Knowledge graph-guided multi-view graph attention for multi-source domain adaptive fault diagnosis

In the field of bearing fault diagnosis, three critical issues severely restrict the improvement of diagnostic performance, which are large data distribution discrepancies across multi-source domains, insufficient discriminability of single-view features, and information loss during multi-source domain feature transfer. To address these challenges, this paper proposes a multi-view knowledge transfer model named Multi-view Knowledge Graph–Graph Attention Network (MVKG-GAT), which integrates multi-view time–frequency-graph collaborative feature extraction, knowledge graph matching rule transfer, and multi-source domain adaptive selection. First, the model achieves accurate capture of global associated features by constructing dynamic K-Nearest Neighbor (KNN) graphs, which serve as the core of multi-view time–frequency-graph collaborative feature extraction. Second, it guaranties the integrity and effectiveness of feature representation via a three-layer heterogeneous knowledge graph matching rule transfer mechanism for fault-pattern-frequency bands, realizing the knowledge graph matching rule transfer function. Third, the adaptive source domain selection strategy based on multi-distance metrics dynamically weights high-quality source domains, suppresses negative transfer, and greatly enhances the cross-domain adaptive ability of the model, thus completing the multi-source domain adaptive selection. Extensive comparative experiments conducted on three public datasets (CWRU, Ottawa, and JNU) demonstrate that the proposed method significantly outperforms traditional approaches in terms of diagnostic accuracy, cross-domain migration performance, and robustness. With high diagnostic precision and reliable feature transfer stability, the model presents promising application potential in practical industrial scenarios.