Fault Modeling to Determine the Reliability Status of Rotating Machines Using Deep Learning Methods Based on Vibrations from Acoustic Emissions from Cooling Fans

Abstract

Modern industrial production acknowledges the increasing significance of maintenance. As of right now, maintenance is seen as a service that aims to maintain the effectiveness of systems and installations while adhering to quality, energy efficiency, and protection standards. An inventive technique to automate rotating machine maintenance procedures has been created in this study. To identify failures and flaws in the motors through their supports, where the fan blades are attached, a technique based on capturing the noises produced by their cooling fans and utilizing deep learning to diagnose problems was investigated. Two operational circumstances were envisioned: the absence of fault and the presence of fault. The machine is correctly powered and running in ideal circumstances when it is not having any issues. In contrast, failures were gradually created purposefully and then documented in order to better understand the faults. Utilizing a pre-trained network (SqueezeNet) built on the ImageNet database, the convolutional neural network (CNN)-based technique was constructed. Applying transfer learning to the spectrograms obtained from the sound emission recordings of our machine's fan in both working modes demonstrated outstanding performance (accuracy = 0.987), confirming the methodology's outstanding quality.