Neural networks are black-box model structures that map inputs to outputs and do not require underlying mathematical models between the two. They are frequently used in the field of system identification, the area that deals with the development of system models based on input-output data. In this work, a hybrid system identification method is implemented with neural networks (NN) and the Minimum Model Error estimator (MME) on different benchmark experimental setups, as well as simulations. The MME algorithm uses a cost function with a covariance constraint to determine smooth state estimates of a system given noisy measurement data and an assumed model. As a byproduct, it generates a vector of unmodeled nonlinear (or linear) system dynamics, which can then be modeled by a neural network. Combining this neural network with the assumed model from MME, a system plant model is obtained. The purpose of neural networks in this research is two-fold: to demonstrate the advantages of combined MME/NN models over some common system identification methods and to investigate the feasibility of using the data stored in the network structure of those models to develop a classification scheme for condition monitoring. The approach to classification that is used in this research does not lead to successful implementation of such a scheme.

Library of Congress Subject Headings

System identification; Process control; Nonlinear systems; Neural networks (Computer science)

Publication Date


Document Type


Student Type


Degree Name

Mechanical Engineering (MS)

Department, Program, or Center

Mechanical Engineering (KGCOE)


Jason Kolodziej


Physical copy available from RIT's Wallace Library at QA402 .E34 2015


RIT – Main Campus

Plan Codes