赤霞珠酿酒葡萄总酚含量的近红外光谱定量分析

酿酒葡萄中的总酚含量是影响葡萄品质的重要指标，也是影响葡萄酒质量的关键因素。为了快速准确地检测赤霞珠葡萄的总酚含量，利用近红外光谱技术结合GA-ELM预测模型对赤霞珠葡萄总酚含量进行预测研究。试验采用5个收获期（每期采集40串，每串取10个）的赤霞珠葡萄，采集200组葡萄的12 500~4 000 cm-1波段范围内的近红外光谱。基于福林酚比色法原理对赤霞珠葡萄的总酚含量进行测定，使用SPXY算法将样品按照3∶1比例分为校正集和预测集，共计150个校正集和50个预测集。分别采用多元散射（MSC）、标准正态变换（SNV）、数据中心化（MC）、移动窗口平滑（MA）和一阶导数+SG方法对原始光谱进行预处理，优选出最佳的预处理方法为MSC。并进一步采用竞争性自适应重加权算法（CARS）、遗传算法（GA）、联合区间偏最小二乘算法（si-PLS）和连续投影算法（SPA）分别对光谱波段进行提取，经对比分析发现CARS提取的69个特征波长数据能有效提高模型的稳定性和预测结果。在MSC预处理和特征波长提取的基础上，引入极限学习机（ELM）算法，建立赤霞珠葡萄总酚含量的预测模型，在总酚含量预测过程中，采用遗传算法（GA）对ELM模型进行优化，并探究了不同的激活函数和隐含层神经元个数对GA-ELM模型预测能力的影响，确定最优的激活函数为Sigmoidal，最优的神经元个数为50个。最后，将ELM和GA-ELM模型的预测能力进行对比，结果显示GA-ELM模型的预测能力高于ELM模型的预测能力，其中MSC+CARS+GA-ELM模型预测能力最好，校正相关系数（R_c）为0.901 7，预测相关系数（R_p）为0.901 3，校正均方根误差（RMSEC）为2.112 4，预测均方根误差（RMSEP）为1.686 8，剩余预测偏差（RPD）为2.308 0。研究结果表明：利用近红外光谱技术结合变量优选建立的GA-ELM模型可实现对赤霞珠葡萄的总酚含量的预测，为赤霞珠葡萄品质的检测奠定了理论基础。

Quantitative Analysis of Total Phenol Content in Cabernet Sauvignon Grape Based on Near-Infrared Spectroscopy

The contents of total phenol in wine grape are an important indicator of grape quality and also a key factor of wine quality directly. To detect the total phenol contents of the cabernet sauvignon grape quickly and accurately, this paper used near-infrared spectroscopy and GA-ELM prediction model to predict the total phenol content of Cabernet Sauvignon grapes. In the experiment, Cabernet Sauvignon grapes were collected in 5 harvest periods (40 bunches were collected in each harvest period, and 10 grapes were acquired in each cluster), and near-infrared spectra information in the range of 12 500~4 000 cm-1 was collected for 200 groups of grapes. The total phenol content of Cabernet Sauvignon grapes was determined based on the principle of Folin-Ciocalteus colorimetry, SPXY algorithm was used to divide the samples into correction sets and prediction sets at a ratio of 3∶1, with a total of 150 correction sets and 50 prediction sets. Multiplicative Scatter Correction (MSC),Standard Normalized Variate (SNV), Mean Centering (MC), Moving Average (MA), and the First Derivative +SG was used to preprocess the raw spectra, MSC was compared as the best pretreatment method.And then, competitive adaptive reweighted sampling (CARS), genetic algorithm (GA), successive projections algorithm (SPA) and synergy interval partial least squares (si-PLS) were extracted the characteristic wavelengths, respectively. The comparative analysis found that the 69 characteristic wavelength variables extracted by CARS could effectively improve the model’s stability and prediction ability. Based on the MSC and different variable optimization methods, the extreme learning machine (ELM) algorithm was introduced to establish the total phenol content prediction model. In predicting total phenol content, a genetic algorithm (GA) was used to optimize the ELM model and the influence of different kernel functions and the number of hidden layer neurons on the prediction ability of the GA-ELM model investigated. The optimal kernel function was Sigmoidal, and the optimal number of neurons was 50. Finally, the prediction capabilities of the ELM and GA-ELM models were compared. The results showed that GA-ELM models were more accurate in predicting than the ELM models, and the MSC+CARS+GA-ELM model was the best with a correlation coefficient of calibration (R_c) of 0.901 7, the correlation coefficient of prediction of 0.901 3, the root mean square error of calibration (RMSEC) of 2.112 4, the root mean square error of prediction (RMSEP) of 1.686 8 and residual prediction deviation (RPD) of 2.308 0. The combination of variable optimization methods and the GA-ELM model was an effective method, which provided a theoretical basis for detecting Cabernet Sauvignon grapes’ quality.