these errors as anomaly scores. Our approach determines the predictive distribution using a deep state-space model, capturing stochasticity in multivariate time series through nonlinear stochastic state transitions. Additionally, we integrate an adaptive graph transformer to capture complex intra- and inter-variable dependencies through a multi-head self-attention mechanism and graph convolutions with a learnable graph, respectively. Experiments demonstrate that our method outperforms current state-of-the-art techniques.