Call it what you will, it's still artificial intelligence

Artificial Intelligence is proving it's worth in process and in practice in plants nationwide. Meanwhile the costs, in time and money, of adding AI capability to a plant keep falling.

By Dana Blankenhorn

Share Print Related RSS
They’re not really using Artificial Intelligence (AI) over at the St. Mary’s Paper plant in Sault Ste. Marie, Ontario, or are they? They are, but paper machine superintendent Dave Dube doesn’t look at it that way. “It’s part of a puzzle that controls the caliper, and the gloss and smoothness of our supercalendared paper,” he says. Sensors measure, in real time, the condition of the product going through manufacturing process, and automatically adjust the sheen in all directions.

This used to be very difficult. "We would spot check a few feet of sheet," Dube recalls, “adjusting the result through the use of steam boxes.” With real-time monitoring, Dube’s system can make adjustments using both caliper coils and steam boxes. "This is what all the engineers were telling us would be the real challenge on this project--not having caliper coils fight the steam boxes," Dube comments.

Such a procedure could only be done effectively in real-time, which meant not just scanning the output, but adjusting the manufacturing process as it happened.

No engineer can do this manually, but software can.

In this case, it’s a state-of-the-art, large-scale, model predictive controller for paper machines which strikes the balance. The software takes input from the sensors and passes instructions to the machines. Honeywell’s Performance CD Multivariable controller is, in essence, a neural network.

Honeywell’s technology, however, is 15 years old, senior consultant Lance Johnson notes. What gave it AI capabilities is Moore’s Law, computer power getting faster and cheaper. That means software can now be embedded “at the base level control layer for single loops,” resulting in “a model predictive controller at an execution interval of ¼ to ½ second,” says Johnson.

Then, by encapsulating multiple systems “in one cohesive horizon,” Johnson adds, you have a system for controlling a plant, not just a machine.

The result is a complete network of sensors, actuators, software and hardware that can manage a plant to its optimum without operator intervention. In other words, Artificial Intelligence.

From the Esoteric World of AI
Many techniques that came from the esoteric world of AI are now proving themselves on the plant floor. Besides neural networks, processors and manufacturers are installing software agents and expert systems. They are often retrofitting existing equipment with new software. But they don’t call it AI.

Instead they just call it improved process control. That’s the way Charles Cutler, president of Cutler Technology Inc. San Antonio, Texas describes what he’s offering.

Cutler, a CONTROL Process Automation Hall of Famer, first retired in 1996 after selling DMC Corp. and its Dynamic Matrix Control technology to AspenTech. "After I retired, I started thinking [about] this problem," he says, and began thinking of ways to improve the DMC algorithm.

According to Cutler it was a revelation. "You put constraints on the system, via a linear program, and it keeps the machine inside those boundaries. When it comes up against a boundary, it makes a decision based on software, rather than [on some] arbitrar[y factor]," he says. Engineers were turning his linear controller into an expert system.

Most major process control vendors are working on some form of expert system or neural network to enhance productivity. These companies are not just getting more production out of each worker, they’re improving quality, extending equipment lives and using all inputs more efficiently.

ARTIFICIAL INTELLIGENCE INSIDE        
         
Although optimized control systems aren’t considered to incorporate “artificial intelligence” per se, the feedback loops, sensors and software management of the inputs paint another picture entirely. 
       
An OSCAR for Water
Michael Eaton lists himself as a neural network specialist at JEA (formerly Jacksonville Electric Authority) in Jacksonville, Fla. Eaton is using the G2 Reasoning Engine from Gensym Corp., which can view information in a web browser and supports major web standards such as CORBA and DCOM objects, Java and ActiveX controls. Eaton’s project is called Optimization System Control of Aquifer Resources (OSCAR). The idea is to react to changes in water demand, rather than just try to keep reservoirs full.

The utility pumps its water from aquifers. Simply building new pumps as demand rises is not only expensive, but could drain the aquifer too fast. “So I first built a pilot system to control four pumps in a single well field, forecasting consumption and then scheduling. all these are inputs to the model, along with water quality and plant quality," says Eaton. That first project reduced the stopping-and-starting of pumps by 50%, and allowed the company to defer drilling a new well, saving $1.4 million.

