Novel Dual-Stream Fusion Graph Convolutional Network for Industrial Process Fault Diagnosis

Fault diagnosis for high-dimensional industrial process data with strong nonlinear coupling remains challenging. Most existing graph convolutional network–based methods rely on static or single graph structures, which limits their ability to jointly capture physical consistency and time-varying data characteristics. In addition, fixed feature fusion strategies lack adaptability to diverse fault patterns. To overcome these limitations, this paper proposes a Dual-Stream Fusion Graph Convolution Network (DSFGCN) that integrates knowledge-driven and data-driven information in a complementary manner. Specifically, a dual-graph construction mechanism is developed, where a knowledge graph incorporates expert prior information to preserve physical interpretability, while a supervised manifold graph based on Discriminative Locality Preserving Projections captures intrinsic nonlinear data structures. Furthermore, an adaptive gated fusion unit is designed to dynamically adjust the contributions of the two graph streams according to operating conditions. Experiments on the Tennessee Eastman Process and Multiphase Flow Facility datasets demonstrate that DSFGCN achieves superior diagnostic accuracy and robustness, particularly for complex nonlinear fault scenarios.