Artificial Intelligence …Within

Perhaps we don't hear much about artificial intelligence (AI) methods used within today's technologies because it's slightly unnerving when computers emulate human thinking. Yet we, and computers themselves, continue to improve the way AI works quietly in the background to optimize, reduce process costs, and improve timing and product quality.

By Frank J. Bartos September 1, 2003

AT A GLANCE

Symbolic computing

Simulate human problem solving

Graphical languages

Apply expert knowledge

Neuro-fuzzy modeling

Online extra included in this article

Perhaps we don’t hear much about artificial intelligence (AI) methods used within today’s technologies because it’s slightly unnerving when computers emulate human thinking. Yet we, and computers themselves, continue to improve the way AI works quietly in the background to optimize, reduce process costs, and improve timing and product quality. For some tough, nonlinear applications, AI may be the only solution.

While AI advances, the basics remain:

“Simply put, artificial intelligence is a technology that works to make computers behave more like humans in executing and solving problems. AI is fundamentally different from conventional programming. It is a more symbolic than numeric process, employing heuristic or rule-of-thumb procedures over explicit ones to arrive at solutions,” as explained in Control Engineering , December 1987.

Multiple technologies

Actually, AI consists of various technologies—expert systems, fuzzy logic, artificial neural networks, and genetic algorithms, among others. The order of listing indicates the extent of symbolic computing in these methods.

An expert system (ES) reasons based on rules and experiences derived from human experts, aided by an inference engine that outputs guidance about operating conditions of a process or system. Fuzzy logic (FL) works with uncertainty and partial truth, hence “fuzzy” inputs, to obtain a crisp result. For a sidebar on “FL Basics,” see this article online at www.controleng.com/issues . Artificial neural networks (ANNs) are pattern computers that apply highly parallel distributed processing to solve problems akin to the workings of neurons in the brain. ANNs can learn by training. Genetic algorithms (GAs) simulate evolution in biological processes. Potential solutions are refined over successive generations through “best fit” functions to produce an optimal result.

AI methods contribute to the high performance of complex, nonlinear processes and automation systems that would not be possible with traditional algorithms or equation-based controls. However, AI methods also complement conventional solutions.

Too many promises made in the 1980s led to exuberance, then to user disappointment with AI. More recently, these methods have made a comeback, albeit with less fanfare. Today, AI is at work in numerous industrial applications and commercial/consumer products, but well in the background.

What’s new?

Gensym Corp. notes three significant areas of advancement for AI technology in recent years: Encapsulation of process-specific knowledge, greater use of graphical languages to represent knowledge, and combining of ANN and ES techniques.

David Siegel, Gensym marketing manager, cites trademarked “Intelligent Objects” as an example of knowledge encapsulation that enables rapid deployment of expert systems for cost-effective operation of specific types of plant equipment—compressors, pumps, sensors, controllers, valves, furnaces, and distillation towers. The “objects” technology was acquired from Gensym partner Key Control, and will be embedded in Gensym’s G2 expert system software later in 2003.

Gensym’s “Intelligent Objects” encapsulate expert knowledge about plant equipment to proactively alert operators of problems, as in this example alert for a fired process heater.

Intelligent Objects incorporate ES and FL to detect equipment or system problems and alert operators to possible environmental concerns, control issues, or loss of efficiency. “These events are typically detected prior to the problem reaching alarm setpoints of a process control system,” says Siegel. The case of impending low-efficiency limit in an intelligent heater object is illustrated .

Use of intuitive graphical languages is on the rise, supplementing traditional rule/procedure- based knowledge representation. “This simplifies the ability to build and maintain application logic…and often enables end-users to participate in the knowledge maintenance,” Siegel continues. The company’s software includes root-cause analysis modeling that associates critical process conditions with their symptoms and diagnostic tests.

One such model is triggered by a “reduced throughput event,” for which various potential fault conditions (root causes) are suggested along with tests to further narrow underlying fault causes. A further model condition—reduced revenue—provides a predicated impact of lesser throughput.

Rockwell Automation notes that few really new developments are ongoing in AI technologies. “These technologies are now rather well understood and are being more accepted in industry as they become implemented,” says Ken Hall, VP of architecture and systems at Rockwell Automation, Advanced Technology. More users are willing to give AI a try today—especially for tough, multivariable processes.

Rockwell uses FL control and other AI methods to solve problems in glass and rubber production, municipal water distribution systems, and other nonlinear applications. Fuzzy building blocks or algorithms run on Logix processors to implement the solution. Hall also mentions the role of ANNs in such areas as “recognizing the state” of certain processes, algorithms for neural self-calibration of photoelectric sensors, and NN-based refraction techniques for torque sensors. In the latter application, neural nets are trained to recognize bifringent patterns exhibited by some plastic materials when subjected to strain (twist), which neural-net methods translate to a measure of torque, explains Hall.

