论文部分内容阅读
针对光谱数据局部效应显著、变量间的严重共线性等特征,实施基于转换权向量约束优化的稀疏偏最小二乘回归新技术。它通过在特征变量提取的优化目标函数中加入转换权向量的罚函数,一并执行光谱波长选择和特征提取,随后再构建特征变量与性质变量间的校正模型。其中,罚函数中的最佳罚因子和校正模型中的最优PLS成分数,经由各自取值范围内一定数量试验水平的两因素全面试验设计与校正模型精度调控的交叉验证方式确定。最后,通过面粉生面团切片的近红外光谱数据的试验应用研究,结果显示该技术光谱数据波长选择和特征提取稳健,去噪明显,并显著提高了光谱数据定量校正模型的预测能力。
Aiming at the significant local effects of spectral data and the serious collinearity among variables, a new sparse partial least squares regression technique based on conversion weight vector optimization was implemented. It adds spectral wavelength selection and feature extraction by adding the penalty function of conversion weight vector to the optimization objective function extracted by feature variables, and then reconstructs the correction model between feature variables and property variables. Among them, the optimal penalty factor in the penalty function and the optimal PLS component in the calibration model are determined by the cross-validation of the two-factor full-scale test design and the calibration model precision control with a certain number of test levels within their respective ranges. Finally, the application of near-infrared spectral data of dough dough slices was studied. The results show that the wavelength selection and feature extraction of the spectral data are robust and the denoising is obvious, and the predictive ability of quantitative correction model of spectral data is significantly improved.