Single Expert Wanted
A single expert system can result in huge savings when it is applied throughout a company’s operations. Dow Chemical senior specialist Jim Sturnfield is using a G2 expert system to save money for Dow’s GulfCoast chemical plants.

Sturnfield first applied G2 to a plant in Victoria, Texas, saving the company $1.25 million per year. Since then he has worked on implementing the same techniques for a plant in Texas City, Texas and one in St. Charles, La, near New Orleans. "These [plants] make a wide range of chemicals," he explains, “and we’re using cogeneration to generate power and steam for the production units.”

Sturnfield describes some of the pitfalls. "The problem you have with a chemical plant, as opposed to a refinery, is [that] you have a very wide variation in the system. You must react to changes caused by one production unit or the other, and you don’t always have an easy way to identify the change, but you must respond to it." That’s the process Sturnfield looked to optimize, transforming an open loop optimizer into a closed loop.

"If you’re not in closed loop, operators are busy trying to make things run and don’t want to run close to the boundaries--the limits of your turbines, even though economically that’s the best way to go,” Sturnfield notes. “If an operator is only going to look at things once a shift, they don’t want to have to be constantly change things. In closed-loop optimization you’re changing values every five minutes, based on numbers and software."

The application Sturnfield helped write runs in the central control for each plant’s energy system and operates the equipment, making fine adjustments to the balance among outputs of a gas turbine, supplemental burning and condensing.

"Before, we didn’t worry about the economics,” he says, “We worried about running the plant. You ran equipment as efficiently as you could, and tried to keep things in a range where you could respond to changes. A lot of things were run by header pressure controls–if it started dropping you’d increase supplemental firing, or decrease condensing turbine flows. That was the main control scheme.”

According to Sturnfield, the company has more flexibility, in terms of how much power it buys and sells to the grid. Sturnfield has been working with what amounts to an expert system since 1996, and says the units generating plants can now react dynamically, not just to changes within processes, but to the market as well.

"When someone does a steady-state model you assume everything has settled down. You look where you want to go and get there. You don’t worry about the dynamics,” he points out. “But the dynamics may affect sensor readings. If I’m not careful about those readings I may fool myself into thinking my boilers are inefficient when they’re just starting to move up. You need intelligence to know what you need in the dynamics," says Sturnfield.
And that’s what his system delivers. "You can optimize a system easily. But the advantage here is being able to implement the results of the optimization. The expert system figures out what information to feed into the model and how to put that into action."

Applied Neural Networks
Over at IMC Agrico in Mulberry, Fla., Tony Lazanowski has the title of chief technology development engineer. He applies neural networks to the production of phosphates.

"We mine phosphate ore here," he explains. "We bring it into a washer to separate the pebble fraction, and the remaining feed goes through a feed tank, an a hydro-sizer. Then we do particle flotation to sink the sand. After that, we recondition again and do an aiming flotation where we sink the phosphate.” The result, says Lazanowski, is concentrated phosphate ore which is then reacted with sulphuric acid to produce fertilizer in other plants.

Lazanowski applied G2 to the problem of optimizing these various processes. He started in the control room of one of IMC’s smallest mines, then "as time went on, we added other features and started getting better results," he says, gradually expanding the program to other locations.

"We had an independent review done by two Ph.Ds who came to the conclusion that overall, we improved recovery 2% and reduced reagent cost by 50 cents per ton of concentrate product, which is pretty significant."

The program is ongoing, so improvements are constantly being made--some slowly and others suddenly--as the virtual analyzer Lazanowski built, based on the program’s neural net technology, came to set reagent flows and controls to desired targets in more areas of the operation.

Meanwhile the costs, in time and money of adding AI capability to a plant keeps falling. Just don’t call it AI.

 

 


Dana Blankenhorn is a freelance writer based in Atlanta and a frequent contributor to CONTROL.
Share Print Reprints Permissions

What are your comments?

Join the discussion today. Login Here.

Comments

No one has commented on this page yet.

RSS feed for comments on this page | RSS feed for all comments