Raphael Golombek, Sebastian Wrede, Marc Hanheide, Martin Heckmann, "A Method for learning a Fault Detection Model from Component Communication Data in Robotic Systems", Proc. 7th IARP Workshop on Technical Challenges for Dependable Robots in Human Environments, 2010.

A promising means to increase the dependability of a robotic system is to equip it with the ability to autonomously monitor it own system state and detect faults. In this contribution we propose a method for fault detection in robotic systems which exploits the concept of anomaly detection and learns a model based on dynamics in the system’s internal exchange of data. Learning a model reduces the need for expert system-knowledge and enables on-line adaptation. Furthermore, communicated data as learning input enables the detection of subtle system failures such as resource starvation. The method in this contribution is applicable during runtime and can be used in an a-posteriori analysis of the system. The evaluate of the method takes place on a mobile robotic platform employed in human robot interaction scenarios.