Development of a hybrid reinforcement learning (hybrid-RL) system for manufacturing resilience

Résumé

In the unpredictable world of modern manufacturing, disruptions are not just challenges but constant threats to efficiency and profitability. This study presents a detailed methodology for a hybrid reinforcement learning (hybrid-RL) system designed to intelligently manage these disruptions. The approach seamlessly blends simulation-driven reinforcement learning—where an agent learns through trial and error in a virtual factory—with offline learning, which extracts valuable insights from historical operational data. The core of the system is an adaptive policy fusion mechanism, allowing us to combine the strengths of both learning paradigms. Crucially, in this study, we wove in multi-layered safety constraints and mechanisms for continuous online adaptation, ensuring that the system operates robustly and safely even when faced with unexpected and severe disruptions. The system leverages Python for building a highly realistic and object-oriented simulation environment that acts as a digital twin, and to handle the heavy lifting of RL implementation, enabling scalable training, evaluation, and eventual real-world deployment. The Python platform allows competence in physical modeling and flexibility in advanced machine learning. The notable improvements achieved in this study include a total reward of 14,964.79 for the proposed RL model outperforming all baseline methods, including random, heuristic, and advanced optimization/control policies. These findings agree significantly with the existing works and trends reported in literature studies. Thus, this study can assist manufacturing industries in utilizing reinforcement learning to achieve effective dynamic scheduling, adaptive and preventive maintenance, and manufacturing resilience and recovery, along with effective decision-making.

Citer ce document

Accès au document

Texte intégral en lecture en ligne, réservé aux abonnés SPHAERO et aux membres de l'institution. Se connecter

Auteur(s)

Statistiques

Consultations : 1

Téléchargements : 0