Disentangled Anomaly Detection For Multivariate Time Series

Anomaly detection in time series that aims to identify unusual patterns has attracted a lot of attention recently. However, the representation of abnormal and normal data is difffcult to be distinguished because they are usually entangled. Recently, disentanglement theory based on variational auto-encoder (VAE) has shown great potential in machine learning and achieved great success in computer vision and natural language processing. In this paper, we propose a novel disentangled anomaly detection approach that adopts VAE-based disentanglement networks for anomaly detection in multivariate time series. The proposed method learns highquality disentangled latent factors in a continuous representation space to facilitate the identiffcation of anomalies from normal data. Extensive experiments demonstrate that our proposed lightweight model DA-VAE achieves state-of-the-art performance.