SIMULATION-DRIVEN EXPLAINABLE AI FOR QUALITY PREDICTION AND CONTROL

Abstract

Manufacturing quality management faces persistent challenges such as data silos, small sample sizes, black-box models, and delayed control. To overcome these issues, this study proposes a closed-loop framework integrating simulation, data, AI, and control. A multi-scale, multi-physics digital twin generates mechanistic data, which are fused with event logs into a dynamic relationship matrix. A physics-constrained Transformer enables interpretable quality prediction, while reinforcement learning implements a predict–diagnose–adjust strategy for adaptive control. Simulation experiments in a CNC milling scenario demonstrate improved prediction accuracy, transparent mechanism tracing, and enhanced robustness against disturbances. The main innovations include: (a) multi-physics digital twin modelling of the full manufacturing process; (b) a discrete–continuous data fusion method; (c) an interpretable Transformer-based model with physical constraints; and (d) a predictive-control closed-loop mechanism. This framework provides a practical pathway for advancing smart manufacturing toward proactive and sustainable quality management.