基于高光谱技术的蛋白粉掺假检测研究

蛋白粉是健身者必备的营养补剂，市场需求在不断增加，一些不法商家为了谋取利益，在蛋白粉中加入廉价的粉末售卖。传统的蛋白粉掺杂的检测方法费时、费力，操作复杂，且成本昂贵。高光谱技术具有易于操作、在不损害实验样本的情况下可快速检测等优点，因此，提出使用高光谱技术以实现蛋白粉掺假检测。在蛋白粉中分别加入质量百分数5%～60%，浓度间隔5%的三类掺假物（玉米粉、大米粉和小麦粉），并采集所有样本的光谱信息。在对蛋白粉中的玉米粉、大米粉和小麦粉三类掺假物进行定性判别时，首先分别采用卷积平滑（SG）、标准化（Normalize）、多元散射校正法（MSC）、基线校正（Baseline）和标准正态变换（SNV）的预处理方法对光谱数据进行处理，然后建立基于主成分回归（PCR）、反向传播神经网络（BPNN）和随机森林（RF）的模型，其中基于全波段光谱MSC预处理方法下建立的RF模型最优，其整体准确率达到了100%，其对应的R_P和RMSEP分别为0.997 9和0.018 9。在对蛋白粉中不同掺假物浓度进行定量分析时，对三类掺假样本的光谱分别进行SG，Normalize，MSC，Baseline和SNV的预处理，并建立LSSVM模型；比较不同预处理方法下的各模型之间的性能，在蛋白粉中掺玉米粉、大米粉和小麦粉的LSSVM预测模型最佳预处理方法分别是无、Baseline和Normalize，然后，采用连续投影算法（SPA）和竞争性自适应重加权算法（CARS）对其筛选，并建立LSSVM模型，三类掺假样本的SPA-LSSVM模型对应的R_P为0.989 0，0.986 0和0.997 9，CARS-LSSVM模型对应的R_P为0.991 0，0.994 6和0.999 1，故三类掺假样本的CARS-LSSVM模型预测效果更佳。研究表明：高光谱技术可以实现对蛋白粉掺假的定性、定量的检测，并且操作简单、检测快速和无损。

Research on Protein Powder Adulteration Detection Based on Hyperspectral Technology

Protein powder is an essential nutritional supplement for bodybuilders, and the market demand is increasing. Some unscrupulous businessmen are adding cheap powder to protein powder for sale to profit. The traditional protein powder adulteration detection method is time-consuming, laborious, complicated and expensive. Hyperspectral technology has the advantages of easy operation and rapid detection without damaging the experimental sample. Therefore, this paper proposes the use of hyperspectral technology to achieve protein powder adulteration detection. In the experiments, three types of adulterants (corn flour, rice flour and wheat flour) with 5%~60% mass percentages and 5% concentration interval were added to the protein powder, and the spectral information of all samples was collected. In the qualitative discrimination of the three types of adulterants (corn flour, rice flour and wheat flour) in the protein powder, the spectral data were firstly processed using the pre-processing methods of convolutional smoothing (SG), normalization (Normalize), multiple scattering correction (MSC), baseline correction (Baseline) and standard normal transformation (SNV), and then the spectral data were established based on principal component regression ( PCR), backpropagation neural network (BPNN), and random forest (RF) models, among which the RF model built under the MSC preprocessing method based on full-band spectra is the best, and its overall accuracy reaches 100%. Its corresponding R_P and RMSEP are 0.997 9 and 0.018 9, respectively. In the quantitative analysis of different adulterant concentrations in protein powder, the spectra of the three types of adulterated samples were pretreated with SG, Normalize, MSC, Baseline and SNV, respectively, and LSSVM models were established. The performance between the models under different pretreatment methods was compared. The best LSSVM prediction models were used for corn flour, rice flour and wheat flour adulterated in protein powder preprocessing methods were None, Baseline and Normalize, and then, the continuous projection algorithm (SPA) and competitive adaptive reweighting algorithm (CARS) were used to screen them and build LSSVM models. The RP corresponding to the SPA-LSSVM models for the three types of adulterated samples were 0.989 0, 0.986 0 and 0.997 9, and the R_P of the CARS- LSSVM model corresponds to R_P of 0.991 0, 0.994 6 and 0.999 1, so the CARS-LSSVM model for the three types of adulterated samples has a better prediction. Research shows that hyperspectral technology can achieve qualitative and quantitative detection of protein powder adulteration and simple operation, rapid and non-destructive detection.