离散小波变换-遗传算法-交互检验法用于近红外光谱数据的高倍压缩与变量筛选

A Discrete Wavelet Transform-Genetic Algorithm-Cross Validation Approach for High Ratio Compression and Variable Selection of Near-infrared Spectral Data

用遗传算法(GA)与交互检验(CV)相结合建立了一种用于对近红外光谱(NIR)数据及其离散小波 变换(DWT)系数进行变量筛选的方法,并应用于烟草样品中总挥发碱和总氮的同时测定。结果表明:NIR数 据经DWT压缩为原始大小的3.3%时基本没有光谱信息的丢失;有效的变量筛选可以极大地减少模型中的 变量个数,降低模型的复杂程度,改善预测的准确度。

An approach for high ratio compression and variable selection of near-infrared (NIR) spectra is proposed. The informative variables, wavelength points or approximation coefficients of discrete wavelet transform (DWT) of NIR spectra, could be selected by combination of genetic algorithm (GA) and cross-validation ( CV) procedure. These selected Variables were used in the determination of total volatile alkaloids (TVA) and total nitrogen (TN) in tobaccos by partial least squares (PLS) method. It is proved that there is almost no loss,of information when the spectral data are compressed to 3.3% of its original size. The method can significantly reduce the number of variables used in the prediction model, decrease the complexity of the model, and improve the predictive accuracy.