【作者】 夏俊芳；

【导师】 李小昱； 李培武；

【作者基本信息】 华中农业大学 ， 农产品加工与贮藏工程， 2008， 博士

【摘要】 为了提出脐橙内部品质近红外光谱快速无损检测方法,预测和控制脐橙可溶性总糖、总酸、维生素C、可溶性固形物、糖酸比等内部品质贮藏特性,收集了420个脐橙样品,比较并优选了11种常用光谱消噪预处理方法,确定了小波消噪的最优分解尺度,提出了近红外光谱消噪的最优小波基,建立了贮藏脐橙内部品质近红外光谱无损检测定量模型,建立了脐橙内部品质贮藏特性的BP神经网络模型。主要研究结果如下:1.可溶性总糖含量近红外光谱常规最佳预处理方法是直线差值(SLS)法;多品种PLS校正模型和验证模型预测值的相关系数R分别为0.9487和0.877,内部交叉验证均方差RMSECV分别为0.776%和0.6992%;单品种PLS校正模型和验证模型预测值R分别为0.961和0.9626,RMSECV分别为0.767%和0.7769%。总酸含量近红外光谱常规最佳预处理方法是附加散射校正法(MSC);多品种PLS校正模型和验证模型预测值R分别为0.9268和0.894,RMSECV分别为0.0355%和0.0407%;单品种PLS校正模型和验证模型预测值R分别为0.9663和0.9813,RMSECV分别为0.0328%和0.01705%。维生素C含量近红外光谱常规最佳预处理方法是一阶导数+向量归一化(FD+VN);多品种PLS校正模型和验证模型预测值R分别为0.9306和0.8689,RMSECV分别为5.07mg/100g和3.888mg/100g;单品种PLS校正模型和验证模型预测值R分别为0.9392和0.9717,RMSECV分别为2.02mg/100g和1.8356mg/100g。可溶性固形物含量近红外光谱无损检测的常规光谱最佳预处理方法是一阶导数(FD);多品种PLS校正模型和验证模型预测值R分别为0.9654和0.8953,RMSECV分别为0.316%和0.4262%;单品种PLS校正模型和验证模型预测值R分别为0.9737和0.94,RMSECV分别为0.282%和0.36%。2.可溶性总糖近红外光谱小波消噪最佳分解尺度为6,PLS模型预测值R达到0.9231,RMSECV为0.672%。总酸度近红外光谱小波消噪最佳分解尺度为3 PLS模型预测值R为0.9371,RMSECV为0.0334%。维生素C近红外光谱小波消噪最佳分解尺度为3,PLS模型预测值R为0.9632,RMSECV为2.78mg/100g。可溶性固形物含量近红外光谱小波消噪最佳分解尺度为5,PLS模型预测值R为0.9791,RMSECV为0.292%。3.小波包变换是脐橙可溶性总糖、总酸、维生素C、可溶性固形物近红外光谱消噪的有效方法。可溶性总糖近红外光谱噪声效果最好的小波基是db6,其PLS模型预测值R为0.9431、RMSECV为0.373%。总酸近红外光谱消噪效果最好的小波基是db4,其PLS模型预测值R为0.9507、RMSECV为0.0336%。维生素C近红外光谱消噪效果最好的小波基是db5,其PLS模型预测值R为0.9427、RMSECV为2.02mg/100g。可溶性固形物近红外光谱消噪效果最好的小波基是db5,其PLS模型预测值R为0.968、RMSECV为0.344%。4.内部品质与贮藏时间的BP人工神经网络模型中,可溶性总糖模型某些优化隐含层神经元数目为60,贮藏时间校正模型预测值R为0.864,验证模型R为0.88。总酸模型优化隐含层神经元数目为50,贮藏时间校正模型预测值R为0.984,验证模型R为0.9814。维生素C模型优化隐含层神经元数目为50,贮藏时间校正模型预测值R为0.82,验证模型R为0.8648。可溶性固形物模型优化隐含层神经元数为30,贮藏时间校正模型预测值R为0.933,验证模型R为0.9343。糖酸比模型优化隐含层神经元数为60,贮藏时间校正模型预测值R为0.89,验证模型R为0.90。与贮藏时间关联最显著的品质指标是总酸度。5.可溶性总糖、总酸、维生素C、可溶性固形物和糖酸比等5个指标建立的多因素随贮藏时间变化的BP人工神经网络模型优化隐含层神经元数为8,校正模型贮藏时间预测值R为0.98,验证模型R为0.99。多因素模型比单因素模型预测效果精确,应该采用多因素模型预测贮藏时间和贮藏寿命。更多还原

