Interpretable Quality Defect Traceability to Production Process Parameters Via Knowledge Graphs and Graph Attention Networks

ABSTRACT Intelligent algorithms have propelled quality control from passive post‐event detection toward proactive real‐time intervention. However, their inherent “black‐box” nature often restricts analysis to surface‐level mappings between process parameters and product quality, obscuring the underlying causal mechanisms and impeding effective traceability of quality defects to their root causes. To address this limitation, this paper proposes a novel traceability framework that integrates deep learning with explainable artificial intelligence (XAI) and knowledge graph (KG) reasoning to enable transparent, inference‐driven defect source identification and traceability. The study proceeds through three key stages, each advancing traceability from a different angle: (1) A predictive model based on one‐dimensional convolutional neural network (1D‐CNN) is first established to capture the complex nonlinear mapping from production process parameters to final product quality attributes, providing the predictive foundation upon which traceability can be built. (2) To demystify the predictive model's decisions, Shapley values—derived from game‐theoretic XAI—are computed to quantitatively disentangle and rank the contribution of each individual parameter toward specific quality defects. These quantified contributions are subsequently transformed into structured triples (parameter‐influence‐defect), forming a specialized quality knowledge graph that explicitly encodes causal feature‐effect relationships, thereby rendering each defect traceable to the parameters that contributed to it. (3) Leveraging this graph, a Graph Attention Network (GAT)‐based link prediction model is developed to learn and infer latent relationships within the KG. Critically, this GAT model enables the forecasting of Shapley values for new or unseen process scenarios, thereby dynamically tracing emerging defects back to their most likely responsible parameters. This integrated approach effectively converts an opaque predictive model into an interpretable, knowledge‐driven traceability tool, offering a robust and generalizable pathway for real‐time, fine‐grained quality defect traceability in complex manufacturing environments.