Neural network-assisted optimisation of production rescheduling under incomplete data
提出一个集成神经网络与混合整数线性规划的决策支持系统，用于单阶段多生产线多产品动态制造环境下的生产调度与重调度，在数据缺失时预测换线时间，显著提升求解速度与调度质量。
This work presents a novel decision-support system integrating Neural Networks (NNs) and a Mixed-Integer Linear Programming (MILP) model to enhance production scheduling and rescheduling in single-stage, multi-line, multi-product dynamic manufacturing facilities. Production engineers frequently face unexpected challenges requiring immediate schedule adjustments and regular product changeover operations which increase downtime and waste. To tackle these challenges, a novel framework is developed for quick generation of flexible production schedules. A NN is trained to embed products into a space where distances reflect changeover durations, facilitating their clustering by production line while minimising internal changeover times. In the case of absent data, another NN predicts unknown changeover times based on products’ characteristics. The predictions and allocations produced are incorporated into an MILP scheduling model, tightening the formulation and improving computational performance. The predictive accuracy of the NN is evaluated against standard methods, whereas the proposed scheduling framework against a conventional monolithic MILP approach. Several cases concerning two real industrial plants show that the proposed system achieves better production schedules in significantly faster solution times. A weekly production horizon with daily disruptions is studied, further highlighting the flexibility and robustness of the approach, making it a promising tool for real-time decision-making in production environments.