The Holy Grail in “intelligent” manufacturing has been to replicate human decision makingâ”this notion of emulating reasoning of people as they assess, diagnose, and respond to unusual operating systems or as they seek to optimize operations, to make decisions to get more with less resources,” says David Siegel, director of marketing for Gensym Corp. (Burlington, MA), a provider of software for “expert operations management.” And there’s the rub. Despite all the promises and research and vulture capital doled out in the 1980s, artificial intelligence (AI) and expert systems were poor substitutes for “liveware”âpeople, especially the experts that expert systems were to replicate.
Yet AI has continued to progress in such areas as representation (natural languages, domain modeling, and knowledge engineering), inference (heuristic reasoning and matching techniques), control (neural networks, search algorithms, and scheduling), and decision support/problem solving (rules-based, constraint-based, and object-oriented). “These are the fundamentals that make up AI,” says Sal Spada, research director for ARC Advisory Group (Dedham, MA). “You mix and match these into various applications.” Some of those applications include speech recognition, autonomous guided vehicles, vision, robotic and machine tool control, planning and scheduling systems, and too-numerous-to-count domain-specific expert systems.
Be that as it may, many of those I talked to have not seen much AI applied in the automotive industry. However, says Siegel, “expert systems are embedded in a lot of different applications and a lot of people don’t know it. This is not only true with spelling checkers and tax programs, it’s also true with industrial applications. An embedded reasoning engine is all part of the overall infrastructure. The average operator doesn’t see it so much as an expert system as just another software application that’s a little smarter than the average software application, a little more functional.”
Automotive applications of AI can only be guessed at because the automakers and their suppliers who actually use them are not talking. Nevertheless, here are some applications of AI learned through the grapevine.
Several simulation companies tout their intelligent, agent-based software infrastructures for modeling design and production constraints when laying out new workcells and production lines, or when generating production schedules. These simulation systems, really intelligent design and analysis software packages, juggle production line parameters without violating any material or machine constraints. The result is better production, measured in terms of production speed, physical space requirements, implementation cost, or production flexibility, to name a few metrics.
Robots seem to have gained the most from AI, including machine vision, force sensing (touch), and advanced servo motion (such as power assistance, motion guidance, and line tracking).
One application exemplifying AI-imbued robots is at Visteon Chassis Div. facility, where intelligent assist devices (IAD) help employees move 4- x 3-ft., 42-lb. catalytic converters from a turntable to a gage fixture for testing. While IADs cost 25% more than conventional materials handing equipment and robots, labor productivity is 100% greaterâwith greatly reduced risk of ergonomic injuries.Â
AI techniques are already used in teaching robots to move about. For instance, software within painting robots can determine the best path to spray paint onto metal automotive parts. Likewise, robots have the wherewithal to optimally apply (and apply in the right amounts) adhesives and sealants onto, into, and around work-in-process.
In the old days, namely the 1980s, robots were programmed to pick up objects known to be in a certain place. By adding vision systems, namely pattern recognition, robots can adapt to some variability in part placement. The location of objects need not be explicitly programmed into the robots, nor must the tolerances in the parts carriers be so tight. Instead, the robot need only detect where the object is, and then it can figure out the restâautomatically re-orienting itself as required to pick up that object.
The concept of “adaptive algorithms” has been repackaged and quietly embedded into servo drive systems. Also called “model reference control” or “self-identifying systems,” these servos can adjust automatically to deviations and can self-optimize to move within a preset tolerance. These deviations may show up in the different weights being carried, the variations in materials being machined, and even in cutting paths.
For instance, explains Spada, when cutting a square pocket in a piece of metal, you want to cut as fast as possible, and with minimal error. To ensure a square cut rather than a curve, the cutting tool must accelerate on one axis and de-accelerate along the other when entering a corner. Adaptive controllers in the machine tool “look” and “plan” ahead, slowing the forward motion of cutting tool enough to minimize error in the corner, while maximizing cutting speed. This yields parts closer to spec and less work in the finishing processes.
Of all the adaptive algorithms, one stands out. “Fuzzy logic modules are embedded in many PLCs and controllers,” says Kevin Prouty, research director for AMR Research (Boston, MA), so “people are often using [AI] without realizing it.” (Note: Fuzzy logic controls the elevators in most modern tall buildings today, gently smoothing the stops and starts the elevator makes at the desired floors.) Fuzzy logic lets machine controllers make decisions based on different variations of “on” and “off” rather than strictly either “on” or “off” as in the past.
