Detecting anomalies in robot time series data using stochastic recurrent networks

Abstract

This thesis proposes a novel anomaly detection algorithm for detect-ing anomalies in high-dimensional, multimodal, real-valued time se-ries data. The approach, requiring no domain knowledge, is based on Stochastic Recurrent Networks (STORNs), a universal distribution approximator for sequential data leveraging the power of Recurrent Neural Networks (RNNs) and Variational Auto-Encoders (VAEs). The detection algorithm is evaluated on real robot time series data in order to prove that the method robustly detects anomalies off- and on-line.