Hybrid systems with neural nets and fuzzy logic are controlling complex manufacturing processes
Lawrence Gould
Machines break. Chemicals react. Devices get stuck. Feed stocks change. Face it: Manufacturing is nonlinear. Interruptions occur randomly.
Controlling -- let alone scheduling -- processes in the face of such nonlinearities taxes conventional manufacturing control systems. At the same time, products and processes are becoming increasingly complex, cycle times and time-to-market are contracting, and product quality requirements get more demanding every day.
"Using classic linear-control techniques does not yield adequate results, especially in this era of extreme competition," says Mohamad Ali, director of new business development at Neural Applications Corp. "This justifies moving awa
y from traditional linear algorithms and looking toward novel, intelligent strategies capable of coping with such nonlinearities."
Enter artificial intelligence. Vendors of process-control equipment are developing
hybrid AI systems
that bundle a variety of AI techniques, including fuzzy logic, neural networks, genetic algorithms, and expert systems. AI is being applied to situations that have resisted control by conventional approaches that use binary logic or proportional, integral, derivative (PID) control, or both.
AI Meets Traditional Systems
The conventional approach to process control uses PID components. These components compare measurement and set-point (desired) values. The difference between measured and desired -- error -- is the input to the controller. The proportional or integral components respond to the error, while the derivative component usually responds directly to the measurement. The proportional component varies the output perce
ntage, depending on the amount of deviation from a set point. The integral component checks for these offsets and then compensates for them by shifting the proportional band up or down. The derivative component increases or decreases the output based on the rate of change of the controlled variables. The sum of the individual P, I, and D coefficients yields the control output.
Now, AI is complementing, and sometimes replacing, PID control. According to Howard Rosenof, manager of process and utilities marketing for Gensym Corp., there are two broad uses of AI in manufacturing. In control and optimization, the plant is working correctly, but the process-control engineer is looking to increase production, speed up operations, and cut costs. In diagnostics, the process-control engineer wants to know, at the earliest instant, when the plant is not operating correctly and how to resolve the problem quickly.
The right AI approach depends on the specific application. Diagnostics typically make use of ba
ckward chaining searches (reasoning from the conclusion backward, using subgoals) in expert systems. Prediction uses forward chaining (reasoning from the known toward a solution). Rule- and case-based logic (expert systems) are usually not suitable to combinatorial problems, such as planning and scheduling problems. Human knowledge is not broad enough for such huge problems, so the resulting expert systems are too slow.
Fuzzy logic mathematically models the world in the vague, subjective way popularized by human beings: It can handle "hot," "cold," "early," "late," and shades of gray, then convert them into numbers supporting conclusions. According to Glenn Anderson, engineering services manager for Omron Electronics Inc., fuzzy logic is well-suited for applications requiring tracking (e.g., set-point control in noisy, nonlinear, and time-variant systems), tuning (handling conflicting constraints), and interpolating (dealing with multiple-input, multiple-processing levels).
Neural networks are a
step up from fuzzy logic systems. Neural nets are based on mathematical models that not only collect information but "learn" (adapt to changes) from actual system operations. Neural networks help to identify patterns: If a process engineer knows
what
works but not necessarily
why
it works, neural networks can help. Neural network applications include forecasting, quality control, and production control.
Then there are genetic algorithms. They not only adapt, they optimize. Genetic algorithms are good "for tasks where training data is not available at each step and where it is not feasible to analytically derive a control rule, such as in an unstable system," says Casey Klimasauskas, product manager for NeuralWare Inc. "It is valuable for back-propagation when gradient information is not available at each feed-forward pass, and it is applicable to networks with unorthodox architectures, for example, cascaded connections."
Fuzzy Chips
Silicon AI comes as
dedicated microprocessors (custom ASICs) with, typically, fuzzy logic in firmware. These chips are fast. For example, Omron Electronics used to sell a fuzzy processor with reasoning speeds of about 10 megaflips (10 million fuzzy logic processes per second) -- 10,000 times faster than a conventional 8-bit microprocessor. (Omron Electronics no longer markets circuit-level components in the U.S.)
The NLX22x family of fuzzy logic controllers from NeuroLogix covers all the bases in manufacturing control. The NLX220, for example, has four 8-bit analog inputs and four 8-bit analog outputs. It also has six types of membership functions, 111 fuzzy variables, and up to 50 rules. These customizable microprocessors can directly perform such calculations as derivatives and integrals.
Embedded fuzzy systems can enhance PID control. Omron's E5AF temperature controller, for example, is a hybrid device containing two modules: a conventional, feed-forward PID controller and an Omron fuzzy processor. The output o
f the E5AF is the sum of the PID and fuzzy outputs.
