A graph attention network for rotating machinery fault diagnosis driven by multi-channel vibration signal

To address the issue of missing feature associations in traditional single-channel analysis methods under multi-sensor monitoring systems, this paper proposes a graph attention network fault diagnosis model based on meta-path guided graph topology construction and hierarchical attention mechanism. This model structurally screens and organises the correlations between channels from the multi-channel signal graph by defining meta-paths, and constructs a meta-graph topology with clear physical meanings. Based on this, the model adopts a hierarchical attention mechanism: intra-graph attention mechanism achieves deep fusion of dual-channel local features within the meta-graph, while inter-graph attention mechanism completes cross-channel global feature enhancement between meta-graphs, forming a hierarchical feature aggregation paradigm of meta-graph construction, intra-channel fusion, and inter-channel enhancement. Experiments on the XJTU and SZ datasets show that the proposed method is significantly superior to the existing methods in terms of classification accuracy and anti-noise robustness. The post-hoc interpretability analysis on the CWRU dataset further verified the model’s effective ability to focus and extract fault frequency domain features.