Classification of System Dynamics Using Neural Networks

Abstract: The thesis is a part of a project which aims at developing highly autonomous controllers. The task of rough classification of system responses is considered. Neutral networks are trained to classify system dynamics from step- and inpulse responses. Two types of networks are discussed, the backpropagation net and Kohonen's self-organizing feature map. The theory behind these algorithms are presented. A method for normalization of the inputs in time and space is given. This is essential for robust classification. Different net configuration, training methods, and noise sensitivity are investigated. It is shown that the networks can be used as classifiers. Rules of thumb for choosing net structure and parameters are given.