【作者】 张益波；

【导师】 滕利荣； 孟庆繁； 逯家辉；

【作者基本信息】 吉林大学 ， 微生物与生化药学， 2008， 硕士

【摘要】 本文完成了应用近红外光谱结合径向基神经网络模型,对三种药用真菌云芝、虫草和松茸的活性成分含量建模分析。经过光谱的预处理和径向基神经网络模型输入节点,隐含节点和径向基宽度(SC)等参数的优化,得到了最佳的三种药用真菌活性成分分析的近红外光谱径向基模型分别为:云芝多糖含量分析的最优径向基神经网络模型是6尺度分解的WPT-NIRS-RBFNN(7-12-1,3.2)、云芝蛋白含量分析的最优径向基神经网络模型是6尺度分解的WPT-NIRS-RBFNN (12-10-1,3.0);虫草腺苷的最优分析模型为小波包分解的5尺度WPT-NIRS-RBFNN(6-6-1,3),此时模型的RMSECV为0.002541,模型的校正集相关系数为0.901162、虫草蛋白含量分析的最优的模型为小波包分解的4尺度WPT-NIRS-RBFNN(9-10-1,3.2),此时模型的RMSECV为0.013678,模型的校正集相关系数为0.951925;松茸蛋白的最优分析模型为径向基神经网络模型是5尺度分解WPT-NIRS-RBFNN(10-18-1,3.0),此时模型的RMSECV为0.00543,模型的校正集相关系数为0.99335、松茸多糖含量分析的最优的模型为径向基神经网络模型是5尺度分解WPT-NIRS-RBFNN(12-10-1,2.6),此时模型的RMSECV为0.00825,模型的校正集相关系数为0.99055。同时各个模型的预测性能很好,满足快速无损分析的要求。更多还原

【Abstract】 Along with the modern science technology’s swift development, the analysis work’s meter mechanized, the automation to obtain the unprecedented development, the request drug analysis toward the green, fast, highly effective and so on directions has also developed.Near infrared spectroscopy technology (NIRS) is a novel technology applying to rapid determination of single component or multicomponents in materials depend on their optical characteristic in near infrared spectroscopy region. The equipment now analysts can rapidly and precisely acquire a large amount of measured data. As the NIRS is very complicate and serious overlap, it is difficult to quantitative analyze multi-components sample depend on the height of peaks in spectra, however, the analysts have to face this problem: how to extract more meaningful chemical information from the measured data, it is necessary to apply a chemometric to develop a multiple regression quantitative model using for analysis and prediction.The radial basis function neural network (RBFNN) has been developed for quantitative analysis of drugs during the last decade. Because of its powerful nonlinear projection ability and local approximation, it is very suitable for the solution of the inherent relationship from mass. The topologic structure of a typical RBF networks is characterized by input nodes number, spread constant and hidden layer neurons number.Wavelet Transform (WT) and Wavelet Packet Transform (WPT) have been proved to be a powerful tool for compressing analytical data. It transforms the raw measured data into the wavelet domain. The information contained in the raw data can be represented by the wavelet coefficients. Because of the WT property, there are many wavelet coefficients with very small amplitude, which can be regarded as uninformative; these can be removed without substantially affecting the useful information. WPT is the generalization of WT, it can decompose the data with approximation coefficient and detail coefficient with different scales for a more precise signal extraction。In this paper, I have studied on the application of NIRS combination with RBFNN on the active constituent of Coriolus versicolor, Cordyceps sinensis and Tricholoma matsutake for non-destruction determination. And then I studied on the influence of pretreatment methods of NIRS spectra, such as Savitzky-Golay smoothing, Fast Fourier Transform (FFT), derivative, and Wavelet Transform (WT) and Wavelet Packet Transform (WPT) on establishing quantitative analysis model, discussed the optimal method and opportune factors influences in the models.1. The optimal quantitative analysis model for Polysaccharide and Protein of Coriolus versicolor: for Polysaccharide the optimal model is 6 scales decompose of WPT, WPT-NIRS-RBFNN(7-12-1,3.2),and RMSECV is 0.009897, Rc is 0.98357; RMSEP is 0.00909;Rp is 0.98283。For Protein the optimal model is 6 scales decompose of WPT, WPT-NIRS-RBFNN(12-10-1,3.0),and RMSECV is 0.00524, Rc is 0.99426; RMSEP is 0.00998;Rp is 0.98246。2. The optimal quantitative analysis model for Adenosine and Protein of Coriolus versicolor: for Adenosine the optimal model is 5 scales decompose of WPT, WPT-NIRS-RBFNN(6-6-1,3.0),and RMSECV is 0.002541, Rc is 0.901162; RMSEP is 0.001969;Rp is 0.950076。For Protein the optimal model is 4 scales decompose of WPT, WPT-NIRS-RBFNN(9-10-1,3.2),and RMSECV is 0.013678, Rc is 0.951925; RMSEP is 0.017243;Rp is 0.950056。3. The optimal quantitative analysis model for Protein and Polysaccharide of Tricholoma matsutake: for Protein the optimal model is 5 scales decompose of WPT, WPT-NIRS-RBFNN(10-18-1,3.2),and RMSECV is 0.00543, Rc is 0.99335; RMSEP is 0.00943;Rp is 0.98292。For Polysaccharide the optimal model is 5 scales decompose of WPT, WPT-NIRS-RBFNN(12-10-1,2.6),and RMSECV is 0.00825, Rc is 0.99055; RMSEP is 0.00956;Rp is 0.98638。Each ingredient’s optimization model’s reproduction quality and the returns-ratio experimental result further indicated that this method is good to for medicinal purposes fungus active constituent’s modelling determination reproduction quality, the returns-ratio is high. In the experiment simultaneously had also indicated the WPT in spectrum pretreatment method has under the suitable scale compared to a convention processing method more superior effect. The results indicate that it is feasible to apply NIRS combine with RBFNN to non-destruction determination of the active constituent of Coriolus versicolor, Cordyceps sinensis and Tricholoma matsutake. It can be generalized to on-line and real-time quality control in pharmacy.更多还原