PengYu Chen, Shang Wan, Xiaohou Shi, Yuan Chang, Yan Sun, Sajal K. Das

Researchers propose VAN-AD, a framework that adapts visual Masked Autoencoders pretrained on ImageNet for time series anomaly detection in IoT systems. The approach addresses two core challenges when transferring vision models to time series: overgeneralization and limited local perception, using an Adaptive Distribution Mapping Module and Normalizing Flow Module to improve detection accuracy. Testing on nine real-world datasets shows consistent improvements over existing methods, suggesting vision foundation models may offer better generalization for anomaly detection across diverse datasets with limited training data.