Robustness and stability of FL control have been challenged over time by traditionalists. The nature of fuzzy logic is such that rigorous mathematical proof of complete robustness can’t be shown. “While stability of FL control can’t be guaranteed under all conditions, we can do so over a limited but practical input range,” adds Hall.

Basic artificial intelligence functionality is incorporated into the PID Control Toolset within National Instruments’ core software product LabView. The toolset’s fuzzy-logic functions help optimize various control algorithms. Real-Time Module, another LabView component, also can be used to implement basic AI methods by running LabView code in a real-time operating system. LabView’s graphical programming flexibility extends the basic AI functions of these tools, allowing users to implement more advanced AI applications, explains Mike Trimborn, LabView product manager.

Combining techniques

AI methods often work in concert, adding synergy. One advanced application cited by Trimborn is optimization of fuzzy controllers, which is complicated by numerous degrees of freedom and nonlinear characteristics of the process being controlled. Prof. Norbert Dahman at HS Niederrhein University of Applied Sciences (Krefeld, Germany) has used genetic algorithms to develop optimization methods for fuzzy-logic controllers. Begun in 2000, the ongoing project applies GA’s evolutionary mechanisms to optimize control parameters.

Developed at HS Niederrhein University in Germany and implemented in National Instruments’ LabView, the genetically optimized fuzzy-PI controller outperforms many controllers using standard optimization routines. The lower the value of ITAE (integral of time and absolute error) the better.

A typical GA process flows from mutation , or a random set of parameters, to propagation , or reproduction of events, to selection , where only parameter variants that best fit specific optimization criteria are retained to populate the next evolutionary step or generation. Dahman has optimized setting parameters of fuzzy-PI controllers for a third-order, closed-loop control circuit (see Optimization Results diagram). To do that, his genetic algorithm runs in LabView.

Increasingly, neural networks are being combined with expert systems to provide comprehensive process prediction and performance applications, according to Gensym. The neural network benefits from the expert system’s “safety net” that ensures validity of incoming data. Also, the ES can reason about predicted process conditions to detect potential problems and alert operators through use of the ANN, explains Siegel.

Neural nets and expert systems applied together can improve process performance management. MinnovEX Technologies (Toronto, Canada), an engineering company servicing mineral processing applications, combined ANN and ES at Ok Tedi Mining Ltd. in Papua New Guinea. Despite highly variable ore conditions, throughput improved 5.3% using a real-time expert system alone; a further 2% improvement was obtained by adding a neural-network model, says Gensym.

Siemens has employed AI methods in software products for steel rolling mills and other industry-specific solutions. A new offering, called Siflot, now seeks to optimize papermaking—specifically flotation or deinking of recycled paper.

Measuring brightness

Pulp brightness is critical as it directly affects production costs. Insufficient brightness means poor paper quality (or possible product rejection), while excess brightness can be more costly due to extra chemicals needed to treat the pulp. Drastic changes of waste paper characteristics over time, even in the same mill, add to the control difficulty. A further complication is that brightness of finished paper must be predicted from a largely liquid state (slurry), explains James Mantooth, segment manager, pulp and paper industry, at Siemens Energy & Automation.

Previously, expertise of plant operators and extensive lab analyses were used to regulate the process. This wasn’t cost-effective, since operators tended to apply conservative control and laboratory procedures were time consuming. Traditional modeling and computing methods could not provide adequate control.

Neuro-fuzzy modeling delivered the solution, according to Mantooth. Siflot resulted from research and input from operators and paper-industry experts. It measures brightness as a process variable to accurately control production. Siflot consists of four modules:

Case Constructor—collects data from online measurements, synchronizes lab data, and validates sensors.

Soft sensors—located before and after the flotation stage use existing sensors, known process parameters, and software to calculate a stable brightness value.

Neuro-fuzzy model—analyzes data to calculate final brightness and recommends process improvements.

Economic model—provides an overview of flotation-cell optimization. Siflot integrates a real-time cost function, advising operators of process changes needed to obtain desired paper brightness at optimal product cost.

A prototype Siflot system has been in service at Lang Papier (Ettringen, Germany) since September 2001, produced jointly by Siemens, the University of Darmstadt, and Lang. Based on annual paper production of 200,000 metric tons, Siflot saved $250,000 in bleaching chemicals. Siemens also lists cost reductions of up to $700,000 with higher brightness processes, $50,000 for reduced lab analysis, and $400,000 for less reject pulp and related disposal costs—for total potential annual savings of $700,000 to $1.150 million.

Humorous beginnings

The humor around AI methods remains part of the challenge… such as the light-hearted “fuzzy jokes” that followed Yokogawa’s introduction of fuzzy logic to industrial automation in 1989. Customers, skeptical of what FL could do for temperature-controller performance, changed their minds after trying out Super Control with FL and experiencing dramatic improvement in their temperature-control applications, says Clayton Wilson, product manager at Yokogawa Corp. of America.