Such decisionmaking gives controllers the ability to quickly compensate for production disturbances or to minimize overshoots during recovery procedures, or both. Fuzzy logic can also make the controller more sensitive to error and process oscillations, decreasing the response time for the control process to reach a set value.
Adaptive algorithms show up in other areas of machine tools (and robots). For example, temperature sensors on bearings and ball screws can adjust the tool as heat makes the tool spindle expand. Such adjustments improve positional accuracy and, thus, cutting accuracy.
Online neural-network models can predict, control, and optimize complex, non-linear processes. These models help optimize the output signals to control production processes, thereby improving production efficiency and product quality. Explains Siegel, real-time applications for neural-network models include soft-sensing for predicting product quality, mode-based sensor validation, set-point optimization and diagnosis, inventory management, inspection data filtering, quality control alarming, and detecting fault-level fluctuations.
A neural network process model for predicting process behavior begins by capturing and storing “interesting” production events while online. Tests are run to evaluate the predictive accuracy of this model. When deemed correct and accurate, this model is then used for controlling production, whether a simple workcell or an entire production line.
In operation, the control system computes optimal steady-state setpoints (targets) based on user needs (such as time factors, material and machine availability, economics). A path optimization algorithm drives the process from its current state toward the target, accounting for production disturbances along the way. Last, error feedback manages prediction errors during sampling, as well as dynamically provides closed-loop, stable process control.
One engine manufacturer, for example, uses such a system to check that the right engine parts are selected and in place for machining and assembly, as well as to perform various WIP and in-situ tests that confirm the proper clearance between parts, basic installation and the orientation or positioning of critical components, and a final cold test of the fully assembled engine. This final test is as close to a complete running, functional test of the engine as is possible, without putting fuel in it. The different waveforms generated by the testâsuch as from exhaust and oil pressure, vacuum, and accelerationâare scrutinized to verify the proper installation and functioning of different components. Compared to conventional testing, this cold test monitors far more parameters, accelerates the engine far more quickly, and takes the engine to twice former RPM test levels. This yields a more severe, but more extensive test that is more sensitive to detecting production faults.
Some automotive companies are using expert systems for work process management (such as work order routing and production sequencing). Nissan and Toyota, for example, are modeling material flow throughout the production floor that a manufacturing execution system applies heuristics (rules) to in sequencing and coordinating manufacturing operations. Many automotive plants use rules-based technologies to optimize the flow of parts through a paint cell based on colors and sequencing, thus minimizing spray-paint changeovers.
At least one APS vendor years back touted its use of genetic algorithms to generate optimized schedules. Alas, the vendor is no more, but the rudiments of using a rules-based, Darwinian (“survival of the fittest”) approach to generating optimized schedules lives on. Such scheduling systems have a model of assembly, as well as standard rules and rates of production. For instance, sunroof installation takes X-amount of time, so don’t schedule the production of two cars with sunroofs back-to-back. In operation, the scheduling system first generates a constraint-based baseline schedule. Then it goes through an optimization routine to see what materials and machines are available, as well as what supplier constraints exist. The system finally generates a realistic production schedule that accounts for the vagaries in manufacturing, customer orders, and business strategies.
AI can be found in analyzing warranty claims to root out post-production problems. Using data mining and pattern recognition, automakers can relate warranty information back to production. Case in point: Tire blowouts indicating problems in the manufacture of tires.
Diagnosing production problems, says Siegel, is difficult using conventional software, such as C, Java, and SQL. Expert systems, on the other hand, are more suited to applying more sophisticated forms of reasoning (rule-based logic) and capturing the knowledge of an expert.
DON’T SAY “AI”
But AI still has a “stigma,” says Tom Fiske, ARC’s senior analyst, Enterprise Applications, which is why factory automation vendors are loathe to label themselves as AI companies or to admit they use AI-embedded technology. Instead, these vendors pitch the features, functions, and benefits their systems provide.
That’s the way it should be. In the final analysis, it’s all “just software.” And all software, including AI, points out Spada, is “really just a bunch of if-then statements.”