Automatisches Training eines Variational Autoencoder für Anomalieerkennung in Zeitreihen

This paper addresses automatic anomaly detection without a machine learning expert and instead builds on readily available computing power. Therefore, we train variational autoencoders, a machine learning model based on deep neural networks, for anomaly detection automatically, without expert knowledge on the data source using Bayesian optimization. This model works with typical input types (binary, discrete & continuous, time dependency) and is thus applicable in all standard environments, which we test in ten different industrial time series. We conclude that the approach is suitable for practical and automatic anomaly detection.