基于光谱波段优化的鸡肉冻干粉粗蛋白近红外定量预测模型研究

Quantitative Prediction Model of the Crude Protein Content in the Chicken Freeze-dried Powder Based on the Optimizing Spectral Region of Near Infrared Spectroscopy

摘要:

本文通过将鸡肉蛋白质近红外光谱特性与组合间隔偏最小二乘法（SiPLS）、遗传算法(GA)相结合筛选校正模型的最佳建模光谱区域，旨在提高鸡肉冻干粉粗蛋白近红外定量预测模型的预测精度和模型稳健性。以260个鸡腿肌冻干粉为研究对象，提取其中100个样品的蛋白质，在999.7~2502.3 nm扫描腿肌冻干粉和腿肌提取蛋白冻干粉的NIRS，比较两光谱异同，根据腿肌提取蛋白冻干粉NIRS光谱特征及主成分分析（PCA）结果将全谱划分为10个建模光谱区，采用PLS建模，比较全谱建模与特征光谱组合区建模的优劣，筛选出基于鸡肉蛋白特征光谱的建模光谱组合区，应用FiPLS和BiPLS在全谱和优选出的光谱区再次进行建模光谱区域筛选，接着使用GA和FBiPLS进行第三次建模光谱筛选。结果表明：在999.7~1850.4 nm波长上采用FBiPLS法优选出的建模光谱区1811.6~1794.0 nm、1756.2~1722.4 nm、1704.4~1688.9 nm、1594.4~1580.8 nm、1510.8~1485.7 nm、1472.1~1424.3 nm、1222.2~1057.6 nm、1051.2~1008.7 nm所建模型最优。研究显示，为保证校正模型的精确性和稳定性，在筛选最佳建模波长时，应将样品预测成分的光谱特征与光谱筛选数学算法相结果，才能获得更好的建模结果。

Abstract:

This paper is trying to obtain the optimum spectra to build the calibration model by the spectral characteristics of the chicken protein in near infrared region, interval partial least-squares regression (iPLS) and geneti calgorithm (GA), and aims to improve the prediction accuracy and robustness of the near infrared spectroscopy quantitative prediction model of the crude protein content in the chicken freeze-dried powder.Taking the muscles powders of 260 freeze-dried legs as the research object, the proteins from the muscles powders of 100 freeze-dried legs were extracted, the near infrared reflectance spectra (NIRS) from the freeze-dried leg muscle powder samples and the protein samples of the freeze-dried leg muscle powder were scanned in the 999.7~2502.3 nm wavelength region. The differences between two NIRS were studied. The NIRS of the freeze-dried leg muscles powder was divided into 10 spectral regions based on spectral characteristic and principal component analysis (PCA). The partial least squares regression (PLS) was used to build the quantitative prediction model. First, the modeling results based on full-spectrum and combining spectral regions of the characteristic spectra were compared to select the optimum spectral regions. Next, the PLS (FiPLS) and backward interval PLS (BiPLS). It is the third time the modeling spectra were filtrated by GA and forward and backward regions and the optimum combining region based on the characteristic spectra by forward interval PLS (FiPLS) and backward interval PLS optimum spectral regions were extracted from the full-spectrum (BiPLS). Then, the modeling spectra were filtrated by GA and forward and backward interval partial least squares (FBiPLS). The result showed that the most optimum modeling spectral ranges were 1811.6~1794.0 nm、1756.2~1722.4 nm、1704.4~1688.9 nm、1594.4~1580.8 nm、1510.8~1485.7 nm、1472.1-1424.3 nm、1222.2~1057.6 nm and 1051.2~1008.7 nm by using FBiPLS. In conclusion, to ensure the accuracy and robustness of the calibration model, the selection method of optimum spectral regions was the combination of the spectral characteristics of sample composition and mathematical algorithm of wavelength selection.