基于主成分分析和小波神经网络的近红外多组分建模研究

Near Infrared Multi-component Prediction Model Based on Principal Component Analysis and Wavelet Neural Network

将小麦叶片原始光谱经过预处理后,采用主成分分析(PCA)对数据进行降维,取前3个主成分输入小波神经网络,建立了基于主成分分析和小波神经网络的近红外多组分预测模型(WNN);进一步研究了小波基函数个数的选取(WNN隐层节点数)对小波神经网络模型性能的影响,并将WNN模型与偏最小二乘法(PLS)和传统的反向传播神经网络(BPNN)模型进行了比较。结果表明,所建立的WNN模型能用于同时预测小麦叶片全氮和可溶性总糖两种组分含量,其预测均方根误差(RMSEP)分别为0.101%和0.089%,预测相关系数(R)分别为0.980和0.967。另外,在收敛速度和预测精度上,WNN模型明显优于BPNN和PLS模型,从而为将小波神经网络用于近红外光谱的多组分定量分析奠定了基础。

A new near infrared model for multi-component prediction was established based on principal component analysis（PCA） and wavelet neural network（WNN） methods.First,original near infrared spectra from wheat leaves were pre-processed,and their principal components were extracted by PCA,which could reduce the dimensionality of spectrum data.The first three principal components were taken as inputs of wavelet neural network（WNN）,and the influence of neuron number in the hidden layer of WNN on the properties of model was further analyzed.The results indicated that the WNN model can be applied to the simultaneous determination of total nitrogen and soluble sugar contents of wheat dry leaves.The root mean square errors of prediction（RMSEP） of total nitrogen and soluble sugar contents of wheat leaves by the WNN model were 0.10% and 0.09%,with correlation coefficients（R） of 0.980 and 0.967,respectively.In addition,the WNN model was superior to the methods of back-propagation neural network（BPNN） and partial least squares（PLS） on convergence speed and prediction precision.This study provided an approach for quantitative analysis of multi-components based on near infrared spectrum.