摘要： 为了满足生鲜肉品质参数无损检测领域，对轻便式、低成本设备的开发需求，提出一种基于多光谱漫反射技术的生鲜肉品质检测方法。首先根据漫反射近似理论，结合牛肉样品散射系数、吸收系数及折射率等参数，在无线细垂直光束的蒙特卡洛仿真的基础上，对具有一定发散角度LED光源进行了初始化的校正，分别从光源照射位置概率分布、不同角度的照射概率分布、仰角、方向角的概率分布、不同角度光线入射样品时反射引起能量损失及对光子权重的影响，得到在LED光源发散角情况下，不同源探距下的漫反射率与检测深度，确定了光源与检测器之间的最佳距离为15 mm，然后根据此距离，搭建了多光谱漫反射检测平台，检测平台由8组中心波长为470，535，575，610，650，720，780和960 nm的LED光源组成，与所要检测的生鲜牛肉品质参数相对应。同时利用LED光源的发散角，确定了光源到样品表面的垂直距离与每个光源的安装位置，保证光源照射到样品的区域是均匀的。样品的漫射光强经由信号采集与放大电路的处理后传至上位机，并在上位机完成建模与分析。最后为验证该检测系统的性能，以生鲜牛肉新鲜度参数中的颜色(L*，a*，b*)与pH值为指标，利用60个样品进行了试验，分别得到8个光源下的原始光强值与校正后的反射率值，然后将牛肉样品按照3∶1比例分为校正集与预测集，针对原始光强值与反射率值，分别利用多元线性回归(multiple linear regression, MLR)，偏最小二乘回归(partial least squares regression, PLSR)与偏最小二乘支持向量机回归(partial least-squares support vector machine, LS-SVM)三种方法，建立各个参数在原始光强与反射率数据两种情况下的预测模型，并得到最佳模型结果。结果表明，利用反射率数据建模结果均好于光强数据结果，其中参数L*，a*，b*的MLR建模结果优于PLSR与LS-SVR，其预测集相关系数分别为0.983 2，0.907 2及0.935 9，预测集误差分别为1.00，2.14及0.67。参数pH值的LS-SVR建模结果优于PLSR与MLR，其预测集相关系数为0.942 0，误差为0.19。最后利用未参与试验的20块牛肉样品对模型进行了验证，颜色L*，a*，b*及pH参数的预测值与实测值的相关系数均大于0.85，结果证明，利用多光谱漫反射技术以及所搭建的多光谱漫反射检测系统对生鲜牛肉品质参数检测是可行的，该方法能够为设计便携式或微型化生鲜牛肉品质的无损检测仪器提供参考与依据。
关键词：多光谱漫反射检测；蒙特卡洛仿真；牛肉品质参数；光源系统；预测模型
Abstract：In order to meet the development requirement of portable and low cost equipment in the field of non-destructive detection of fresh meat quality parameters, a new method based on multi spectral diffuse reflectance technology for fresh meat quality detection is proposed. Based on the diffusion approximation theory and combined with the sample scattering coefficient, absorption coefficient and refractive index of beef and other parameters, based on Monte Carlo simulation of thin vertical beam on the radio, on a certain divergence angle of LED light source are initialized with correction respectively from the light source position probability distribution and different angles of the irradiation probability distribution, angle, direction angle the probability distribution and different incident light angle sample reflection caused by the energy loss and influence on the photon weight, the LED divergence angle under different source detector diffuse reflectance and depth of detection distance, the optimum distance between the light source and the detector is 15 mm, then according to the distance, to build a multi spectral diffuse reflection detection platform, multi spectral detection platform by 8 groups of 470, 535, 575, 610, 650, 720, 780, 960 nm LED. The source composition corresponds to the quality parameters of fresh beef to be detected. At the same time, according to the 8 LED light source, the light source design layout structure of probe point symmetry, 8 light sources inside the probe to the detector as the center, symmetric distribution, while using LED light source divergence angle, determine the installation position of the light source to the sample surface and the vertical distance of each light source, to ensure the light source to the sample area is uniform. In addition, the probe embedded within the design of signal acquisition, amplification and transmission components, signal acquisition part uses the spectral response range of 400～1 100 nm light intensity detector, sample diffuse intensity after processing to the host computer through the signal acquisition and amplification circuit, and the software finished modeling and analysis. Finally in order to verify the performance of the detection system, with fresh beef freshness in the color parameter (L*, a*, b*) and the pH value as the index was tested using 60 samples, 8 light source under the original light intensity value and corrected reflectance values respectively, and then the beef samples according to 3∶1 the proportion is divided into set and prediction set correction, for the original value of light intensity and reflectivity values, respectively, using multiple linear regression (MLR), Multiple Linear Regression Partial Least Squares Regression partial least squares regression (PLSR) and partial least squares support vector machine regression Partial Least-Squares Support Vector Machine (LS-SVM) three methods, model parameters in the original light intensity and reflectivity data of the two cases, and get the best results. The results show that the results of modeling using reflectivity data are better than those of light intensity data. The MLR modeling results of parameters L*, a* and b* are better than those of PLSR and LS-SVR, and their correlation coefficients of prediction set are 0.983 2, 0.907 2 and 0.935 9, respectively, and the prediction set errors are 1.00, 2.14 and 0.67, respectively. The LS-SVR modeling results of parameter pH value are better than that of PLSR and MLR, and the correlation coefficient of the prediction set is 0.942 0 and the error is 0.19. Finally, using 20 pieces of beef samples which did not participate in the test to validate the model, the color of L*, a*, b* and pH parameters of the prediction value of the correlation coefficient and the measured value is greater than 0.85, the results proved that using multispectral diffuse reflection technology and building the multispectral reflectance detecting system are feasible for the detection of fresh beef the quality parameters, this method can provide reference and basis for the nondestructive testing instrument design of portable or micro fresh beef quality.
