Deep dynamic high-order graph convolutional network for wear fault diagnosis of hydrodynamic mechanical seal

Abstract

The hydrodynamic mechanical seal (HDMS) in the reactor coolant pump of third-generation nuclear power units is vulnerable to failure due to prolonged operational periods and inevitable wear. However, traditional fault diagnosis methods are not robust to noise and can not leverage both the topological relationships among samples and local features. To resolve these challenges, in this paper, we propose a novel graph convolutional network (GCN) for wear fault diagnosis of HDMS called deep dynamic high-order graph convolutional network (DDHGCN). A dynamic graph learning module is designed to control the connectivity and sparsity of the iterated graph and thus eliminate errors and redundancies caused by noise. A high-order GCN module is proposed to effectively model the correlations between nodes, capturing contextual information and mutual influences among them. A residual convolutional module is applied to extract local features hidden in individual samples to further improve the classification performance. All three modules are jointly optimized for reliable wear fault diagnosis of HDMS. Experimental results demonstrate that our DDHGCN can achieve higher performance when compared with the state-of-the-arts.