Weakly Augmented Variational Autoencoder in Time Series Anomaly Detection

TL;DR

This paper introduces a novel VAE-based framework with weak augmentation and self-supervised learning to improve time series anomaly detection, especially under data scarcity conditions, by enhancing latent space robustness.

Contribution

It proposes a new generative model combining VAEs and SSL with weak augmentation to address latent holes in data-scarce anomaly detection tasks.

Findings

Improved robustness in anomaly detection under limited data.

Enhanced latent space continuity and reconstruction quality.

Better detection accuracy compared to existing methods.

Abstract

Due to their unsupervised training and uncertainty estimation, deep Variational Autoencoders (VAEs) have become powerful tools for reconstruction-based Time Series Anomaly Detection (TSAD). Existing VAE-based TSAD methods, either statistical or deep, tune meta-priors to estimate the likelihood probability for effectively capturing spatiotemporal dependencies in the data. However, these methods confront the challenge of inherent data scarcity, which is often the case in anomaly detection tasks. Such scarcity easily leads to latent holes, discontinuous regions in latent space, resulting in non-robust reconstructions on these discontinuous spaces. We propose a novel generative framework that combines VAEs with self-supervised learning (SSL) to address this issue.