RE: BISC_SIG_ES, Smart Controller, Open Discussion

masoud nikravesh (nikraveshl@msn.com)
Tue, 17 Jun 1997 15:10:20 +0200


Dear Biscers;

Now the follwoing subjects are open to discussion.

****Smart Controllers
****Neuro-Fuzzy and Neuro-Geometric Control Strategy
****Design of Smart Controllers for Petroleum Industries

BISC_SIG_ES is accepting Abstract (Max. 1 Page), Extended Abstract (Max. 2
Pages) and Short Paper (Max. 5 Pages).
Also, If you would like to be considerd as part of our multi-objectives
proposal and project, please send an extra page (Max. 2 pages) describing
type of contribution to BISC_SIG_ES Smart Controller Project.

Your comments and suggestions are appreciated. Please feel free to contact
us.

Please use " BISC_SIG_ES Smart Controller" as the subject of the mail, and
include the
following information in the body if possible:

Name:
Title:
Affiliation:
Field of Interest:
Mailing address:
Phone:
Fax:
Preferred E-mail address:
Http address:

There is a homepage which you could find out more information about the BISC.

http://bisc.lbl.gov/ , http://128.3.20.27/ , http://godunov.lbl.gov/ ,
http://128.3.21.156/

Regards
Masoud Nikravesh, Ph.D.
Chairman, BISC Special Interest Group In Earth Sciences
BISC is an acronym for the Berkeley Initiative in Soft Computing.

Earth Sciences Division, MS 90-1116, Lawrence Berkeley National Laboratory
and Electrical Engineering and Computer Science Department, BISC Program, Soda
Hall,
University of California at Berkeley
Berkeley, CA 94720

Email: MNikravesh@lbl.gov
Email: nikraves@cs.berkeley.edu
Email: nikraves@eecs.berkeley.edu
Email: nikraves@uclink2.berkeley.edu
Email: nikraveshl@msn.com

http://bisc.lbl.gov/
http://godunov.lbl.gov/
http://128.3.20.27/
http://128.3.21.156/
http://core.lbl.gov/
http://128.3.15.116/

Tel: (510) 486-7728
Fax: (510) 486-5686
Fax: (510) 642-3805

=========================================

Controller Design

Process Model: The process model has been developed based on historical data.

In this study, a neural network model of the form

y(k+1)= h { u(k-n) u(k-n+1), ..., u(k), y(k-m), y(k-m+1), ..., y(k)}

is used.

Here y is controlled variable, u is the manipulated variable, and h{., ..., .}
is a smooth nonlinear function. The preceding model can easily be written in
state-space [1] form by setting

x(k)= {u(k-n) u(k-n+1), ..., u(k-1), y(k-m), y(k-m+1), ..., y(k)} (2)
u(k)=u(k)
y(k)=y(k)

The model of (1) can be rewritten in the following form:

x(k+1)=f{x(k)}+g{x(k),u(k)},
y(k) = x[m+n+1] (k) (3)

with,

f{x(k}= [x2(k) x3(k) ... x[n](k) 0 x[n+2](k) ... x[m+n+1](k) 0 ]T

g{x(k), u(k)}= [0 0 ... 0 0 u(k) 0 0 ... 0 0 h{x(k), u(k)} ]T (4-a)

or,

x1(k+1)=x2(k)
.
.
.
x[n](k+1)=u(k)
x[n+1](k+1)=x[n+1](k)
.
.
. (4-b)
x[n+m](k+1)=x[n+m+1](k);
x[n+m+1](k+1)=h{x(k),u(k)};
y(k)=x[m+n+1](k);

Once the process has been transformed into the state space form of Equation
(3) can be used to design and analyze the controller performance.

Nonlinear Controller: A neural network model can be used for controller
synthesis directly or indirectly. In this study, the indirect method is used.
In this case, the inverse of the process model at each sampling time is
calculated numerically. We consider an objective function of the form:

E(k)=[r(k)-h{., ..., .}]2 (5)

where r(k) is a reference value.

