基于核特征近邻指标的批过程监视及仿真研究

Nearest Neighbors Index of Kernel Features Based Batch Processes Monitoring and Simulation Research

批次生产过程为了不断满足快速增长的市场需求,经常进行生产装置的重组、负荷大小的调整以及品种的更换等,使批次过程的采样数据呈现出非线性、非高斯、多工况等复杂特征。针对这些特征,提出在批次过程的核特征空间利用核主元的k个近邻构建故障检测统计量指标进行过程监视的方法。将批次过程正常工况的数据按批次方向展开并进行标准化。采用核方法将其投影到核特征空间,确定出核主元,组成新的建模样本数据。通过相似性原理找到每个样本的k个近邻,构建k近邻的距离平方和故障检测指标,并采用核密度估计法确定出正常工况指标的统计控制限。利用SPE统计量对核残差空间同时进行监视。通过一个半导体生产过程的仿真实验结果表明了所提方法的有效性。

In order to meet the increasing market demand, batch processes often carry out recombination of producing device, adjustment of load size and replacement of variety, and so on. So, the characteristics of sampling data of batch processes appear nonlinear, non-Gaussian and multi-modes. A process monitoring method based on k nearest neighbor index of kernel feature space was proposed for batch processes. Firstly, the normal mode data of batch process was unfolded along batch direction and standardization was implemented. Secondly, after kernel method was used to put the raw data into the kernel feature space and the kernel principal components were extracted, a new modeling data set was formed. Then k nearest neighbors of each modeling data were found, k nearest neighbor distance square sum was constructed and calculated, meanwhile, they were treated as the monitoring statistics index. The control limit of the statistics index was determined by the kernel density estimation method. Moreover, SPE of the kernel feature space was used as the other statistic index for process monitoring. Finally, the results of simulation experiment of a semiconductor production process show the effectiveness of the proposed method.