Fault detection and isolation in robotic systems via artificial neural networks

Abstract

Faults in robotic manipulators can cause economic losses and serious damages. In the paper, two artificial neural networks are employed to provide FDI to robotic manipulators. The first is a multilayer perceptron trained with backpropagation utilized to reproduce the dynamic of the manipulator and, so, generate the residual vector. The second is a radial basis function network employed to classify the residual vector and, thus, generate the fault isolation. As the system model is not employed, false alarms due to modeling errors are avoided. Two different algorithms are employed to train the last network. The first employs ridge regression (a regularization type) and the second uses forward selection (an algorithm for subset selection). Simulations in a two link manipulator evince that the FDI system can detect and isolate correctly faults that occur in nontrained trajectories.