Abstract: Most research on apparel workshop scheduling assumes that each workstation has the same production load. In reality, during the initial scheduling phase prior to production in apparel manufacturing enterprises, managers indeed consider the production efficiency of each worker to be the same. Although this relaxation model facilitates the ease of constructing of objectives, handling constraints, and improving algorithmic efficiency across various production scenarios, it deviates from reality, resulting in low and unstable actual production line balance rates despite low theoretical production line loss rates. In response to the heterogeneity of workers' processing time, this paper proposes an integrated method based on existing theoretical research, adopting a load coefficient prediction model and the grey wolf optimization (GWO) algorithm to address the imbalance in  trouser-hanging production lines caused by worker differences. The objective is to to minimize the smoothness index (SI) to optimize the overall balance of the hanging line.
The optimization process comprises two main modules: a neural network-based load coefficient prediction module and a process balancing module based on the GWO algorithm. Prior to prediction, the factors affecting worker efficiency in existing research were expanded to include apparel-specific factors, constructing a framework for workers' processing time heterogeneity factors. In the prediction module, leveraging RFID and IoT technologies, a dataset was constructed from the perspective of personalized influencing factors, focusing on apparel-related factors and some collectible physiological factors. The neural network was trained using Bayesian optimization to achieve optimal parameter settings. The evaluation index MAE of the optimized model in predicting the workstation load coefficient reached 0.091, indicating an acceptable prediction accuracy. 
With the predicted workstation load coefficients as constraints on workers' processing time, the GWO  algorithm was adopted to optimize the problem model. The optimization results indicated that the GWO algorithm demonstrated superior algorithmic performance and stability. This data-driven, concise, and comprehensive intelligent decision-making model can effectively address the issue of production line balancing caused by varying processing time in garment manufacturing. Post-optimization validation results revealed that the balancing index of workstations decreased from 66.4 to 13.8. Therefore, this model significantly enhances the actual balance rate of the production line.
This study conducts static scheduling with the load coefficient of the entire order as the prediction target. Dynamic scheduling and static scheduling are not contradictory, and the results can provide a reasonable initial schedule for dynamic scheduling. Additionally, the real-time collection and storage of worker efficiency through RFID and IoT technologies lay the foundation for adopting this model for dynamic scheduling. The neural network prediction function module and optimization module are independent, thus possessing strong generality and integrability, allowing scholars to construct prediction datasets and choose more suitable process balancing optimization algorithms based on the problem's characteristics.

Key words: perturbation factors, processing time heterogeneity, GWO algorithm, Bayesian optimization, actual balance rate

摘要： 服装生产线工人易受到各种扰动和非扰动因素影响，导致加工时间存在较大异质性，进而引发生产线不平衡问题。针对该问题，提出了集成神经网络和灰狼算法的裤装吊挂流水线平衡优化模型。首先，分析了工人加工时间异质性的影响指标，并构建了相应的数据集；其次，使用神经网络对工人的生产负荷系数进行预测；最后，将通过贝叶斯优化后预测的工位负荷系数作为约束条件，使用灰狼算法对吊挂流水线进行全局优化。结果表明：经过优化，裤装吊挂流水线的均衡指数从66.4降低至13.8，生产线实际平衡率显著提高。该裤装吊挂流水线平衡优化模型可为服装智能调度和柔性生产相关研究提供参考。

关键词: 扰动因素, 加工时间异质性, 灰狼算法, 贝叶斯优化, 实际平衡率