Key words：Multispectral diffuse reflectance detection; Monte Carlo simulation; Beef quality parameter; Light source system; Prediction model
魏文松，彭彦昆，郑晓春，王文秀. 基于多光谱漫反射的牛肉品质参数检测方法研究[J]. 光谱学与光谱分析, 2019, 39(04): 1177-1185.
WEI Wen-song, PENG Yan-kun, ZHENG Xiao-chun, WANG Wen-xiu. Research on Detection of Beef Freshness Parameters Based on Multi Spectral Diffuse Reflectance Method. SPECTROSCOPY AND SPECTRAL ANALYSIS, 2019, 39(04): 1177-1185.
[1] Elmasry G, Sun D W, Allen P. Journal of Food Engineering, 2012, 110(1): 127.
[2] Trinderup C H. Meat Science, 2015, 102(102C): 1.
[3] Kamruzzaman M, Makino Y, Oshita S. Meat Science, 2016, 116: 110.
[4] Stuart O J C, Sascha M K, Victoria P, et al. Meat Science, 2017, 134: 14.
[5] Jia B. Journal of Food Engineering, 2017，208: 57.
[6] Sharifzadeh S, Clemmensen L H, Borggaard C. Engineering Applications of Artificial Intelligence, 2014，27(1): 211.
[7] Iqbal A, Sun D W, Allen P. Journal of Food Engineering, 2013，117(1): 42.
[8] MA Shi-bang，TANG Xiu-ying, XU Yang, et al(马世榜, 汤秀映，徐 杨，等). Transactions of the Chinese Society of Agricultural Engineering(农业工程学报), 2012，28(18): 263.
[9] Wu J, Peng y, Li y et al. Journal of Food Engineering, 2012，109(2): 267.
[10] CHEN Kun-jie, SUN Xin, LU Qiu-yan(陈坤杰, 孙 鑫, 陆秋琰). Transactions of The Chinese Society of Agricultural Machinery(农业机械学报), 2009，40(4): 173.
[11] WEI Wen-song, PENG Yan-kun(魏文松，彭彦昆). Transactions of The Chinese Society of Agricultural Machinery(农业机械学报), 2016，47(s1): 324.
[12] HOU Rui-feng，HUANG Lan, WANG Zhong-yi, et al(侯瑞锋, 黄 岚，王忠义，等). Transactions of the Chinese Society of Agricultural Engineering(农业工程学报), 2005，21(9): 12.
[13] PENG Dong-qing, LI Hui, XIE Shu-lin(彭东青, 李 晖, 谢树森). Journal of Optoelectronics·Laser(光电子·激光), 2003，14(1): 107.
[14] Fang Q, Boas D A. Optics Express, 2009, 17(22): 20178.
[15] LI Xiao-qin(李晓勤). Journal of Agricultural Mechanization Research(农机化研究), 2009，31(7): 104.
[16] Pu H, Kamruzzaman M, Sun D W. Trends in Food Science & Technology, 2015，45(1): 86.
[17] YANG Qi-chang, XU Zhi-gang, CHEN Hong-da, et al(杨其长, 徐志刚，陈弘达，等). Journal of Agricultural Science and Technology(中国农业科技导报), 2011，13(5): 37.
[18] LIAO Yi-tao, FAN Yu-xia, CHENG Fang, et al(廖宜涛, 樊玉霞，成 芳，等). Transactions of the Chinese Society of Agricultural Engineering(农业工程学报), 2010，26(s1): 379.
[19] HU Yao-hua, GUO Kang-quan, SUZUKI Taka-ri，et al(胡耀华, 郭康权，铃木孝介，等). Journal of Northwest Sci-Tech University of Agriculture and Forestry(西北农林科技大学学报: 自然科学版), 2008，36(7): 160.
