Application of Wavelet Packet Transform-Radial Basis Function Neural Network in NIR Spectroscopy for Non-destructive Determination of Cordyceps Gunnii Mycelia Powder

A novel calibration model has been proposed for synchronous, rapid and non-destructive determination the content of adenosine and protein in the Cordyceps gunnii mycelia powder by near infrared spectroscopy. This model is a combination of wavelet packet transform (WPT) data disposal with multi-scale analysis and radial basis function neural network (WPT-RBFNN). Via using principal component analysis (PCA) method for analyzing these reconstructed spectra matrix, the anterior 20 PC scores of principal components (PC) were obtained, which were used as input data in RBFNN. The network parameters including number of input nodes, number of hidden layer neurons and spread constant (SC) were investigated. WPT-RBFNN model which reconstructed the spectra data, removed the noise and the interference of background, and reduced the randomness, the capabilities of prediction is well optimized. The root mean square errors of prediction (RMSEP) for determination of adenosine and protein obtained from the optimum WPT-RBFNN model are 0.5987 and 0.0169, which are superior to those that obtained by the optimum RBFNN models with origin spectra. Regression coefficient (Rp) between NIR predicted values and actual values for adenosine and protein are 0.9385 and 0.9614. It is verified that WPT-RBFNN model with multi-scale analysis is a suitable approach to deal with NIR spectroscopy and model this complex non-linearity. The proposed method is convenient, rapid, no pretreatment and non-destructive.