Aiming at the limitations of traditional graph convolutional networks (GCNs) in dynamic feature capture and multi-dimensional feature fusion, this paper proposes a novel method called multi-level adaptive fusion and edge-enhanced attention GCN (ME-GCN). The ME-GCN integrates the multi-level dynamic collaborative fusion mechanism, which combines feature-space and topology-space graph convolution modules, enabling comprehensive extraction of local features and global topological relations. By introducing the edge-enhanced graph attention mechanism, the association weights between nodes are dynamically adjusted, thereby enhancing the modeling capability of complex fault patterns. Moreover, the adaptive fusion strategy dynamically balances and optimizes multi-module features, achieving efficient and robust fault representation. Experimental validation on both public datasets and real industrial datasets demonstrates that ME-GCN achieves superior diagnostic performance with strong robustness and generalization ability under different operating conditions.
Hu et al. (Thu,) studied this question.