论文部分内容阅读
将近红外光谱(NIRS)与三层径向基神经网络(RBFNN)结合,建立药用真菌云芝中活性成份多糖和蛋白的快速无损分析模型(NIRS-RBFNN)。采用卷积平滑、傅里叶变换、一阶变换、二阶变换、多尺度小波变换和小波包变换对原始光谱进行预处理。对处理后的光谱进行主成份的提取,以前15个主成份得分作为径向基神经网络的输入节点选择范围。对网络相关的参数(输入节点数、中间神经元数、径向基宽度常数)进行了优选。得到了最佳的云芝多糖分析模型的条件为:小波变换6尺度重构光谱,模型参数为WPT-NIRS-RBFNN(7-12-1,3.2),此时模型的交换验证方根误差(RMSECV)为0.009897,校正集相关系数Rcv=0.98357,此模型对预测集的预测远离方根误差(RMSEP)为0.00909,其相关系数Rp=0.98283;对云芝蛋白的最佳分析模型的条件为:对小波变换6尺度重构光谱,模型参数为WPT-NIRS-RBFNN(12-10-1,3.0),此时模型的RMSECV为0.005240,Rcv=0.99426,此模型对预测集的RMSEP为0.00998,Rp=0.98246。结果表明模型具有很好的稳健性和精确度。对实现药用真菌的无损快速分析有重要的意义。
Near infrared spectroscopy (NIRS) was combined with RBFNN to establish NIRS-RBFNN (Rapid Non-Destructive Analysis Model). The original spectra were pre-processed by convolution smoothing, Fourier transform, first-order transform, second-order transform, multi-scale wavelet transform and wavelet packet transform. The main components of the processed spectra were extracted, and the score of the former 15 principal components was chosen as the input node of RBF neural network. Network-related parameters (number of input nodes, number of intermediate neurons, radial basis width constant) are optimized. The optimal conditions for the polysaccharide analysis were as follows: wavelet transform 6-scale reconstruction spectrum, the model parameters WPT-NIRS-RBFNN (7-12-1,3.2), the root-mean-square error RMSECV) was 0.009897, and the correlation coefficient of calibration set was Rcv = 0.98357. The prediction of the prediction set was far away from root mean square error (RMSEP) of 0.00909 and the correlation coefficient Rp was 0.98283. The optimal analysis model for the samples was: For wavelet transform 6-scale reconstructed spectra, the model parameters are WPT-NIRS-RBFNN (12-10-1, 3.0). The RMSECV of this model is 0.005240 and Rcv is 0.99426. The RMSEP of this model to the prediction set is 0.00998. = 0.98246. The results show that the model has good robustness and accuracy. The realization of non-destructive rapid analysis of medicinal fungi is of great significance.