Self-optimising systems in factories

Abstract

Self-optimizing systems in manufacturing integrate real-time sensing, digital twins, machine learning, and control theory to autonomously tune production processes for higher throughput, lower energy use, and reduced downtime. This article surveys the architectures and methods enabling factory self-optimization, including reinforcement learning schedulers, model predictive control (MPC), digital twin-based optimization, predictive maintenance, multi-agent cooperation, and hybrid RL-MPC systems. We analyze trade-offs between throughput, energy consumption, defect rates, and operational resilience. A synthetic study compares representative methods across normal, high-demand, and maintenance scenarios, providing tabular results and charts that illustrate throughput gains and energy profiles. The discussion highlights challenges — data quality, sim-to-real transfer, safety constraints, human-in-the-loop control, and standards for verification — and outlines research directions such as federated learning for cross-plant adaptation, causal reinforcement learning, and scalable digital twin orchestration. Practical recommendations cover phased deployment, KPI alignment, and governance for trustworthy autonomous optimization