Doctoral Dissertations

Junghui Chen

Date of Award

5-1995

Degree Type

Dissertation

Degree Name

Doctor of Philosophy

Major

Chemical Engineering

Major Professor

Duane D. Bruns

Committee Members

Charles F. Moore, Tse-Wei Wang, J. Douglas Birdwell, David A. Schoenwald

Abstract

The WaveARX network, a new artificially neural network, is proposed along with a systematic design synthesis procedure. Its development is motivated by the opportunity to capitalize on recent research results that allow some shortcomings of the traditional artificial neural net (ANN) to be addressed. ANN has been shown to be a valuable tool for process identification but suffers from slow convergence and long training time due to the global activation function. The structure of ANN is derived from trial and error procedures and the trained network parameters are often strongly dependent on the random selection of the initial values. The design procedures or guidelines on how to select the number of neurons are poorly formulated and have little theoretical basis. Also, few identification techniques are available for distinguishing linear from nonlinear contributions to a process behavior. In order to overcome these drawbacks, WaveARX integrates the wavelet function and the traditional AutoRegressive eXternal input model (ARX) into a three layer feedfor ward network. In order to reduce computation when the wavelet function is extended to multidimensional space, the norm is incorporated into the wavelet function. The WaveARX formalisms provide a systematic design synthesis for the network architec-ture, training procedures and good initial values for the network parameters. The new structure also isolates and quantifies the linear and nonlinear components of the training data sets. These characteristics are demonstrated through several simulation examples and compared with some widely used linear and nonlinear identification techniques. The simulation results also suggest that the WaveARX network with less neurons and training time can have a better or, at least similar, approximation ability as other techniques. The proposed wavelet network is compared with two existing wavelet-based neural networks using literature example. The possible advantages of the proposed wavelet networks are illustrated. The WaveARX network is modified and extended to an on-line identification method called the adaptive WaveARX network. With a rectangular data window that is moved at each sampling time, multiresolution and the classical Gram-Schmidt algorithm are used to determine when and where a new neuron should be generated and/or an old one should be removed. By implementing these training procedures, the shape of the network architecture is not fixed. With the flexible structure, this adaptive and recursive technique can adjust the number and location of neurons, as well as, the parameters according to the changes of the system behavior. Four simulation cases are presented, including a static nonlinear process, a chaotic system, a nonlinear time series with noise and a pH CSTR model. All of them lead to favorable results. Comparisons with other neural network algorithms are also discussed.

Download
Files over 3MB may be slow to open. For best results, right-click and select "save as..."

Share

COinS