“Over the years, Yokogawa has embedded fuzzy logic into all of its single-loop temperature controllers—FA-M3 PLC and CX1000/2000 Network Control Station,” explains Wilson.

Super Control with FL is designed to eliminate shortcomings inherent to PID controllers, which can be tuned for various conditions, though not perfectly for all situations. When tuned to be sensitive to process upsets, PID controllers recover quickly, but tend to overshoot excessively on start-ups, setpoint changes, or recovery from major disturbances. Conversely, when tuned to reduce overshoot, a PID controller will “hunt” around its setpoint and provide slow response to disturbances. “Most people compromise the tuning parameters, which results in mediocre control,” continues Wilson.

Paper brightness value calculated by Siemens’ Siflot tool is indicated to process operators or used by the automatic control system as a manipulated variable.

Yokogawa applies fuzzy logic to a controller through two basic parts of Super Control: Setpoint Modifier (SM) and Setpoint Selector . Acting as an “expert operator,” SM models a specific process, looking at the process knowledge base, while recognizing limitations of the system’s dynamics and nonlinearities. “[SM] tricks the system into performing perfectly by feeding artificial target setpoints into the PID block,” he adds. Setpoint Selector passes SM input to the controller.

One example of temperature control using FL is at Dolco Packaging Corp. (Decatur, IN), a maker of pressed Styrofoam food cartons (photo). On a press with 28 temperature zones, fuzzy logic in Yokogawa FA-M3 controllers stabilized temperature under a variety of load conditions and materials, keeping overshoot to under 1 degree.

Not only for processes

Many, but not all AI applications focus on the process industries. IEC 61131, the standard for programmable controller languages, includes FL in part 7, Fuzzy Control Programming. This section provides the basics on how to integrate fuzzy control into the PLC languages covered in the much-used part 3 of the standard. One feature of IEC 61131-7 permits modeling only part of the controller via FL. An example is ability to “zoom in” on a specific controller function that needs improvement.

NASA employs neural networks in aerodynamic design to cut computation costs and optimize procedures. In developing turbine airfoils at Ames Research Center a generic shape was evolved into an optimized one, meeting target pressure distribution and other parameters. Response surface methodology is also used in the now-patented procedure.

Motion control has also tuned in to the possibilities of AI methods for faster, less costly development of motor controls and for supervisory functions. See the Control Engineering May 1999 issue online for a look at some of these developments.

Artificial intelligence technologies are hard at work behind the scenes, and in some cases the technology is being kept there purposely. Certain industrial applications are so closely guarded that some companies don’t even want to be known as users of AI technologies for fear of losing competitive advantage. That’s not good news for promoting the benefits of advanced technology. More importantly, if you think that’s protection, think again. Your competitor is already quietly using artificial intelligence, within.

Online Extra to September 2003 Control Engineering article on‘Artificial Intelligence’

Artificial intelligence is a relatively new term in the dictionary of control engineers, reportedly coined during a technical conference at Dartmouth University in 1956. Today, various technologies reside under the umbrella of Artificial Intelligence (AI)— notably fuzzy logic, artificial neural networks, experts systems, and genetic algorithms, among others.As mentioned in the main article, fuzzy logic (FL) provides a means for finding crisp conclusions from vague and imprecise inputs similar to the way problems occur in everyday life. It offers a simpler method that can eliminate rigorous equations and the totally numeric logic flow of traditional computing.

Normally FL is used an alternative tool for applications or systems too difficult to model by other means. Rockwell Automation (https://www.rockwell.com) mentions a non-intuitive approach that involves modeling via fuzzy logic to build a better “model” of the process itself. “The resulting nonlinear system model may be a fuzzy model or a more traditional one, and the final system control may or may not be fuzzy, depending on the application,” says Ken Hall, VP of architecture and systems at Rockwell Automation, Advanced Technology.

Still, numerous calculations are required in a real FL solution, for example, to compute degrees of membership, fuzzy-set operations, etc. (see FL Basics sidebar). To make the method practical, special chips and microprocessors have been developed by the technology suppliers to do the number crunching. Chip and microprocessor suppliers include Intel (https://www.intel.com), Motorola (https://www.motorola.com), and Texas Instruments (https://www.ti.com).Artificial neural networks (ANNs) have the ability to learn, given initial training with appropriate inputs of data. Genetic algorithms (GAs) go a step further: Existing knowledge directly incorporated into an initial “generation” or “population” of possible solutions eliminates the need for initial training and often results in faster convergence to a best-fit solution. GAs find wide application in computational intelligence, even involving the design and training of ANNs.

