Novel LPP-Constructed Topology Integrated With Knowledge-Based GCN for Fault Diagnosis

Graph convolutional networks (GCNs) have gained popularity in industrial fault diagnosis due to their ability to model complex relationships between nodes. However, when dealing with high-dimensional and strongly coupled data in industrial processes, the performance of GCNs is constrained, particularly in scenarios where the graph structure is incomplete, making it challenging to effectively capture the underlying relationships between fault patterns. To address the limitations of GCN in industrial fault diagnosis, this article presents an innovative approach that integrates locality preserving projection (LPP) from manifold learning for constructing a topological graph (TG). This approach, referred to as LPP-constructed topology integrated with knowledge-based GCN for fault diagnosis (LTIKGCN), is designed for fault diagnosis. First, a knowledge graph (KG) based on prior domain knowledge is constructed, and TG is built using manifold learning through LPP. The KG and TG are then fused and input into the GCN. This approach enables the effective capture of the complex relationships between system components and fault patterns. Simulations were conducted on the multiphase flow facility (MFF). The experimental results demonstrate that the proposed method achieves higher accuracy and robustness in handling fault diagnosis tasks in complex industrial processes compared to other approaches.