采用近红外光谱(NIRS)结合偏最小二乘(PLS)方法, 实现对饲料中粗脂肪和粗纤维的快速定量分析。 采用Norris-Williams平滑求导(NW)和多元散射校正(MSC)方法对光谱进行预处理; 蒙特卡罗无信息变量消除法(MCUVE)、 变量组合集群分析法(VCPA)和区间变量迭代空间收缩法(iVISSA)用于光谱变量选择和优化; PLS用于光谱校正模型的建立, 采用校正集相关系数( R_c)、 交互验证均方根误差(RMSECV)、 预测集相关系数( R_p)和预测集均方根误差(RMSEP)评价模型。 光谱预处理中经MSC处理后的光谱模型优于其他预处理方法, 其RMSECV和RMSEP值都减小, R_c和 R_p值都增大。 脂肪定量分析中, 原始光谱模型的RMSECV和 R_c为0.21和0.87, RMSEP和 R_p为0.20和0.88, 变量数( V_n)为1 501; 经MCUVE方法选择变量后建立的定量模型, 其RMSECV和 R_c为0.17和0.92, RMSEP和 R_p为0.19和0.89, V_n为400个; 经VCPA选择变量建立PLS定量模型, 其RMSECV和 R_c为0.21和0.87, RMSEP和 R_p为0.25和0.81, V_n为12; 经iVISSA选择变量后的模型, 其RMSECV和 R_c为0.21和0.86, RMSEP和 R_p为0.20和0.87, V_n为20。 粗纤维定量分析中, 原始模型的RMSECV和 R_c为0.28和0.91, RMSEP和 R_p为0.25和0.95, V_n为1 501; 经MCUVE选择后的模型, 其RMSECV和 R_c为0.23和0.95, RMSEP和 R_p为0.23和0.94, V_n为740; 经VCPA选择变量后的模型, 其RMSECV和 R_c为0.27和0.91, RMSEP和 R_p为0.30和0.91, V_n为11; 经iVISSA选择后变量的模型, 其RMSECV和 R_c为0.29和0.90, RMSEP和 R_p为0.27和0.93, V_n为20。 结果表明, MSC方法可以有效提高光谱质量, 消除光谱平移误差; MCUVE变量选择方法可以简化模型提高模型精度和稳定性, 建立最优模型。 在粗脂肪的定量分析模型中, MSC处理后的光谱经过MCUVE选择后剩余400个变量, R_c和 R_p相较于全谱模型提高了0.05和0.01, RMSECV和RMSEP分别降低到了0.17和0.19; 经VCPA和iVISSA选择变量的模型其结果与全谱模型相似, 但其变量分别只有12和20个。 在粗纤维模型中, 经MCUVE选择后740个变量用于建立PLS模型, 其 R_c和 R_p为0.95和0.94, RMSECV和RMSEP分别为0.23和0.23; VCPA和iVISSA分别运用11和12个变量建立回归模型, 但结果都比MCUVE模型差。 利用饲料近红外光谱建立MSC-MCUVE-PLS模型可以有效对饲料粗脂肪和粗纤维进行定量分析。

Near infrared spectroscopy (NIRS) combined with partial least squares (PLS) method was used to achieve rapid quantitative analysis of crude fat and corase fiber in feed. The norris-williams derivation (NW) and multiplicative scatter correction (MSC) methods were used to pretreat the spectrum, and the monte carlo based uninformative variable elimination (MCUVE), variables combination population analysis (VCPA) and interval variable iterative space shrinkage approach (iVISSA) were used to select and optimize the variables of the spectrum. PLS was used for the establishment of the spectral calibration model, and the parameters of calibration set correlation coefficient ( R_c), root mean square error of cross validation (RMSECV), prediction set correlation coefficient ( R_p) and root mean square error of prediction (RMSEP) were used to evaluate the models. Compared with other pretreatment methods, the RMSECV and RMSEP values of the spectral model after MSC treatment decreased, while the R_c and R_p values increased. In the crude fat quantitative analysis model, the RMSECV and R_c of the original spectral model were 0.21 and 0.87, RMSEP and R_p were 0.20 and 0.88, and the number of variables ( V_n) was 1501. After selecting variables by MCUVE method, RMSECV and R_c were 0.17 and 0.92, RMSEP and R_p were 0.19 and 0.89, and V_n was 400. For VCPA-PLS model, the RMSECV and R_c were 0.206 and 0.87, RMSEP and R_p were 0.25 and 0.81, and V_n was 12. For iVISSA-PLS model, the RMSECV and R_c were 0.21 and 0.86, RMSEP and R_p were 0.20 and 0.87, and V_n was 20. In the corase fiber model, the RMSECV and R_c of the original PLS model were 0.28 and 0.91, RMSEP and R_p were 0.23 and 0.95, and V_n was 1 501. After MCUVE selection, the RMSECV and R_c of the model were 0.23 and 0.95, RMSEP and R_p were 0.25 and 0.94, and V_n was 740. After VCPA selection, the RMSECV and R_c of the model were 0.27 and 0.91, RMSEP and R_p were 0.30 and 0.91, and V_n was 11. After iVISSA selection, the RMSECV and R_c of the model were 0.29 and 0.90, RMSEP and R_p were 0.27 and 0.93, and V_n was 20. The results showed that the MSC method could effectively improve the spectral quality and eliminate the spectral translation error; the MCUVE variable selection method could simplify the model to improve the model accuracy and stability, and establish the optimal model. In the crude fat quantitative analysis model, after the MSC-processed spectrum was selected by MCUVE, the remaining 400 were used to establish the PLS model, R_c and R_p were improved by 0.05 and 0.01 compared to the full-spectrum model, and the RMSECV and RMSEP were reduced to 0.17 and 0.19; The model selected by VCPA and iVISSA had almost the same result as the full-spectrum model, and its greatest feature was that only 12 and 20 variables were selected. In the corase fiber model, 740 variables selected by MCUVE were used to establish the PLS model with R_c and R_p of 0.95 and 0.94, RMSECV and RMSEP of 0.23 and 0.23, respectively; VCPA and iVISSA used 11 and 12 variables to establish the regression model, but its model results were all worse than the MCUVE model. The establishment of MSC-MCUVE-PLS quantitative analysis model using feed near-infrared spectroscopy could effectively quantify crude fat and corase fiber in feed.