Sobre o desempenho de algoritmos de aprendizado de máquinas na detecção de falhas em motores de indução trifásicos: um estudo comparativo

Abstract

This dissertation aims at the detection of short-circuit incipient fault condition in a threephase squirrel-cage induction motor fed by a sinusoidal PWM inverter. In order to detect this fault, a test bench is used to impose different operation conditions to an induction motor, and each sample of the data set is taken from the line currents of the PWM inverter aforementioned. For feature extraction, the Motor Current Signature Analysis is used. The detection of this fault is treated as a classification problem, therefore different supervised algorithms of machine learning are used so as to solve it: Ordinary Least Squares, Singlelayer Perceptron, Multi-layer Perceptron, Extreme Learning Machine, Support-Vector Machine, Least-Squares Support-Vector Machine, the Minimal Learning Machine, and Gaussian Classifiers. Together with Reject Option technique, these classifiers are tested and the results are compared with other works that use the same data set. Maximum accuracy rates of 100% with Support-Vector Machine and Least-Squares Support-Vector Machine classifiers suggest that, in near future, an embedded system can be developed with these algorithms.