Past, Present, and Future of Automation – Part 3

In the early 1990s the chemical company I was at for decades decided the way to capture and put online process and control expertise was to get corporate licenses to the most powerful real time expert system available for real time process control applications. Some of the most experienced people at the plants developed applications for industrial and agricultural chemical production. We even had a well attended and enthusiastic user’s group meeting centered about the biosphere. In retrospec the biosphere story itself was prophetic about how unaddressed non-idealities and chaos can make a seemingly good idea impractical.

The expert system allowed the entry of rules without any sort of order or overall logic. What we ended up with was a hodgepodge of rules whose interrelationship and execution was difficult to decipher. It was hard to demonstrate that the advisory messages resulted in quantifiable benefits or an increase in knowledge. Even when the expert system did correctly diagnose fault, operations said they had also identified the problem. When the expert behind the expert system moved on, the expert systems fell into disuse.

There was an exception. An online real time expert system was developed using first principal calculations and rules to detect the onset of flooding in a distillation column that was being pushed beyond its design limit to maximize production. When flooding was eminent, the expert system initiated online automatic actions to override the control system. The expert system was a success and routinely prevented flooding problems and is in service to this day.

You might judge another expert system to detect measurement problems to be a success based on the fact it was widely employed at the plants and eventually commercialized by a third party application software company. However, the system did not endure the test of time in the control room. The utility of the system was severely diminished from the attempt to provide inferential flow measurements based on the position of control valves without positioners or with pneumatic positioners whose calibration was questionable. Even if the valve positions were accurately known, the lack of pressure measurements meant the installed valve characteristic was not accurately modeled. There were false alarms. I don’t think any of these systems are really being used.

There were also attempts at using neural networks and Fuzzy logic. Networks lead to some knowledge discovery in terms of previously unsuspected process relationships and was able to predict batch cycle time in some cases but these did not stay online and had no quantifiable benefits. A fuzzy logic system for optimization of reagent usage in a waste treatment system is still automatically trimming upstream pH loop set points and achieves a reduction in reagent cost. However, operations can not figure out the actions or diagnose any possible problems. In other words, the logic is fuzzy. Also it is a one of a kind. A model predictive controller has been demonstrated to be able to accomplish the same goal with much more definable and understandable functionality. http://www.controlglobal.com/articles/2007/385.html

In a separate case, the highly touted use of fuzzy logic controllers and its special tuning of scaling factors for temperature loops are now in question because if you increase the gain of a standard PID controller and add set point filtering, you can do about as well.

So what are the lessons? Most importantly the technology needs to take corrective automatic action without operator intervention and provide benefits that are definable and observable. Second, the system must be able to be maintained and be adjusted without the presence of the author. Third, operations needs to understand and appreciate the actions taken by the system. Training and simplicity are keys here. Fourth, the solution should use standard supported tools wherever possible. Fifth, the performance realized and effort required greatly depends upon the scope and technology of the measurements and valves and is only as good as the inputs and the feedback correction. Finally, the system must use documented algorithms, logic, and calculations based on solid process and control principles. If we can do this, we can make advanced control a bigger part of the future.