A novel fault diagnostic system for rolling element bearings using deep transfer learning on bispectrum contour maps

Transfer learning using Convolution Neural Networks (CNNs) has improved the state of the art results in many research studies. Rolling element bearing fault diagnosis is a domain that has been researched extensively using different data mining and machine learning techniques. In this paper, we prove that deep CNNs, when trained on Bispectrum images of fault signals using transfer learning, provide highly accurate and reliable results for fault diagnosis that are at par with the state of the art results. These transfer learning based models are able to quickly learn patterns from visual features of a vibration signal’s bispectrum, eliminating the need for manual feature extraction. In this paper, four pretrained networks – Alexnet, VGG-19, GoogLeNet, ResNet-50 – have been fine tuned on bispectrum images prepared from vibration signals of machine ball bearing elements. Each network has been trained with 3 optimizers – Stochastic Gradient Descent (SGD), Adaptive Moment Estimation (Adam) and Adamax. These models are able to obtain high classification accuracy within a few epochs. We also visualise and analyze the feature maps associated with intermediate convolution layers for one of these models.