Industrial quality inspection decision optimization faces critical limitations, including loose inter-module coupling and weak multi-agent collaboration when deploying Large Language Models (LLMs) and Deep Reinforcement Learning (DRL). To achieve accurate inspection, intelligent decision-making, and efficient scheduling in intelligent manufacturing scenarios, this study puts forward an improved Large Language Model-driven multi-agent deep reinforcement learning decision optimization framework. The framework uses YOLOv8 to build a perception layer for defect detection and feature extraction. It uses LLaMA to build a semantic layer for defect analysis and knowledge injection. It utilizes the QMIX algorithm for developing the decision layer of many agents for making decisions in collaboration with each other. Furthermore, the framework incorporates Temporal Convolutional Networks (TCN) for temporal feature extraction, K Means clustering for defect grading, and Ant Colony Optimization (ACO) for dynamic workstation scheduling to construct the optimization layer collectively and develops a combined collaborative system encompassing perception, semantics, decision, and optimization. Experimental results show that the proposed framework achieves a mAP@0.5 of 97.82% on the fused dataset combining the public MVTec AD dataset and the private automotive parts quality inspection dataset. The quality inspection decision accuracy and defect grading accuracy reach 96.35% and 95.78%, respectively, outperforming conventional DQN and VDN based multi agent DRL baselines by 13.00% and 9.97%, respectively. The scheduling path length decreases by 23.60%. The framework also shows better convergence speed and load balancing performance than the comparison methods. Although the overall computational overhead imposes challenges for direct deployment on resource-constrained industrial edge terminals, the proposed YLQTAK framework provides an effective technical solution for edge side near real time collaborative decision making (under cloud edge collaborative mode) of multi-workstation quality inspection in flexible production lines and points out a feasible lightweight edge-cloud collaborative deployment direction for follow-up research.
