Time series anomaly detection with Variational Autoencoder using Mahalanobis distance

Abstract

Two themes have dominated the research on anomaly de- tection in time series data, one related to explorations of deep architec- tures for the task, and the other, equally important, the creation of large benchmark datasets. In line with the current trends, we have proposed several deep learning architectures based on Variational Autoencoders that have been evaluated for detecting cyber-attacks on water distribution system on the BATADAL challenge task and dataset. The second research aim of this study was to examine the impact of using Mahalanobis distance as a reconstruction error on the performance of the proposed models.