期刊
TALANTA
卷 167, 期 -, 页码 563-572出版社
ELSEVIER SCIENCE BV
DOI: 10.1016/j.talanta.2017.02.047
关键词
Milk nutrient analysis; Vis/NIR spectroscopy; Space-resolved measurement; Three-way PIS regression; Variable selection; Optical sensor
资金
- Russian Ministry of Education and Science [4.7001.2017/BP]
- AiF Projekt GmbH
- Bundesministerium fur Wirtschaft und Thechnologie (BMWi) [KF2784803BN2]
New technique of diffuse reflectance spectroscopic analysis of milk fat and total protein content in the visible (Vis) and adjacent near infrared (NIR) region (400-995 nm) has been developed and tested. Sample analysis was performed through a probe having eight 200-um fiber channels forming a linear array. One of the end fibers was used for the illumination and other seven for the spectroscopic detection of diffusely reflected light. One of the detection channels was used as a reference to normalize the spectra and to convert them into absorbance equivalent units. The method has been tested experimentally using a designed sample set prepared from industrial raw milk standards with widely varying fat and protein content. To increase the modelling robustness all milk samples were measured in three different homogenization degrees. Comprehensive data analysis has shown the advantage of combining both spectral and spatial resolution in the same measurement and revealed the most relevant channels and wavelength regions. The modelling accuracy was further improved using joint variable selection and preprocessing optimization method based on the genetic algorithm. The root mean square errors of different validation methods were below 0.10% for fat and below 0.08% for total protein content. Based on the present experimental data, it was computationally shown that the full-spectrum analysis in this method can be replaced by a sensor measurement at several specific wavelengths, for instance, using light-emitting diodes (LEDs) for illumination. Two optimal sensor configurations have been suggested: with nine LEDs for the analysis of fat and seven for protein content. Both simulated sensors exhibit nearly the same component determination accuracy as corresponding full-spectrum analysis.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据