Intelligent Fault Diagnosis In Manufacturing Systems Via A Risk-Level Knowledge Graph And Genetic Algorithm Enhanced GNN

Abstract

Industrial fault diagnosis commonly faces two critical challenges: low training efficiency and limited model interpretability. To address these issues, this study presents a hybrid framework that integrates a Risk Level Knowledge Graph (RLKG) with a Graph Neural Network (GNN), further optimised using a Genetic Algorithm (GA). The RLKG is constructed through a novel Risk Level modelling approach that encodes structured domain knowledge into a knowledge graph aligned with key characteristics of industrial systems. This structured prior is leveraged to initialise node features and sparsify GNN connectivity, thereby improving both training efficiency and model interpretability. The GA is employed to fine-tune the hyperparameters of the GNN, resulting in the RLKG-GA-GNN framework. Simulation results on benchmark industrial datasets demonstrate that the proposed method improves convergence speed
by 35% and achieves a fault classification accuracy of 97.9%, outperforming standard GNN-based approaches by 6.5%. Moreover, over 89% of the attention weights in the model can be directly mapped to physical system components, offering clear insights into fault propagation and enabling actionable engineering decisions. This work contributes a scalable, interpretable, and high-performance solution for intelligent fault detection and classification in industrial systems.