Simplicial complexes graph convolution networks with higher-order features learning for limited samples diagnosis

With the advancement of industrial automation, there is an increasing focus on research concerning limited fault samples. Although meta-learning and other methods can address this issue, they often necessitate the incorporation of additional data and are unable to directly diagnose faults using only unlabeled data along with a small amount of labeled data. In response, this article proposes the use of simplicial complexes graph convolutional networks for fault diagnosis, which simultaneously account for both higher-order and lower-order topological structures among samples. This approach effectively addresses the challenge of limited samples by extracting relevant information from unlabeled data without the need to introduce new knowledge. Initially, simplices of varying dimensions are employed within a constructed simple graph to represent different relationships among samples. Subsequently, the simplicial complexes convolutional network is introduced to capture the higher-order information, while the graph convolutional network is utilized to obtain the lower-order information. The combined feature information is then input into a classifier for fault diagnosis. Finally, experiments conducted on two datasets characterized by small sample sizes or imbalanced samples demonstrate the method’s commendable diagnostic performance, as well as its robustness and practicality.