针对食用油中单不饱和脂肪酸(MUFA)和多不饱和脂肪酸(PUFA)含量的快速检测问题,研究探索应用拉曼(Raman)和近红外(NIR)光谱以及特征层数据的融合,结合化学计量学分析,建立食用油MUFA和PUFA含量预测模型.重点研究各种预处理算法对模型预测能力的影响.应用竞争性自适应重加权采样(CARS)提取Raman和NIR光谱的特征波长,应用网格搜索(GS)算法选取支持向量机回归(SVR)模型的参数组合(C,g)值,分别建立基于拉曼和近红外光谱的特征波段的SVR预测模型;建立基于特征层的多源光谱融合的SVR预测模型.试验表明,基于特征层融合建立的Raman-NIR-SVR模型能够实现食用油MUFA和PUFA含量的快速预测,且预测效果更优.其中预测MUFA含量的SG15-ALS-Nor-CARS-MSC-CARS-SVR模型的预测集决定系数R2为0.977 3,与单光谱中最优含量预测模型相比增加了2.43％;而预测PUFA含量的MA11-airPLS-Nor-CARS-MSC-CARS-SVR模型的预测集R2为0.993 0,比较最优单光谱数据建立的SVR模型增加了2.57％.结果表明,采用特征层融合方法建立的含量预测模型的综合性能优于基于单光谱数据建立的模型.%It was explored in this paper that characteristic fusion of Raman-NIR was combined with Chemometrics to establish monounsaturated fatty acid (MUFA) and polyunsaturated fatty acids (PUFA) content prediction model to solve the problem of their content rapid prediction.It focused on what effects various pretreatment algorithms had on prediction model.Competitive adaptive reweighted sampling (CARS)was used to extract the characteristic wavelength of Raman and NIR spectra,grid search (GS)algorithm to select the parameter combination (C,g).The author established SVR model based on characteristic wavelength of Raman and NIR spectra respectively,and based on characteristic fusion did SVR model.The experiment showed that based on feature fusion of Raman-NIR the SVR model was better to realize fast prediction.For MUFA content,prediction set determination coefficient R2 of SG15-ALS-Nor-CARS-MSC-CARS-SVR model was 0.977 3,compared with the optimal prediction model of the single spectra,which increased by 2.43 ％;and for MUFA,MA11-airPLS-Nor-CARS-MSC-CARS-SVR model R2 was 0.995 8,comparing the optimal single spectra SVR model increased by 2.57％.Result showed that comprehensive performance of the prediction model established by feature fusion method was better than the model based on single spectral data.
展开▼
