基于Vis/NIR光谱传感的鲜食葡萄糖度检测系统

Vis/NIR Based Spectral Sensing for SSC of Table Grapes

糖度是影响鲜食葡萄品质与风味的关键因素,对其可溶性固形物SSC的检测具有切实需求。近年来,随着芯片级光谱传感器的生产技术趋于成熟,具有高精确性与稳定性的片上光谱传感器为可见/近红外检测技术开辟了新的途径。设计、搭建、测试了一套体积小、易操作、低成本的用于鲜食葡萄糖度无损检测的光学系统。系统以两块搭载芯片级光谱分析技术的新一代可见/近红外光谱传感器AS7263(美国AMS半导体公司)为核心元件。每个AS7263传感器具有6个集成了纳米光干涉滤波器的数字光谱通道和一个可通过单芯片准确控制电流(1~100 mA)的LED光源。传感器光谱通道的中心波长范围610~860 nm;两个LED光源的中心波长分别为730和850 nm,半峰全宽(FWHM)为50 nm。首先,运用此原型在避光环境下采集276颗巨峰葡萄浆果的光谱信息;用手持式PAL-1糖度仪检测样本SSC(°Brix)并计算基于t分布的样本糖度真值SSC_(t):SSC_(t0.9)与SSC_(t 0.95)。其次,针对样本原始光谱数据,采用PCA提取主成分,根据得分因子分布,剔除了16个位于置信区间外的异常样本;进一步采用一阶导数First Derivative(FD)、归一化Normalization(0,1)与标准化Standardization(0,1)3种方式做数据预处理,求取样本在12个通道下的吸光度A或Kubelka-Munk函数值F(R)。针对可见/近红外光谱自变量之间具有多重相关性、光谱信息与糖度信息之间非线性相关的特点,建立PLS-BP神经网络糖度预测模型(自变量为吸光度A或F(R)值,因变量为SSC_(t))。结果显示,当t分布的置信概率为0.95、光谱预处理方式为Standardization(0,1)、光谱信息指标为吸光度A时所建立的预测模型精度最高:决定系数r_(p)^(2)为0.93、均方根误差RMSE_(P)为0.181、预测集偏差Bias为-0.01、残留预测偏差RPD为3.78,可认为模型具有较高精度与较好适应性对葡萄SSC做出预测。最后,结合实验结果,作了葡萄浆果SSC光谱检测原理的分子尺度分析:在各分子振动类型中,O—H键伸缩振动的3倍频、4倍频,O—H键剪式振动与伸缩振动3倍频、4倍频的合频,C=O键伸缩振动的8倍频、9倍频为可见/近红外光谱检测的有效振动频率。该研究为未来工业与消费领域在线质量检测设备的高精度化、便携化、低成本化提拱了技术参考。

Detecting table grapes’s soluble solids content(SSC)is a crucial issue since berry quality and flavor are directly related to it.In recent years,as the technology of chip-level spectral sensors is becoming more and more advanced,on-chip spectral sensors with high accuracy and stability have blazed a new trail for spectral detection.In this work,a small,user-friendly,and cost-effective optical device that can detect the SSC of table grapes nondestructively has been designed,built,and tested.New generation Vis/NIR spectral sensor AS7263(sensor 1,2)with the capacity of chip level spectral analysis,does the key work for the system.Each sensor has six digital spectral channels with an integrated Gaussian filter and anLED with the programmable current(1~100 mA).The central wavelength of the spectral channel increases uniformly from 610 to 860 nm.Moreover,LEDs can emit light at 730 or 850 nm with fullwidth half max(FWHM)of 50nm.Firstly,this optical prototype collected a spectrum from 276 grape berries in a dark room.PAL-1 was used to detect SSC,and then we calculated the SSC_(t)based on t-distribution:SSC_(t0.9)and SSC_(t 0.95).Secondly,for the original spectral data,PCA was used to extract the principal components,and 16 abnormal samples located outside the confidence interval were excluded according to the distribution of the score factors.Besides,First Derivative(FD),Normalization(0,1)and Standardization(0,1)were used to preprocess the data.After that,we calculated the absorbance or KubelkaMunk function value F(R)for the samples at 12 channels.According to the multiple correlations between the independent variables of the Vis/NIR spectrum and the nonlinear correlations between the spectrum and SSC_(t),a PLS-BP neural network prediction model was developed for the grape SSC detection.The results showed that whenβwas 0.95,the preprocessing method was Standardization(0,1),the Parameter of the spectrum was absorbance(A),and the prediction model worked best:R_(p)^(2)=0.93,RMSE P=0.181,Bias=-0.01,and RPD=3.78,which can be considered that the model has high accuracy and better adaptability to predict the SSC of table grapes.Finally,on the one hand,referring to the experimental results,a very interesting molecular scale principle analysis is obtained for the grape absorption spectrum:among numerous molecular vibration types,the 3x,4x frequency of O—H bond(3x means stretching vibration at a triple fundamental frequency),the 3x+C,4x+C frequency of O—H bond(3x+C means the combination of scissoring vibration at a fundamental frequency and stretching vibration at a triple fundamental frequency),the 8x,9x of C=O bond are the effective vibration frequencies for Vis/NIR spectral detection.On the other hand,the prototype provides a technical reference for future online quality inspection equipment that is high-precision,portable and low-cost.