数据驱动下的半导体生产线动态调度方法

Data-driven dynamic scheduling method for semiconductor production line

中文关键词  半导体生产线  动态调度  数据驱动  特征选择  分类算法

英文关键词  semiconductor production line  dynamic scheduling  data-driven  feature selection  classification method

中文摘要

      本文研究了一种数据驱动下的半导体生产线调度框架, 该框架基于调度优化数据样本, 应用机器学习算法,获得动态调度模型, 通过该模型, 对于半导体生产线, 能够根据其当前的生产状态, 实时地定出近似最优的调度策 略. 在此基础上, 利用特征选择和分类算法, 提出一种生成动态调度模型的方法, 并且具体实现出一种混合式特征选择和分类算法的调度模型: 先采用过滤式特征选择方法对生产属性进行初步筛选, 然后再采用封装式特征选择和分类方法生成模型以提高模型生成的效率. 最后, 在某实际半导体生产线上, 对在所提出的框架上采用6种不同算法实现的动态调度模型进行测试, 并对算法性能数据和生产线性能据进行对比和分析. 结果表明, 数据驱动下的动态调度方法优于单一的调度规则, 同时也能满足生产线调度实时性要求. 在数据样本较多的情况下, 建议采用本文所提出的方法.

英文摘要

      This paper investigates a framework of data-driven scheduling method for semiconductor production line. Based on optimized scheduling data samples, this framework applies machine learning algorithm to obtain the dynamic scheduling model. Using this model, we determine the approximate optimal scheduling strategy for the production line under the current production status. On this basis, we propose a method based on feature selection and classification to generate a dynamic scheduling model, and provide a realization by means of a hybrid algorithm of feature selection and classification as such: production attributes are primarily selected by using a filter selection algorithm, and then the wrapped selection and classification algorithm is employed to produce the scheduling model. On a real semiconductor production line, the proposed framework is tested by using the dynamic scheduling models realized by 6 different algorithms. The performance data of algorithms and the production performance using different scheduling methods are compared and ana- lyzed. The data shows that the data-driven dynamic scheduling methods are superior to the method with simple dispatching rule, and satisfy the real-time scheduling requirements for the production line. In the case with large size of sample data, the application of the proposed method is even preferable.