An Intelligent Implementation Aid for

Industrial Process Control Systems

James H. Taylor & Cheney Chan
Department of Electrical & Computer Engineering
University of New Brunswick
Fredericton, NB CANADA E3B 5A3
Internet: [email protected] & [email protected]

Abstract The specic diculty faced by both dcs vendors and

their customers is that many line engineers lack the
We present the plan for an extended expert- knowledge and experience to take full advantage of
system \front end" or Design Advisor for Implementing the advanced capabilities of industrial control systems
Systems (dais) for use in conjunction with a commer- equipment such as the infi 90 system. From a vendor's
cial digital control system environment, e.g., the Elsag perspective, it seems that many customers exploit only
Bailey infi 90 System. The objective of dais is to a small percentage of the algorithms available; from the
make it substantially easier for applications engineers customers' perspective, either they are buying a system
to make eective use of the broad spectrum of capabil- that seems to provide a lot of unnecessary functional-
ities of this and similar hardware and software systems ity, or else there is a frustration that they can't take
for industrial controls implementation. This concept is advantage of functionality that they believe they need
of quite general applicability for industrial controls en- but cannot use eectively.
vironments.
The extensions are primarily focussed on handling Based on these considerations, we decided to create a
multi-input/multi-output (mimo) systems. This should Design Advisor for Implementing Systems (dais), using
greatly enhance the value of dais in modern process an \expert-aided cace" paradigm [3, 4, 5] and a suitable
control applications. cace environment that combines numerical and sym-
bolic evaluations, namely Pang's expert-system frame-
work and software medal [6, 7]. More specically, dais
1 INTRODUCTION was conceived to elicit a denition of the problem (e.g.,
simplied or more detailed form of process model, quali-
One of the main goals of computer-aided control en- tative and/or quantitative performance objectives, con-
gineering (cace) is to facilitate the design and imple- straints), and then recommend a solution or outline a
mentation of control systems for practical applications. decision-making process so that the user could reach
While control-theoretic considerations are important, his/her own conclusions. The \solution" is in the form
they do not provide all the answers needed by eld en- of one or more controller designs, with simulated perfor-
gineers in carrying out this task. This leaves a substan- mance plots and support for implementation using the
tial gap between the capabilities of well-known control- infi 90. A preliminary prototype [8] incorporated rules
theoretic software environments such as matlab [1] and of thumb for determining controller type and simple
matrixx [2] and more practical problems associated tuning rules for parameterizing the compensator (e.g.,
with chosing algorithms, tuning them, and implement- see [9, 10]); in the current phase, we are adding similar
ing systems. We emphasize that there is little or no but extended approaches for multivariable systems (e.g.,
gap between the systems that can be implemented on [11, 12]). In addition, we are working to implement
a modern distributed control system (dcs) such as the a more seamless interface between dais and the infi
infi 901 and those that can be designed using mod- 90 and demonstrating it using real industrial process
ern control theory and packages such as matlab and control equipment (a liquid ow, pressure and tempera-
matrixx { what is missing is support for more down- ture control laboratory test stand built from commercial
to-earth concerns such as those outlined here. components contributed by Honeywell Canada).
1 We would like to acknowledge the generous gift of an infi 90 The prototyping phase of this project is complete [8].
system from Elsag Bailey (Canada) Inc., Burlington, Ontario. The \alpha" version of dais provides the basic func-
tionality outlined above, for single-input /single-output medal [6, 7] is particularly well suited, since it seam-
(siso) systems. The system is being tested at the Uni- lessly integrates a rule-based expert system for analysis
versity in conjunction with an industrial control systems and design heuristics with matlab-like numerical capa-
course, where dais suggests how to design a controller bilities.
using matlab, and, in the associated laboratory, where A high-level description of this idea is outlined be-
students are faced with a realistic industrial problem, low using two approaches: working through a partial
i.e., \here is a process (the above-mentioned Honeywell \scenario" using dais, and then sketching the knowl-
equipment), here are some requirements and specica- edge and decision-making framework of the environ-
tions, here is the Elsag Bailey hardware and software ment. The scenario deals primarily with a typical siso
for implementation; use dais to solve the problem". In problem, for simplicity, with discussion of extensions
addition, a beta test program will be undertaken with to the mimo case representing planned beta-version ca-
industrial partners, in the summer of 1999. At the end pabilities. This paper will then conclude with a brief
of the current phase, we will have rened, built and restatement of the goals and signicance of the project
tested a substantially more functional \beta" version of and future plans.
dais and applied it in the eld.
