Abstract Near infrared spectroscopy (NIRS) combined with a partial least squares (PLS) algorithm was utilized as a rapid alternative analytical method for estimating the four main oxides in cement raw meal samples. An algorithm, known as cross-validation-absolute-deviation-F-test (CVADF), was proposed to eliminate the outliers existed in the calibration set. 5, 6, 2 and 2 out of 76 samples were identified as outliers for CaO, SiO2, Al2O3 and Fe2O3, respectively. The correlation coefficient of prediction (Rp) increased from 0.7773, 0.7877, 0.8894 and 0.6357 to 0.9075, 0.8572, 0.9038 and 0.6400, while the root mean square error of prediction (RMSEP) decreased from 0.2493, 0.2331, 0.0832 and 0.0449 to 0.1664, 0.1949, 0.0779 and 0.0447, respectively, indicating that the outliers are accurately identified and that the prediction performance of the PLS models established by the remaining samples was significantly improved. Some common outlier elimination methods, leverage diagnostic (LD), Euclidean distance diagnostic (EDD), Mahalanobis distance diagnostic (MDD) and principal component scores diagnostic (PCSD) were used for comparison. The results show that the proposed method is very promising with good results for the prediction capability.

中文翻译：

---

通过近红外光谱 (NIRS) 和交叉验证绝对偏差 F 检验 (CVADF) 提高水泥生料中氧化物测定准确度的新方法

摘要 近红外光谱 (NIRS) 结合偏最小二乘法 (PLS) 算法被用作一种快速替代分析方法，用于估算水泥生料样品中的四种主要氧化物。提出了一种称为交叉验证绝对偏差 F 检验 (CVADF) 的算法来消除校准集中存在的异常值。76 个样品中的 5、6、2 和 2 个分别被确定为 CaO、SiO2、Al2O3 和 Fe2O3 的异常值。预测的相关系数（Rp）从0.7773、0.7877、0.8894和0.6357增加到0.9075、0.8572、0.9038和0.6400，而预测的均方根误差（RMSEP）从0.2403、0.8894和0.6357减少到.0.2403、4093和4093。 ，分别为 0.0779 和 0.0447，表明异常值被准确识别，剩余样本建立的PLS模型的预测性能得到显着提高。一些常见的异常值消除方法，杠杆诊断（LD），欧几里德距离诊断（EDD），马氏距离诊断（MDD）和主成分评分诊断（PCSD）被用于比较。结果表明，所提出的方法非常有前途，具有良好的预测能力。