<![CDATA[The machine learning architectures for autonomous, smart manufacturing systems must guarantee data privacy across geographically distributed production cells and be able to continuously adapt to the non-stationary process conditions. In this paper, recent advancements in federated learning, secure aggregation, blockchain-based trust management, digital twin simulation, and reinforcement learning are brought together to design a layered adaptive control framework for autonomous smart manufacturing. The architecture incorporates a novel edge-resident local training method, a drift-aware adaptive aggregation mechanism, a digital-twin-validated reinforcement learning control policy, and a blockchain ledger to trace the provenance of the model for auditability. Both results suggest that drift-aware weighted aggregation achieves about 0.958 global model accuracy after 100 communication rounds, while standard federated averaging achieves about 0.887 global model accuracy in the same number of rounds. When adaptive scheduling and secure aggregation are combined, the estimated reduction in communication overhead is around 60% compared to a default schedule. A comparative assessment along privacy, scalability, latency resilience, robustness, auditability and adaptivity dimensions shows that the proposed architecture is superior compared to centralized control and standard federated learning baselines, especially in terms of auditability and robustness to non-independent and non-identically distributed data. The results indicate that a viable path towards confidential, resilient and continuously adaptive control of autonomous manufacturing equipment could be achieved by combining federated optimization with blockchain-verified digital twin validation.]]>
Copyrights © 2025
