Exploiting graph neural network with one-shot learning for fault diagnosis of mechanical equipment

Abstract

For the classification of mechanical fault diagnosis, a graph neural network (GNN) method with one-shot learning is proposed. Convolutional Neural Network (CNN) is used to extract the feature vectors and One-Hot coding from images of Fault diagnosis of mechanical equipment. Inputting feature vectors and One-Hot coding into GNN, according to the Adjacency Matrix between vertices in the Graph, and is used for classification and inference. The method with one-shot learning is used for fault diagnosis classification. Through the fault classification for the industrial robot RV reducer and public data set CWRU pictures, the effectiveness of the method is verified. Five categories are used for fault diagnosis and classification in RV Reducer of the industrial robots. 80 categories are used in the public data set CWRU, and 55 categories are used as the training set. GNN is employed to spread the label information from the supervised sample of the unlabeled query data. The large-scale dataset can then be used as baseline classes to learn transferable knowledge for classifying novelties with one-shot samples. The one-shot learning with graph neural network GNN significantly improves the classification accuracy. The results show that the proposed method is superior to other similar methods and has a substantial potential for improvement in Fault diagnosis of mechanical equipment.