Parallel feature fusion networks for rotating machinery fault diagnosis with multi-sensor information

Rotating machinery constitutes fundamental components within vital industrial systems, including power generation and transport networks, where accurate malfunction detection is crucial for upholding the safety and efficiency of operations. Multi-sensor data fusion serves as a crucial strategy to improve diagnostic accuracy. However, effectively fusing heterogeneous sensor information still poses challenges. This is mainly attributed to the inadequate exploration of physical correlations between sensors and the limited integration of spatio-temporal characteristics in practical industrial scenarios. To overcome these limitations, this study introduces an innovative parallel feature fusion framework for fault diagnosis of industrial rotating machinery with heterogeneous multi-sensor data. The framework adopts a dual-branch feature extraction strategy: (i) a physics-informed graph isomorphism network that captures spatial and physical correlations among sensors, and (ii) a network constructed by global attention bidirectional gated recurrent unit that extracts global temporal patterns from real-world sensor time series. A cross-attention module is then employed to achieve synergistic fusion of spatial and temporal features. Extensive testing on industrial datasets confirms the approach’s efficacy, highlighting its applicability for identifying faults in rotating machinery.