局部偏最小二乘回归建模参数对近红外检测结果的影响研究

报道了在局部加权(LWR)回归方法基础上，自主改进的更简单、实用的局部偏最小二乘回归(LPLS)的原理和方法。并以云南优质烤烟为实验材料，在国产光栅漫反射型近红外仪器上，研究了主成分数以及局部建模样品数对检测结果的影响。结果表明：应用交叉验证方法推荐的尼古丁组分模型主成分数并不是最优，通过适当降低主成分数可提高检测效果;局部建模样品数为30～50个时总糖、总氮、尼古丁预测准确度的提高幅度可分别达7%，14%，10%以上。该方法能有效提高近红外数学模型的预测准确度，是建立具有高度适应性近红外数学模型的有效方法。

Influence of LPLS Algorithm Parameters on NIR Veracity

The theory of local partial least square (LPLS) algorithm was described based on locally weighted regression algorithm (LWR). The influence of data processing parameters, such as principal component numbers and local set-up sample number in LPLS mode, on the NIR veracity was studied with homemade grating diffuse NIR instrument using Yunnan flue-cured tobacco. Results showed that for nicotine model, the principal component number decided by cross validation was not the best choice, and better results could be achieved by reducing the principal component number; using 30-50 samples to set up NIR model, the veracity of total sugar, total nitrogen, and nicotine could be improved by 7%, 14% and 10%, respectively. So, LPLS algorithm can effectively improve NIR model's veracity, and is a good method to set up robust NIR models.