Novel Cosine Distance Based Semi-Supervised Learning Using Discriminant Graph Convolutional Network for Industrial Fault Diagnosis

Semi-supervised learning can effectively utilize limited labeled data and large amounts of unlabeled data to achieve fault diagnosis on process industries. This paper proposes a novel cosine distance based semi-supervised learning using discriminant graph convolutional networks (CD-GCN) at node-level Firstly, the CD-GCN method uses the discriminant information to pull the training sample features of different classes farther away from each other. Secondly, CD-GCN replaces Euclidean distance with Cosine distance as the distance metric in the original samples space and feature space. With Cosine distance and discriminant information, CD-GCN better considers spatial structure information and improve the whole graph by the dual effects of graph convolution and nearest neighboring with these moving training sample features. Finally, a real industrial simulation case is carried out to verify the performance of the proposed method. Compared with other related and classic methods, the simulation results show that the CD-GCN method achieves the best performance in diagnostic accuracy.