Diagnosis the Bearings Faults through Exercising Deep Learning Algorithms

Abstract

Rolling bearings are vital components in process industries, and their faults can disrupt industrial processes. This paper presents a novel approach to diagnose the health state of rolling bearings, combining time-frequency domain signal analysis with deep learning models, namely ResNet-50 and DenseNet-121. Utilizing a dataset from the CWRU bearing datacenter containing various bearing health conditions, including normal and faulty states, this research addresses the limitations of conventional statistical feature-based fault detection methods. These scalograms are converted into grayscale images to optimize the learning process. The final grayscale CWT images are fed into the deep learning models for fault classification. Results indicate that the proposed framework, particularly the combination of CWTSV and DenseNet-121, yields promising outcomes of 95.25% and 99.77% respectively, surpassing existing methods for rolling bearing fault diagnosis. This approach holds potential for significantly enhancing industrial maintenance practices and ensuring process reliability.