基于近红外光谱的小麦品质参数快速检测

发布时间：2018-09-07 17:10

【摘要】：为了实现对小麦品质参数的快速检测,本文提出了基于近红外光谱结合定量算法建立数学模型,以全国各地的160份小麦颗粒和小麦面粉为对象,采集其漫反射吸收光谱,并以国标法检测得到小麦的蛋白质含量、水分值和面筋值作为参考值,将所得样品均匀划分成训练集和验证集,对所建模型进行定标验证,并使用最佳模型检测小麦其它成分,具体研究内容和结果如下:1、以BP神经网络为基础,结合不同的优化算法建立数学模型通过神经网络定量分析方法,结合小波算法、去趋势算法、一阶导数算法、主成分分析法等预处理算法,对得到的小麦蛋白质光谱数据进行预处理,结果表明:神经网络通过主成分分析和去趋势算法得到的结果最佳R值高达0.98,RMSEP为0.26%,其他数学模型线性结果较佳,R值在0.95到0.98之间,RMSEP在0.26%到0.30%之间。2、以偏最小二乘回归PLSR为基础,结合不同的优化算法建立数学模型通过偏最小二乘法定量分析方法,结合小波算法,一阶导数,二阶导数,去趋势算法建立数学模型,采用交叉验证留一法取得最佳主成分个数。对小麦蛋白质光谱数据进行处理,结果比较,数据经过交叉验证留一法结合小波变换对光谱数据进行预处理,得到的结果最佳R值为0.92,RMSPCV为1.71%,而其他数学模型线性效果一般。3、比较分析上述两个模型通过两个数学模型进行分析比较得出,神经网络的预测能力明显高于偏最小二乘法,不管是R还是RMSEP的数值大小都说明神经网络更适合处理非线性的问题。4、基于BP神经网络模型,对小麦水分和面筋含量的检测使用已有的神经网络数学模型对小麦的水分和面筋值含量进行检测,通过分析比较,面粉面筋在BP-ANN结合Detrended优化算法下,得到的分析结果最佳,其中R值高达0.98,RMSEP值仅为0.24%,而小麦水分在BPP-ANN结合小波变换的算法下得到的结果最佳,其中R为0.96,RMSEP值为0.31%。
[Abstract]:In order to detect wheat quality parameters quickly, a mathematical model based on near infrared spectroscopy (NIR) combined with quantitative algorithm is proposed. 160 wheat grains and wheat flour from all over the country are taken as objects, and their diffuse reflectance absorption spectra are collected. The protein content, water value and gluten value of wheat were measured by national standard method as reference values. The samples were divided into training set and verification set, the model was calibrated and the best model was used to detect the other components of wheat. The specific research contents and results are as follows: 1. Based on BP neural network, mathematical model is established by combining different optimization algorithms. Quantitative analysis method of neural network, wavelet algorithm, trend removal algorithm, first-order derivative algorithm are used. Pretreatment algorithms such as principal component analysis (PCA) were used to preprocess the wheat protein spectral data. The results show that the optimum R value of neural network obtained by principal component analysis and de-trend algorithm is as high as 0.26, and that of other mathematical models is 0.95 to 0.98. The optimum R value of neural network is 0.26% to 0.30%, which is based on partial least square regression (PLSR). Combined with different optimization algorithms, mathematical models were established by partial least square quantitative analysis method, wavelet algorithm, first derivative, second order derivative, de-trend algorithm. The best number of principal components is obtained by using cross validation method. The spectral data of wheat protein were processed. The results showed that the data were preprocessed by cross validation and wavelet transform. The optimum R value is 0.92g RMSPCV (1.71V), while the linear effect of other mathematical models is normal. The comparison and analysis of the above two models show that the prediction ability of the neural network is obviously higher than that of the partial least square method. The magnitude of either R or RMSEP indicates that neural networks are more suitable for dealing with nonlinear problems. 4, based on the BP neural network model, The existing neural network mathematical model was used to detect the water content and gluten content of wheat. Through the analysis and comparison, the results of wheat flour gluten analysis under BP-ANN and Detrended optimization algorithm were the best. Among them, the R value was as high as 0.98g RMSEP was only 0.24, while the wheat moisture content was the best under the BPP-ANN and wavelet transform algorithm, where R was 0.96rMSEP = 0.31.
【学位授予单位】：中国计量学院
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：S512.1;TN219

【相似文献】