Intelligent Fault Diagnosis of Bearings based on Convolutional Neural Network using Infrared Thermography

Abstract

Bearings, as a key part of rotating machinery, are prone to failure due to fatigue and aging resulting from their long-term and high-load operation. To ensure stability of the mechanical equipment, monitoring bearing health is helpful to guarantee smooth operation of machinery and increasing machinery availability. This article puts forward an intelligent non-invasive thermal images-based fault diagnostic approach to periodically monitor condition of the rolling contact bearings in respect of their deterioration due to defects on the inner race, outer race and balls/rollers. Thermal images of four bearing conditions, including one healthy and three faulty states, have been considered followed by a performance classification based comparative analysis using Support Vector Machines (SVM) and Convolutional Neural Network (CNN). The CNN consists of many tools under its cap but for this work, the AlexNet architecture is used which has proved to be more effective than SVM. The experimental findings reveal that non-contact infrared thermography has enormous potential for automatically identifying problems and detecting early warning, regardless of speed, resulting in negligible shutdowns of the system due to bearing failure.

Towards Deploying DNN Models on Edge for Predictive Maintenance Applications

“…Different structural formulations of CNNs have been experimented with bearing datasets for predictive maintenance applications. Different variations of AlexNet [14][15][16][17], ResNet [18][19][20][21][22] and LeNet [23,24] are used in different research works for fault diagnosis of rolling bearing elements. In [14,15], the authors proposed an one dimensional model slightly deeper than the original AlexNet model to enhance the bearing fault classification and compared their model with AlexNet.…”

“…Instead of this, [17] shows that a AlexNet model can be used by just retraining the fully connected (FC) layers at the end to identify bearing defects. Thermal images were used as input to AlexNet in [16] in order to detect bearing faults.…”

“…Deleting a neuron produces some structural instability in the model which can lead to a broken DNN network. Removing a neuron from the ith FC layer can be explained by function D i FC in Equation (16).…”

A CNN-Based Methodology for Identifying Mechanical Faults in Induction Motors Using Thermography

“…Recently, infrared thermography (IRT) has emerged as an interesting alternative to monitoring IM machines, due to its measurement method being: (1) contactless, (2) noninvasive, and (3) the sensor can be located at a distant location [30]. In this sense, different applications have been reported [31][32][33][34]. Li et al [31] proposed a support vector machine (SVM)-based methodology for diagnosing the condition of gearboxes.…”

“…One of the detected failures was a broken bar, obtaining a range of accuracy between 95 and 100%, depending on the detected failure. Sharma et al [33] and Choudhary et al [34] used IRT fused with a Convolutional Neural Network (CNN) to determine if a bearing was damaged. They reported accuracies ranging from 90 to 99%, depending on the level and ubication of the bearing damage.…”