Automate the routine

 Auto1
By R. Russell Rhinehart 

Mostly, we use proportional-integral (PI) feedback control, which makes one ask, "It's 2006. Why are we so primitive?"

I think the answer is in the K.I.S.S (keep it simple stupid) principle. PI is simple. Well, maybe not. 

It seems to me that PI is on the edge of what K.I.S.S. allows. 

Sometimes we use derivative (D) action. In addition, relatively infrequently, we use the classical advanced techniques that became commercial standards of the 1940s-cascade, ratio, feedforward, override, and decouplers. These are all powerful techniques, but the complexity of controllers talking to controllers is over the edge of the resident skill base.

In the 1960s, anticipating the move to digital from pneumatic and electronic analog control, researchers began looking at digital control techniques for model predictive control.

In the 1970s, we widely accepted the computer for process control and installed DCS and PLC systems. We also created in-house teams of control system engineering specialists to implement advanced control in our plants. 

However, we held onto the legacy of control strategies and control algorithm types, because they worked. If users do not have the skills to understand what they are using, they keep what works. Mainly, computers gave us more, and instantaneous data to observe, which helped us manage the assets.

In the 1980s, we accepted linear model predictive control (MPC) as a practicable commercial tool. However, skills and costs required for implementation and maintenance relegate it to very few applications where the economic incentive of multivariable, future-considering, constraint-handling control justifies the expense. 

The digital computers, however, allow new data processing strategies. Nonlinear control of nonlinear processes promises once-tuned-always-tuned. We can do control to an economic optimum, data reconciliation, and plant-wide coordinated control. However, can something that requires MS engineers with a control specialty become part of widespread, 24/7 plant practice? 

No-however there are a few success stories of sustained nonlinear control applications based on both first-principles and neural network models.

It appears the costs to develop and sustain personnel skills to maintain advanced systems mostly exceeded the benefit of advanced control. We dispersed our advanced control teams, often out-source control system maintenance, and stick to PI algorithms. 

Academic control research, however, is still predominately pursuing control perfection through mathematical joy. 

While vendors are seeking to capitalize on the power of the computers, the directions are in the use of technology (wireless, Internet, protocol, security, data validation, speed, etc.), but not in substantive advances in strategies that will improve process management in a way that will affect the operating company "bottom line."

I believe there are opportunities to develop algorithms for process automation that are simple enough to implement and sustain, and that will bring positive benefit to the process owner. To find the opportunities, look at the human level of automation. If analysis and decisions of the operational staff are routine, automate them. 

For example, operators view all control charts, on schedule, to determine whether a loop is having a problem. If the computer were programmed to observe the relevant loop characteristics, the computer could flag a poor controller at the beginning (for quick operator recognition and remediation) and not waste the operators time required to view each loop (release operator time for higher-level duties).

We continually seek to find strong models of process behavior in the historical data. Models are important for process understanding. Rather than initiating model building, on occasion, by process engineers, the computer could continually observe data relations, test new observations with historical ones, and report both new "knowledge" and change of behavior.

I believe each of these system automations can have appeal to the variety of participants. They automate necessary process management actions that are a part of the bottom-line. The algorithms can standardize to a one-function-block-fits-all offering. Moreover, the new algorithms open research opportunities for academics.

ABOUT THE AUTHOR

R. Russell Rhinehart (rrr@okstate.edu) is Head of the School of Chemical Engineering at Oklahoma State University. He has 13 years of industrial experience and 21 years in academia.