Anomaly detection in self-organizing industrial systems using pathlets

In this paper, we present a novel anomaly detection method which addresses the main challenge of self-organizing industrial systems: the state space explosion. In particular, the flexibility and dynamic nature of such systems result in an exponentially growing number of possible execution plans. To handle this problem, we propose to learn the underlying topology, instead of storing whole paths a work-piece can take through the factory. Therefore, we use the concept of pathlet learning. With it, the topology is represented by a pathlet dictionary, which contains significant sub-paths which have been extracted in a pre-processing step from a training data set. These sub-paths can then be used to evaluate at runtime the incoming trajectories. We show that with this approach we are able to detect both, global anomalous events, like the fail of a production station, as well as single anomalous trajectories, e.g. work-pieces which moves out of the known paths.