Semi-supervised Bearing Fault Diagnosis Using Improved Graph Attention Network under Time-varying Speeds

Abstract

Recent researches on semi-supervised bearing fault diagnosis based on Graph Neural Network (GNN) still have some problems, such as insufficient label information mining and relatively ideal diagnosis scenarios. In engineering practice, bearings are often operated under time-varying speeds such as startup and shutdown, and fault label samples become increasingly expensive. In response to the above challenges, a new method called semi-supervised bearing fault diagnosis using improved Graph ATtention network (GAT) under time-varying speeds is proposed. Based on K-Nearest Neighbor (KNN) algorithm and Smoothing Assumption (SA), the pseudo-label propagation strategy is designed to spread the label information to the neighborhood samples with similar distribution along the edge, so that the label information hidden in the limited samples can be fully utilized. Each vibration spectrum sample is considered as a node, and a semi-supervised learning model based on node-level GAN is constructed to explore further representative bearing fault features through the attention mechanism. The proposed method is applied to analyze two sets of bearing fault experimental data under time-varying speed, and the results show that the proposed method is able to diagnose accurately different fault modes of bearings at low label rates of no more than 2%, which is better than other commonly used semisupervised learning methods of GNN.