Nolinear multi-component spectroscopy analysis based on evolutionary construction optimazation

Spectroscopy has been widely used to evaluate product quality or to predict components. To deal with the nonlinearity of spectral data, artificial neural networks (ANN) are widely used. One weakness of ANN is we have no accurate analytical method to design a optimal network structure. A multivariate component prediction method based on optimized neural network combined with evolutionary algorithm (EA) for spectral analysis is proposed in the paper. For the proposed method, ANNs are combined with nonlinear adaptive evolutionary programming algorithm (NAEP) to evolve ANNs architecture including the number of hidden nodes and the number of hidden layers. And the root-mean-squares error of cross-validation (RMSECV) is the fitness function of NAEP. In order to present the effectiveness of this method, back propagation neural network (BP) and ANN with genetic algorithm (ANN-GA) methods were also used for component predicting models. An application research has been demonstrated with spectral data which is recorded in an experiment of meat content determination. Results indicate that our method has the ability to design the best ANN structure to predict more accurate and robust as a practical spectral analysis tool.