反向传播神经网络结合紫外-近红外融合光谱对“互助”青稞酒的判别研究  

Discriminative Study on Huzhu Qingke Liquor by Back Propagation Neural Network Combined With Ultraviolet-Near Infrared Fusion Spectroscopy

作　　者：赵玉霞 张明锦 王茹 张世芝 殷博 ZHAO Yu-xia;ZHANG Ming-jin;WANG Ru;ZHANG Shi-zhi;YIN Bo(College of Chemistry and Chemical Engineering,Qinghai Normal University,Xining 810016,China;College of Chemistry and Chemical Engineering,Qinghai Minzu University,Xining 810016,China;Qinghai Key Laboratory of Advanced Technology and Application of Environmentally Functional Materials,Xining 810016,China)

机构地区：[1]青海师范大学化学化工学院,青海西宁810016 [2]青海民族大学化学化工学院,青海西宁810016 [3]青海省环境功能材料先进技术与应用重点实验室,青海西宁810016

出　　处：《光谱学与光谱分析》2025年第5期1290-1299,共10页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(22363010);青海省自然科学基金项目(2022-ZJ-769)资助。

摘　　要：“互助”青稞酒作为保护地理标志产品,对其准确评价分类具有重要意义。紫外光谱(UV)和近红外光谱(NIR)技术具备快速、准确、无损检测、无需样品预处理等优势,在食品等领域已广泛应用。本研究采用UV、NIR及紫外-近红外中级数据融合光谱(UV-NIR)结合反向传播神经网络(BPNN)法建立了快速、无损、高效的“互助”青稞酒判别分类模型。由于光谱特征峰叠加干扰,未经优化的光谱受到噪声和基线漂移等影响,采用标准正态变量变换(SNV)、Savitzky-Golay平滑(SG)、一阶导数(1D)和二阶导数(2D)4种预处理方法对光谱进行去噪处理。相对单一光谱,融合光谱能够互补多元化学信息,提高分类模型性能,通过竞争自适应重加权采样(CARS)、连续投影算法(SPA)、主成分分析(PCA)、变量投影重要性分析(VIP)和变量组合集群分析(VCPA)5种变量筛选方法选择特征变量,达到优化模型性能及融合两种光谱有效信息。选择最佳方法建立单一光谱和融合光谱的BPNN模型。结果表明,UV光谱经SNV预处理以SPA选择30个特征变量建立的分类模型识别效果最好,分类准确率为100%,MSE值、R_(P)^(2)、R(Train)、R(Validation)、R(Test)和R(All)分别为0.0180、1、0.9283、0.9587、0.9130、0.9297;NIR和UV-NIR经SG预处理后以PCA分别选择84和106个特征变量建立的分类模型识别效果最好,NIR光谱分类准确率为100%,MSE值、R_(P)^(2)、R(Train)、R(Validation)、R(Test)和R(All)分别为0、1.000、1.000、1.000、1.000、1.000;UV-NIR光谱分类准确率为100%、MSE值、R_(P)^(2)、R(Train)、R(Validation)、R(Test)和R(All)分别为0.0057、1.000、1.000、0.9871、0.9913、0.9964;与单一光谱建模相比,融合光谱可明显提高分类模型的预测能力和稳健性,实现“互助”青稞酒的快速、无损分析。Chinese Huzhu Qingke Liquor is a protected geographical indication product,and it is of great significance for its accurate evaluation and classification.Due to the advantages of ultraviolet(UV)and near-infrared(NIR)spectroscopy,such as fast,accurate,non-destructive detection and no sample pretreatment,are widely used in food and other fields.In this study,a fast,nondestructive,and efficient discriminative classification model for Huzhu Qingke Liquor was established based on UV,NIR,and UV-NIR intermediate data fusion spectroscopy(UV-NIR)combined with theback-propagation neural network(BPNN)method.Since the unoptimized spectra are affected by noise and baseline drift due to the superimposed interference of spectral eigenpeaks,the spectra are denoised using four preprocessing methods,namely,standard normal variable transform(SNV),Savitzky-Golay smoothing(SG),first-order derivative(1D)and second-order derivative(2D).Further,relative to a single spectrum,the fused spectrum can complement the diversified spectroscopic information and improve the performance of the classification model,so the feature variables are selected by five variable screening methods,namely,competitive adaptive reweighted sampling(CARS),successive projection algorithm(SPA),principal component analysis(PCA),variable projection importance analysis(VIP),and variable combinatorial clustering analysis(VCPA)to achieve the optimization of model performance and the purpose of fusing the effective information of two spectra.Finally,the best method for establishing the BPNN model for single and fused spectra was selected.The results show that the classification model established by selecting 30 feature variables by SPA after SNV preprocessing for UV spectra has the best recognition effect,with a classification accuracy of 100%.The MSE value,R_(P)^(2),R(Train),R(Validation),R(Test)and R(All)were 0.0180,1,0.9283,0.9587,0.9130,and 0.9297,respectively;PCA selected the NIR and UV-NIR after SG preprocessing with 84 and 106 The classification model built by feat

关 键 词：“互助”青稞酒 紫外光谱 近红外光谱 光谱融合 变量筛选 反向传播神经网络(BPNN)模型 

分 类 号：O657.3[理学—分析化学]