Data Selection to Improve Anomaly Detection in a Component-Based Robot

Abstract

The rise in complexity of robotic systems usually leads to an increase in failures of such systems. To improve the maintenance of this type of systems and thus reducing economic costs and downtime, present paper addresses anomaly detection in a component-based robot. To do so, the problem of anomaly detection is modelled as a classification problem, being Support Vector Machine (SVM) the selected classifier. It is applied to a publicly-available and recent dataset containing useful information about the performance of the software system in a component-based robot when certain anomalies are induced. Different preprocessing strategies and data sources are compared to get the best scores for some classification metrics through cross-validation.