A hybrid approach for fault detection in autonomous physical agents

One of the challenges of fault detection in the domain of autonomous physical agents (or Robots) is the handling of unclassified data, meaning, most data sets are not recognized as normal or faulty. This fact makes it very challenging to use collected data as a training set such that learning algorithms would produce a successful fault detection model. Traditionally unsupervised algorithms try to address this challenge. In this paper we present a hybrid approach that combines unsupervised and supervised methods. An unsupervised approach is utilized for classifying a training set, and then by a standard supervised algorithm we build a fault detection model that is much more accurate than the original unsupervised approach. We show promising results on simulated and real world domains.