Artificial intelligence (AI) techniques are alive and well. No, it's
not another suggestion of machines taking over the world, but of
advanced controls better able to complement classical techniques for the
new industrial enterprise that continually seeks lower costs and shorter
product development cycles.
AI technologies include artificial neural networks (ANNs), expert
systems (ESs), fuzzy logic (FL), genetic algorithms (GAs), and others
still to come. Their underlying methods differ from conventional
computing and include concepts that simulate the way humans solve
problems or how processes work in nature. For example, ANNs learn by
training, ESs reason based on rules and experiences of experts, and FL
works with uncertainty and partial truth. They often work in concert,
adding terms like neurofuzzy and neurogenetic to our vocabulary.
These techniques are intended to tackle highly complex, nonlinear
problems beyond the capability of conventional methods, or too costly or
time consuming for them. AI has had many successes, though not always in
the role of an exclusive solution. AI methods are also controversial due
to their nature and have not been fully accepted by some proponents of
conventional approaches. Today, however, AI methods are closer to the
mainstream as an extra tool for the control engineer. Separate sidebars
provide some insight on the various AI techiques. Fuzzy developments
Fuzzy logic is progressing on several different fronts: in software
tools, in controllers, and embedded in chips. Controllers with FL
capability include models from Fuji Electric, Klckner-Moeller,
Rockwell-Automation/Allen-Bradley, Siemens, and Yokogawa, among others.
Controller manufacturers teaming up with software suppliers is another
route for expansion.
As applications widen, “the challenge becomes to make FL widely
available on many hardware and software platforms,” says Constantin von
Altrock, director of Fuzzy Logic Technologies Div., Inform GmbH of
Aachen, Germany (Inform Sofware Corp., is the firm’s U.S. entity in a
suburb of Chicago).
In some ways fuzzy logic is already an “accepted” technology. The
International Electrotechnical Commission standard IEC 1131,
Programmable Controllers, has a proposed extension on Fuzzy Logic
Programming. Now at voting stage, IEC 61131-7 CD1 (committee draft 1.0,
Dec. 6, 1996) is intended to be Part 7 of IEC 1131. Mr. von Altrock is
confident that it will be in effect before 1998.
Inform’s fuzzyTech FL design and user interface software is already
based on IEC 1131-7. Object-oriented fuzzyTech integrates fuzzy logic
functions with standard PLCs (A-B, Modicon, and Siemens) and with
popular process control software, such as Citect (Ci Technologies), the
FIX (Intellution), and InTouch (Wonderware). The latest version of
fuzzyTech (5.0) is optimized for Windows NT. It was shown at the recent
Wonderware and Inform have also developed an ActiveX extension for
InControl, based on the fuzzyTech package. Shown in preliminary version
at ISA ’96 in Chicago, this FL ActiveX product is due to ship this
As chipmakers increase integration of FL at the silicon level (e.g.,
Motorola’s 68HC12 microcontroller, see CE, July 1996, pp. 47-50), design
of FL systems will benefit. “Fuzzy computation speeds can become as
efficient as in traditional logic,” says Mr. von Altrock, “making
possible control loops as fast as fractions of a millisecond.”
Neural networks in control
Among trends noted by NeuralWare Inc. (Pittsburgh, Pa.) is the move
toward hybrid systems that combine neural network technology with other
methods (not necessarily of the AI variety). Casey Klimasauskas, cto and
director of NeuralWare’s Advanced Technology Group, mentions partial
least squares (PLS, see first sidebar) and principal component analysis
(PCA) as valuable add-ons to neural networks. The resulting hybrid ANNs
are effective in system modeling and batch monitoring.
Batch processes, with strong nonlinearities and complex control, are
especially attractive to these methods. “A variety of approaches looks
at how statistical techniques, such as PLS and PCA, can be combined with
neural networks to do run-to-run control of batch processes,” says Mr.
An evolving application for ANNs is to enhance plantwide
first-principle models (FPMs). It includes two areas:
- Modeling specific nonlinear processes for which FPMs don’t exist
or are not very accurate, or would be very costly to develop
- Providing a correction term for the FPM
NeuralWare attributes improved plantwide optimization and other
incremental benefits to FPM enhancement through neural nets.
Sensor and data validation represent another new solution area for
ANN technology. This is important especially for model-predictive
control. How does it work? The idea is to use redundancies in measured
data to identify correlations among key variables. “When the structure
of these correlations changes, it can indicate a potential problem,”
explains Mr. Klimasauskas.