The controller's response is based on size-of-error information and the error's rate of change, which can be altered by adjusting three fuzzy parameters (see the figure
"Combined Effects of Fuzzy Adjustments"
).
Soft AI
The more common form of AI is in manufacturing software running on general-purpose computers, especially for the many production processes involving low-level dynamics that are not well understood, such as catalytic reactions in distillation columns. These multivariable, nonlinear processes run continuously, but no analytic model fully describes their underlying dynamics. Neural nets can formulate the underlying connections needed to create a robust model of these production processes. With such models, users can intelligently change an operating variable in order to reach some process objective.
Neural-based control systems are available for specific manufacturing application
s. For example, Texaco uses the Neural Control and Optimization Package (NeuCOP) it developed with NeuralWare to generate petrochemical products, cut costs, and meet environmental standards.
NeuCOP's identification subsystem captures and stores "interesting" process events while on-line. These events go into a database that becomes the training file for the secondary neural-network model.
The control subsystem has three modules:
target optimization
computes optimal steady-state set-points (targets) based on economic and time factors;
path optimization
drives the process from its current state toward the target, while rejecting disturbances; and
error feedback
manages prediction errors during sampling.
NeuCOP uses the G2 Real-Time Expert System from Gensym as the controller's operator interface to provide dynamic testing and on-line monitoring. G2 also acts as a diagnostic tool for when NeuCOP can't solve a problem effectively -- because the limits predefined in
NeuCOP are too tight or because the problem is impossible for NeuCOP to solve.
G2 models heuristic and neural-network reasoning in the form of rules, procedures, objects, and relationships between objects. You write G2 rules in a structured natural-language syntax. The rules can be specific or generic, applying either to a particular object or to an entire range of objects within an object class. Moreover, G2 rules can be event-driven (through forward chaining) to automatically respond whenever new data arrives. They can also be data-seeking (through backward chaining) to automatically invoke other rules, procedures, or formulas. Rules can determine the values of referenced variables, or values checked at regular user-specified time intervals can trigger rules.
G2 uses object-oriented technology: Graphical objects representing production components can inherit properties and behaviors from multiple classes. Object libraries help quickly generate graphs, charts, dials, and tables of real-time dat
a. Generic rules and heuristic procedures represent knowledge (e.g., the "Acidity Rule") that applies to all objects of the same class. New instances of these objects automatically inherit the specified behavior. The AI system is built on top of a client/server architecture that can invoke access privileges to the application for various levels of developers and users.
Interprocess communications between NeuCOP, G2, and other plant-wide information, data-collection, and control systems is through the G2 Standard Interface (GSI), a separate process from G2. The resulting API manages protocol handling, data buffering, initial communications handshaking, and restoring after break.
AI Can Do It
Besides all the benefits to the manufacturing application itself, such as increased throughput, optimized production, reduced waste, and faster response, advances in AI benefit overall system implementation. At one steel plant, the engineers wanted their Intelligent Arc Furnace (IAF)
controller, from Neural Applications, to adapt to an incremental change by adding new hydraulic back-pressure inputs. "The inputs were simply wired in, and the system adapted quickly on-line to the new inputs," says Neural Applications' Ali. "No changes were necessary in the system hardware or software configurations."
So, are you ready to hand over control of your factory to a bunch of algorithms? With manufacturing processes getting more complex, you might have no choice. But don't worry. AI is proving it can handle the job.
WHERE TO FIND
Gensym Corp.
Cambridge, MA
(617) 547-2500 ext. 241
Neural Applications Corp.
Coralville, IA
(319) 626-5000
NeuralWare Inc.
Pittsburgh, PA
(412) 787-8222
NeuroLogix
San Jose, CA
(408) 383-7200
Omron Electronics Inc.
Schaumburg, IL
(708) 843-7900
illustration_link (8 Kbytes)
Omron's E5AF temperature controllers combine advanced PID (proportional, integral, derivative) control with fuzzy-logic control. User-controlled fuzzy parameters improve the response to external process disturbances. Fuzzy intensity governs the magnitude of the fuzzy-logic effects on the final output. Fuzzy scale 1 governs how big the "error" range is.
illustration_link (12 Kbytes)
A hybrid system to control manufacturing processes employs classical methods and a variety of AI techniques, including neural networks, expert systems, and genetic algorithms.
Lawrence Gould specializes in advanced manufacturing technologies. You can reach him at
2541345@mcimail.com
.
If AI Ran The Zoo