2.1 Illustrative DAIS Scenario
2 GENERAL CONCEPT dais has been implemented using a simple graphical
The preceding discussion motivates the usefulness of a user interface that gives the applications engineer rea-
Design Advisor for Implementing Systems (dais) sonable exibility in carrying out the activities enumer-
for the Elsag Bailey infi 90 and other digital control ated above. One direct way to proceed in developing
implementation environments of a similar class. The this idea is to display some screens that arise in a sce-
following conceptual outline provides the basis for this nario using such a system:
\front end" environment. dais is designed to support The dais system start-up screen is depicted in Fig. 1.
the industrial applications engineer in the following ar- It provides the primary menu, supporting the six ar-
eas: eas of activity enumerated above, plus a standard set
1. Denition of plant characteristics (description of of options stop, skip, and why, with obvious meaning.
the process To begin work, a user would select a menu item and
to be controlled) proceed.
2. Denition of performance objectives for the con-
trolled process
3. Denition of implementation, operational and
other constraints
4. Selection of control scheme(s)
5. Design / tuning of controller(s)
6. Implementation of the control system
These activities form the core of computer-aided con-
trol engineering (cace), and involve problem-solving
approaches that combine knowledge of both theory and
\heuristics" or experience-based \rules of thumb". In
light of this, it has been observed that articial intelli-
gence (AI) can provide useful contributions in \diagnos-
ing the plant model, setting up a realistic design prob-
lem, selecting appropriate design methods, performing
trade-os, validating the design, implementing the con-
troller, and using conventional cace software" [3]. This Figure 1: dais' Start-up Menus
is because: \Heuristics are certainly a major factor in a
human expert's ability to formulate a well-posed design Normally, a user will begin by executing Step 1 in the
problem." (same citation). From these considerations, process, denition of plant characteristics. Respond-
the use of a rule-based expert system is a natural choice ing to the rst menu with a 1 produces the second
for implementing dais. Pang's expert-system package menu shown in Fig. 1, which supports several exact and
approximate approaches for accomplishing this crucial domain characteristics derived from specications in the
step. At present, dais supports entering a plant model time-domain, using classical heuristics [16] to make this
in transfer function or state-space form, or supplying translation. In brief summary, this technique uses this
less detailed plant characteristics in either the time or ideal closed-loop response specication to determine an
frequency domain, or providing input/output data se- ideal (but generally unrealizable) controller, then de-
quences for model identication. nes an optimization problem to determine the actual
Options 3 and 4 represent the most practically-oriented controller parameters (e.g., pid gains) that approximate
approaches, especially for process industry appplica- the frequency response of the ideal controller, using
tions, as long as the problem is siso. The alpha ver- a frequency-dependent weighting scheme that accounts
sion of dais also supports option 5, the loading of for the sensitivity of the closed-loop system to changes
input /output data sequences for a siso plant. dais in the compensator. This approach is systematic, with
then carries out a simple least-squares model identica- simple heuristics, and is thus ideal for implementation
tion procedure to generate a plant model, of the form in dais using medal's expert system and numerical ca-
K exp(?s )=(1+sT ) or a rst-order lag with time delay.
pabilities.
Option 3 has been illustrated in [8] and not repeated To continue the example from Fig. 1, and assuming that
here. For mimo problems we are planning to incorpo- the user provided the following information:
rate adaptx, a commercially proven package for iden-  parameters for a simple qualitative model are: de-
tication of time-series models for multivariable plants lay time = 6 sec, rise time (after delay) = 2 sec and
[14, 15]. steady-state gain = 10.5;
Elicitation of performance objectives and operational
and other constraints (if any) is accomplished in a sim-  desired closed-loop settling time (after delay) is 5
ilar fashion, working down from Steps 2 and 3 on the sec and tolerable % overshoot is 0%: and
rst menu (Fig. 1) and using a similar menu/screen-  the constraint on the input to the plant is that u
based interface. The user is then ready to progress to should not exceed 0.24 \units" in magnitude,
control scheme selection (Step 4) and design (Step 5).
If the applications engineer has supplied a suitable the Smith Controller is recommended (Block 160 in the
problem denition (plant characterization, performance infi 90) and the parameters are determined by a simple
specication, constraints), then the rule-based system tuning algorithm (see the rst two lines in Fig. 2). Us-
will select a suitable control scheme or set of candi- ing the internal plant model and parameterized Smith
date schemes and support the user in selecting one (if controller, simulated step-response plots are generated;
more than one candidate exists) and completing the de- in this case a warning is raised that the constraint is
sign (e.g., tuning controller parameters). If the problem violated (continuation of Fig. 2) and the step-response
specication is not adequate for controller selection and plots are displayed as in Fig. 3. A recommendation is
design, then dais will provide guidance on how to rec- made that the user relax the settling time specication
tify the situation (see Section 2.2 for further detail). to 7 seconds (bottom of Fig. 2); of course, the user is
Once a control scheme is selected, the design/tuning free to take that advice, or make any other change to
step can be undertaken. The exact nature of this part the problem denition that might remove this violation,
of the process will depend on the detail and quality of or simply decide that the controller is \good enough".