【Abstract】 In order to research rapid&non-destructive measurement method of the quality characteristics about navel oranges using near-infrared spectroscopy, predictive and control quality characteristics of total soluble sugar, total acidity, vitamin C, soluble solids in navel orange while storing, 420 navel orange samples were chosen as a sample collection, the denoising effect of 11 different popular spectroscopy pretreament approaches were compared, turned out the best decomposing levels of wavelet denoising, put forwarded the best means of wavelet of near-infrared spectroscopy denoising, established non-destructive near-infrared spectroscopy quantitative analysis model of quality characteristics about navel oranges in storage, established BP artificial neural network model of quality characteristics about navel oranges. The results are as follows:1. The best conventional pretreatment methods using near-infrared spectroscopy of total soluble sugar content is straight line subtraction (SLS); The predictive values correlation coefficient R of PLS calibration and validation model about different varieties were 0.9487 and 0.877, the root mean square error of cross-validation variance(RMSECV) were 0.776% and 0.6992%; while the R of single variety were 0.961 and 0.9626, RMSECV were 0.767% and 0.7769%; The best conventional pretreatment methods using near-infrared spectroscopy of total acidity content was multiplication scattering correction (MSC), the R of different varieties were 0.9268 and 0.894, RMSECV were 0.0355% and 0.0407%, and the R of single variety were 0.9663 and 0.9813, RMSECV were 0.0328% and 0.01705%; The best conventional pretreatment methods using near-infrared spectroscopy of vitamin C (Vc) content is first derivative+vector normalization (FD+VN), the R of different varieties were 0.9306 and 0.8689, RMSECV were 5.07mg/100g and 3.888mg/100g, and the R of single variety were 0.9392 and 0.9717, RMSECV were 2.02mg/100g and 1.8356mg/100g; The best conventional pretreatment methods using near-infrared spectroscopy of total soluble solids content is first derivative (FD), the R of different varieties were 0.9654 and 0.8952, RMSECV were 0.316% and 0.4262%, and the R of single variety were 0.9737 and 0.94, RMSECV were 0.282% and 0.36%.2. The best decomposing levels using near-infrared spectroscopy wavelet denoising of total soluble sugar was 6, PLS model predictive value R was 0.9231, RMSECV was 0.672%. The best decomposing levels using near-infrared spectroscopy wavelet denoising of total acidity was 3, PLS model predictive value R was 0.9371, RMSECV was 0.0334%. The best decomposing levels using near-infrared spectroscopy wavelet denoising of vitamin C is 3 PLS model predictive value R was 0.9632, RMSECV was 2.78mg/100g. The best decomposing levels using near-infrared spectroscopy wavelet denoising of soluble solids was 5, PLS model predictive value R was 0.9791, RMSECV was 0.292%.3. Wavelet packet transform was an effective method to denoise near-infrared spectroscopy of total soluble sugar, total acidity, vitamin C, soluble solids in navel orange. The best near infrared spectroscopy denoising wavelet of Vitamin C was db5, its R of PLS model predictive value was 0.9427, RMSECV was 2.02 mg/100g. The best near infrared spectroscopy denoising wavelet of soluble solids was db5, its R of PLS model predictive value was 0.968, RMSECV was 0.344%. The best near-infrared spectroscopy denoising wavelet of total soluble sugar was db6, its R of PLS model predictive value was 0.9431, RMSECV was 0.373%. The best near-infrared spectroscopy denoising wavelet of total acidity was db4, its R of PLS model predictive value was 0.9507, RMSECV was 0.0336%.4. In the BP artificial neural network model of quality characteristics and storage time, the hidden layer neurons number of the model for total soluble sugar was 60, the predictive value R of correction model about storage time was 0.864, validation model R was 0.88, The hidden layer neurons number of the model for total acidity was 50, the predictive value R of correction model about storage time was 0.984, validation model R was 0.9814. The hidden layer neurons number of the model for VC was 50, the predictive value R of correction model about storage time was 0.82, validation model R was 0.8648. The hidden layer neurons number of the model for soluble solids contents was 30, the predictive value R of correction model about storage time was 0.933, validation model R was 0.9343. The hidden layer neurons number of the model for sugar-acidity ratio was 60, the predictive value R of correction model about storage time was 0.89, validation model R was 0.90. The most notable quality index related to the storage time was total acidity.5. The hidden layer neurons number which was based on five indexes including total soluble sugar, total acidity, VC, soluble solids, sugar-acid ratio and optimized by the BP artificial neural network model was 8, the model was multi-element and changing with the storage time. Predictive value R about storage time of correction model was 0.98, the same one of validation model was 0.99.The forecasted effect of multi-element model was better than the one of single-element model. It should take the multi-element model to predict the storage time and the storage life.更多还原