牛奶含水率介电谱结合化学计量学检测方法

Detecting Moisture Content of Cow's Milk Using Dielectric Spectra and Chemometrics

为了实现牛奶含水率的快速检测，采用网络分析仪和同轴探头测量了室温（（25±0．5）℃）下20—4500MHz间105个牛奶样品的相对介电常数和介质损耗因子。发现基于单一频率下的介电参数很难预测牛奶的含水率。为此，将介电谱与化学计量学方法相结合预测牛奶的含水率。基于X—Y共生距离法进行了样本集划分，得到校正集样本75个和预测集样本30个。采用连续投影算法从全介电谱中提取出了15个用于预测牛奶含水率的特征变量；建立了基于全介电谱和连续投影算法提取的特征变量预测牛奶含水率（87．28％～91．30％）的广义神经网络、支持向量机和极限学习机模型。结果发现，基于连续投影算法提取的特征变量所建立的极限学习机模型是预测牛奶含水率的最优模型，其预测相关系数、预测均方根误差和剩余预测偏差分别为0．988、0．119％和6．723。研究表明，介电谱结合化学计量学方法可用于检测牛奶的含水率。

To explore a rapid method for detecting moisture content of cow＇ s milk, a network analyzer and an open-ended coaxial-line probe were applied to measure the dielectric properties （relative dielectric constant and dielectric loss factor） of 105 milk samples over the frequency range of 20 - 4 500 MHz at room temperature （25 ± 0.5） ℃. The low linear correlation coefficient between the moisture content and the permittivities at a single frequency of used milk samples showed that it was difficult to predict the moisture content of milk using a single permittivity value. Therefore, the dielectric spectra combined with chemometrics were used to determine the moisture content of milk. All samples were partitioned into calibration set （75 samples） and prediction set （30 samples） by using set partitioning method based on joint X - Y distances. Fifteen characteristic variables that predicting moisture content of cow＇ s milk were selected by successive projection algorithm from full spectra. The generalized regression neural network, support vector machine and extreme learning machine models were established to predict moisture content of milk （87.28% -91.30% ）, based on the original full dielectric spectra and characteristic variables. The results showed that the extreme learning machine model established using the characteristic variables selected by successive projection algorithm was the best model in determining moisture content of milk, with the correlation coefficient of prediction, root-mean-square error of prediction, and residual prediction deviation of 0. 988, 0. 119% , and 6. 723, respectively. The study indicates that the dielectric spectra combined with chemometrics could be used to detect moisture content of milk. The research is helpful to develop a new milk moisture detector which could be used in situ or online detection.