快速、 准确检测酒醅酸度, 可显著提高白酒出酒率和成品酒品质。 近红外光谱(NIR)提供了分子的倍频和合频, 即有机物中含氢基团(C-H、 N-H、 O-H)的振动信息, 通常用于样品中含氢化合物的定性和定量分析。 采用NIR能简单、 迅速的测定酒醅酸度, 克服了传统化学分析方法检测周期长、 试剂消耗大、 人为误差等不足。 由于NIR是一种间接分析技术, 如何建立校正模型是准确检测酒醅酸度的关键。 作为深度学习中的典型模型, 卷积神经网络(CNN)具有局部区域连接, 分享权值等优点, 不仅能从复杂的光谱数据中提取关键特征, 还能减少网络模型的复杂度。 因此, 提出基于CNN和NIR的酒醅酸度定量分析方法, 以某酒企生产线中采集的545个酒醅样本光谱数据作为研究对象, 采用标准正态变换(SNV)、 Savitzky-Golay (SG）滤波和一阶求导（1stD）三种算法相结合对原始光谱进行预处理; 利用无信息变量消除法(UVE)选择光谱数据的特征波长; 使用CNN建立酒醅酸度模型。 结果表明: (1)对光谱数据进行预处理后, 消除了原始光谱中的基线偏移, 噪声等问题; 经过预处理后的光谱数据模型相较于原始光谱建模, 预测集决定系数提升了22.85%, 预测集均方根误差降低了0.049 5, 提高了酒醅酸度与光谱反射率的相关性。 (2)对光谱数据进行波长筛选后所建立的模型相较于全波段建模, 预测集决定系数提升了2.04%, 预测集均方根误差降低了0.004 8。 (3)基于CNN建立的酸度预测模型, 预测集决定系数为0.955 5, 预测集均方根误差为0.039 1。 相较于偏最小二乘回归模型, 预测集决定系数提升了1.03%, 预测集均方根误差降低了0.097 6; 相较于反向传播神经网络模型, 预测集决定系数提升了1.16%, 预测集均方根误差降低了0.099 4。 该方法可实现对酒醅酸度的快速、 准确测量, 为后续酒醅酸度在线检测提供方法支撑。

Rapid and accurate detection of the acidity of fermented grains can significantly improve the yield of Baijiu and the quality of finished liquor. Near infrared spectroscopy (NIR) mainly contains information on octave and ensemble frequencies of molecules, i. e., the vibrations of hydrogen-containing groups (C-H, N-H, O-H) in organic matter. It is usually used for qualitative and quantitative analysis of hydrogen-containing compounds in samples. The NIR can be used to determine the acidity of fermented grains in a simple, rapid overcoming the shortcomings of traditional chemical analysis methods, such as long detection cycles, large reagent consumption, and human errors. As NIR is an indirect analysis technology, establishing a calibration model is the key to accurately detecting the acidity of fermented grains. As a typical model in deep learning, convolutional neural networks (CNN) have the advantages of local area connection and weight sharing. It can not only extract critical features from complex spectral data, but also reduce the complexity of network models. Therefore, a quantitative analysis method for the acidity of fermented grains based on CNN and NIR is proposed in this work. The research object is the spectral data of 545 fermented grains samples collected in the production line of a wine enterprise, and the original spectra are preprocessed using a combination of three algorithms: standard normal variation (SNV), Savitzky-Golay(SG) filtering and first derivative (1stD); uninformative variable elimination (UVE) is used to select the characteristic wavelength of spectral data; CNN is used to establish the acidity model of fermented grains. The results show that: (1) The pre-processed spectral data eliminated the baseline shift and noise problems in the original spectra, increased the prediction set coefficient of determination by 22.85%, and decreased the root mean square error by 0.049 5 compared with the original spectral modeling, which improved the correlation between the acidity of fermented grains and spectral reflectance. (2) The model established after wavelength screening of spectral data increased the determination coefficient of the prediction set by 2.04% and decreased the root mean square error of the prediction set by 0.004 8 compared with full-wavelength modeling. (3) The acidity prediction model based on CNN had a determination coefficient of 0.955 5 and a root mean square error of 0.039 1. Compared with the partial least squares model, the determination coefficient of the prediction set is increased by 1.03%, and the root mean square error of the prediction set is reduced by 0.097 6. Compared with the backpropagation neural network model, the determination coefficient of the prediction set is increased by 1.16%, and the root mean square error of the prediction set is reduced by 0.099 4. The research results can realize the rapid and accurate measurement of the acidity content of fermented grains and provide method support for subsequent online detection of the acidity of fermented grains.