Automatic flat dies gain artificial intelligence.

To the growing list of machines that can "learn" through artificial intelligence, add a new automatic flat-die control system from The Ohmart Corp., Cincinnati. Known chiefly for gauging devices, Ohmart can now integrate its two auto-die systems for cast film and sheet with artificial neural-network software for on-line thickness profile control.

The control cycle begins much like other auto-die systems. Starting with signals from the gauging device, the neural network compares the actual product with the desired target. If necessary, it then adjusts particular diebolts to control thickness on extruded film, sheet or coating lines. Here, however, the similarity with conventional systems ends.

Relying on analog rather than simple yes-no logic, neural networks have a conceptual layout that approximates a system of human neurons. Layers of "nodes" connected through a series of weighted numeric values respond to input stimuli and trigger appropriate outputs. As the system refines the weights that relate one level to another, the system "learns."

In Ohmart's case, the neural network learns in the sense that it bases its future responses to extrusion process upsets on historical experience with a given line. "It becomes the expert on its own process," explains Stephan Wessels, one of the engineers who designed the system. Except this computerized expert never goes home or does anything else. As a result, Ohmart claims improvements in control consistency and speed, as well as hefty reductions in changeover and start-up times.

In its first installation in January '91, the neural-net system produced acceptable product on a biaxial line in about 2 hr rather than the two weeks normally required for a complete auto-die start-up, Wessels notes. Recently that first customer, a major materials supplier, has ordered two more systems to bring its total to four. HPM Corp., meanwhile, has ordered its first system for a large sheet extrusion line.

NEURAL NET WORKINGS

Rather than taking a completely unfamiliar path to auto-profile control, the neural network instead augments a common approach. Like conventional systems, it relies on error tables full of numeric weights that relate the influence of individual diebolts to a given web segment and neighboring segments.

The neural net goes a step further, in that the tables it starts with are only a departure point. Over time the neural network "edits the tables," continuously assigning new weights derived from its "inherent ability to recognize patterns," explains Wessels. "It soon does a better job than the algorithms we input."

By contrast, conventional auto-die systems have fixed tuning that requires an operator to change these values, according to R&D v.p. Daniel Jahns. And deducing the complex relationships between diebolts and the web segments in the first place requires both experience and painstaking experimentation for a human operator. Jahns explains that a person detects the patterns affecting a given process by introducing a large enough error to observe at some point downstream. For instance, a human operator might tune a system by "tweaking a diebolt" in order to measure its effect on the web. The neural network, however, can cull subtle patterns from its inputs by recognizing upsets as small as 0.25%.

In the current configuration, separate error tables correspond to the only three tuning parameters currently interpreted by the neural network--diebolt mapping, cross talk and gain. Wessels notes that the desire to determine mapping, or match up each web segment to the diebolts that influence it, provided the impetus for adopting a neural network. "Mapping has traditionally been one of the industry's black arts," Wessels says, especially on stretch-intensive biax lines. Cross talk, or the influence of a diebolt on its neighbors, and gain were similarly found to benefit from the neural net's ability to find patterns.

Users of the current system can select the neural network's learning rate, or sensitivity to error. And it can also set the system in a mode where it merely makes recommendations. An icon-based graphical user interface guides operators and allows viewing of the error tables. In the near future, Ohmart plans to optionally apply the ANN system to the full range of processing variables, such as temperature and line speed.

Ohmart's new neural-network software integrates with either the company's Series 9000 or ConceptOne gauging systems. Introduced at NPE '91, the newer ConceptOne is a PC-based system that runs on a 32-bit Intel 486 platform using IBM's OS/2 operating system (see PT, Aug. '91, p. 52). The Series 9000 has proprietary operating software.