Abstract: Chemical production process has the characteristics of high dimension and strong nonlinearity. For the deficiency of traditional neighborhood preserving embedding (NPE) algorithm in feature extraction of non-linear data, a Gaussian kernel function is introduced to transform data from non-linear input space to linear feature space. Kernel neighborhood preserving embedding (KNPE) algorithm can extract the non-linear structure of data better on the basis of constructing local spatial feature structure. By a case study on the Tennessee Eastman (TE) simulation process,  and SPE statistics are constructed for fault detection, which proves that KNPE method can detect the occurrence of non-linear faults faster and more accurately than NPE and KPCA methods.

Key words: Chemical failure, manifold learning, kernel neighborhood preserving embedding algorithms, fault detection

摘要： 化工生产过程具有维数高、非线性强等特点。针对传统的邻域保持嵌入(NPE)算法对非线性数据特征提取不足的缺陷，引入高斯核函数，将数据由非线性的输入空间转换到线性的特征空间。核邻域保持嵌入(KNPE)算法在构建局部空间特征结构的基础上，能够更好地提取数据的非线性结构。通过以田纳西-伊斯曼(TE)仿真过程为例，构造 和SPE统计量进行故障检测，证明了KNPE方法比NPE和KPCA方法能够更快更准确的检测出非线性故障的发生。

关键词: 化工故障, 流形学习, 核邻域保持嵌入算法, 故障检测