the internal model elicited from the user. For siso plant Once the preliminary controller behavior is displayed,
models dais already supports pid and Smith predictor as in Fig. 3, the user is allowed to tune the controller
control schemes. For mimo processes we have evaluated by perturbing its parameters by either  40% (\coarse
a number of controller synthesis approaches, to nd a tuning") or  10% (\ne tuning"); comparative step-
suitable technique that will t into the dais paradigm. response plots are displayed for the user to select the
Based on industrial applications and preferences, we fo- nal design. Finally, the beta version of dais will auto-
cussed on frequency-domain methods that can be used matically generate the infi 90 conguration le for the
to synthesize simple controllers. In addition, such a control system.
technique should be applicable to simple mimo process
models, including those with time delay. The best can- 2.2 DAIS Knowledge Framework
didate we found was the multivariable controller syn-
thesis method faster of Engell and Muller [11, 12]. The knowledge and decision-making framework of dais
To implement this approach in dais, we are limiting is organized along the lines suggested in [3]. The basic
the controller type to pid and the performance objec- idea is that there are two foci of attention, called the
tive to be nearly decoupled loops with simple frequency- Problem Frame (pf) and the Solution Frame (sf). The
 or capability can be represented in the sf by a \tem-
plate" dening the pf information needed (or slots that
must be lled) to permit its application. The more ba-
sic control algorithms such as pid may have a few basic
information slots specied in their templates; more ad-
vanced schemes or functions may have more extensive
pf data requirements (templates). Each template, then,
would be comprised of a set of slot labels specifying the
pf information that must be provided for a successful
application of the corresponding function or capability.
The decision-making in dais thus involves several com-
ponents:
1. checking that the data in the pf is consistent and
well-posed, e.g., seeing if the performance require-
Figure 2: Screen with Controller Recommendation ments are sensible given the plant characteristics
specied, seeing that the constraints and perfor-
mance requirements are not incompatible, etc., and
2. matching the entries in the pf with the templates
specied in the sf { if a template is satised by the
user's information in the pf, then the correspond-
ing function or capability is said to be \available".
In the alpha version of dais these features are quite
limited. However, the beta system will not stop with a
simplistic `yes' or `no' assessment with respect to avail-
ability; rather it will also check templates that are nearly
satised and advise the user that additional functions
or capabilities would be available if further information
were provided or if the plant characteristics, perfor-
mance requirements and/or constraints were modied
slightly. dais will tell the user how to obtain and sup-
ply additional information or make modications in as
many instances as possible.
To illustrate this, the process of adding to the list of
available controllers will proceed as follows: The expert
Figure 3: Predicted closed-loop Time-Responses system will work through the \unavailable" controllers'
missing-data tags, rst selecting those with one item
of information missing, etc., to determine those that
information in the pf is elicited from the applications appear to be easiest to satisfy. The result of this analysis
engineer, as suggested in Section 2.1, or derived from { a prioritized list of pf slots to try to ll { will be used
information thus supplied. At the present time, the to initiate a dialog with the user to increase the number
contents of the sf are dictated by the functions and ca- of pf slots that are lled, in a systematic manner. It
pabilities of the infi 90 system, and the corresponding will analyze the missing-data tags, see which of them are
problem-denition information needed to apply these most likely to be easily ascertained, and interact with
functions and capabilities eectively. the user to accomplish the job. The outcome of such
The information in the pf is gathered through the an iteration will usually be a request for an additional
straight-forward use of menus, screens and prompts, as piece of information, e.g.,
illustrated. Once the user signals that problem-solving  \If you can state that the plant is stable you will
is to commence (by choosing item 4 on the main menu, be in a position to apply <list of additional control
Fig. 1), this data is processed to see if it is an adequate schemes>."
problem denition, and if so what implementation op-
tions are available. If we think of the information in the  \If you can determine that the plant gain margin
pf as being stored in \slots", then each infi 90 function is innity you will . .. "
Once a control scheme has been selected, the expert
system attempts to evaluate controller parameters to
4 REFERENCES
[1] matlab User's Guide, The MathWorks, Inc., Natick,
achieve satisfaction of the performance criteria and con- MA 01760.
