果酒发酵中的多酚是引起果酒口感、 颜色变化的重要因素。 为保证果酒品质, 有必要开发一种快速监测发酵过程中多酚含量变化的技术。 收集不同批次成熟期的蓝莓、 桑葚为原料, 分别碾压成汁, 同时按比例混合二者, 于小型发酵罐进行发酵。 通过离线收集不同发酵时段的发酵液于离心管, 高速离心后取上清液置于棕色瓶保存, 共计得到48个果酒发酵样本。 将上清液置于三个平行样比色皿, 以傅里叶快速变换近红外光谱仪（FT-NIR）采集其透射光谱, 取平均值作为该样本的光谱信号。 然后将棕色瓶内的发酵液以国标法（即以标准液的吸光度值制定标准曲线）测定各样品的总酚含量, 以duplex法计算样本光谱之间的距离且按2∶1的比例划分为训练集和预测集。 采用间隔偏最小二乘法（iPLS）将训练集样本的透射光谱与总酚含量之间构建定量模型, 间隔数从2依次变化到60个。 该研究创新之处是使用共识方法融合多个已构建好的iPLS成员模型, 按一定的共识规则分配权系数。 通过各成员模型交互验证的残差及其残差之间的相关性来优化各成员模型的线性组合, 以拉格朗日乘数法求解各成员模型的权系数, 使间隔偏最小二乘-共识模型（consensual iPLS, C_iPLS）的交互验证均方根误差最小。 相比于全局PLS模型、 划分不同间隔数量时的iPLS模型, C_iPLS均具有较小的预测误差。 当划分39个间隔时由三个iPLS成员模型（即14th, 16th, 18th）组成的共识模型误差最小为124.2, 交互验证相关系数为0.944, 对预测集样本的预测均方根误差为163.4, 预测相关系数为0.931, 预测性能均优于PLS和iPLS模型。 另外, 作为对比选用连续投影算法与无信息变量剔除法来优化光谱模型, 其预测性能均不及本文提出的共识模型。 分析各iPLS模型预测残差之间的相关性, 发现共识模型主要是融合那些具有较高预测性能且模型间较低相关性的成员模型。 结果表明, 光谱分析结合共识方法可提高回归模型的预测精度、 减少建模所需变量数, 能够用于果酒总酚含量的离线快速检测。

Polyphenol is one of important factors that cause the changes of taste and color in fruit-wine. To ensure the quality of fruit-wine, it is necessary to develop a fast measurement that monitors the change of polyphenol content during the fermentation. The ripe blueberry and mulberry were collected from different harvest batches. They were crushed respectively into juice, and their mixed juice was also mixed in certain ratio for fermentation in the small fermentation tanks. Those fermenting liquors from the different fermenting periods were collected through the off-line sampling access. The supernate was obtained by centrifugation pretreatment and totally 48 fermenting samples were preserved in the brown bottles for later use. The supernate were injected into three paralleled cuvettes, whose transmission spectrums were scanned by FT-NIR spectrometer, and their repeated readings were averaged for the spectral signals. Then, the total phenol content was measured by the national standard method (i.e. the standard curve was established between the absorbance value and the standard solution), and all samples were divided into the calibration and prediction set in a ratio of 2∶1 by duplex algorithm, which was used to calculate the spectral distance from the divided sample to the center of the rest samples. Interval partial least square (iPLS) was used to construct series of quantitative models between the transmission spectra and the total phenol contents in the training set, and the number of intervals was successively changed from 2 to 60. The innovate point in this work was that the consensual rule was used to integrate the calibrated member models (here referring to the iPLS model) into a consensus model and distribute the weighting coefficients. The linear combinations of member models were optimized to minimize the mean squared error (MSE) in the consensus model through the residual errors from the cross validation and their correlations. The weighted coefficient of each member model was solved by Lagrange multiplier method, so as to minimize the root mean square error of the consensus model. Compared with the global model of partial least squares (PLS), interval partial least-squares (iPLS) model with different number of spectral intervals, the consensual iPLS (C_iPLS) model commonly obtained a better performance. When the full spectra were divided into 39 intervals, the C_iPLS model, composed of three iPLS members models (those were 14th, 16th, 18th iPLS model respectively), got the minimum root mean squared error of cross validation (RMSECV) of 124.2, as well as the correlation coefficient of cross validation (Rcv) of 0.944, and the samples in prediction set were tested well with root mean square of prediction (RMSEP) of 163.4, as well as the correlation coefficient (Rp) of prediction of 0.931. In addition, the successive projection algorithm and the uninformative variable elimination were used to optimize the spectral model, but the predictive performances were not better than the proposed consensus model. By analyzing the correlation between the predicted residuals of each iPLS model, it was found that the consensus model commonly screened these member models featured with high prediction performance and low correlation between member models. Results showed that the spectral analysis technology combined with the consensus method could improve the prediction accuracy of the regression model, reduce the modeling number of variables, and could be employed off-line for the rapid detection of total phenol content in fruit wine.