No shortage of applications AI methods are applied across a mosaic of industries, with special appeal to tough, nonlinear processes. Examples here include production of cement, detergent powders, pulp and paper, control of electric kilns, and active suspension systems for vehicles.Artificial neural networks have enjoyed particularly wide usage, reaching into fields such as oceanography and meteorology. Siemens (https://www.siemens.com) mentions extensive work and success with ANNs in hot rolling mills for strip-steel production.

FL has been implemented in a variety of commercial and consumer products. Top-of-the-line washing machines are perhaps the best-known examples in the Western world. In Southeast Asia, use of “fuzzy” rice cookers is legend, while room air conditioners from Mitsubishi Electric (https://www.mitsubishielectric.com) sold there are embedded with intelligent FL control to allow temperature adjustment precisely to room conditions.

The NASA aerodynamic design application mentioned in the main article is another example of ANNs working to optimize procedures. More information about this patented evolution of airfoil shapes is available at https://www.nasatech.com/tsp under the Information Sciences category, or contact Ames Research Center (Patent Counsel) at +1 650/604-5104 (Reference ARC-14281).

Forward-looking developments for FL include the sensors arena. One example of self-verifying temperature sensing via fuzzy-logic control involves the Idaho National Engineering and Environmental Laboratory (INEEL, https://www.inel.gov) in collaboration with AccuTru International Corp. Kingwood, TX (AIC, https://www.accutru.com). The novel sensor uses optical and resistive methods to measure high-temperature environments like molten glass with an accuracy of 4%. Two temperature measurement methods serve to make the sensor immune to common-mode failure. INEEL supplied the optical part and AIC the thermocouple (resistive) part of the sensor.

Human operator expert knowledge, captured in a so-called, rule-based fuzzy observer, arbitrates the two temperature measurements and confidence values, says INEEL. “The fuzzy observer uses linguistic rules to emulate the rule-of-thumb thought process used by humans to‘control’ process parameters. For example, one rule set would state the trust an operator places in a new temperature sensor versus the trust the operator places in a temperature sensor that has been used for several months,” according to an INEEL representative.

Further reading There is no lack of references from which to enlarge your knowledge about the umbrella technologies of AI. Here are a few recent sources.

“Computational Intelligence, The Experts Speak,” edited by David B. Fogel and Charles J. Robinson, ISBN 0-471-27454-2, 282 pp., IEEE Press/Wiley-Interscience (2003).

“Neural and Fuzzy Logic Control of Drives and Power Systems,” by M.N. Cirstea, A. Dinu, J.G. Khor, ands M. McCormick, ISBN 0 7506 55585, 399 pp., Newness, an imprint of Elsevier Science (2002).

“Constraint Processing,” by Rina Dechter, ISBN 1-55860-890-7, 480 pp, Morgan Kaufmann Publishing, (June 2003). Available from https://www.elsevier.com

“Artificial Intelligence Applications in Manufacturing,” edited by A. Fazel Famili, Dana S. Nau, and Steven H. Kim, ISBN 0-262-56066-6, 469 pp. Soft cover price is $49.95

“Artificial Intelligence and Soft Computing” (Behavioral and Cognitive Modeling of the Human Brain), by Amit Komar, ISBN 0-8493-1385-6, 750 pp., CRC Press (1999).

“The International Dictionary of Artificial Intelligence,” by William Raynor, ISBN 1-88899-800-8, 380 pp., CRC Press (1999).

“Fuzzy Logic Control, Advances in Applications,” edited by Hen Verbruggen and Robert Babuška, ISBN 981-02-3825-8, 340 pp., World Scientific Publishing Co. (1999).

Also, search for appropriate keywords at https://www.controleng.com/bookstore.

Fuzzy logic basics: Clarity from vagueness

Fuzzy logic (FL) brings reasoning to vagueness found in unclear boundaries of physical processes or in human experiences. For example, even a simple concept such as “normal height” is difficult to express by digital-type computing, yet easily described by a characteristic function (or curve) with continuous gradation between 0 and 1. This “fuzzy set” or membership function takes on various shapes (bell curve, trapezoidal, triangular, etc.). Triangular shape is often used.

Stages of fuzzy processing are shown in the diagram for a simple case of two variables (U, dU), combined by only three FL rules to form the result, DV. In real-world problems, expertise would be described through many such linguistic rules.Variables and results are not single-valued, hence they’re described by the fuzzy sets. Overlapping labels indicate the degree of change or control needed. A seven-label gradation&M>from positive large to negative large&M>is commonly used in fuzzy controllers. This example uses four labels.

Inputs from sensors are compared to membership functions and the lower of two conditions is elected (taking the minimum). Output of all rules is combined in a logical sum. Then, in a defuzzification step, the one most valid control value is output, based on variable input. Center of gravity or other averaging method is used to arrive at the single solution.