基于局部保持投影–加权k近邻规则的 多模态间歇过程故障检测策略

Fault detection strategy based on locality preserving projections-weighted k nearest neighbors in multimodal batch processes

中文关键词  局部保持投影  权重k近邻规则  间歇过程  故障检测

英文关键词  locality preserving projections  weight k nearest neighbor rule  batch process  fault detection

中文摘要

      针对多模态间歇过程故障检测问题，本文提出一种基于局部保持投影—加权k近邻规则(Locality Preserving Projections—Weighted k Nearest Neighbors, LPP-WkNN)的故障检测策略。首先，应用局部保持投影(Locality Preserving Projections, LPP)方法将原始数据投影到低维主元子空间；接下来，在主元子空间中，应用样本第k近邻的局部近邻集确定每个样本的权重并计算权重统计量DW；最后，应用核密度估计方法确定Dw控制限并进行故障检测。本文方法应用LPP对过程数据进行维数约减，既能够降低训练过程离群点对模型的影响，又能够降低在线故障检测的计算复杂度。同时，WkNN方法通过引入权重规则能够使得过程故障检测统计量分布具有单模态结构。相比传统的kNN统计量，本文引入的权重统计量具有更高的故障检测性能。通过数值例子和半导体蚀刻过程的仿真实验，并与主元分析(Principal Component Analysis, PCA)、kNN、WkNN、LPP-kNN等方法进行比较，实验结果验证了本文方法的有效性。

英文摘要

      Aiming at fault detection in multimodal batch process, fault detection strategy based on locality preserving projections-weighted k nearest neighbors (LPP-WkNN) in multimodal batch processes is proposed in this paper. First, raw data are projected into low dimensional principal component subspace using LPP. Then, apply the local nearest neighbor set of the k-th nearest neighbor of the samples to determine the weight of samples and construct the weighted statistics DW. Finally, apply kernel density estimation to determine control limits of DW to detect faults. Dimensionality reduction using LPP is capable of not only eliminating the influence of outliers on the model, but also reducing the computational complexity of fault detection. At the same time, the WkNN method can make the statistics of samples have a single model structure by introducing weight rules. Compared with the traditional kNN statistics, the weight statistics introduced in this paper have higher fault detection performance. The efficiency of the proposed strategy is implemented in a numerical case and in the semiconductor etching process. The experimental results indicate that the proposed method outperforms principal component analysis (PCA), LPP, kNN, WkNN and LPP-kNN.