面粉吸水率是评价面粉质量和预测面制品加工特性的重要品质性状。 面粉吸水率的测定主要参照国际或国家标准利用粉质仪进行, 其测定方法费时费力。 基于此, 提出利用可见近红外光谱分析技术结合多元统计分析进行面粉吸水率快速、 无损检测。 参照国标法测定150份小麦面粉样品的吸水率, 面粉吸水率变幅为53.10%~74.50%。 利用可见近红外分析仪采集面粉样品的光谱信息, 有效光谱范围为570~1 100 nm。 采用偏最小二乘回归(PLSR)、 主成分回归(PCR)和支持向量机回归(SVR)将光谱信息和面粉吸水率进行关联, 分别建立面粉吸水率的定量分析预测模型, 筛选最优的建模方法。 在优选的建模方法的基础上, 采用竞争性自适应重加权(CARS)、 区间随机蛙跳(iRF)、 迭代保留信息变量(IRIV)和连续投影(SPA)算法提取特征波长, 筛选最优的特征波长提取算法。 基于最优的建模方法和最优的特征波长提取算法提取的特征波长, 采用标准化(NL)、 一阶求导(1st Der)、 基线校正(BL)、 标准正态变换(SNV)和去趋势化(DT)5种光谱预处理方法对特征波长的光谱进行预处理, 筛选最优的光谱预处理方法。 结果表明, 采用NL光谱预处理方法对CARS算法提取的24个特征波长(仅占原始波长的2.26%)的光谱进行预处理后建立的PLSR模型性能最佳, 预测集相关系数(R2p)、 预测集均方根误差(RMSEP)和预测相对分析误差(RPD)分别为0.889 4、 1.458 5和2.641 3。 采用CARS算法提取的特征波长所建的模型不仅能提高模型的性能, 还很大程度提高模型运算效率、 降低仪器制造成本和光谱仪微型化的难度, 从而为面粉吸水率可见近红外无损、 快速检测研究奠定了基础。

The water absorption rate of flour is an important quality parameter for evaluating flour quality and predicting the processing characteristics of flour-based products. Determining the water absorption rate is mainly conducted using a gluten analyzer according to international or national standards, which is time-consuming and labor-intensive. Therefore, this study proposes using visible near-infrared spectroscopic analysis technology for rapid and non-destructive detection of the water absorption rate of flour. The water absorption rates of 150 wheat flour samples were determined according to the national standard method, and the value rang from 53.10% to 74.50%. The spectral information of the flour samples was collected using a visible near-infrared spectrometer, with an effective spectral range from 570 to 1 100 nm. Partial least squares regression (PLSR), principal component regression (PCR), and support vector machine regression (SVR) was used to correlate the spectral information with the water absorption rate of flour. Quantitative analysis prediction models for the water absorption rate were established, and the optimal modeling methods were selected. Based on the selected modeling methods, competitive adaptive reweighted sampling (CARS), interval random frog leaping (iRF), iterative variable selection using the retained informative variables (IRIV), and successive projections algorithm (SPA) were employed to extract feature wavelengths and select the optimal feature wavelength extraction algorithm. Five spectral preprocessing methods, including normalization (NL), first derivative (1st Der), baseline correction (BL), standard normal variate (SNV), and detrending (DT), were applied to preprocess the spectral data of the feature wavelengths. The optimal spectral preprocessing method was determined. The results showed that the PLSR model built after preprocessing the spectra of the 24 feature wavelengths (only 2.26% of the original wavelengths) extracted by the CARS algorithm using the NL spectral preprocessing method achieved the best performance. The correlation coefficient (R2p), root mean square error of prediction (RMSEP), and relative prediction deviation (RPD) for the prediction set were 0.889 4, 1.458 5, and 2.641 3, respectively. The model built using the feature wavelengths extracted by the CARS algorithm not only improved the models performance but also significantly increased the computational efficiency, reduced instrument manufacturing costs, and alleviated the challenges of miniaturizing the spectrometer. This study provides a foundation for the non-destructive and rapid detection of the water absorption rate of flour using visible near-infrared spectroscopy.