Expert system views
Gensym Corp. (Cambridge, Mass.) is active in several AI sectors,
among them expert systems. These methods are incorporated in its main
product, called G2, a graphical object-oriented package “for creating
intelligent systems.” Newly introduced version 5.0 of G2 makes major
upgrades in user interfaces and connectivity to host systems. It’s good
news especially for expanding expert system capabilities. For example,
new “rule triggers” that speed up response to events in real time are
part of the G2 enhancement.
Robert Moore, president of Gensym, stresses the vital need to
properly present knowledge of the process and plant to an operator from
the mass of information available. “An expert system must bring
time-critical decision support to an operator clearly and rapidly.
Enhanced graphical functions make the management of faults and abnormal
situations more intuitive,” he remarks.
Enhancements to G2 include Internet-related products (ActiveX, Java,
etc.) and are scheduled for second-half 1997 introduction.
Indicating real growth for industrial uses of AI, an April ’97
meeting of the Gensym Users’ Society in Paris drew over 150 worldwide
attendees and presented 55 papers. Applications ranged from gas
production and pipeline safety to batch management and scheduling, with
a sprinkling of topics in industrial machinery, power plants, and
Comdale Technologies (Canada) Inc. (Toronto, Ontario) looks at
intelligent system development, including ESs, in a streamlined way.
Wayne Thompson, vp of marketing, thinks a lower cost “component expert
system” can be easily embedded into any system, such as a SCADA system
or DCS. “Through object-oriented technology, OLE, and ActiveX, it can be
deployed without substantial AI experience, in a fraction of the time
[compared to older ESs],” he says. Embedding is the key element; it
allows use of existing control software and host system.
A product example in this category is the Comdale Expert Optimizer.
It operates in either advisory or control mode, and contains a FL
element for determining the magnitude of control actions needed based on
the user’s confidence about the process variables.
“Data glut” affecting system operators and users is another concern
at Comdale. One solution ap-plies alarm detection and management within
an expert system. Such functions are part of Comdale Smart Alarm, where
an embedded inference en-gine filters, changes priority, or suppresses
repetitive and nuisance alarms by looking at current and historical
data. A still higher level of alarm management called “meta-alarms” (or
alarms about alarms) promises to further reduce unnecessary information
flow. It provides advice to the operator about root causes of alarms and
recommends actions to take even in a process upset. Now in beta test, a
demo of this unnamed product is slated for the October Anaheim, Calif.,
ISA show, says Mr. Thompson.
From the field
User experiences with AI remain quite closely guarded. A recent
success for fuzzy logic comes from the Rheinbraun Corp.’s plant at
Huerth, Germany. The complex process to stabilize coal gasification—with
coal of varying quality—uses a fuzzy logic supervisory control system
for setpoints of eight underlying PID controllers. FL control strategy
is defined by 115 fuzzy rules and uses 24 input variables. The plant is
automated with equipment from Foxboro/ Eckardt; the FL controller is
implemented in Inform’s fuzzyTech software and uses USDATA’s FactoryLink
FL Runtime Module. Besides rapid implementation time (6 months), “the
fuzzy controller adapted the plant’s operation point to changing coal
quality more than twice as fast as previous human operators,” states
NeuralWare cites an application of its system integrator in Poland,
Transition Technologies, to predict NOx emissions and also optimize
performance of pulverized-coal boilers at the Ostroleka Power Plant. The
three-unit plant (600 MW) has Westinghouse WDPF control systems.
NeuralWare’s new NeuCOP II controller (on a Solaris 2.5 platform) is
completely integrated with the WDPF system. A complex control problem is
compounded by special burners and nozzles that distribute the combustion
process over a larger volume to reduce NOx emissions. However, the
process raises the amount of unburned carbon, thereby lowering boiler
ANNs in NeuCOP II model the system, predict NOx emission values, and
change secondary air flow to compensate for burning efficiency.
Preliminary results of modeling indicate that NOx reduction of 15% or
more is possible with increased boiler efficiency, as well. Start up of
ANN control in an advisory mode is complete; full on-line operation is
expected by mid-July. “NeuCOP II was chosen as it is one of the most
advanced and reliable neural-network controllers, and for its
applicability to large scale systems,” comments Konrad Swirski, director
Inferential sensing in NeuCOP II can also predict NOx emissions
off-line based on boiler control parameters. “Such ‘soft monitoring’
methods have been allowed by the U.S. EPA to replace or defer the
purchase of costly, high maintenance monitoring equipment,” says
NeuralWare’s Mr. Klimasauskas.
