A Neural Network Approach for Subspace Decomposition and Its Dimension Estimation

In this paper a novel method for subspace decomposition and its dimension estimation based on principle components analysis (PCA) neural network is proposed. This method use an improved Sanger PCA network model which can directly process the array data to obtain its signal subspace and does not involve any estimation of the covariance matrix or its Eigen decomposition. Meanwhile, this method can estimate its dimension with the network outputs by AIC criterion. Computer simulation results demonstrate its effectiveness.