Novel Method of Generating Feature Set Based on the Neural Network and Principal Component Analysis

The feature set, which can make the classification linear and simple, has important meaning to pattern classification. But it is always difficult to generate such feature set because of the limitation and fixity of the calculation methods of feature set. A novel method generating above-mentioned feature set is proposed, which is based on the neural network and principal component analysis(PCA). In this method, the neural network is used to carry out nonlinear mapping of the existing feature set, thereby generating a new feature set, and then PCA is used to reduce the dimension of the new feature set. On first principal component direction, if the information reservation rate is greater than 85%, the performance of linear and simple classification of the projection data will be evaluated. Then the evaluation method is designed, and a training approach of neural network using a novel GA algorithm whose crossover operation is asymmetric is proposed. The numerical simulation and experiments show that the performance of the novel method is stable, the classification is accurate and the generalization capability is remarkable.