基于偏最小二乘法建立大株红景天片素片硬度近红外光谱预测模型

Near infrared prediction model for hardness of Rhodiola grandiflora Tablets based on partial least square method

目的基于近红外光谱(near infrared spectrum,NIRS)技术,建立一种快速预测大株红景天片(Rhodiola grandiflora Tablets,RGT)素片硬度的方法。方法采集共600个生产样本与自制样本的NIRS,通过比较不同光谱预处理方法与不同特征变量筛选条件下模型的优劣,建立偏最小二乘(partial least-square,PLS)算法模型,另采集120个样本的NIRS对模型进行外部验证,预测RGT素片硬度。结果建立的素片硬度PLS模型中,无预处理的光谱采用全波段建模的模型效果最佳,预测模型的校正集相关系数(correlation coefficient of training set,R_(cal))与验证集相关系数(correlation coefficient of verification set,R_(pre))分别为0.9719与0.9887,相关性良好,预测均方根误差(root mean square error of prediction,RMSEP)为2.03 N,性能偏差比(ratio of performance to deviation,RPD)为6.68,预测相对偏差(relative standard to deviation,RSEP)为4.24%,模型内部验证的平均相对预测误差为2.82%,外部验证的平均相对预测误差为4.59%,均<5%,对不合格素片的检出率高达97.33%。结论NIRS分析技术结合PLS算法建立的RGT素片硬度预测模型具有良好的模型性能与预测能力,为RGT素片硬度的无损检测提供了一种新方法。

Objective To establish a rapid method for predicting the hardness of Rhodiola grandiflora Tablets(RGT)based on near infrared spectroscopy(NIRS).Methods The NIRS of 600 production samples and self-made samples were collected.The partial least square(PLS)algorithm model was established by comparing the advantages and disadvantages of the models under different spectral pretreatment methods and different characteristic variable screening conditions.The NIRS of 120 samples were collected for external verification of the model to predict the hardness of RGT.Results Among the PLS models for plain tablet hardness,the full band model was the best for the spectrum without pretreatment.The correlation coefficient of training set(Rcal)and the correlation coefficient of verification set(R_(pre))of the prediction model were 0.9719 and 0.9887,respectively.The root mean square error of prediction(RMSEP)is 2.03 N,the ratio of performance to deviation(RPD)is 6.68,and the relative standard to deviation(RSEP)is 4.24%.The average relative prediction error of the internal validation of the model is 2.82%,and the average relative prediction error of the external validation is 4.59%,both of which are less than 5%.The detection rate of unqualified tablets is as high as 97.33%.Conclusion The hardness prediction model of RGT established by near-infrared spectroscopy combined with partial least squares algorithm has good model performance and prediction ability.This study provides a new method for nondestructive testing of the hardness of RGT.