straints. For siso systems, with pid or Smith predictor [2] matrixx User's Guide, Integrated Systems, Inc., Santa
control, this involves execution of known tuning strate- Clara, CA.
gies [8]. For mimo plants the multivariable pid control,
parameters will be synthesized using faster [11, 12]. [3] J. H. Taylor and D. K. Frederick, \An Expert System
Architecture for Computer-Aided Control Engineering"
Finally, the expert system will aid in the validation of (invited), Proceedings of the IEEE, Vol. 72, December
the nal design produced under dais's guidance and in 1984.
moving into implementation and testing. Design valida- [4] J. H. Taylor, \Expert-Aided Environments for CAE of
tion might involve setting up and executing simulation Control Systems", Plenary Lecture, Proc. CADCS '88
studies using a more detailed nonlinear model of the (Fourth IFAC Symposium of CAD of Control Systems),
plant, for example { however, this is beyond the scope Beijing, PR China, 23 August 1988.
of the beta version currently being developed and tested. [5] J. H. Taylor, J. R. James and D. K. Frederick, \Expert-
Presently we are focussing on how dais will aid the user Aided Control Engineering Environment for Nonlinear
in interfacing the infi 90 with the process, setting up Systems", Proc. IFAC World Congress, Vol. 6, pp. 363-
the selected controller, tuning it on line, and activating 368, Munich, FRG, 31 July 1987.
it { beyond automatic generation of infi 90 congura- [6] G. K. H. Pang, \An Intelligent Front End for a Control
tion les, we have to work out the details with Elsag System Design and Analysis Package", Proc. CADCS
Bailey applications engineers. '88 (Fourth IFAC Symposium of CAD of Control Sys-
tems), Beijing, PR China, 23 August 1988.
[7] G. K. H. Pang, \A Matrix and Expert System De-
velopment Aid Language", Proc. CACSD'92, IEEE
Computer-Aided Control Systems Design Conference,
3 SUMMARY AND CONCLU- Napa, CA, pp. 218{224, March 17{19, 1992.
SION [8] J. H. Taylor and P. Seres, \An Intelligent Front End for
Control System Implementation", Proc. IEEE Interna-
tional Symposium on Computer-Aided Control System
Our intention in creating an \intelligent front end" for Design, Dearborn, Michigan, September 1996.
the Elsag Bailey system is to increase the eciency [9] K. Astrom, Ziegler-Nichols Auto-Tuners, Department
and satisfaction of industrial controls engineers in us- of Automatic Control, Lund Institute of Technology,
ing powerful, state-of-the-art distributed control sys- May 1982.
tems (dcss) such as the infi 90. The present eort in [10] K. Astrom and B. Wittenmark, Computer Controlled
developing the \beta" version of dais represents a ma- Systems { Theory and Design, Prentice Hall, 1984.
jor step towards achieving this goal. The overall frame- [11] S. Engell and R. Muller, \Fast and Ecient Selection of
work, functionality and approach are well dened, but Control Structures", Proc. ESCAPE-1, Computational
a few details concerning the dais-dcs interface are still Chemical Engineering, Vol. 16, pp. 157-164, 1992.
somewhat tentative and subject to revision as we ex- [12] S. Engell and R. Muller, \Multivariable Controller De-
periment with realistic industrial applications and work sign by Frequency Response Approximation", Proc.
in collaboration with Elsag Bailey personnel to better ECC, Groningen, June 1993.
understand requirements; these areas will be resolved in [13] D. W. Clarke and P. J. Gawthrop, \Implementation
time for beta test this coming summer. and Application of Microprocessor-Based Self-Tuners",
Automatica, Vol. 17, 233, 1981.
We do believe, based on our progress to date and the [14] W. E. Larimore, \The Optimality of Canonical Variate
reception from Elsag Bailey Canada and a few indus- Identication by Example", Proc. 10th IFAC Sympo-
trial users, that the benets of dais will be substantial. sium on System Identication, Copenhagen, Denmark,
From an industrial point of view, making it easier for July 1994.
an applications engineer to employ a broader range of [15] W. E. Larimore, \Optimal Order Selection and E-
a dcs's capabilities has a clear and signicant pay-o. ciency of Canonical Variate Analysis System Identica-
From an academic standpoint, the research is impor- tion", Proc. 13th IFAC World Congress, San Francisco,
tant in its own right (advancing the state of the art in CA, June 1996.
computer-aided control engineering), and the potential
for using dais in classroom and laboratory settings to [16] J. J. D'Azzo and C. H. Houpis, Feedback Control System
bring industrial approaches and solutions to the fore is Analysis and Synthesis, McGraw-Hill Book Co., New
York, NY, 1960.
an additional attraction.