基于多尺度动态核主元分析的化工过程故障检测

Fault Detection of Chemical Process Based on Multiscale Dynamic Kernel Principal Component Analysis

针对化工过程数据的多尺度性和非线性特性,提出一种多尺度动态核主元分析方法。使用小波变换分析数据的多尺度特性,借助核函数来解决非线性映射问题,同时解决了噪声和干扰造成的各变量数据具有时间序列动态性问题。在此基础上,提出一种基于矩阵相似度量的核函数参数选优方法。将上述方法应用于TE模型的故障检测过程中,仿真结果表明,该方法提高了过程性能监视和故障检测的准确性,优于线性主元分析法的检测效果。

According to muhiscale and nonlinear properties of chemical process data, a multiscale dynamic kernel principal component analysis method was presented. Used wavelet transform to analyze multiscale characteristic of data,used kernel function to solve nonlinear problem and time series＇ dynamic of every variable data caused by noise and disturbance. A method based on matrixes similarity measures to choose parameters of kernel function best was presented. The simulations on the Tennessee Eastman （TE） process indicate that the performances of process monitoring and fault diagnosis by the presented method are superior to that by linear principal component analysis.