Visual MAE with Normalizing Flow for Time Series Anomaly Detection
A new method, VAN-AD, adapts a visual Masked Autoencoder (MAE) pretrained on ImageNet to time series anomaly detection (TSAD). The approach addresses the limited generalization of existing TSAD models, which require per-dataset training and struggle with cross-modal gaps or in-domain heterogeneity. By leveraging large-scale vision models, VAN-AD aims to improve performance in scenarios with scarce training data. The paper, arXiv:2603.26842, investigates the applicability of vision models to TSAD, proposing a normalizing flow to bridge the modality gap.
Key facts
- VAN-AD adapts a visual MAE pretrained on ImageNet to TSAD.
- Existing TSAD methods require training one model per dataset.
- Foundation models are a promising direction for TSAD.
- Current approaches repurpose LLMs or build large time series datasets.
- Cross-modal gaps and in-domain heterogeneity remain challenges.
- The paper is arXiv:2603.26842.
- The method uses normalizing flow.
- The approach targets scenarios with scarce training data.
Entities
Institutions
- arXiv