Multi-Layered Graph Convolutional Network-Based Industrial Fault Diagnosis with Multiple Relation Characterization Capability

Abstract

Fault diagnosis of industrial equipments is extremely important for the safety requirements of modern production processes. Lately, deep learning (DL) has been the mainstream fault diagnosis tool due to its powerful representational ability in learning and flexibility. However, most of the existing DL-based methods may suffer from two drawbacks: Firstly, only one metric is used to construct networks, thus multiple kinds of potential relationships between nodes are not explored. Secondly, there are few studies on how to obtain better node embedding by aggregating the features of different neighbors. To compensate for these deficiencies, an advantageous intelligent diagnosis scheme termed AE-MSGCN is proposed, which employs graph convolutional networks (GCNs) on multi-layer networks in an innovative manner. In detail, AE is carried out to extract deep representation features in process measurement and then combined with different metrics (i.e., K-nearest neighbors, cosine similarity, path graph) to construct the multi-layer networks for better multiple interaction characterization among nodes. After that, intra-layer convolutional and inter-layer convolutional methods are adopted for aggregating extensive neighbouring information to enrich the representation of nodes and diagnosis performance. Finally, a benchmark platform and a real-world case both verify that the proposed AE-MSGCN is more effective and practical than the existing state-of-the-art methods.