Bispectrum and Convolution Neural Network Based Bearing Fault Diagnosis

Abstract

The roller bearings are the heart of any rotating machine. The bearing's health is monitored at regular intervals to keep the machine running. In the traditional fault diagnostic approach, features are obtained from the vibration data, and then fault classification is performed. In this paper, a more refined method based on Convolutional Neural Networks (CNN) and vibration bispectrum has been developed to increase the effectiveness and accuracy of fault diagnostics of bearings. As bispectrum enables nonlinear feature augmentation and noise reduction, it is preferred as the input for the CNN model rather than using raw vibration signals. Although many pretrained CNNs are available, in this paper a standard CNN model is shown to be very effective. In this work, bispectrum analysis of the vibration signal is carried out, and these images are used as input to the standard Convolution Neural Network (CNN) for classification. With this method, the training and testing accuracy of 100% percent has been achieved for the classification of bearing faults.