DHFGAT: A dynamic heterogeneous fusion graph attention network for early gearbox fault detection

• DHFGAT method for dynamic heterogeneous sensor fusion and early fault detection. • RDA algorithm build information-rich graphs by adaptive feature selection. • A novel GNN architecture integrating SNE and SGAT is designed. • Effectiveness is validated by extensive gearbox test rig experiments. Accurate detection of early faults in industrial gearboxes is a critical but challenging task, particularly under dynamic operating conditions. Converting multi-sensor signals into graph-structured data via Graph Neural Networks (GNNs) offers a framework based on signal correlation modeling. However, the efficacy of this approach hinges on the quality of the graph structure, which is generally compromised by suboptimal construction methods and the inclusion of redundant features. To address these challenges, this study proposes a heterogeneous data fusion framework named Dynamic Heterogeneous Fusion Graph Attention Network (DHFGAT). Specifically, the current signal is first preprocessed to align its data morphology with the vibration signal. Then, to construct an information-rich graph structure, the Regularized Dual Averaging (RDA) algorithm adaptively selects the key feature subsets of the signal, thereby suppressing interference from redundant features. Subsequently, a training architecture that integrates the Spiking Neural Encoder (SNE) with the Simplified Graph Attention Network (SGAT) is adopted to achieve accurate and robust feature fusion and fault classification. Finally, extensive experiments were conducted on a test rig with eight fault types. The results demonstrate that DHFGAT achieves maximum average detection accuracies of 99.59% and 98.36% under constant and variable speed conditions, respectively. The proposed method provides a robust, reliable solution for early gearbox fault detection and offers insights into the optimal sensor combination for complex industrial scenarios.