一种基于分级输入训练神经网络的非线性主元分析

Nonlinear Principal Component Analysis Based on Hierarchical Input-training Neural Network

摘要: 基于输入训练神经网络的非线性主元分析（PCA）能够有效地提取过程变量的非线性主元，但是存在主元的个数不能通过网络训练确定，且各个主元重要程度在神经网络中无法区分等缺点，本文提出一种分级输入自调整神经网络，并进一步提出基于此网络的非线性PCA，通过多级输入自调整神经网络，将主元按顺序找出，且根据主元对过程数据的预测误差定量地确定出主元的个数，克服了上述缺点．

Abstract: Nonlinear principal component analysis(PCA) based on neural network with inputs training can effectively extract nonlinear principal components(PCs) from process variables,but the number of PCs can not be decided by training network,and the order of PCs can not be distinguished.In order to overcome these defaults,a hierarchical input-training neural network is proposed,and a nonlinear PCA based on this kind of network is presented,which can orderly find nonlinear PCs and quantitatively determine the number of PCs according to the prediction error of process data based on PCs.