An Intelligent Fault Diagnosis for Rolling Bearing Based on Adversarial Semi-Supervised Method

Abstract

Intelligent fault diagnosis of rolling bearing issues have been well addressed with the rapid growth of data scale. However, the performance of most diagnostic algorithms heavily depends on sufficient labeled samples. How to ensure the fault diagnosis accuracy with scarce labeled samples is always a great challenge in real industrial scenarios. To alleviate this issue, a deep adversarial semi-supervised (DASS) method based on the prototype learning network is proposed for rolling bearing fault diagnosis in this study. First, the nine data augmentation methods are investigated to improve the ability of the network to extract discrepant features. Second, a cosine similarity-based classifier architecture is exploited to estimate prototypes of sample features. Third, the cross entropy and the adversarial entropy are introduced to improve diagnostic accuracy and reduce intra-class variation. Extensive experiments on two rolling bearing datasets verify that the proposed method owns a remarkable identification ability and great robustness under scarce labeled samples as training date. At the same time, the diagnostic performance under different number of labeled samples is extensively evaluated in noisy environment with different signal-to-noise ratio.INDEX TERMS Intelligent fault diagnosis, deep neural network, adversarial semi-supervised learning, rolling bearing.

Compound Fault Diagnosis of Aero-Engine Rolling Element Bearing Based on CCA Blind Extraction

“…The substitution of equation ( 14) and equation ( 15) into equation ( 12) leads to 17). If the current gradient is small enough, the iteration is stopped, otherwise the iteration is continued; c) compute extraction vector w according to equation (19) and equation (10); d) compute search direction of the next iteration according to formula equation ( 7) and equation (9).…”

“…Take equation (10) into consideration, the closed form solution for step size η t cannot be obtained by equation (19) because H t+1 also depends on step size η t . To overcome this and simplify the optimization with respect to step size, let us recall that a fixed small step size for gradient learning still performs well.…”

“…To overcome this and simplify the optimization with respect to step size, let us recall that a fixed small step size for gradient learning still performs well. So we propose to firstly update the blind extraction vector w according to equation (10) with a small step size, so as to obtain ŵt+1 temporarily, and then we substitute ŵt+1 into equation (20) to get H t+1 , finally substituting H t+1 into equation (19) leads to the optimal step size. Without losing generality, we simply use η t = 0, then we have ŵt+1 = w t and H t+1 = H t .…”

“…Without losing generality, we simply use η t = 0, then we have ŵt+1 = w t and H t+1 = H t . This can not only ensure the convergence of the algorithm, but also avoid the tedious iterative operation between equation (10), equation (20) and equation (19).…”

“…However, this kind of method depends heavily on the selection of features, and it is very difficult to select features with obvious differentiation under complex conditions. Deep learning methods [19]- [25] such as generative countermeasure network and convolutional neural network (CNN) avoid feature extraction based on artificial experience. The network automatically finds features in massive data with better generalization.…”