Printer Friendly

'Artificial intelligence' already taking many forms in plastics processing.

|Artificial Intelligence' Already Taking Many Forms in Plastics Processing

Software "experts," "smart" machines, and even entire "smart factories" are coming to plastics processing. They're coming real soon--and some are already here. You may already be using one of these "expert systems" without knowing it, since some suppliers are hesitant to publicize their use of little-understood new technologies. They don't want to scare off processors who may think they or their employees "are not ready for it."

But the good news is that the sophistication that underlies the new technologies is designed to be invisible to the user. Intelligent systems are meant to be easier to run than "dumb" ones, because they require less action and less expertise from the user. And the other good news is that increased computer sophistication need not be proportionately more expensive. Suppliers of the new systems say that "fuzzy logic" chips, "neural network" chips and software-development packages, and expert-system "shell" programs are available today off-the-shelf and at reasonable cost.

Information from several hardware and software suppliers indicates that you'll be hearing a lot more about artificial intelligence, or "AI," as well as such exotic terms as fuzzy logic and neural networks, in the next few months. A number of significant developments were previewed late last year at the Plastics Manufacturing Automation '90 conference and exhibit in Chicago, sponsored by Innovative Expositions, Inc., sub. of Schotland Business Research, Inc., Princeton, N.J. One of the AI products unveiled there has since become commercially available, and PLASTICS TECHNOLOGY learned that more will be introduced at the SPE ANTEC in Montreal in May and at NPE '91 in Chicago this June. (There will also be a second Plastics Manufacturing Automation show and conference--PMA '91--in Detroit, November 12-14.)


Speaking at the PMA conference, Dr. Perry W. Thorndyke, a software consultant based in Atherton, Calif., helped clarify the significance of AI. He said that in the age of computers, the paramount unmet need in manufacturing is no longer information processing--indeed, we're drowning in data--but knowledge processing, which he defined as interpretation of information. What are needed now, he suggested, are new types of thinking machines called knowledge-based systems (KBS). Unlike a traditional computer, which manipulates numbers by means of arithmetic calculations and algebraic formulas, a KBS manipulates symbols and concepts by means of rule-based logic ("if...then...") and experienced-derived rules of thumb. Thorndyke further defined an expert system as "a KBS that functions at a level comparable to a human expert." Industry's task for the 1990s, Thorndyke said, is to "shift knowledge processing from people to machines."

One of the newer tools to accomplish this is so-called fuzzy logic. At PMA '90, Dr. Dean Reber, director of product development at Cincinnati Milacron's U.S. Plastics Machinery Div., Batavia, Ohio, explained it this way: "As you know, the information we get isn't always exact. Sometimes, an approximation is all that is available. With fuzzy logic, a computer control system can take inexact inputs and translate them into exact outputs."

A practical example of what this means was given at the same meeting by Gary Bartholomew, executive v.p. and general manager of Buhl Automatic Inc., Guelph, Ontario. Taking vague input information from the operator about a molding problem, such as "I have short shots" or "The part has flash," a fuzzy-logic controller makes logical inferences on the basis of preprogrammed rules. Then a process of "defuzzification" occurs to result in a concrete process correction.

Bartholomew also spoke of another key new AI technology, neural networks. He defined this as a type of parallel processing that allows computers to process large amounts of data from multiple inputs very quickly. It also allows a computer to learn--for example, to sort out which process variables have most influence on part quality. Said Milacron's Reber, "Neural networks really are systems capable of learning and organizing their data much more like we human beings think. This capability will ultimately give machines the ability to learn by evaluating information while they are in operation."

The conceptual layout of a neural-net system really is analogous to a system of human neurons, explains Denes Hunkar, president of Hunkar Laboratories, Inc., Cincinnati, which previewed an application of this technology at PMA '90. As shown in the accompanying schematic, the network consists of "layers" of "nodes," each of which receives inputs ("stimulation") and delivers outputs (which may be "stimulation" for the next layer of nodes) calibrated in strength from zero to 10. A key feature of the network is that each node is linked to every other node in the adjacent layers. After a node receives and processes an input, it weights the outputs to the next layer of nodes differently (from zero to 10), so as to reinforce or inhibit the "connection strength" to each subsequent node. This weighting of outputs is a form of analog logic, in contrast to the binary yes-or-no logic of today's digital computers.


