豆粕是大豆浸提取豆油后经适当干燥和热处理所得副产品, 是制作禽畜类饲料的主要原料, 其品质决定营养价值。 针对现有豆粕品质检测方法存在着有毒化学试剂使用多、 操作复杂、 分析时间长、 无法满足实际生产线快速检测及调控需求等问题, 提出一种基于近红外光谱分析的豆粕品质多组分检测方法, 以期用于产品质量在线检测及调控。 从大豆油脂加工生产线上采集豆粕样品449个, 利用105 ℃烘箱法、 凯氏定氮法和索氏提取法分别测定样品的水分、 蛋白质和脂肪化学值, 采用瑞士BuchiNIRMaster傅里叶变换近红外光谱仪采集样品漫反射光谱。 首先利用马氏距离法剔除异常样本, 然后用多种方法对光谱数据进行降噪处理, 对比分析发现小波去噪效果最优。 分别采用KS和SPXY两种算法确定豆粕不同组分的最佳样本分集。 为了探讨豆粕组分的近红外吸收特性, 剔除光谱冗余信息, 降低模型计算复杂度, 采用区间偏最小二乘法(iPLS)对4 000~10 000 cm^-1全谱进行特征提取, 优选出水分、 蛋白质和脂肪的特征吸收波段分别为4 904~5 200, 4 304~4 600和4 304~4 600 cm^-1。 最后建立豆粕组分含量的广义回归神经网络(GRNN)预测模型。 为了减少网络的输入变量, 缩小网络规模, 提高运行速度, 采用PLS对光谱数据降维, 提取主因子得分作为GRNN输入变量。 通过交叉验证循环法优选网络参数光滑因子spread值, 建立豆粕多组分含量PLS-GRNN预测模型, 并与经典的PLS和BP模型对比, 发现PLS-GRNN模型效果更优, 其水分、 蛋白质和脂肪的预测集 R²分别为0.976 9, 0.940 2和0.911 1, RMSEP分别为0.091 2, 0.383 4和0.113 4, RSD分别为0.79%, 0.83%和8.53%。 虽然脂肪的预测误差相对较大, 但也在模型评定标准可用范围之内。 实验表明基于PLS-GRNN的近红外光谱分析用于豆粕品质检测是可行的, 能够用于实际生产过程中的品质监控。

Soybean meal is a by-product of soybean oil extracted from soybean after proper drying and heat treatment. It is the main raw material for making livestock feed, and its quality determines the nutritional value. There are many problems with existing soybean meal quality detection methods, such as the use of toxic chemical reagents, complex operation, long analysis time, so they cannot meet the needs of rapid detection and control in the production process. This paper proposes a multi-component detection method of soybean meal quality based on near infrared spectroscopy for on-line detection and control of product quality. 449 soybean meal samples were collected from the soybean oil processing line. The chemical values of moisture, protein and fat were determined by 105 ℃ oven method, Kjeldahl nitrogen determination method and Soxhlet extraction method, respectively. The diffuse reflectance spectra of samples were collected by the Swiss Buchi NIRMaster Fourier Transform near-infrared spectrometer. Firstly, the Mahalanobis distance method was used to remove abnormal samples, and then the spectral denoising was processed by various methods. The results show that the wavelet denoising effect is the best. KS and SPXY algorithms were used to determine the optimal sample partition of different components. In order to investigate the NIR absorption characteristics of soybean meal components, eliminate spectral redundancy and reduce the computational complexity of the model, interval Partial Least Squares (iPLS) was used to extract the features from the whole spectrum of 4 000~10 000 cm^-1. The optimized characteristic absorption bands of moisture, protein and fat were 4 904~5 200, 4 304~4 600 and 4 304~4 600 cm^-1, respectively. Finally, a Generalized Regression Neural Network (GRNN) model was established to predict the component contents of soybean meal. In order to reduce the input variables and the network scale improve the operation speed, PLS was used to reduce the dimension of spectral data, and the principal factor score was extracted as the input variable of GRNN. The PLS-GRNN prediction models of soybean meal multi-component contents were established by optimizing the smooth factor spread through the cross-validation and compared with the classical PLS and BP models. The results show that the PLS-GRNN models are good, the prediction determination coefficients ( R²) of moisture, protein and fat are 0.976 9, 0.940 2 and 0.911 1, the Root-Mean-Square Errors of Prediction (RMSEP) are 0.091 2, 0.383 4 and 0.113 4, the Relative Standard Deviations (RSD) of prediction are 0.79 %, 0.83 % and 8.53 %, respectively. Although the prediction error for fat is relatively large, it is also within the available range of the model evaluation criteria. The results show that the near infrared spectroscopy analysis based on PLS-GRNN is feasible to detect soybean meal quality and can be used for quality monitoring in the actual production process.