Variational Autoencoder based Novelty Detection for Real-World Time Series

There are numerous applications that deal with data captured over time making them potential subject to time series analysis. Detecting unknown events and anomalies in time series data is challenging due to the presence of noise, seasonalities and long-term trends. Data-driven methods applied to identify such patterns are called anomaly detection. Typically, the amount of available abnormal data, e.g. failure states in manufacturing plants, is not sufficient to construct an explicit model. Novelty detection is a special form of anomaly detection which detects when a data point differs from the majority of data. In this work, a novel approach to detect anomalous patterns and events in real-world time series data is proposed. The novelty detection approach is based on deep generative learning and utilizes natural properties of the variational autoencoder to create a novelty indicator. Therefore, a wide range of deterministic and stochastic novelty scores calculated in the latent and original data space are combined. The combination of these novelty scores leads to a novelty indicator that accurately detects novel events in real-world time series data. An experimental study evaluates the method on data collected at an electrocoating plant which was affected by internal and external disturbances. The proposed method is benchmarked against other state of the art methods and achieves highly competitive results.