One of the most basic applications of AI, and the first to be used commercially in plastics, represents an attempt to remedy the chronic shortage of human experts by "canning" the know-how of the most experienced people in industry. An example of this is the Plastics Education and Troubleshooting System (PETS) from GE Plastics, Pittsfield, Mass. This personal computer system incorporates the knowledge of GE experts to aid in troubleshooting molding problems with the company's resins via a question-and-answer dialogue; it also contains a textbook of processing information (see PT, Sept. '89, p. 14). The system recently was expanded to include troubleshooting guides for Cycolac ABS and Cycoloy ABS/PC, in addition to all other GE resin families except Supec PPS.

PETS was previously sold only with a personal computer for around $5000. Now it can be purchased for $1200 as a mass-storage device with software for molders who already have compatible computers. Single resin-specific discs are offered for $100, and a somewhat simplified version of PETS is available free of charge to customers on GE's on-line Engineering Design Database. (CIRCLE 13)

Hunkar Laboratories recently introduced a new AI product that makes use of both fuzzy-logic and neural-net technologies. It's an injection molding troubleshooting advisor that runs on a personal computer; a version for extrusion blow molding will be presented at the upcoming ANTEC. Says company president Denes Hunkar, "The neural-network technology employed by our system is, by nature, a fuzzy-logic system that is not designed to provide precise answers and solutions to every problem. It is a tool that attempts to classify patterns according to other patterns it has learned and to give the most reasonable answer. It is not guaranteed to always give an absolutely |correct' answer." Rather, he says, the program's outputs are expressed in terms of "confidence levels," a larger number signifying greater confidence that the answer is correct.

The program offers the user a menu screen of 16 molding problems (see opening photos). The user can indicate which problems are being experienced (unlike most human engineers, the program can address more than one problem at a time), and apply to each one a weighting factor from zero to 10, according to the seriousness of the problem or the priority level of importance the user places on it. Then the program provides a screen showing the confidence levels of probable solutions. This is in the form of a Pareto-type bar chart, with the bar length corresponding to the numerical confidence level (also displayed) for that potential solution; and the solutions are arranged in order of priority, from higher to lower confidence levels. This suggests an order in which the user might implement the solutions.

The fuzzy logic behind this software enables it to give conflicting recommendations, such as to both raise and lower barrel temperature, in cases where human experience is inconclusive. Different confidence levels may indicate that one is more likely than the other in this case.

This software illustrates how an expert system may be based on human experience and knowledge, but not necessarily limited by it. The vastly more rapid data-processing capacity of a computer may enable it to recognize patterns that a person might miss. Hunkar cites the example of a case where a molder may be experiencing both flashing and short shots. In this apparently paradoxical circumstance (illustrated in the photos), the computer recommends increasing overall injection speed, which would increase shear, thereby lowering the melt viscosity and allowing complete fill without excessive injection pressure to cause flashing.

This IBM-compatible software, previewed in Chicago, is now commercially available on a disc for $300. Hunkar has also developed versions for thermoset injection, transfer molding, and electroplating, and is now working on one for automotive exterior paint lines. (CIRCLE 14)

One injection molding troubleshooting advisor on the market allows the user to modify or increase the "knowledge base" of the system. For example, he can add information on what solutions worked for certain problems in his particular plant; or he could rule out certain solutions as impractical in his circumstances. This software is called SASS (Situation Analysis and Solution System), and was introduced at the K '89 exhibition in Dusseldorf by Plastics & Computer, Inc., Montclair, N.J. (see PT, Jan. '90, p. 68). It's designed to complement the firm's TM-Concept CAE software. Other details on SASS were presented in a paper at the PMA '90 conference. (CIRCLE 15)


