Cyber-physical internet enabled hierarchical attention network based reinforcement learning for order dispatch in fast fashion manufacturing

Fast fashion platforms such as SHEIN coordinate thousands of small, nearby garment factories to fulfil large numbers of small-lot, fast-switch orders. Dispatching each order operation to the most suitable factory while respecting process routes creates a Flexible Job Shop Scheduling Problem with sequence-dependent setup times (FJSP-SDST), which is NP-hard and must be solved repeatedly as new batches arrive. Current end-to-end deep reinforcement learning (DRL) schedulers either reschedule every batch from scratch or disregard setup costs, which undermines efficiency. We introduce a Cyber-Physical Internet (CPI) scheduling framework that provides routers for every factory and the company, enabling them to cache solved schedules and broadcast real time factory states. This approach skips redundant computations and supplies fresh setup time data. Within this framework, we have developed a Hierarchical Attention Network based Reinforcement Learning (HANRL) scheduler to model the interactions between orders and factories, as well as factory competition and setup costs. Experiments on synthetic and public benchmarks demonstrate that HANRL reduces makespan and improves generalization over state-of-the-art DRL baselines, all while retaining sub-second decision times. This proves the suitability of HANRL for large scale social manufacturing environments.