猪配合饲料多品质近红外光谱关键变量筛选与模型建立

Key variables selection and models development based on near-infrared spectra for the multi-qualities in formula feedstuff for swine

中文关键词:  猪配合饲料  近红外光谱  蒙特卡罗-无信息变量消除-连续投影算法

英文关键词:formula feedstuff for swine  near-infrared spectroscopy  Monte-Carlo-uninformative variable elimination- successive projection algorithm

中文摘要:

      目的 基于饲料近红外光谱数据筛选影响猪配合饲料主要品质指标的关键波长变量, 从而建立饲料品质无损快速定量校正模型, 进而提高饲料品质无损快速检测效率。方法 采集饲料样品近红外光谱数据并获取水分、粗蛋白、粗脂肪、粗纤维参考值数据; 剔除异常值后采用基于联合X-Y距离样本集划分法(sample set partitioning based on joint X-Y distance, SPXY)划分校正集和外部验证集; 基于校正集数据采用蒙特卡罗-无信息变量消除-连续投影算法分别针对4个品质指标筛选25、20、15、10、5个关键变量, 分别建立校正模型并对外部验证集进行预测。结果 针对饲料水分、粗蛋白、粗脂肪、粗纤维所选关键变量个数分别为15、25、15、15, 模型维数分别为9、11、10、9, 测定系数分别为0.8288、0.8605、0.9338、0.8327, 校正均方根误差分别为0.17、0.81、0.31、0.22, 交互验证均方根误差分别为0.19、0.93、0.34、0.23, 相对预测性能分别为2.79、2.38、4.01、2.89。结论 通过变量筛选结合外部验证结果表明, 在保证模型准确度的前提下, 所选关键变量数明显少于全谱变量数, 可为提高饲料多品质无损快速定量检测工作效率提供一定的参考。

英文摘要:

      Objective Based on the near-infrared spectra of feedstuff samples for swine, to select key wavelength variables for the main quality indices in feedstuff, so as to develop the non-destructive rapid quantitative calibration models of quality indices of feedstuff, and thereby to improve the efficiency of non-destructive and rapid detection. Methods The near-infrared spectra data of feedstuff samples were collected and the specified values of moisture, crude protein, crude fat, and crude fiber were obtained. After outlier elimination, sample set partitioning based on joint X-Y distance (SPXY) algorithm was used to divide the data set into the calibration set and external validation set. Based on the calibration set data, Monte-Carlo-uninformative variable elimination-successive projection algorithm was used to select 25, 20, 15, 10, and 5 key variables for the 4 quality indices, respectively. Based on the key-variables data, the calibration models were developed and the external validation sets were predicted. Results The number of key variables selected for feed moisture, crude protein, crude fat, and crude fiber were 15, 25, 15, and 15 respectively, the number of the model factors were 9, 11, 10, and 9 respectively, the determination coefficients were 0.8288, 0.8605, 0.9338, and 0.8327 respectively, the root mean square errors of calibration were 0.17, 0.81, 0.31 and 0.22 respectively, the root mean square errors of cross-validation were 0.19, 0.93, 0.34 and 0.23 respectively, and the ratio performance deviations were 2.79, 2.38, 4.01 and 2.89 respectively. Conclusion It is demonstrated by the results of key variables selection combined with the prediction of external validation set that the numbers of the key variables selected are less than that of the full spectra obviously for the 4 quality indices when the accuracy of the models are ensured, which can provide a certain reference for improving the efficiency of non-destructive and rapid quantitative detection of the quality of feedstuff.