Plastics & Computer expects to add a new fuzzy-logic expert module to its FA-1 flow-analysis package in time for NPE. Its function, says company founder Giorgio Bertacchi, who is based in Milan, Italy, will be to "make a rational choice in cases where there's not only one solution." The new module is being developed in response to customer requests for an expert advisor to tell users how to evaluate the shear-stress output of the flow-analysis program. Their frequent question is "How much shear stress is too much?" The new module will recommend a maximum "safe" level of shear stress for cross sections that are above a certain wall thickness and are molded above a certain melt temperature. Since shear stress relates to dimensional stability, the solution will also depend on whether the user identifies the application as precision molding or one with a looser level of tolerances. (CIRCLE 16)

A newly commercial TMConcept software module from Plastics & Computer, called CseCad, is an expert system that uses the results of the firm's CSE computerized shrinkage analysis module to go back and automatically redimension the mold cavity to correct for shrinkage. This does not involve simply mathematical application of nominal "data-sheet" shrinkage values for the material. Rather, the AI element comes from applying an extensive, experimentally derived knowledge base and inferences from related data. One example is interpretation of the packing phase of the flow analysis (which should precede shrinkage analysis) to predict stress levels molded in during the last increments of flow. In addition, the program asks the user to characterize the overall tightness or looseness of dimensional tolerances required, which are the critical dimensions of this particular part, the degree of process variation that the user normally experiences with an individual molding machine, and the level of tolerances the user can expect from his moldmaker. Also, the program considers not only the orientation of each dimension in the mold relative to flow, but also whether a dimension occurs on the male or female side of the mold from a "steel-safe" point of view. (CIRCLE 17)


Some degree of AI is also being applied to costing/quoting/scheduling software. For instance, Data Technical Research, Jacksonville, Fla., has programmed its system to produce the most efficient machine/job schedule for injection molding and other plastics operations. It keeps a record of which machine has run a given part most efficiently, and it "knows," for example, not to schedule a white part after a black one on the same machine, instead of vice versa. Also, the user can designate a job for "Just-in-Time" production, which gives it priority to "pull things through the system to meet it," explains Mitch Timberlake, regional sales manager. Other jobs, he says, can be designated for MRP treatment based on assigned lead time, or for simple First-In/First-Out scheduling. (CIRCLE 18)


The next step, industry observers agree, will be to package expert systems in machine controls and to give them on-line functions, rather than just off-line as at present. That is, future generations of process-control systems will be able to pick up the output of the expert advisor, or "knowledge engine," and automatically adjust a processing machine to remedy the problem--perhaps asking a human operator for confirmation before doing so.

In fact, such systems are already appearing. A new system of this type from Japan was previewed at the International Programmable Controllers expo in Detroit last spring, and introduced at JP 90 in Tokyo in November (see PT, July '90, p. 14). Omron Corp.'s Fuzzy Molding Machine Assisting System (not available in the U.S.) is designed to overcome molding problems without the aid of a skilled machine operator. It consists of an Omron ES1000 analog process controller, C1000H programmable logic controller (for sequence control), and the new C500-FZ001 Fuzzy Unit.

The operator merely enters the nature of the molding problem--such as a short shot--and the degree of the problem ("a little short," "very short," etc.). The Fuzzy Unit employs 60 rules and what Omron calls "membership functions" for executing "fuzzy inference operations." As an example, for the conditions of flash, short shots, and internal voids, the Fuzzy Unit will generate instructions to the machine controller to execute adjustments to two stages of injection speed and speed changeover position, injection and holding pressure, velocity/pressure changeover position, and plasticating stop position. (CIRCLE 19)

Another on-line expert system, which may arrive in the U.S. in time for NPE, was developed by Fanuc Ltd. in Japan, Cincinnati Milacron's partner in developing the ACT line of all-electric, servo-driven injection machines. An expert system designed specifically to solve molding problems on ACT machines was shown at JP fairs in Osaka and Tokyo in 1988 and 1990 (see PT, Jan. '89, p. 63; Jan. '91, p. 60). Milacron is hoping to introduce it here soon. (CIRCLE 20)

As explained by Dr. Reber at PMA '90, Milacron is also developing an expert molding advisor to be included in its top-of-the-line CAMAC XTL machine controller. This on-screen troubleshooting guide will allow the operator to select the problem from a menu and then provide suggested corrective actions, prioritized in the order they should be tried.

