Redes neurais convolucionais para classificação de sinais de emissão acústica de falhas em motores de indução trifásicos

Abstract

Three-phase induction motors (TIM) play a fundamental role in the industry due to their simple construction, uncomplicated maintenance, and economical operation. However, due to wear and tear, the environments in which they are used, and power quality issues, these motors are prone to failures, which can result in unexpected shutdowns. Given the consequences of these failures in industrial processes, it is crucial to identify them to minimize damage. This work proposes an alternative methodology to detect faults in MIT, using acoustic emission (AE) signals and convolutional neural networks (CNN). The methodology encompasses the classification of faults related to the rotor, stator, and bearing, as well as tests related to the classification of electrical unbalance faults. Different CNN topologies were compared, with the topology performing best achieving a validation accuracy of 98.67% in the multiclass classification related to rotor, stator, and bearing faults.