Inproceedings,
Qualitative Bayesian Failure Diagnosis for Robot Systems
International Conference on Intelligent Robots and Systems, AI and Robotics, page 1 -- 6. Chicago, IEEE, (2014)
Abstract
Reliability is a key challenge for intelligent robot systems. In order to address this challenge, runtime failure detection and diagnosis (FDD) is an essential task to maintain autonomous operation. The complexity of fully fledged robot systems and the included noise in system observations complicate this task. In this paper, we present our Qualitative Bayesian Failure Diagnosis (QBFD) for precise and robust failure estimation. Our approach uses a Dynamic Bayesian Network to model uncertainties of the measurements while considering temporal relations. Instead of detailed a priori knowledge of system dynamics, our approach models cause-effect relations. These relations are, in practice, more intuitive to specify. As a consequence, we reduce the level of needed system knowledge and therefore increase the practical applicability. We evaluate the quality in respect to two reference approaches in extensive simulations. Due to our results, we are confident that our proposed approach provides comparable, if not superior, estimation quality, while simultaneously reducing the level of needed model details. Furthermore, we provide evidence that, given a proper system decomposition, high quality estimates are possible using general observations, like the resource usage.