Also in development is a Materials Expert System for the CAMAC XTL, designed to assist in first-time setup with new material and/or mold. From a database that currently includes 40 materials (including even liquid-crystal polymers), the system will recommend barrel and mold temperatures, shot size (based on part size and shrinkage), and other parameters. Both the materials advisor and troubleshooting guide are planned for introduction at NPE. (CIRCLE 21)

Reber said Milacron hopes eventually to develop intelligent self-diagnostics for the machine's electrical and hydraulic circuits. Today, the CAMAC XTL controller can display a schematic of those circuits, and even zoom in on certain areas. But in the future, the diagnostic screen might automatically display the problem area with a flashing highlight on the precise site of the malfunction.


One way in which machines can show off their intelligence is by analyzing situations with multiple input and output variables. Typical controls today--even multi-loop models--are fundamentally collections of single-loop controllers, each with one input and one output variable--such as one barrel-zone temperature and one heater on/off duty cycle. Future smart control systems will juggle several input variables in order to control one or more complex process phenomena.

An example is Milacron's development effort on closed-loop viscosity control in an injection machine. As explained by Reber at PMA '90, real-time melt-viscosity measurements could be acquired by sensing pressure drop through the nozzle together with injection speed. The key R&D challenge is to determine how to close the loop by feeding back adjustments to barrel temperature, screw rpm, back-pressure, injection rate, or some combination of these. (Dynisco in Sharon, Mass., and the Plastics Engineering Dept. at the University of Lowell, Mass., are also said to be working on this problem.)

The ultimate goal, Reber explained, is to develop a computer model of the behavior of an injection machine to complement the existing models of melt behavior in a mold. Combined together into a machine control system, these models could allow the molder to define a multi-dimensional "quality space" for a part. This parameter space would be defined in terms of cavity pressure, melt temperature, transfer pressure, viscosity and so on. "The machine-mold model would then be used to control the machine and mold parameters so the part would remain always within the acceptable parameter space," said Reber. "Moving a step beyond that, we could look at ways to optimize such functions as energy consumption or cycle time while always staying within the limits of our quality parameter space."


Also at PMA '90, Frank H. Dyke, manager of the Plastics Machinery Automation Dept. of Allen-Bradley, Highland Heights, Ohio, shed new light on the firm's 1771-QD Injection Control Module introduced in 1989 (see PT, May '89, p. 77). Dyke revealed much new information about the system, including what he called the "first AI routine to continuously retune the loop" for injection control.

Dyke explained that "loop instability problems" of oscillating speeds and pressures occur unpredictably and intermittently in injection molding, which often make it impossible for machine operators and technicians to retune the loop successfully. Dyke said there are at least four "detuning influences" that produce these instabilities:

* Changing melt viscosity within the shot and between shots. This can result from water absorption by the resin, stratification of additives or regrind, inhomogeneous melt temperature throughout the shot, and variations in regrind content.

* Hydraulic-oil temperature changes. This may be observed on start-up or after prolonged cycling; from variations in cooling-water temperature; and from changes in pressure in the manifold.

* Changing mechanical elasticity of the machine system. Causes may include changing elasticity of piping and hoses, or thermal expansion of machine components.

* Changes in hydraulic valve response in the injection circuit. Not only do valves wear during their service life, but valve response varies over the broad pressure ranges that occur in the course of an injection cycle. A single response curve thereby has only approximate validity.

To compensate for these detuning influences, Dyke said that the 1771-QD closed-loop controller uses "Autolearning" to continuously observe the characteristics of the system and retune itself to match them. Starting with manually entered tuning constants on the first shot, the AI routine in the software generates small incremental corrections over five to 10 shots. And it continues doing so whenever it detects a change in system response for whatever reason. For instance, it will continuously retune to account for a valve that is wearing or for its different response characteristics at high and low pressures within a cycle.

