Identification of Nonlinear Dynamic Models with Partially Connected Neural Networks Trained Using Orthogonal Least Square Estimation

抄録

A methodology for identifying nonlinear dynamic systems using NARX model structure and three layered feed forward neural networks is presented. The neural network, which maps between regressor vector of the NARX model and the output, is viewed as comprising of number of linear models, each of which is represented by a hidden neuron. An algorithm that uses orthogonal least squares based regressor selection procedure for the estimation of the structure and weights of the hidden neurons is developed for training partially connected neural networks. The use of regressor selection algorithm facilitates implicit identification of unknown dead times and dynamic orders by the neural network during the training process. The conventional error back propagation is used for online adaptation of the identified neural network model. The validity of the proposed approach is demonstrated by applying it to obtain the model of a wastewater pH neutralization process.