Rolling Bearing Fault Diagnosis Based on Improved GAN and 2-D Representation of Acoustic Emission Signals

Abstract

Bearing fault diagnosis is essential in manufacturing systems to avoid problems such as downtime costs. Convolutional neural network (CNN) models have enabled a new generation of intelligent bearing fault diagnosis methods for smart manufacturing owing to their capability to extract features for 2-dimensional (2D) representations, such as signals represented in the time-frequency domain. Nevertheless, the cost and time required to collect sufficient training data tend to result in a lack of data and data imbalance in real fault diagnosis scenarios. This inevitable consequence leads to a high misclassification rate in conventional CNN models. In this study, to address this problem, we propose a novel effective generative adversarial network (GAN)-based method for rolling bearing fault diagnosis in early-stage and low rotational speeds based on data enhancement, which uses acoustic emission (AE) as a monitoring signal. In the proposed approach, generator, discriminator, and fault classifier models are trained simultaneously with the proposed strategy for updating parameters to avoid the gradient vanishing problem and outperform conventional methods. The fault classifier was developed based on CNN models which are compatible with 2-D signal representations represented by a constant-Q transform. The results of experiments conducted with unbalanced compound fault datasets verify the capabilities of the proposed method in various diagnosis scenarios compared with traditional methods, including SVM, CNN, and DCGAN models.