CAJ | 학술논문

采集了来自全国20种单植物源和其它多植物源的101份的蜂蜜样品,分别运用傅立叶型近红外光谱仪采用光纤透反射(800～2500 nm,2 mm光程)和透射(800～1370 nm,20 mm光程)采集方式获得近红外光谱,来预测蜂蜜中结构和含量都很相近的果糖和葡萄糖含量.结果发现,两种测量方式下果糖、葡萄糖的预测准确度存在着较大的差异.为了分析这种差异产生的原因,采用支持向量机分析其非线性信息,采用遗传算法分析其特征波长,结果表明:这种差异主要来自两种糖分特征波长分布不同所导致.通过对两种糖分的检测方案进行优化,得出在利用近红外光谱技术检测蜂蜜中葡萄糖成分含量时应尽量采集短波区、长光程的光谱,或者对全谱区、短光程的光谱,进行特征波长的提取,避开水分的干扰,从而提高其预测精度;而对于果糖,则应尽量采集全谱区、短光程的光谱;采用常用线性定量建模方法PLSR就可以得到很好的预测模型,非线性的支持向量机模型未能明显提升模型性能.
채집료래자전국20충단식물원화기타다식물원적101빈적봉밀양품,분별운용부립협형근홍외광보의채용광섬투반사(800～2500 nm,2 mm광정)화투사(800～1370 nm,20 mm광정)채집방식획득근홍외광보,래예측봉밀중결구화함량도흔상근적과당화포도당함량.결과발현,량충측량방식하과당、포도당적예측준학도존재착교대적차이.위료분석저충차이산생적원인,채용지지향량궤분석기비선성신식,채용유전산법분석기특정파장,결과표명:저충차이주요래자량충당분특정파장분포불동소도치.통과대량충당분적검측방안진행우화,득출재이용근홍외광보기술검측봉밀중포도당성분함량시응진량채집단파구、장광정적광보,혹자대전보구、단광정적광보,진행특정파장적제취,피개수분적간우,종이제고기예측정도;이대우과당,칙응진량채집전보구、단광정적광보;채용상용선성정량건모방법PLSR취가이득도흔호적예측모형,비선성적지지향량궤모형미능명현제승모형성능.
A total of 101 honey samples that originated from 20 different unifloral honey and other multifloral honey samples were collected from China.FT-NIR spectrometer were applied to determinate the content of fructose and glucose of honey with two different modes: transflectance (800-2500 nm, 2 mm optical path length) and transmittance (800-1370 nm, 20 mm optical path length).It was found that the prediction accuracy of fructose and glucose had significant difference with the two modes.In order to analyze the reason of this difference, support vector machine (SVM) was used to analyze the non-linear information, and genetic algorithm (GA) was used to analyze the characteristic wavelengths.The result indicated that the detection difference of fructose and glucose was originated from their different characteristic wavelengths.Through the optimization of detection method, it was found that for the determination of glucose, short wavelength and long optical path length should be used, on the other side, the whole wavelength region and short wavelength, with selecting the characteristic wavelength to avoid the disturb of water can also be used.For the determination of fructose, whole wavelength region and short optical path length should be used.Linear regression methods such as PLSR could obtain good results, and non-linear methods such as SVM did not improve the model performance.