Li Yan, Wang Jun-de, Chen Zuo-ru, Zhou Xue-tie, Huang Zhong-hua
Laboratory of Advanced Spectroscopy, Nanjing University of Science and Technology, Nanjing 210014, China.
Guang Pu Xue Yu Guang Pu Fen Xi. 2002 Oct;22(5):758-60.
The concentration determination abilities of four multivariate calibration methods--classical least squares (CLS), partial least squares (PLS), kalman filter method (KFM) and artificial neural network (ANN) were compared in this paper. Five air toxic organic compounds--1,3-butadiene, benzene, o-xylen, chlorobenzene, and acrolein--whose FTIR spectra seriously overlap each other were selected to compose the analytical objects. The evaluation criterion was according to the mean prediction error (MPE) and mean relative error (MRE). Results showed that PLS was superior to other methods when treating multicomponent analysis problem, while there was no comparable difference between CLS, KFM and ANN.
本文比较了四种多元校准方法——经典最小二乘法（CLS）、偏最小二乘法（PLS）、卡尔曼滤波法（KFM）和人工神经网络（ANN）的浓度测定能力。选择了五种空气中有毒有机化合物——1,3-丁二烯、苯、邻二甲苯、氯苯和丙烯醛，它们的傅里叶变换红外光谱严重重叠，以此构成分析对象。评价标准依据平均预测误差（MPE）和平均相对误差（MRE）。结果表明，在处理多组分分析问题时，PLS优于其他方法，而CLS、KFM和ANN之间没有显著差异。
