Modeling, normality characterization and anomaly detection in manipulator robots

---

This projects aims at developping methodology for fault detection and RUL estimation for manipulator robots. The approach mixes knowledge based model with data-driven sub-models in order to characterize the normal beehaviour under a wide range of context of use. The data driven model are of two kinds: Recurrent neural networks (GRU) on the one hand and piece-wise multivariate polynomial on the other hand. The features used in the model incorporate the temperature that showed very critical in the derivation of sufficiently precise model. The results are promizing on the rather artificial set of incidents and failures we incorporated. The collection of real failures data is undergoing with the large set of Staubli’s robots customers.