Return to Home Page

Neural Network Applications

Neural Network Applications

Gary M. Scott, Ph.D.


Faculty of Paper and Bioprocess Engineering
Empire State Paper Research Institute

State University of New York, College of Environmental Science and Forestry

gscott@esf.edu

Summary

Artificial Neural Networks (ANNs) are a rapidly growing facet of artificial intelligence, using a collection of simple processing units that are massively interconnected in order to produce meaningful behavior. We applied ANNs to the task of process modelling of nonlinear chemical systems. The MANNIDENT (multivariable artificial neural network identification) algorithm was developed as a method of using a traditional modelling paradigm to determine the topology and initial weights of a network. The use of linear models in this way eliminates network design problems, such as the choice of network topology (e.g., the number of hidden units) and reduces the sensitivity of the network to the initial values of the weights, which are traditionally set to random values. Furthermore, the initial configuration of the network is closer to its final state that it would normally be in a randomly configured network. Thus, the MANNIDENT networks perform better and more consistently than the standard, randomly initialized three-layer approach. This work presented a paradigm for including previously known knowledge into a neural network model of the process. The benefits of such a technique include a better understanding of the final model (the model is less of a ``black box'' and more amendable to interpretation), ease in network design (the topology and starting weight values are determined by the initial linear model), and quicker training of the model (since the starting point is closer to the final model).

Highly complicated and nonlinear processes can be modelled through the use of artificial neural networks. Building on the MANNIDENT system of modelling, We developed the MANNCON (Multivariable Artificial Neural Network Control) algorithms for incorporating knowledge-based ANNs into traditional model-based controller paradigms for the control of nonlinear processes. Many processes involve nonlinear relationships, which can be handled by the nonlinear nature of the ANNs. Also, using in place of linear models in these controllers improves the performance of the controller over a wide range of conditions. To achieve these goals, We produced the algorithms for using the ANNs in such model-based controllers as Internal Model Control and Direct Synthesis Control. The ANN-based controller was incorporated as a supervisory controller to a traditional PID (Proportional-Integral-Derivative) controller, resulting in an adaptable controller that capitalized on the robustness of a PID controller. We demonstrated the effectiveness of these methods using a highly nonlinear process as an example and outlined the steps necessary for implementation of these controllers industrially. Technology transfer efforts have included the implementation of computer software incorporating these process control models.

Co-Workers

Dr. W. Harmon Ray
Professor
Department of Chemical Engineering
University of Wisconsin
Madison, WI
   
Dr. Jude W. Shavlik
Professor
Department of Computer Sciences
University of Wisconsin
Madison, WI

Key Publications

  1. Scott, Gary M.; Shavlik, Jude W.; and Ray, W. Harmon (1992). "Refining PID Controllers using Neural Networks." In J.E. Moody, S.J. Hanson, and R.P. Lippmann (Eds.), Advances in Neural Information Processing Systems 4, San Mateo, CA: Morgan Kaufmann Publishers. pp. 555-562.
  2. Scott, Gary M.; Shavlik, Jude W.; and Ray, W. Harmon (1992). "Refining PID Controllers using Neural Networks." Neural Computation, 4(5), pp. 746-757. (PDF)
  3. Scott, Gary M. and Ray, W. Harmon (1993). "Creating Efficient Nonlinear Neural Network Process Models That Allow Model Interpretation." Journal of Process Control, 3(3), pp. 163-178. (PDF)
  4. Scott, Gary M. and Ray, W. Harmon (1993). "Experiences with Model-Based Controllers Based on Neural Network Process Models." Journal of Process Control, 3(3), pp. 179-196. (PDF)
  5. Scott, Gary M. and Ray, W. Harmon (1994). "Neural Network Process Models Based on Linear Model Structures," Neural Computation, 6 (4), pp. 718-738. (PDF)

Copyright 2001, Gary M. Scott. All rights reserved.