最小角回归结合竞争性自适应重加权采样的近红外光谱波长选择

Least Angle Regression Combined With Competitive Adaptive Re-Weighted Sampling for NIR Spectral Wavelength Selection

近红外光谱分析技术对检测样品无损伤且检测速度快、精度高,因此被广泛应用在了药品检测、石油化工等领域,尤其近年来机器学习和深度学习建模方法的深入应用使其具备了更准确的检测性能。然而,样品的近红外光谱数据具有比较高的维度且存在谱间重合、共线性和噪声等问题,对近红外光谱模型的性能产生消极影响,此时样品有效特征波长的筛选极为重要。为了提高近红外光谱定量和定性分析模型的准确性和可靠性,提出了一种近红外光谱变量选择方法,其结合了最小角回归(LAR)和竞争性自适应重加权采样(CARS)的优点,具有更优的性能。该方法利用LAR初步筛选样品全谱区的特征波长,接着利用CARS对筛选出来的特征波长进一步选择,从而有效去除无关特征波长。为验证该方法的有效性,从定量和定性分析两个方面评价该方法。在定量分析实验中,以FULL,LAR,CARS,SPA和UVE作为对比方法,以药品样品数据集为实例建立PLS回归分析模型,经LAR-CARS筛选出的变量建立的PLS模型在药品数据集表现出较高的预测决定系数和较低的预测标准偏差。在定性分析实验中,以SVM,ELM,SWELM和BP作为对比方法、不同比例训练集的药品数据集为实例建立分类模型,经LAR-CARS筛选出的变量建立的SVM分类模型精度最高达100%。从实验结果可见,LAR-CARS可有效的筛选出表征样品特征的波长,利用其筛选出的波长建立的定量、定性分析模型具有更好的鲁棒性,可用于样品光谱的特征波长筛选。

Near-infrared spectroscopy is widely used in drug detection,petrochemical industry,etc.,because it has no damage to the samples,and the detection speed is fast,and the accuracy is high.In particular,it has more accurate detection performance with the in-depth application of machine learning and deep learning modeling methods in recent years makes.However,the NIR spectral data of the sample has relatively high dimensions and has problems such as spectral overlap,collinearity and noise,which will negatively impact the performance of the NIR spectral model.In this case,the selection of effective characteristic wavelength points of the sample is extremely important.In order to improve the accuracy and reliability of the quantitative and qualitative analysis models of NIR spectra,a variable selection method for NIR spectra is proposed,which combines the advantages of Least Angle Regression and Competitive Adaptive Re-weighted Sampling,and has better performance.In this method,LAR was used to preliminarily screen the characteristic wavelengths in the whole spectrum of the sample,and then CARS was used to further select the selected characteristic wavelengths to effectively remove the irrelevant characteristic wavelengths.In order to verify the effectiveness of the method,the method was evaluated from two aspects of quantitative and qualitative analysis.In the quantitative analysis experiment,PLS regression analysis model was established using FULL,LAR,CARS,SPA and UVE as comparison methods and drug sample data set as example.PLS model established by variables screened by LAR-CARS showed higher predictive determination coefficient and lower predictive standard deviation in drug data set.In the qualitative analysis experiment,the classification model was established with SVM,ELM,SWELM and BP as comparison methods and drug data sets with different proportions of training sets.The accuracy of the SVM classification model established by the variables screened by LAR-CARS reached the highest 100%.From the experimental results,it can be seen that LAR-CARS can effectively select the wavelength points that the characteristics of the sample,and the quantitative and qualitative analysis model established by using the selected wavelength points has better robustness and can be used for the characteristic wavelength screening of the sample spectrum.