Unsupervised Anomaly Detection in Time Series with Convolutional-VAE

Emanuele La Malfa¹, Gabriele La Malfa²

1. University of Oxford, Oxford, UK.
2. EMLYON Business School, Paris, FRANCE

Abstract: We propose an unsupervised machine learning algorithm for anomaly detection that exploits self-learnt features of monodimensional time series. A Variational Autoencoder, where convolution takes place of dot product, is trained to compress each input to a low-dimensional point from a normal distribution, detecting an anomaly as low probability and high density sequence. We validate our work on different public datasets, obtaining results that shed new light on Variational Autoencoders applied to anomaly detection.

Pages: 103 – 108 | Full PDF Paper