VQ-GCAE: Vector-Quantized Graph Convolutional Autoencoder for Early Anomaly Detection in Multivariate Time Series

GwanPil Kim, Jason J. Jung

Abstract: Early detection of anomalies in multivariate time series is essential for ensuring the stability and safety of industrial systems. However, in environments where sensor interactions are complex and normal states are irregular and non-periodic, existing methods often fail to effectively model dynamic dependencies or distinguish subtle anomalies. To address this issue, this paper proposes VQ-GCAE, a graph convolutional autoencoder that jointly learns sensor relationships and patterns of normal behavior. The proposed model estimates dynamic inter-sensor dependencies from data through a trainable adjacency matrix, enabling it to sensitively capture structural anomalies. In addition, vector quantization is applied to compress the latent representations of normal states into a discrete codebook, which clarifies the boundary between normal and abnormal patterns and amplifies minor deviations as reconstruction errors. This design improves both the sensitivity and accuracy of early anomaly detection. Experiments on the SKAB benchmark demonstrate that VQ-GCAE outperforms existing methods, achieving F1-score and NAB improvements of approximately 2.07% and 9.94% respectively, and reducing FAR by approximately 5.47%, confirming its effectiveness under complex conditions.

External IDs:dblp:conf/ideal/KimJ25