At Amoco’s Texas City oil refinery, time delays from on-stream
analyzers can take over 30 minutes. To rein in this quality control
problem, Gensym’s G2 Diagnostic Assistant and NeurOn-Line software
packages predict real-time “soft sensor” values for the composition of
five petroleum products being refined. “The result is more than 95%
closed-loop [plant] utilization, and estimated savings of over $500,000
per year in recovered product,” according to Amoco.
Many large automation and process control companies develop an
internal AI element for their plant control systems. At Fisher-Rosemount
Systems (Austin, Tex.), the approach has an added intent: Make AI
methods available to the routine user in the plant by full integration
with overall controls. “The perception by many users that AI
applications are just too difficult to use and require consultants to
implement is no longer the case,” says promotions manager, David Holmes.
Glen Mertz, a manufacturing technologist at Monsanto Co. (St. Louis,
Mo.) uses F-R Systems’ ANN software. “The neural network software is
almost intuitive. I can sit down and in less than two hours be building
a model with plant data.” remarks Mr. Mertz.
Ease of use and installation was found to be helpful in another F-R
Systems application—FL controllers for process water temperature
control. Ken Wildman, senior engineer at Eastman Kodak (Rochester, N.Y.)
states, “As a result of greater startup stability with fuzzy logic
control, time-consuming manual startups are avoided and demand is
quickly met for process water at a precise temperature throughout a vast
Motor and motion control continues to be a growing area for AI
techniques. As an example, at the 1st IEEE International Electric
Machines and Drives Conference, May 18-21 in Milwaukee, Wis., some 20
papers included fuzzy logic, neural nets, and genetic algorithms (or
their combination) as problem solving tools.
Two major upcoming seminars in Europe will further highlight FL, NN,
and GA technologies. The 5th European Congress on Intelligent Techniques
& Soft Computing (Sept. 8-12) and the 1st European Symposium on
Applications of Intelligent Technologies (Sept. 9-10) will run in
parallel in Aachen, Germany.
From chaos to order
Above the various AI technologies are more advanced adaptive systems
that resist definitions of state. They’re loosely termed here as chaos
systems (CSs), but some people prefer to call them complex systems or
emergent systems. “Complex system behaviors are emergent in the chaos
regime, rather than explicit; they’re deterministic, not random,”
explains Dick Morley, president of Flavors Technology Inc. (Manchester,
In a CS, behaviors of individual elements (agents) work together to
form system behavior that cannot occur in the elements. Examples include
the swarming of bees and wolfpack hunting tactics. Applications closer
to CE’s audience are predictive scheduling, raw material allocation, and
transportation needs. “Neural nets, fuzzy sets, logic statements, and
expert systems are, in truth, subsets of behavioral identities we call
agents,” says Mr. Morley.
He sees growing use of the chaos approach in several areas of
manufacturing and process industries. The goal is to dramatically cut
the software effort for solving highly complex problems. And results
look very promising: reduction factors of 10 to 100 in software
development effort, according to Mr. Morley. How is all this possible?
Because the CS works with behavioral rather than data aspects of
objects, only the various types of entities (but relatively few) need to
be described to yield a system solution—not the behavior of the system
or individual entities. “In effect, it’s a bottom-up approach that
starts with objects and lets the system evolve.”
He thinks engineers can make do with classical control methods for
the near future, but in just a few years “the average control person had
better know how to use this powerful system for solving problems such as
MES, ERP, MRP, and predictive scheduling and diagnostics.” Tool sets for
chaos system control are still limited. Paracell, from Flavors
Technology, is one of them.
There is a downside, of course. A change in mind-set will be needed
for people to accept chaos methods. “The technique is seemingly
unbelievable, both as to benefits and the efficiency of solutions,” adds
Mr. Morley. Some danger of hype always exists, but over the next decade,
chaos system design might become the basic advancement in the control
engineer’s tool box that it’s meant to be.
Looking ahead, the combination of AI methods can build strengths and
offset weaknesses. Summarizes Inform’s Mr. von Altrock, “In the future,
these technologies will merge together. This trend, dubbed ‘soft
computing,’ exploits the fact that each AI technology has its strengths
and weaknesses. Thus, a clever combination of AI methods can be good
Artificial Intelligence Suppliers
|Bios Group L.P.
|ANN - Neural nets FL - Fuzzy logic
ES - Expert sytems GA - Genetic algorithms CS
- Chaos systems