MEN-GAE Based Intelligent Fault Diagnosis for Rotating Machinery under Very Low Labeling Rate and Variable Speed Working Condition

During the process of graph construction, existing Graph Neural Networks (GNN) suffer from information redundancy between nodes. This leads to insufficient mining of neighbor node information and insufficient network feature extraction capability. In industrial production, rotating machinery typically often works under variable speed conditions, and the cost of labeling samples is high. Aiming at the above two problems, the Maximum Effective Neighborhood-Graph Autoencoder Network (MEN-GAE) is proposed in this paper, and is applied to semi-supervised fault diagnosis of rotating machinery under variable speed conditions and very low labeling rate. Firstly, the entropy value is used for determining the similarity of neighboring nodes in the GNN, and the similarity between nodes is used as the basis for the existence of edges between nodes. Second, two different KNN graphs are constructed using two different entropy values. Next, compare the structures of the two KNN graphs constructed with two different entropy values, and the edges with information redundancy are removed with the proposed MEN to obtain higher precision graph structure data. Finally, the high precision graph structure data are input into the constructed GAE for feature extraction. The proposed method is validated on bearing and gearbox datasets under variable speed conditions. The experimental results show that the proposed method can accurately diagnose different types of faults of rotating machinery with a label rate not exceeding 1%, and its performance is superior to other commonly used GNN.