Research on fault differentiation methods for similar faults based on multiple Gated memory graph convolutional networks (M−GM−GCN)

Abstract

In nuclear power plant accidents, many fault modes exhibit similar characteristics, making their effective differentiation highly challenging. To address this issue, this paper proposes a Multiple Gated Memory Graph Convolutional Network (M-GM-GCN) model. The traditional Graph Convolutional Network (GCN) is improved by embedding Gated Memory Units, forming a Gated Memory Graph Convolutional Network (GM-GCN). Given the similarity in the impact scope and propagation paths of similar faults within the system, specific fault propagation graphs are designed for each fault type. These graphs are then embedded into corresponding GM-GCN modules, creating multiple parallel Gated Memory Graph Convolutional Networks. The proposed method can effectively extract discriminative information among similar faults by the proposed model from the fault-related data. Features extracted from each GM-GCN module are fused using concatenation or multi-head attention mechanisms to generate the final fault diagnosis results. Experimental results demonstrate that the proposed method significantly improves fault detection accuracy.