基于密度标准误差的局部保持投影故障检测策略

Fault detection strategy based on density standard error associated with locality preserving projections

中文关键词  局部保持投影  k近邻  多模态过程  主元分析  故障检测

英文关键词  locality preserving projections  k nearest neighbor  multimodal process  principal component analysis  fault detection

中文摘要

      针对协方差结构具有显著差异的多模态过程故障检测问题, 本文提出一种基于密度标准误差的局部保持 投影故障检测策略(LPP–DSE). 首先, 根据样本距离矩阵确定样本截止距离; 接下来, 应用截止距离计算每个样本的 本质密度及其前k近邻样本的估计密度; 最后, 通过样本的密度误差及其k近邻密度的标准差构建统计量并完成过 程监控. 本文方法通过应用局部保持投影(LPP)对过程数据进行维数约减可以保证过程监控的及时性; 同时, 通过设 计密度标准误差(DSE)统计量可以有效提高多模态过程的故障检测率. 此外, 本文给出基于贡献图的诊断方法能够 准确识别故障发生的原因. 通过数值例子和半导体工业实例测试, 并与主元分析、邻域保持嵌入、局部保持投影、 k近邻故障检测等方法比较, 实验结果进一步验证了LPP–DSE方法的有效性.

英文摘要

      Aiming at the fault detection of multimodal processes with significant variation in covariance of each mode, a fault detection strategy based on density standard error associated with locality preserving projections (LPP–DSE) is proposed in this paper. Firstly, calculate the cutoff distance according to sample distance matrix. Secondly, calculate respectively the intrinsic density and the estimated density of a sample through cutoff distance. Finally, build a new statistic to accomplish process monitoring. In LPP–DSE, the timeliness of process monitoring is guaranteed by using locality preserving projections (LPP); meanwhile, the fault detection rate of a multimode process is improved through using density standard error (DSE) statistic. Moreover, the proposed fault diagnosis method based on contribution chart is able to identify accurately the abnormal variables. Compared with principal component analysis, neighborhood preserving embedding, LPP, k nearest neighbor rule and other methods, the effectiveness of LPP–DSE is verified by a numerical case and semiconductor etching process.