Classificador não supervisionado baseado em curvas principais para detecção de falhas em motor de indução

Abstract

Electric motors are highly versatile equipment, possessing an immense range of industrial applications. Therefore, they are extremely important in any industrial plant and their maintenance is crucial for their quality operation in production, safety to employees and without environmental damages. In this paper a method for fault detection is proposed by means of vibration analysis, based on structural health monitoring. Vibration signals of a three-phase induction motor were collected using a 3-axis accelerometer sensor controlled by an Arduino microcontroller. After the collection, the extraction of features through the 2nd, 3rd and 4th order cumulants with lag zero was performed. Finally, a classifier is designed using principal curves. Principal curves are a non-linear generalization of Principal Component Analysis and have the advantage of presenting good data representation capability in one dimension. A one-class learning method based on Principal Curves is proposed to generate a decision border where the data within it represents the data of non-failed motor and out of it, the data referring to the failed motor. The method presented low computational cost and high detection rates, reaching up to 100% using real data from an induction motor.