基于异常样品剔除的酒醅近红外定量分析模型的精度提升

Accuracy improvement of near infrared quantitative analysis model for fermented grains based on abnormal sample removal

中文关键词:  近红外光谱  异常点检测  酒醅  马氏距离  主成分马氏距离  蒙特卡罗交叉验证法

英文关键词:near infrared spectroscopy  outlier detection  fermented grains  Mahalanobis distance  principal component analysis Mahalanobis distance  Monte Carlo cross validation method

中文摘要:

      目的 利用异常识别算法识别出原数据集中存在的奇异点, 以建立预测精度更高的酒醅定量分析模型。方法 采用集群分析思维, 利用马氏距离、主成分马氏距离、蒙特卡罗交叉验证法对108个样本进行异常样品识别及剔除, 以光谱-理化值共生距离算法进行样品集的划分, 划分比例为3:1。结果 酒醅水分近红外定量分析模型经马氏距离处理后预测精度达到最高, 预测相关系数上升了0.43%, 预测均方根误差下降了6.94%; 酒醅酸度近红外定量分析模型经马氏距离处理后预测精度达到最高, 预测相关系数上升了0.02%, 预测均方根误差下降了0.20%; 酒醅还原糖近红外定量分析模型经蒙特卡罗处理后, 预测相关系数上升了8.74%, 预测均方根误差下降了42.14%; 酒醅淀粉近红外定量分析模型经蒙特卡罗处理后预测精度达到最高, 预测相关系数上升了2.81%, 预测均方根误差下降了57.80%。结论 经过验证, 剔除异常样品可建立出预测精度更高的酒醅定量分析模型。

英文摘要:

      Objective To establish a quantitative analysis model of fermented grains with higher prediction accuracy, the singularity existing in the original data set is identified by anomaly recognition algorithm. Methods The cluster analysis thinking was adopted, the Mahalanobis distance, principal component Mahalanobis distance and Monte Carlo cross validation method were applied to identify and eliminate the abnormal samples of 108 samples, and the sample set were divided with sample set partitioning based on joint X-Y distance sampling, with the division ratio of 3:1. Results The prediction accuracy of the near infrared quantitative analysis model of fermented grains moisture reached the highest after Mahalanobis distance treatment, with the increase of the predicted correlation coefficient 0.43%, and the decrease of predicted root mean square error 6.94%; after Mahalanobis distance treatment, the prediction accuracy of the near infrared quantitative analysis model of fermented grains acidity reached the highest, with the increase of the predicted correlation coefficient 0.02%, and the decrease of the predicted root mean square error 0.20%; after Monte Carlo treatment, the predicted correlation coefficient of the near infrared quantitative analysis model of fermented grains reducing sugar increased 8.74% and the predicted root mean square error decreased 42.14%; the accuracy of the near infrared quantitative analysis model of fermented grains starch reached the highest after Monte Carlo treatment, with the increase of the predicted correlation coefficient 2.81%, and the decrease of the predicted root mean square error 57.80%. Conclusion After verification, eliminating abnormal samples can establish a quantitative analysis model of fermented grains with higher prediction accuracy.