FORCE: A Federated Deep Reinforcement Learning Framework for Scalable and Privacy-Preserving Process Control in Smart Manufacturing

• A federated reinforcement learning framework is developed for distributed manufacturing control. • Hypergraph neural encoding captures complex multi-machine and multi-resource relationships. • Dhole-optimized DDPG improves training stability, robustness, and convergence speed. • FORCE increases throughput up to 96% while reducing energy and communication overhead. • The framework ensures data privacy by sharing only encrypted model updates across nodes. The increasing complexity and interconnections inherent in modern manufacturing systems have increased a demand for integrated methodologies to boost productivity, quality and streamline operations across the entire system. Existing isolated process-level or system-level optimization approaches are no longer enough to attain global efficiency. These traditional methods focus on the optimization of individual processes, neglecting their interdependencies at the system level and causing lower throughput, spoil product quality, and reduce responsiveness in uncertain situations. To overcome these drawbacks, a novel Federated Orchestrated deep Reinforcement learning for process Control and Execution (FORCE) model has been proposed in this paper. The proposed model uses a federated reinforcement learning model which combines federated learning with the adaptive reinforcement learning model for efficient process control. The proposed model uses a Hypergraph Neural State Encoder for converting raw inputs into high-order relational representations and statistical refinement for generating compressed and denoised features. The study also leverages an adaptive control system integrated with a Deep Deterministic Policy Gradient (DDPG) model optimized using a Dhole optimization algorithm to enable real-time production control that sequences training tasks to accelerate convergence and improve policy robustness. The proposed model has been evaluated using various metrics such as accuracy, precision, f1-score, specificity, sensitivity, throughput, and Mathew’s correlation coefficient. The proposed FORCE model achieves the highest accuracy of 97%, while the previous techniques such as DDQN-PER, SATD3, and GCNN-DRL achieves the accuracy of 95%, 93%, and 90% respectively.