Neural Network to Failure Classification in Robotic Systems

Abstract

A robotic system is a reconfigurable element, and in
its programming, an algorithm can be implemented in order to
detect and classify failures. This is an important step to ensure
that errors in actions do not cause damage or bring risks.
Considering this, a Neural Network Multi Layer Perceptron
(MLP) was used, in order to classify a set of failures in robot
actuators, present in a database. This purpose is to analyze if
robotic failures could be classified by MLP. The raw data are
divided in a temporal progression manner and torque in x, y and
z axes. In total, five MLP neural networks were implemented for
each type of failure classification, using two different topologies.
The number of neurons in the hidden layer is in accord with the
criteria of Kolmogorov and Weka, being the latter the best
topology for such application. In comparison to an algorithm
(SKIL) using the same set of data, the MLP obtained the best
performance in any topology of classification, with hit rates in
80 to 90%.