VACE: Self-Supervised Anomaly Detection for Multivariate Time Series
VACE (Velocity-Aligned Channel Embeddings) is a self-supervised anomaly detection method for multivariate time series, introduced in a paper on arXiv (2605.23504). The method addresses the challenge of learning a characterization of normality without labels, using contrastive approaches to embed temporal patches into a latent space. Unlike standard contrastive methods that shape the space indirectly through pair-sampling heuristics, VACE provides explicit control over geometric structure, ensuring tight grouping of normal representations and directionally meaningful distances. This improves distance-based scoring for anomaly detection. The paper is a cross-type announcement, indicating it may be submitted to multiple venues.
Key facts
- VACE stands for Velocity-Aligned Channel Embeddings.
- It is a self-supervised anomaly detection method for multivariate time series.
- The paper is available on arXiv with ID 2605.23504.
- Announcement type is cross.
- The method uses contrastive learning to embed temporal patches.
- VACE provides explicit control over geometric structure of latent space.
- It aims to improve distance-based anomaly scoring.
- The approach addresses the challenge of rare, unlabeled anomalies.
Entities
Institutions
- arXiv