用小波包变换广义回归神经网络法处理硝基苯胺异构体重叠紫外吸收光谱

Processing of Overlapping Ultraviolet Absorption Spectra of Nitroaniline Isomers Using a Wavelet Packet Transform Based Generalized Regression Neural Network

开发了一种小波包变换广义回归神经网络(WPTGRNN)法,用于处理对硝基苯胺、邻硝基苯胺和间硝基苯胺重叠的紫外吸收光谱,达到不经预先化学分离进行同时测定的目的。WPTGRNN法结合小波包变换和广义回归神经网络(GRNN)的特点,改进除噪质量和预测能力。通过最佳化实验,选择了小波函数、小波包分解水平及GRNN的平滑因子。实验结果表明,WPTGRNN法的预测标准误差为1.08μg/mL,预测相对标准误差为4.20%,与小波变换广义回归神经网络、广义回归神经网络和主组分回归3种方法进行比较,WPTGRNN法优于其他3种方法。

A wavelet packet transform based generalized regression neural network （WPTGRNN） was developed to process overlapping ultraviolet absorption spectra of p -nitroaniline, o -nitroaniline and m -nitroaniline. The determination was simultaneously carded out without any chemical separation steps in advance. This method combines wavelet packet transform with generalized regression neural network （GRNN） for improving quality of denoise and enhancing its ability of prediction. By optimization, wavelet function, decomposition level and smoothing factor of GRNN were selected. Experimental results showed standard error prediction and relative standard error prediction of WPTGRNN methods are 1.08 μg/mL and 4.20%, respectively. The WPTGRNN method performed better than other three, namely wavelet transform GRNN, GRNN and principal component regression.