Eventos Anais de eventos
COBEM 2023
27th International Congress of Mechanical Engineering
FAULT DETECTION IN BEARINGS BASED ON CONVOLUTIONAL NEURAL NETWORKS USING LOW-COST MEMS ACCELEROMETERS
Submission Author:
Caique Movio Pereira de Souza , SP
Co-Authors:
Lucas Almeida Willenshofer, Caique Movio Pereira de Souza, Rogério Daniel Dantas, Rene Oliveira, Vanessa Seriacopi, Wilson Carlos Silva Junior
Presenter: Lucas Almeida Willenshofer
doi://10.26678/ABCM.COBEM2023.COB2023-1271
Abstract
The detection of bearing fault signals plays a vital role in the industrial field, directly affecting the performance and reliability of mechanical equipment. In most cases, the acquisition of fault signals is done using expensive commercial accelerometers, making it unfeasible for applications that can easily be implemented in serial equipment. Convolutional neural networks (CNN) have emerged as an emerging method for fault detection, but this method does not work well with one-dimensional data. Therefore, this work proposes to evaluate and compare the use of two low-cost MEMS accelerometers for obtaining the temporal vibration signal, which will be converted into 2D images and applied to a CNN in order to learn more subtle characteristics of bearing signals, thereby improving the accuracy of the signal classification. To this end, an experimental platform was built in which the individually encapsulated MPU 6050 and ADXL 345 accelerometers were attached to the selected bearing that presented normal condition scenarios, outer race defect, and rolling element defect. The images of the vibrational signatures of the measurements taken were input into the network with three different dimensions: 16x16, 22x22, and 28x28, with the intention of evaluating the influence of the image dimension on algorithm accuracy. In the scenarios evaluated for the MPU6050 sensor, the proposed algorithm achieves an accuracy of 97% for 16x16 images, 98% for 22x22 images, and 99% for 28x28 images. For the ADXL345 sensor, the proposed algorithm achieves an accuracy of 98% for 16x16 images, 99% for 22x22 images, and 99% for 28x28 images.
Keywords
Bearings, Convolutional neural networks, MEMS Accelerometers
