Semi-supervised bearing fault diagnosis based on label propagation multi-receptive field graph convolutional neural network

Abstract Accurately diagnosing bearing faults is very important for improving the reliability of rotating machinery. Recently, data-driven fault diagnosis methods have been widely used. However, in practical applications, acquiring enough labeled samples is challenging, and it is difficult for traditional methods to fully exploit the intrinsic relationships between samples so that a large amount of unlabeled data cannot be fully utilized. To solve these problems, a label propagation multi-receptive field convolutional neural network (LP-MRF-GCN) based on semi-supervised learning is proposed for fault diagnosis with limited label samples. Firstly, the K-nearest neighbor relationships are determined by frequency domain analysis, and the labeled and unlabeled data are constructed as graph structures. Then, a pseudo-label propagation mechanism is designed to propagate the label information along the edges of the graph to similar neighborhood samples with the help of the smoothing assumption to mine the potential information of unlabeled data fully. Finally, a semi-supervised learning model is constructed based on the node-level multi-receptive field graph convolutional neural network. Combined with the attention mechanism, an adaptive feature fusion layer is introduced to realize information integration at different neighborhood scales. The experimental results show that on two bearing datasets, the proposed method can accurately diagnose different fault modes with a low label rate (no more than 2%), and the performance is better than the semi-supervised learning method of the traditional graph neural network.