将多种单分类器模型融合, 并用融合后的模型对不同品种干红葡萄酒进行判别分析。 用BRUKER MPA傅里叶变换型近红外光谱仪采集170个干红葡萄酒样品的近红外透射光谱, 选取PLS-DA, SVM, Fisher和AdaBoost作为单分类器建模方法, 分别建立葡萄酒品种判别模型, 通过差异性度量值对单分类器进行筛选, 得到差异性较大的四个单分类器作为基分类器, 其中基分类器对测试集葡萄酒品种判别准确率最高为88.24%, 最低为81.18%。 然后通过加权投票机制对基分类器进行融合, 融合后的模型对测试集葡萄酒品种判别准确率提高至92.94%, 误判样品个数由单分类器最少的9个降为6个。 实验结果表明多分类器融合所建立的模型优于传统近红外光谱定性分析一般采用单分类器模型结果, 提高了葡萄酒品种判别的准确性, 采用基于近红外光谱的多分类融合方法对葡萄酒种类判定具有可行性。

The conventional qualitative analysis of near infrared spectroscopy (NIR) commonly uses one single classification model. This paper focused on the fusion of multiple classifiers based on different single classifiers by using the fused classifier to determine different varieties of red-wines. NIR spectra of 170 red-wine samples were collected by using Fourier transform near-infrared spectrometer. Red-wine classification models were established respectively, based on PLS-DA, SVM, Fisher and AdaBoost. Then these models were selected to obtain some different base classifiers according to Diversity Measure Feature Selective (DMFS). The highest accuracy rate of determining different varieties of red-wine test samples of four single base classifiers was up to 88.24%, and at the same time the lowest discriminant accuracy rate was 81.18%. At last, we got the fused classifier, which combined four base classifiers with weighted voting principle, and determined its test set again by using the fused classifier. The final classification accuracy rate for red-wine varieties increased to 92.94%, In contrast with one single classifier, the lowest misjudged number of fused classifiers decreased from 9 to 6.These results suggested that the performance of fused classifier is superior to one single classifier. It is feasible to use fused classifier combined with near infrared spectroscopy to determine different varieties of red-wines.