小波变换近红外光谱结合径向基神经网络快速分析异福片

Application of Wavelet Transform-Radial Basis Function Neural Network in NIRS for Determination of Rifampicin and Isoniazide Tablets

应用小波变换(WT)处理近红外漫反射光谱结合径向基神经网络(RBFNN)建立快速分析异福片中利福平和异烟肼含量的模型(WT-RBFNN)。用小波变换的低频系数作为RBFNN的输入节点,研究了网络参数包括隐含层神经元数和径向基宽度(SC)对模型的影响。与经典的RBFNN和PLS相比较表明,WT-RBFNN模型压缩了原始光谱,除去了噪音和背景的影响,拟合效果很好。优选的WT-RBFNN模型对校正集样品中利福平、异烟肼的交互验证均方根误差(RMSECV)分别为0.006 04和0.004 57;对预测集样品预测均方根误差(RMSEP)分别为0.006 39和0.005 87。同时预测集样品中利福平和异烟肼与RP-HPLC测定结果的回归系数(r)分别为0.995 22和0.993 92,相对误差在2.300%以下。这些结果显示了该方法建模的稳健性和模型的预测精度均很高,同时此方法具有非破坏、无污染、可在线检测等优点,对替代常规药物分析方法有重要的意义。

A calibration model （WT-RBFNN） combination of wavelet transform （WT） and radial basis function neural network （RBFNN） was proposed for synchronous and rapid determination of rifampicin and isoniazide in Rifampicin and Isoniazide tablets by near infrared reflectance spectroscopy （NIRS）. The approximation coefficients were used for input data in RBFNN. The network parameters including the number of hidden layer neurons and spread constant （SC） were investigated. WT-RBFNN model which compressed the original spectra data, removed the noise and the interference of background, and reduced the randomness, the capabilities of prediction were well optimized. The root mean square errors of prediction （RMSEP） for the determination of rifampicin and isoniazide obtained from the optimum WT-RBFNN model are 0. 006 39 and 0. 005 87, and the root mean square errors of cross-calibration （RMSECV） for them are 0. 006 04 and 0. 004 57, respectively which are superior to those obtained by the optimum RBFNN and PLS models. Regression coefficient （R） between NIRS predicted values and RP-HPLC values for rifampicin and isoniazide are 0. 995 22 and 0. 993 92, respectively and the relative error is lower than 2. 300%. It was verified that WT-RBFNN model is a suitable approach to dealing with NIRS. The proposed WT-RBFNN model is convenient, and rapid and with no pollution for the determination of Rifampicin and Isoniazide tablets.