Allen-Bradley has over 150 systems in the field and has used the system itself on eight machines for over two years. In its own operations, Allen-Bradley reportedly has seen as little as one bad cycle in 250,000. Many customers report that the self-tuning feature eliminates machine warmup time, Dyke says. For example, instead of taking one hr to get back on track after lunch break, they obtain good parts on the first shot. Dyke said the system also helps adjust for moisture changes in hygroscopic materials and minimizes the effects of lot-to-lot material variations (it's said to be especially helpful with thermosets). For example, on one HIPS part that was normally prone to flashing, the system allowed a change of resins over a course of eight successive shots, involving a threefold increase in flow rate, with no flashing. (CIRCLE 22)


Bartholomew of Buhl Automatic propounded a comprehensive vision of AI at PMA '90 that applies not just to machine controls, but to every department of plant operations. The ultimate realization of this, he predicted, will be the "AI plant," which he described in terms of a high-speed peer-to-peer network of distributed databases and "knowledge bases" that are instantly accessible by every "node" on the network. Thus, problem solving at any one node may be aided by querying knowledge bases and databases at other nodes.

This relatively low-cost architecture is achievable today, he said, with off-the-shelf software, and standard local-area network operating systems and hardware standards. It will present numerous opportunities for AI implementation not only in machine control, but in scheduling and production planning, purchasing (e.g., vendor selection based on past delivery, quality and price performance), preventive-maintenance scheduling (based on trend analysis and production optimization), quality control (including "automated Taguchi analysis"), design engineering (such as for manufacturability), and corporate financial and other planning (including manufacturing needs forecasting and product time-to-market analysis).

Realization of this scheme may not be far distant. Bartholomew hinted that Buhl will make a dramatic introduction at NPE of a plantwide AI system for injection, extrusion and blow molding. We'll have first details on that system next month. (CIRCLE 23)

PHOTO : Hunkar Laboratories' new injection molding advisor is an "expert system" that employs both "fuzzy logic" and "neural network" technology. Neural network concept (left) is a form of parallel processing with analog logic, involving different "stimulation levels" and "connection strengths" between nodes. Above, two problems selected from a menu of 16 are shown above the dotted line, together with an importance or priority level from zero to 10 for each. Possible solutions to this apparently paradoxical pair of molding problems appear as a Pareto-type bar chart below the line, arranged in order of "confidence level" for each suggested solution. Fuzzy logic permits contradictory answers to be displayed (raising and lowering the same variable), along with corresponding probabilities that either one is correct.

PHOTO : New CseCad software from Plastics & Computer applies artificial intelligence to CAE. Program takes calculated shrinkage values for each part dimension and automatically adjusts cavity dimensions, taking into account the molder's process variability, toolmaking tolerances, "steel-safe" considerations, and quality level required.

PHOTO : Situation Analysis and Solution System from Plastics & Computer, Inc. offers "expert" advice on solving injection molding problems. Its creator says truly "intelligent" software asks the right questions in order to give sensible answers.
COPYRIGHT 1991 Gardner Publications, Inc.
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 1991, Gale Group. All rights reserved. Gale Group is a Thomson Corporation Company.

Article Details
Printer friendly Cite/link Email Feedback
Author:Naitove, Matthew H.
Publication:Plastics Technology
Date:Mar 1, 1991
Previous Article:Welding & bonding.
Next Article:Novel low-VOC paint technology.

Related Articles
New technologies emerge in medical AI.
Antibiotic polymer prostheses.
Artificial intelligence and natural confusion.
Possible applications of neurocomputing in defense.
AI at work: from prisons to autos to space.
Replacing humans with machines: the insurance industry has begun to leverage artificial intelligence to cut costs and improve efficiency. But...
Our Molecular Future: How Nanotechnology, Robotics, Genetics, and Artificial Intelligence Will Transform Our World.
AI learning aids Java programming. (IT News).
Artificial Intelligence for Computer Games.
Using AI to learn about algorithms.

Terms of use | Copyright © 2017 Farlex, Inc. | Feedback | For webmasters