Abstract:

Kernel distance-based support vector data description (SVDD) for batch process monitoring exhibited poor monitoring precision by setting control limit from the largest kernel distance in historical process dataset but ignoring hyperspherical irregularity in high dimensional space. A kernel similarity based SVDD monitoring method was proposed for batch process monitoring. Kernel similarity was taken as kernel function value between support vectors and data samples for testing. The weighted summation of kernel similarity and distance of support vectors at various time points was utilized to set dynamic control limit for data samples of batch process to be monitored. Batch process monitoring was achieved by judging if kernel distance of test sample exceeded the dynamic control limit. This monitoring method considered irregularity of hypersphere, local distribution characteristics of process dataset in high dimensional space, and spontaneity of data samples, so that it could improve accuracy in batch process monitoring. Method effectiveness was demonstrated by numerical simulation and metal etching process in semiconductor manufacturing.

Key words: kernel similarity, support vector data description, dynamic monitoring, batch process

摘要：

基于支持向量数据描述的间歇过程监测方法选择历史过程数据中最大的核距离作为控制限，忽略了高维空间中超球体的不规则性，导致基于该方法的过程监测精度不高。针对上述问题，提出了一种基于核相似度支持向量数据描述的间歇过程监测方法，将间歇过程数据待监测样本与支持向量之间的核函数值作为相似度权重，利用该相似度对不同时刻的支持向量球心距加权求和，得到待监测间歇过程数据样本的动态控制限，通过判断待监测样本的球心距是否超过其动态控制限，实现间歇过程监测。所提方法综合考虑了超球体的不规则性和过程数据在高维空间分布的局部特性，以及间歇过程数据待监测样本的时变性，提高了间歇过程监测的准确性。利用数值仿真实验和半导体金属刻蚀实验验证了该方法的有效性。

关键词: 核相似度, 支持向量数据描述, 动态监测, 间歇过程