Abstract: Aiming at the strong nonlinearity and over-high dimension features of the monitoring data in chemical system,a new improved Local Fisher Discrimination Analysis (LFDA) method was proposed based on local linear analysis of labeled data and global analysis of training samples.The label information was used to redefine boundaries between classes according to the minimum Euclidean distance from every sample to other samples in different classes,and local between-class scatter matrix was reconstructed.Total scatter matrix of training samples was introduced to preserve global structure and overcome the disadvantage of computing only local scatter matrix.The simulation on Tennessee-Eastman(TE) process data and a compressor's monitor data showed that the proposed method had better capability of processing nonlinear data by comparing with methods of other nonlinear discriminant analyses,and the more accuracy rate on abnormal identification was obtained.

Key words: feature extraction, status monitoring, manifold learning, local linear discriminant analysis, Tennessee-Eastman process

摘要： 针对化工系统监测数据呈现出的强非线性、数据高维等特点,将标注样本的局部线性分析与训练样本的全局分析相结合,提出一种改进的局部线性判别分析方法。利用训练样本标签信息,以异类样本点间的最小欧式距离重新定义异类样本之间的边界,构建了新的局部类间离散度矩阵;引入全局离散度矩阵强化训练样本全局分析,克服了只计算局部离散度矩阵的缺点。在田纳西—伊斯曼过程数据和某企业压缩机组监测数据上进行了仿真实验,结果表明所提方法与局部线性判别分析等若干种非线性分析方法相比,具有更好的非线性处理能力,可以获得更高的异常状态识别准确率。

关键词: 特征提取, 状态监测, 流形学习, 局部线性判别分析, 田纳西—伊斯曼过程