基于便携式近红外技术的生鲜乳品质现场评价

生鲜乳作为乳制品生产的基本原料，其质量是保证乳制品食用安全、维护人类健康的基础。可见/近红外光谱技术结合化学计量学方法，构建生鲜乳品质指标的数学模型，实现生鲜乳品质的现场评价。在不同年份，收集88份来自不同奶牛个体的生鲜乳样品。便携式光谱仪采集生鲜乳漫透射光谱(500~1 010 nm)，二阶导数和卷积平滑进行光谱预处理，以消除脂肪球引起的光散射和高频噪声。变窗宽移动窗口偏最小二乘法(CSMWPLS)和遗传偏最小二乘法(GAPLS)用于筛选信息区间，并构建预测模型。CSMWPLS与GAPLS模型的预测性能相当，脂肪、蛋白质、干物质和乳糖的预测标准误差(RMSEP)分别为0.115 6/0.103 3，0.096 2/0.113 7，0.201 3/0.123 7和0.077 4/0.066 8，相对预测误差(RPD)分别为8.99/10.06，3.53/2.99，5.76/9.38和1.81/2.10。同时构建了生鲜乳品质指标的多元线性回归(MLR)方程，采用的最优变量数分别为8，10，9和7。采用外部数据集检验，MLR预测性能与PLS相近甚至更优，脂肪、蛋白质、干物质和乳糖模型的RMSEP分别为0.107 0，0.093 0，0.136 0和0.065 8;相对预测误差(RPD)分别为9.72，3.66，8.53和2.13，可用于现场准确测量。结果显示，便携式近红外光谱仪结合MLR模型可实现生鲜乳品质的现场快速评价，为生鲜乳按质论价收购提供了一种新方法，同时为便携式乳品近红外专用仪器设计提供技术参考。

On-Site Evaluation of Raw Milk Qualities by Portable Vis/NIR Transmittance Technique

To ensure the material safety of dairy products, visible (Vis)/near infrared (NIR) spectroscopy combined with chemometrics methods was used to develop models for fat, protein, dry matter (DM) and lactose on-site evaluation. A total of 88 raw milk samples were collected from individual livestocks in different years. The spectral of raw milk were measured by a portable Vis/NIR spectrometer with diffused transmittance accessory. To remove the scatter effect and baseline drift, the diffused transmittance spectra were preprocessed by 2nd order derivative with Savitsky-Golay (polynomial order 2, data point 25). Changeable size moving window partial least squares (CSMWPLS) and genetic algorithms partial least squares (GAPLS) methods were suggested to select informative regions for PLS calibration. The PLS and multiple linear regression (MLR) methods were used to develop models for predicting quality index of raw milk. The prediction performance of CSMWPLS models were similar to GAPLS models for fat, protein, DM and lactose evaluation, the root mean standard errors of prediction (RMSEP) were 0.115 6/0.103 3, 0.096 2/0.113 7, 0.201 3/0.123 7 and 0.077 4/0.066 8, and the relative standard deviations of prediction (RPD) were 8.99/10.06, 3.53/2.99, 5.76/9.38 and 1.81/2.10, respectively. Meanwhile, the MLR models were also calibrated with 8, 10, 9 and 7 variables for fat, protein, DM and lactose, respectively. The prediction performance of MLR models was better than or close to PLS models. The MLR models to predict fat, protein, DM and lactose yielded the RMSEP of 0.107 0, 0.093 0, 0.136 0 and 0.065 8, and the RPD of 9.72, 3.66, 8.53 and 2.13, respectively. The results demonstrated the usefulness of Vis/NIR spectra combined with multivariate calibration methods as an objective and rapid method for the quality evaluation of complicated raw milks. And the results obtained also highlight the potential of portable Vis/NIR instruments for on-site assessing quality indexes of raw milk.