A Novel Adaptive Mechanism-Data Fusion Graph Embedding Network for Fault Diagnosis

Abstract

Due to the complex information transfer between devices, sensor signals exhibit complex interactions in large-scale industrial process, making fault diagnosis a challenging task. Despite the effectiveness of existing methods, models that lack the guidance of priori mechanisms are extremely data-dependent, and most methods ignore the fact that information interacts differently between strongly and weakly correlated variables. Therefore, a novel adaptive mechanism-data fusion graph convolutional network (AMDF-GCN) is proposed. The approach uses the mechanism graph structure to guide the model to aggregate strongly correlated variable features. However, due to the lack of complete mechanism knowledge of the process, a mutual information patching mechanisms graph strategy is designed to jointly construct the mechanism-data adjacency matrix. The local-global feature aggregation module is then designed to mine potential weak correlations between variables. Furthermore, to compensate for the coupling problem caused by deep mining, a node information preserving module is designed to maintain the original information of the variables. Finally, the fused features are used for category recognition. The superiority of the AMDF-GCN method is verified through a typical industrial case and an actual industrial case.