基于扩散K近邻距离的间歇过程故障诊断
Fault detection of batch process based on diffusion K-nearest neighbors distance

针对间歇过程多模态、变量非线性、非高斯分布等特征, 提出一种基于扩散K近邻距离的故障诊断方法. 该方法首先在样本集完全图中应用马尔科夫随机游走定义带有分量权重的扩散距离, 可以有效提取数据样本的关 联信息和统计特征, 然后应用K近邻规则方法对样本数据进行故障诊断. 这种应用扩散距离替换传统K近邻规则欧 式距离的统计方法, 既可以提升对数据样本关联性信息的有效提取能力, 又可以使得K近邻规则处理非线性、多模 态检测问题的性能得以保持. 通过在半导体蚀刻批次过程中的仿真应用, 与传统线性、非线性方法的对比分析, 实 验结果验证了方法的有效性.
According to the multi-mode, variable nonlinear and non-Gaussian distribution characteristics of the batch process, we propose a new fault detection method based on diffusion K-nearest neighbor distance (FD–DDKNN). In this work, the diffusion distance with component weight which can effectively fetch correlation information and statistical characteristics in data samples is defined through Markovian random walk in complete graph of the samples set. Then, the adapted K-nearest neighbor rule (KNN) method is applied to data samples to detect faults. This method that replaces diffusion distance to conventional Euclidean distance in K-nearest neighbor rule, not only raises the ability of fetching relevance information in data samples, but also improves the performance of dealing with nonlinear and multi-mode characteristics based on KNN rule for detection problem. By the simulation application in the semiconductor etching batch process, compared with the traditional linear and nonlinear methods, the experimental results validate the effectiveness of the proposed method.