小麦粉的粉质特性决定了小麦粉的品质以及最终用途, 粉质特性受到小麦的品种, 产地, 以及加工工艺等多个因素的影响, 重要的粉质参数包括4个: 吸水率、 形成时间、 稳定时间、 弱化度。 近红外光谱广泛应用于小麦粉成分参数的检测, 如水分、 蛋白质、 灰分和湿面筋含量, 其中大多直接应用线性回归算法建立模型, 预测的精确度较低, 且检测粉质特性的研究较少, 研究结果也受到样本丰富度不足的影响。 该研究收集了968份来自不同国家和地区的小麦粉粉质特性数据及近红外光谱, 通过分类模型和回归模型的结合来提高粉质特性预测的精确度。 采用包括标准正态变换（SNV）、 线性去趋势（Detrend)、 多元散射矫正（MSC）和Savitzky-Golay一阶求导的方法对光谱数据进行预处理, 并通过交叉验证选择最佳预处理方法。 在建模方法上, 首先尝试了经典的线性回归方法, 即偏最小二乘回归（PLSR) 和主成分回归（PCR)。 发现两种方法的精确度大致相当, PCA模型的校正均方根误差（RMSEC）分别为2.186、 1.838、 4.037、 21.693, PLSR模型为2.039、 1.837、 3.968、 21.252, PLSR模型比PCR所需的因子更少。 其次, 使用该文提出的二阶段回归模型, 即先用高斯过程回归（GPR)的结果作为分类器对样本进行分类, 在不同类别的样本簇中分别建立PLSR模型进行粉质特性的预测, 再使用Sigmoid函数对PLSR模型进行融合。 这种建模方法对粉质特性预测的精确度有较大提高, 在不同粉质特性指标上的RMSEC分别为1.876、 1.160、 2.459、 14.449。

The farinograph characteristics of wheat flour determine the quality and the end use of wheat flour. The farinograph characteristics of wheat flour are influenced by wheat variety, origin, and milling process technology. There are four important farinograph parameters: water absorption, development time, stability time and degree of softening. Near-infrared spectroscopy (NIR) is widely used to determine wheat flour composition parameters, such as moisture, protein, ash and wet gluten content. Most of them directly use linear regression algorithms to establish models, which has low prediction accuracy, and there are few studies on detecting farinograph characteristics, and the results are also affected by the lack of sample richness. In this study, 968 samples of wheat flour from different countries and regions were collected, and an ensemble method of classification model and a regression model was proposed to improve the prediction accuracy of farinograph characteristics. Spectral preprocessing methods, including standard normal variation (SNV), linear detrending, multiplicative scatter correction (MSC) and Savitzky-Golay first-order derivative, were applied to the spectral data, and the best preprocessing method was selected with cross-validation. As for the modeling methods, the classical linear regression methods, i.e., partial least squares regression (PLSR) and principal component regression (PCR), were explored. The accuracies of the two methods are approximately equivalent. The root mean squared error of calibration (RMSEC) on farinograph parameters (i.e. water absorption, development time, stability time, and degree of softening) of the PCA model were 2.186, 1.838, 4.037, 21.693 and 2.039, 1.837, 3.968, 21.252 for PLSR correspondingly. The PLSR model requires fewer factors than PCR. Secondly, the two-stage regression model proposed in this paper was explored. Gaussian process regression (GPR) results were used as the classifier to cluster the samples, PLSR models were established in different clusters to predict the farinograph characteristics, and the sigmoid function was used to fuse the PLSR models. This modeling method can significantly improve the prediction accuracy of farinograph characteristics. The RMSEC on the predictions of farinograph parameters is 1.876, 1.160, 2.459 and 14.449 correspondingly.