Detecção de barras quebradas em motores de indução usando imagens geradas a partir das perturbações de fluxo no entreferro

Abstract

During the last few decades, induction motors have been used in a variety of industrial and commercial applications, particularly due to their robust construction and relatively simple maintenance. Today, induction motors are responsible for a large amount of energy consumption in industries (about 85%) and capable of moving many loads during energy conversion processes. Although three-phase induction motors are reliable machines, various types of failures can affect their structure or their operating conditions, such as bearing failures, air gap eccentricity, stator winding failures and broken rotor bars. It is known that, normally, failures in the rotor bars are related to 10% of the total failures in the induction motor, and this type of operating condition leads to a reduction in the life cycle of the motor. Furthermore, when this type of failure occurs, the consequences can be disastrous, not only for the machine, but also for the production process associated with it, and for this reason, the present study focused on the development of an approach capable of detecting the rupture of rotor bars, to avoid high intervention costs and maximize adjustments in the preventive maintenance schedule. To this end, the purpose of this research is quantitative, post-positivist, exploratory in its beginning and later evolved into an experimental and applied phase, to achieve the proposed objectives. For this study, a Hall effect sensor was invasively used on the stator coils so that variations in the magnetic flux density could be monitored, with the occurrence of a defect, mainly on the rotor bars. To extract the defect and healthy features from the images generated from the signal coming from the sensor, different computer vision techniques and machine learning techniques with convulsive neural networks were used. Tools such as TensorFlow and the Keras API were used to develop convolutional neural networks diligently and accurately. The computational model obtained at the end was able to detect the referred defect, with good accuracy, for different operating conditions of the engine, from its no-load condition to the full load condition, in many cases without the need to apply the erosion technique for image generation.