基于L1-L2联合范数约束的中药近红外光谱波长选择

A WAVELENGTH SELECTION METHOD FOR NEAR INFRARED SPECTROSCOPY OF CHINESE MEDICINE BASED ON L1-L2 NORM SIMULTANEOUS CONSTRAINT

针对近红外光谱分析中存在的高维数据降维、多重共线性及模型稀疏性问题,提出一种基于正则偏最小二乘RPLS(Regularization Partial Least Squares)的近红外光谱波长变量选择方法。该方法在偏最小二乘回归模型中同时引入L1和L2范数罚正则项,使模型产生稀疏性,通过交替迭代算法求解主成分载荷系数的稀疏解,实现光谱数据降维和重要波长变量的自动选择。对当归近红外光谱进行正则偏最小二乘波长选择实验。结果表明,与CARS(Competitive Adaptive Reweighted Sampling)随机蛙跳等变量选择方法相比,正则偏最小二乘方法在选择变量数及模型的预测能力方面均具有一定的优势。

In the view of the problems of dimensionality reduction, multiple collinearity and model sparsity existed in the near infrared spectroscopy analysis, we proposed a wavelength variable selection method for near infrared spectroscopy based on regularization partial least squares （RPLS）. The L1 and L2 norm penalty regularization terms were introduced in the partial least squares regression （PLSR） model to make the model sparse. We utilized the alternative and iterative algorithm to obtain the sparse solution of principal component load coefficients so as to achieve dimensionality reduction of the spectral data and the automatic selection of the important wavelength variables. Regular partial least squares wavelength selection experiments were conducted for near infrared spectra of Angelica sinensis. The experimental results show that compared with variable selection methods such as CARS and random-frog, RPLS method has certain advantages in terms of selected variable number and prediction accuracy of model.