漫透射近红外光谱的便携式面粉品质安全检测仪的设计

基于近红外漫透射光谱分析技术，设计了便携式面粉品质安全检测仪，该检测仪主要包括光谱采集模块、光源控制模块、处理与显示模块以及电源模块。其中漫透射检测附件不仅可以实现光谱补偿功能，还可以有效避免外界杂散光的干扰，设计了控制光源开关的电路，通过实验确定样品的最佳厚度。选用树莓派4B作为核心处理器，选用可充电锂电池供电，仪器可持续供电2 h，仪器大小为250 mm×170 mm×300 mm。以去除麸皮后由小麦磨成的面粉为研究对象，总共180份样品，每份样品再分三份，分别为黄色、红色和蓝色。对所有的红色样品使用波长为900~1 870 nm的近红外漫透射光谱进行光谱信息采集并记录，对所有的黄色样品进行湿度值的测量并记录，对所有的蓝色样品进行DON含量的测量并记录，三种样品需要同时进行测量。利用箱线图剔除光谱两端的噪声和一个异常样本，最终选取1 048～1 747 nm波段光谱进行建模。利用多元散射校正(MSC)、S-G卷积平滑和标准正态变换(SNV)对原始光谱数据进行预处理，分别建立了面粉湿度的偏最小二乘回归预测模型和DON含量超标与否的PCA-逻辑回归分类模型。所建湿度的最优PLSR预测模型的建模集和预测集相关系数分别为0.883和0.853，均方根误差分别为0.382%和0.286%，残差预测偏差RPD为2.5；所建DON含量超标与否的PCA-逻辑回归分类模型的预测集ROC曲线下的AUC值为0.927，混淆矩阵显示未超标样本的预测准确率为96%，超标样本的预测准确率为89%。基于PyQt5设计GUI界面，运用Python语言编写了面粉品质实时检测系统，该检测软件可以实现PLSR、PCA-逻辑回归模型的训练、保存和加载。利用外部验证集试验验证了便携式面粉多品质检测仪的精确性和稳定性。结果显示面粉湿度的外部验证集相关系数和均方根误差为0.876和0.21%，最大相对误差为2.89%。面粉DON含量超标与否的识别准确率为90%，表明该仪器可以对面粉的湿度和DON含量超标与否进行无损检测分析。

Design of Portable Flour Quality Safety Detector Based on Diffuse Transmission Near-Inf rared Spectroscopy

Based on near-infrared diffuse transmission spectrum analysis technology, a portable flour quality safety detector is designed, which mainly includes a spectrum acquisition module, light source control module, processing and display module and power supply module. Among them, diffuse transmission detection accessories can be lifted freely to facilitate the placement of samples and effectively avoid the interference of external stray light. The circuit for controlling the switch of the light source is designed, and experiments determine the optimal thickness of samples. Raspberry Pi 4B is selected as the core processor, and a rechargeable lithium battery is selected for the power supply. The instrument can continuously supply power for 2 hours, and its size is 250 mm× 170 mm×300 mm. A total of 180 samples were taken from the flour ground from wheat after bran removal. Each sample was divided into three parts: yellow, red and blue. For all red samples, use the near-infrared diffuse transmission spectrum with the wavelength of 900~1 870 nm to collect and record the spectral information, measure and record the humidity value of all yellow samples, and measure and record the Don content of all blue samples. The three samples need to be measured at the same time. The noise at both ends of the spectrum and an abnormal sample are eliminated by box diagram, and finally, the spectrum in 1 048~1 747 nm band is selected for modeling. The multivariate scattering correction (MSC), S-G convolution smoothing and standard normal transformation (SNV) were used to preprocess the original spectral data, and the partial least squares regression prediction model of flour humidity and PCA-logistic regression classification model of DON content exceeding the standard were established respectively. The correlation coefficients of the calibration set and prediction set of the optimal PLSR prediction model for humidity are 0.883 and 0.853, the root mean square error is 0.382% and 0.286%, and the residual prediction deviation RPD is 2.5. The AUC value under the ROC curve of the prediction set of the PCA-logistic regression classification model is 0.927. The confusion matrix shows that the prediction accuracy of samples not exceeding the standard is 96%, and that of samples exceeding the standard is 89%. The GUI interface is designed based on PyQt5, and the flour quality real-time detection system is written by Python language. The detection software can realize the training, saving and loading of PLSR and PCA-logistic regression models. The accuracy and stability of portable flour multi-quality tester were verified by external verification set test. The results showed that the correlation coefficient and root mean square error of the external verification set of flour humidity were 0.876 and 0.21%, and the maximum relative error was 2.89%. The recognition accuracy of flour DON content exceeding the standard is 90%, indicating that the instrument can be used for nondestructive detection and analysis of flour humidity and DON content.