Answers are grounded in this paper and cite their sources.
Fault diagnosis of rotating machinery under complex operating conditions requires the joint use of diverse signal representations, operating-condition information, and their dependencies. However, existing methods still provide insufficient differentiated modeling of such heterogeneous information. To address this limitation, we propose an attention-enhanced temporal heterogeneous graph neural network (ATHGNN) that provides structured modeling of frequency representations, operating conditions, and graph relations. First, two differently preprocessed vibration branches employ cosine-distance-based and Euclidean-distance-based Shapelet matching, respectively, together with one-dimensional convolutional block attention modules (1DCBAMs) with different attention orders, to extract local fault patterns. Meanwhile, exogenous operating variables are encoded by a multilayer perceptron (MLP) and explicitly injected into node features as condition information. Next, feature similarity and K-nearest-neighbor (KNN) sparsification are used to construct a heterogeneous graph containing intra-type and cross-type relations. Finally, to address the modeling requirements of different graph relations, a graph convolutional network (GCN) aggregates information from nodes of the same type, while multi-head attention constrained by a heterogeneous adjacency mask learns node dependencies. The resulting representations are fused for fault classification. Experiments on three public datasets, CWRU, XJTU Gearbox, and JNU, yield mean diagnostic accuracies of 99.43%, 99.61%, and 98.67%, respectively, outperforming the six baseline methods compared. Additional experiments assess diagnostic performance under controlled random noise conditions using additive white Gaussian noise (AWGN) perturbations.
No takes yet. Share an insight, caveat, or question.
Liang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: