siPLS-LASSO的近红外特征波长选择及其应用

摘要

近红外技术广泛应用于食品、药品等生产过程和产品质量检测,具有样品无需预处理、成本低、无破坏性、测定速度快等优点.但是,全光谱数据维数高、冗余信息多,直接应用于建模会导致模型复杂性高、稳定性差等问题.siPLS是最常见的光谱数据降维方法,但是难以处理光谱数据的共线性问题.LASSO是一种相对新的数据降维方法,但在小样本应用中具有不稳定性.针对siPLS和LASSO在近红外光谱数据应用中存在的问题,提出了基于siPLS-LASSO的近红外特征波长选择方法,并将其应用于秸秆饲料蛋白固态发酵过程pH值监测.该方法首先采用siPLS算法,实现对光谱波长最佳联合子区间的优选;然后,对优选联合子区间使用LASSO算法进行特征波长选择,在此基础上建立 PLS校正模型.同时,将siPLS-LASSO方法与其他传统特征波长选择方法进行了对比.结果表明:建立在siPLS-LASSO方法优选33个特征波长基础上的PLS模型预测结果更好,其预测方差(RMSEP)和相关系数(Rp)分别为0.0711和0.9808;所提siPLS-LASSO方法有效选取了特征波长,提高了模型预测性能.%Near-infrared spectroscopy(NIR)is widely used in entire production processes and product quality test,especially in food and drug industries.It has many advantages,e.g.no requirement of sample pretreatment,low cost,non-destructive de-tection,and fast determination.However,the application of the whole spectrum data in modeling can lead to complexity and poor stability.The synergy interval PLS(siPLS)is the most common dimensionality reduction method for spectral data.Howev-er,it cannot deal with the collinearity problem of spectral data.Least absolute shrinkage and selection operator(LASSO)is a relatively new method for data dimensionality reduction.However,when it comes to small samples,its instability cannot be ig-nored.For disadvantages of siPLS and LASSO in NIR calibration,a novel wavelength selection method named siPLS-LASSO was proposed.It was validated in a wheat-straw solid-state fermentation process by monitoring pH values.In the method,siPLS was firstly used to selected intervals of NIR spectroscopy.Secondly,LASSO was used to select wavelengths on the selected in-tervals.Finally,the selected wavelengths were used to construct PLS model for prediction.For comparisons,several conven-tional wavelength selection methods were also studied.In the case study,33 wavelengths were eventually selected by the siPLS-LAS-SO method and used for PLS modelling.The RMSEP and Rp of the model were 0.071 1 and 0.980 8 respectively.Results showed that the proposed siPLS-LASSO was an effective method of wavelength selection and can improve prediction performance of models.