基于特征波段的黄酒近红外光谱检测模型递归更新方法

摘要

NIR (near-infrared)spectroscopy is a fast,non-destructive quantitative analysis tool.In order to improve the detection of yellow rice wine,NIR is employed for the quantitative analysis.In the detection,due to the varying factors (e.g. environment,raw material,instrument aging),the performance of model developed by the old samples may deteriorate over time.To guarantee the prediction accuracy,the recursive partial least square (RPLS)method is introduced to update the predic-tion model.However,the whole spectrum used to be involved in the model update,and the number of spectral variables in a whole spectrum is very large,which may result in intensive computation and no obvious improvement in prediction accuracy due to interference information included.Considering the insignificant change of characteristic wavelengths in yellow rice wine pro-duction,a model updating method is proposed in this paper based on characteristic wavelength.The correlation coefficient meth-od is employed to extract the characteristic wavelength,and then the RPLS model is developed by incorporating the new sample information in the method.This method is applied to update the NIR detection model of total acid in yellow rice wine.The corre-lation coefficient r,root mean square errors of prediction (RMSEP)and residual predictive deviation (RPD)are employed to e-valuate model performance.These three indices reach 0.9657,0.1843 and 3.7362 by using the proposed method.Therefore, the proposed method may optimize the model stability,improve the computational efficiency and provide an useful practical refer-ence.%近红外光谱是一种快速、无损的定量分析工具.为了提高黄酒关键参数的检测水平,采用近红外光谱法进行定量分析.检测过程中,由于受环境波动、仪器老化、原料变化等因素的影响,基于旧样品所建的模型的精确度逐渐下降.为保持模型的预测精度,引入递归偏最小二乘(recursive partial least square,RPLS)对模型进行更新.以往此方法多使用全谱信息扩充建模集并进行递归计算,光谱的变量多,且包含环境影响等干扰信息,更新计算量大,且精度的提升效果不明显.考虑到黄酒生产过程中特征波段变化小的特性,提出了一种基于特征波段的黄酒近红外光谱检测模型递归更新方法.先采用相关系数法提取特征波段建立低维模型,在采集到新样品理化值后,再利用其特征波段光谱信息,使用递归偏最小二乘对低维模型进行更新.此方法被应用于黄酒总酸的近红外检测模型更新.模型评价使用相关系数r,预测标准偏差RM-SEP和预测相对分析误差RPD三个指标.结果表明:采用本方法后,模型稳定性显著优化,计算效率有所提升,模型预测效果良好,三个评价指标分别达到0.9657,0.1843和3.7362,较全谱PRLS时分别提高3％,24％和31％,在实际应用中有一定的参考价值.