Pre-Training Graph Neural Network for Fault Diagnosis and Safety Assessment

Fault diagnosis and safety assessment are two pivotal tasks for ensuring the reliable and safe operation of dynamic systems. Constructing separate data-driven models for each task usually involves a large amount of labeled data, which is challenging for practical applications. In this study, the problem that building a unified data-driven model for these two related tasks under the few-shot setting is proposed. A pre-training graph neural network (GNN) scheme is designed. Firstly, the input data is transformed into graph data. Then, the pre-training strategy based on node dropping is employed to initialize the parameters of GNN. Finally, supervised learning is performed using the labeled data for downstream tasks, including fault diagnosis and safety assessment. A case study based on bearing data demonstrates the effectiveness of the proposed scheme.