Real-Time Assembly Line Balancing Optimization Based on Deep Reinforcement Learning

Abstract

The issue of real-time optimisation for assembly-line balancing in large-scale, high-throughput production can be effectively resolved with deep reinforcement learning. The goal of this research is to address the dynamic scheduling problem with varying station restrictions and a fluctuating workload. The concepts of work allocation and the actual restrictions of a system in discrete manufacturing are integrated in a comprehensive mathematical model. In this paper, a digital twin-driven simulation platform constructed from actual industrial data and factory deployment records was used to develop a graph-based neural network with prioritised experience replay. The method outperforms traditional integer linear programming, DQN, and metaheuristic baselines by up to 6.3% in terms of throughput and reduces load imbalance by over 28% under stress, according to experiments conducted on over 60,000 production cycles. It achieves a median task throughput of 0.942 and an average decision latency of 38ms. According to robustness study, the system will continue to function steadily with a throughput loss of less than 5% even in the event of large loads and equipment failures. The resources have been optimised and the response time to environmental changes has been improved. The technical viability of deep reinforcement learning for intelligent assembly line balance has been confirmed based on the aforementioned findings, and real-time, data-driven production optimisation support for industrial deployment is offered.