BEARING FAULT DETECTION VIA CONVOLUTIONAL NEURAL NETWORKS USING LOW-COST MEMS ACCELEROMETERS

Abstract

The study focuses on detecting bearing failures in industrial environments, crucial to equipment performance and reliability. Commercial accelerometers used for fault signal acquisition are expensive, limiting their implementation 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. To address this, the study proposes evaluating two low-cost microelectromechanical systems (MEMS) accelerometers for obtaining vibration signals and converting them into 2D images for CNN analysis. An experimental platform was built, connecting the MPU6050 and ADXL345 accelerometers to bearings with normal and faulty conditions. Vibration signature images were inserted into CNN with dimensions of 16x16, 22x22, and 28x28 to assess their impact on accuracy. The results showed that the proposed algorithm achieved high accuracy: for MPU6050, 97,31% accuracy with 16x16 images, 96,95% with 22x22 and 98,92% with 28x28, for ADXL345, 98,31% with 16x16 images, 99,70% with 22x22 and 99,82% with 28x28. This research demonstrates the effectiveness of low-cost accelerometers and 2D image conversion to improve bearing failure detection.