Research on Generative Design Method for Modular Intelligent Manufacturing Cell Layout based on Diffusion Models
作者
Zhihui Pang,Cheng Li,Wenkai Zhao
出处
期刊: [Institute of Electrical and Electronics Engineers] 日期:2026-03-20卷期号:: 1934-1938
标识
DOI:10.1109/icaace69793.2026.11508947
摘要
Manufacturing cell layout optimization is a critical combinatorial problem that significantly impacts production efficiency and operational costs. Traditional optimization methods and recent deep learning approaches struggle to balance multiple competing objectives while satisfying complex spatial constraints. This paper proposes a novel conditional diffusion model framework for intelligent manufacturing layout generation that integrates generative modeling with explicit multi-objective optimization. The approach develops a multi-channel grid representation encoding geometric, relational, and constraint information, enabling end-to-end learning of layout patterns. A hierarchical conditioning strategy combines global parameters, graph neural network-encoded material flow relationships, and spatial constraints through cross-attention mechanisms for fine-grained generation control. Most importantly, we introduce a gradient-guided sampling mechanism with annealing schedule that explicitly optimizes material handling cost and space utilization during the reverse diffusion process, rather than relying solely on conditioning. Experiments on 800 synthetic layouts and 3 real-world industrial cases demonstrate that the proposed method achieves material handling cost of 2847 units and constraint satisfaction rate of 94.3 %, representing 36.8 % improvement over traditional optimization algorithms and 9.8% over the best deep learning baseline. Ablation studies confirm the gradient guidance mechanism contributes$\mathbf{1 8. 3 \%}$performance gain. The framework provides a scalable solution for intelligent manufacturing design automation with$\mathbf{5 0 - 1 0 0} {\times}$speedup over traditional methods.