基于支持向量机算法的红外光谱技术在奶粉蛋白质含量快速检测中的应用

Application of Infrared Spectroscopy Technique to Protein Content Fast Measurement in Milk Powder Based on Support Vector Machines

蛋白质是奶粉中重要的营养成分,实现对奶粉中蛋白质含量的快速、无损检测十分重要。文章采用近红外及中红外光谱技术检测了不同品种奶粉的蛋白质含量。采用最小二乘支持向量机对光谱透射率值和蛋白质值建模。模型在全波段对样本蛋白质含量预测得到了较好的结果,绝对系数(R2p)达到0.9517,预测误差均方根(RMSEP)为0.520 201。预测结果要优于传统的偏最小二乘回归(PLS)的预测结果,说明红外光谱技术能够实现奶粉蛋白质含量的无损检测,且检测过程比传统化学检测方法简单,操作性强。文章同时还研究了分别基于中红外光谱范围和近红外光谱范围的建模。模型预测结果显示中红外光谱区域的建模效果要优于近红外光谱区域。该研究为今后奶粉蛋白质含量快速无损检测提供了新的方法。

In the present study, the JASCO Model FTIR-4 000 fourier transform infrared spectrometer （Japan） was used, with a valid range of 7 800-350 cm^-1. Seven brands of milk powder were bought in a local supermarket. Milk powder was compressed into a uniform tablet with a diameter of 5 mm and a thickness of 2 ram, and then scanned by the spectrometer. Each sample was scanned 40 times and the data were averaged. About 60 samples were measured for each brand, and data for 409 samples were obtained. NIRS analysis was based on the range of 4 000 to 6 666 cm^-1 , while MIRS analysis was between 400 and 4 000 cm^-1. The protein content was determined by kjeldahl method and the factor 6.38 was used to convert the nitrogen values to protein. The protein content value is the weight of protein per 100 g of milk powder. The NIR data of the milk powder exhibited slight differences. Univariate analysis was not really appropriate for analyzing the data sets. From NIRS region, it could be observed that the trend of different curves is similar. The one around 4 312 cm^-1 embodies the vibration of protein. From MIRS region, it could be determined that there are many differences between transmission value curves. Two troughs around 1 545 and 1 656 cm^-1 stand for the vibration of amide Ⅰ and Ⅱ bands of protein. The smoothing way of Savitzky-Golay with 3 segments and zero polynomials and multiplicative scatter correction （MSC） were applied for denoising. First 8 important principle components （PCs）, which were obtained from principle component analysis （PEA）, were the optimal input feature subset. Least-squares support vector machines was applied to build the protein prediction model based on infrared spectral transmission value. The prediction result was better than that of traditional PLS regression model as the determination coefficient for prediction （Rp^2） is 0. 951 7 and root mean square error for prediction （RMSEP） is 0. 520 201. These indicate that LS-SVM is a powerful tool for spectral analysis. Moreover, the study compared the prediction results based on near infrared spectral data and mid-infrared spectral data. The results showed that the performance of the model with mid-infrared spectral data was better than the one with near infrared spectra data. It was concluded that infrared spectroscopy technique can do the quantification of protein content in milk powder fast and non-destructively and the process was simple and easy to operate. The results of this study can be used for the design of a simple and non-destructive spectra sensor for the quantitative of protein content in milk powder.