Integrating Physics-Based and Data-Driven Models Using Reinforcement Learning and Adversarial Learning for Intelligent Manufacturing

Hybrid models that integrate physics-based (PB) and data-driven (DD) approaches have gained increasing attention in intelligent manufacturing. To address the dual challenges of limited dynamic adaptability in static hybrid models and the difficulty of traditional residual learning in handling unmodeled dynamics, we propose a PB–DD hybrid model using reinforcement learning (RL) and adversarial learning (AL), namely, RA-PBDD, which is different from a DD-centric hybrid approach that often suffers from limited training data representativeness, whereas PB-centric hybrid methods face challenges because of model incompleteness. RL is employed for one-time parameter identification in PB models, overcoming the limitations of conventional stepwise techniques. A residual value model using AL with neural networks is proposed, which achieves improved accuracy through iterative refinement compared with standard learning approaches. Additionally, we analyze the operational principles of the PB, DD, and hybrid PB–DD models from a state–space perspective which is a particular angle to systematically examine and contrast the characteristics of these model types. The effectiveness of RA-PBDD is validated through a case involving a machine tool feed system. The results confirm the superiority of RA-PBDD in both prediction and generalizability accuracy over the stand-alone PB, DD models and other two hybrid models. Specifically, it improves the prediction accuracy by up to 73% from 15.4 to 4.2 μm and machining precision by up to 46.7% from 15.2 to 8.1 μm.