VAN-AD: Visual Masked Autoencoder with Normalizing Flow For Time Series Anomaly Detection
Time series anomaly detection (TSAD) plays a crucial role in ensuring the reliability and security of Internet of Things (IoT) enabled service systems. Traditional methods for TSAD typically require the development of a model tailored to each specific dataset, which significantly limits their generalization capabilities across different scenarios. This limitation poses challenges in effectively detecting anomalies when training data is scarce. Recent advancements in foundation models have opened up new possibilities, yet existing approaches often either adapt large language models (LLMs) or create extensive time series datasets, which can lead to difficulties stemming from cross-modal gaps and in-domain heterogeneity.
In an effort to overcome these challenges, researchers have turned their attention to the potential of large-scale vision models for TSAD applications. In this context, a novel approach known as VAN-AD has been introduced, which leverages a visual Masked Autoencoder (MAE) that has been pretrained on ImageNet, specifically adapted for the TSAD task. However, transferring MAE directly to TSAD introduces two significant challenges: overgeneralization and restricted local perception.
Challenges in Time Series Anomaly Detection
- Overgeneralization: This occurs when the model fails to accurately identify anomalies due to its tendency to generalize too broadly, resulting in missed detections.
- Limited Local Perception: The model may not effectively capture local patterns in the data, which are essential for identifying anomalies in time series.
The VAN-AD Approach
To address the aforementioned challenges, the VAN-AD framework introduces two innovative components. First, the Adaptive Distribution Mapping Module (ADMM) is designed to tackle the overgeneralization problem. This module effectively maps the reconstruction results from the MAE into a unified statistical space, thereby amplifying discrepancies that are indicative of abnormal patterns.
Second, the Normalizing Flow Module (NFM) is implemented to enhance local perception. By integrating the MAE with normalizing flow, this module estimates the probability density of the current time series window in relation to the global distribution, thus improving the model’s ability to detect nuanced anomalies.
Experimental Validation
Extensive experiments conducted on nine real-world datasets have demonstrated that VAN-AD consistently outperforms existing state-of-the-art methods across a variety of evaluation metrics. The results indicate that the innovative components of VAN-AD significantly enhance both the detection accuracy and robustness of time series anomaly detection.
For those interested in further exploring this research, the authors have made their code and datasets publicly available at https://github.com/PenyChen/VAN-AD.
In conclusion, the VAN-AD framework represents a significant advancement in the field of time series anomaly detection, promising improved performance and generalizability across diverse datasets.