Based on the preceedingstate-space model, we obtain the following
discrete-time nonlinear controller:

r(k+1)=alpha [r(k)-y(k)] + h{x(k),u(k)}, r(0)=y(0); (6)

h{x(k),u(k)}=alpha y(k) + (1-alpha) [ e(k)-r(k)];

where 'alpha' is an adjustable parameter such that 0 < alpha< 1 and ym is the
measured value for controlled variable. The smaller the parameter 'alpha',
the faster the closed-loop response. the error e(k)=yset(k)-y(k), where yset
is the controlled variable setpoint. The controller has integral action and
calculates u(k) based on x(k) and yset(k). Since the equation,

h{x(k),u(k)}=alpha y(k) + (1-alpha) [ e(k)-r(k)] (7)

is nonlinear in u, then at each time instant a nonlinear algebraic equation
has to be solved to obtain the value of u(k). The Newton-Raphson method is
used for the numerical inversion. Several authors have employed this
technique. The following equations are used in this method at each iteration j
to calculated the controller action.

u(k)[j]=u(k)[j-1] - E(k)[j-1] / { del(E(k)[j-1])/ del(u(k)[j-1])} (8)

with the initial guess for u(k) as follows,

u(k)[0]=u(k-1) (9)

. In the absence of constraints, the nonlinear controller induces the
following nominal linear input-output closed-loop response:

y(k+1) - alpha y(k) = (1-alpha) yset(k). (10)

Masoud Nikravesh & Masoud Soroush

=========================================

BISC_SIG_ES Smart Controller Proposal and Project

1. Project Title:

Neuro-Fuzzy and Neuro-Geometric Control Strategy for Design of Smart
Controllers

2. Basis for Collaboration

The basis for this collaboration is: to enhance, apply, and transfer
technologies developed within the National Laboratories, Universities, and
Industrial Partners to promote rapid development of new technologies in
modeling and control of complex processes; and to increase industry's ability
at lower cost and with reduced environmental risk.

3. Project Abstract

We present the next generation of "smart" controllers based on
neural-network/fuzzy-logic models and differential-geometric/model-predictive
control methods. The model helps to improve management and the design of
optimal control strategies. In particular, the developed neural network and
fuzzy logic models can be used to control and predict the behavior of complex
systems

4. Reasons for Cooperation

When a change in the process parameters occurs, the controller needs to be
retuned to maintain the conventional controller such as PID controller
performance at a satisfactory level. Retuning the PID controller for nonlinear
and/or non-stationary processes is usually time consuming and requires a
combination of operational experience and trial and-error. In addition,
compared to pilot-scale, industrial application have a more stochastic nature
and their residence-time distributions are often very complex. These features
have made the development of first-principle models for complex industrial
processes very challenging, if not impossible.
Neural networks, however, have been shown to be able to provide
sufficiently-accurate models of these processes using historical process data.
Furthermore, for many of these complex processes, little is known about the
complex dynamics governing these systems, which makes the problem more
challenging.
Despite our incomplete knowledge, neural network models are able to predict
the complex behavior of the nonlinear and nonstationary processes.
Furthermore, the Neuro fuzzy techniques can be used to design the model-based
controllers capable of providing an effective process control.
In this project, we are creating an innovative Neural-Network/Fuzzy-Logic
Model Based Control Strategy (Neuro-Fuzzy Control; NFC) for developing and
maintaining optimal control policy. First, the NFC will continuously acquire
data from the processes. Second, a neural network model is used to "learn" the
complex process and to recognize symptoms of efficient and inefficient control
technique around the operating point of the process. However, since we update
the neural network models continuously around the process operating points,
the suggested neural network models will be accurate over the entire region of
operating condition. Third, feedback from neural network models in
conjunction with differential-geometric control is used for design of a
nonlinear model-based control strategy. The optimal parameters of the
controller can be used for retuning the existing PID controller to achieve a
better control performance.
Fourth, a fuzzy logic model will be used for assisting the existing PID and
the neuro-geometric controller for optimal performance under process
uncertainty. The intrinsic structure of fuzzy models allows the wealth of
operator experience available to be easily integrated into the system. The
fuzzy logic enhanced PID system will also be constructed such that it
considers the tightly coupled nature of the process controller network and
adjusts the controller for all local processes simultaneously.
In addition, for these processes, we will demonstrate that neural networks in
conjunction with a recursive least squares technique can be used effectively
for model identification. It can also been shown that if any of the process
parameters change over time, the network and the controller has the ability to
adapt itself to the new situation based on the new information. The model can
also predict the time at which any of these parameters change occurs, by
examining the weights and bias terms at each sampling time. The final goal of
this project is to bring the advance control technology (i.e. neural-network
differential geometric control and fuzzy-logic technology), to the point of
commercialization.

5. Project Goal

National Laboratories, Universities, and Industrial Partners will work
together to develop a new advanced technology. The work will be carried out
over a three-year period with the objective of bringing advanced control
technologies (i.e neural-network model-based control and fuzzy-logic
technology), to the point of commercialization. In this time period, model and
software development and testing will be completed on specific projects. In
addition, training and education will be provided for the engineers using the
new technology.

Copy Right(C), BISC_SIG_ES