Publikation
Detection of Rare Fault Cases for Mobile Robot Applications
Benjamin Blumhofer; Jonas Weigand; Leonhard Kunz; Pascal Rübel; Achim Wagner
In: Achim Wagner; Kosmas Alexopoulos; Sotiris Makris (Hrsg.). Advances in Artificial Intelligence in Manufacturing. European Symposium on Artificial Intelligence in Manufacturing (ESAIM-2023), Cham, Pages 61-70, ISBN 978-3-031-57496-2, Springer Nature Switzerland, 2024.
Zusammenfassung
Modern production and transport systems consist of many individual mechanical components, sensors, and actuators. Despite the complexity, critical errors must be diagnosed quickly to avoid the propagation of faults and the associated safety risks and production downtime. The problems are incomplete descriptions of the systems and disturbed sensor signals, which makes it difficult to use model-based methods directly. In this article, a method is presented that is based on the description of dynamical systems using Runge-Kutta Neural Networks (RKNN) as nominal models. RKNNs have a good ability to predict dynamic processes and can be trained with comparatively little data. Deviations from the real process behavior to the nominal behavior lead to generic fault symptoms classified by a second neural network. The fault cases are purely modeled with the help of the process data without providing special knowledge of the internal system behavior. The procedure was implemented and tested on a mobile robot for factory-internal transport processes. With little training data, a binary classification for elementary fault detection. The method can easily be adapted to other cases of dynamical systems.