DE-SVDD: dual-encoder SVDD for multivariate time-series anomaly detection

Abstract

Reconstruction-based multivariate time-series anomaly detection often relies mainly on reconstruction error, leaving the learned representation insufficiently constrained for normal-boundary modeling. We propose DE-SVDD, a dual-encoder framework that separates reconstruction-oriented representation learning from one-class boundary modeling. By imposing the SVDD constraint in a dedicated boundary space rather than directly on the reconstruction representation, DE-SVDD reduces the gradient conflict and scale imbalance between the two objectives. The framework re-encodes reconstructed sequences through a lightweight second encoder and a compact projection head, while a multi-scale context block captures temporal dependencies at different scales. Reconstruction error and center distance are then combined for anomaly scoring. Experiments on six public datasets show that DE-SVDD achieves the best F1 scores on MSL, SMAP, SMD, and PSM, reaching 0.5546, 0.5495, 0.5544, and 0.6382, respectively.