Control using Neural network

  • Thread starter Licenio A. Balo Jr.
  • Start date
L

Thread Starter

Licenio A. Balo Jr.

Can anyone pls. tell me the basic procedure or concepts in using neuralnet to control a process?

I am currently having a research on designing a magnetic levitation controller. The controller should be a nonlinear and predictive one. Now, I am thinking of using neural networks. Pls. can anyone tell me at least its basic concepts.
 
Why? The PID control algorithm has lots of life, if you have ever heard of it. Neural networks are basically pattern recognition engines. They "learn" from mistakes as you train them. They are not adaptive or especially non-linear.

The PID control can be adapted with continuous self-tuning, if you expend the effort. The most common control algorithm for your type of application in mechanical systems is fuzzy logic. Try it, you may like it. It's very simple.

Dick Caro
============================================
Richard H. Caro, CEO
CMC Associates
2 Beth Circle, Acton, MA 01720
Tel: +1.978.635.9449 Mobile: +1.978.764.4728
Fax: +1.978.246.1270
E-mail: [email protected]
Web: http://www.CMC.us
============================================
 
P

Pasi Airikka

Hi Licenio!

The neural net alone is not probaply an answer to your pray. Neural nets are a powerful tool for identifying non-linear systems, that is, they are non-linear estimators. However, for controlling your process you need something which relies on negative feedback e.g. fuzzy control or PID control with extensions.

Although fuzzy logic is a quite appealing choice with its capabilities to mimic human reasoning, it is not so obvious to construct and tune fuzzy logic controller. However, it is worth trying.

PID control itself is not appropriate for controlling highly non-linear systems. Therefore, it should be extended with adaptive features such as auto-tuning or self-tuning for non-linear systems. In its simpliest adaptive form, the adaptive PID controller contains the so called "gain scheduling" feature. This means that the non-linear system must be linearized for those operating points where the system is to be controlled. For each linearized point, the PID controller is tuned and its tuning parameters are recorded into the table. Unfortunately, this approach requires additional logics which is used to pick the right tuning parameters from the parameter table for each case (operating point or region) at hand. Still, this approach is also worth trying.

PID control can be easily modified to include the prediction feature. This requires to combine the basic PID controller with a Smith predictor. As a result, the PPID (predictive PID) controller is capable of handling with dead time dominating systems.

You may find plenty of written material on the topics mentioned here from either the text books or the journal papers.

Good luck!

-Pasi A.
 
Top