In this portion of process control trends, let's discuss the need for gradually changing the very basis of our control mentality, which today is still based on the single PID loop. It is time for us to start thinking in terms of multivariable control. Right now, many still approach the control of processes by evaluating the dynamic response of one controlled variable to one manipulated variable (in terms of the dead time, time constant and gain of the response). We treat this single loop, as if it operated in vacuum, as if the opening or closing of a control valve affects only the variable we are controlling.
It is also high time we pay attention to the interaction among all the controlled loops of a process and to consider the relative gains of these interactions. Because the plants we work in do not produce flow, level or temperature; it stands to reason that the control of these variables should not be the ultimate goal. In short, the single loop mentality is wrong. It’s wrong because the ultimate goal is operating safety and production efficiency.
The time has come for the process control profession to start thinking in terms of Unit Operation Controllers. It’s time to develop multivariable systems that encompass and control the total unit’s operation (be it a boiler, a distillation column or a compressor). In such a unit-operations control scheme, controllers manage the interactions between the various controlled and manipulated variables that are built into the control strategy for that process.
Unit Operations Control
All expert systems take advantage of the large memory and the fast data manipulation capability of computers. There are hundreds of expert systems on the market and we should understand what these systems can and cannot do. One common feature expert systems share is that they all serve some form of optimization. What distinguishes these systems is the type of optimization they perform and the methods used to perform it. From this perspective, one can distinguish model-based from model-free methods.
"It is high time we pay attention to the interaction among all the controlled loops of a process and to consider the relative gains of these interactions."
Model-based control (MBC), model-predictive control (MPC) and internal-model control (IMC) are all suited for optimizing well-understood unit processes such as heat transfer or distillation. The performance of these model-based methods is superior to that of the model-free systems, because they are capable of anticipating and responding to new situations. In this sense performance is similar to that of feed-forward control systems, even though the model free systems behave in a feedback manner only.
Model-free expert systems can be compared to the behavior of a tennis player. The tennis player does not necessarily understand
Neural networks, fuzzy logic and statistical process control are all such methods, which can be used without the need for knowing the mathematical model of the process. The major difference between fuzzy logic and neural networks is that the latter can only be trained by data, but not with reasoning. Fuzzy logic is superior from this perspective, because it can be modified both in terms of the gain (importance) and also in terms of the functions of its inputs.
The main limitations of all model free expert systems is their long learning period, (which can be compared to the growing up of a child) and the fact that their knowledge is based solely on past events. Consequently, they are not prepared to handle new situations and therefore, if the process changes, they require retraining, because they are not well suited to anticipation.
Artificial Neural Networks
One of the tools used in building internal models is the Artificial Neural Network (ANN), which can usually be applied under human supervision or integrated with expert and/or fuzzy logic systems. The figure shows a three-layer ANN, which serves to predict the distillate boiling point and the Reid vapor pressure of the bottoms product of a column. Such predictive ANN models can be valuable, because they overcome the limitations of analyzers, which include both availability and dead time.
Model Prediction Tool
Artificial Neural Networks (ANNs) are well-suited
to build internal models. The three-layer ANN shown here
predicts the distillate boiling point and Reid vapor pressure.
The process model’s knowledge is stored in the ANN by the way the processing elements (nodes) are connected and the importance that is assigned to each node (weight). The ANN is "trained" by example and therefore it contains the adaptive mechanism for learning from examples and to adjust its parameters based on the knowledge that is gained through this process of adaptation. During the "training" of these networks, the weights are adjusted until the output of the ANN matches that of the real process. Naturally, these networks do need "maintenance", because process conditions change and when that occurs the network requires retraining. The hidden layers help the network to generalize and even to memorize.
The ANN is capable to learn input/output relationships and inverse relationships and hence it is useful in building IMC based on the ANN constructed plant models and their inverses. In a neural controller the ANN is used in calculating the control signal.
Consider applying model free herding control when a large number of variables are involved. This approach to control can be compared to the herding of sheep, where the shepherd’s dog goes after one animal at a time and does that on the basis of pursuing some overall goal. For example, if the herd is moving too slow, it will urge on the slowest animal to move faster, or, if a change in direction is needed, it will change the direction of the sheep that is farthest away. Because this control strategy was successful in returning herds of Hungarian sheep to the church of each village at exactly the same time every day, I named this strategy the "Hungarian Puli algorithm."
Such herding algorithms were successfully applied in designing the controls of the new headquarters building of IBM ("Lessons Learned" Nov. 2003, p24), which was made self-heating by herding the warm air from the offices which became the heat generators in winter (interior offices) to the perimeter, where heat is needed (offices with windows). This was done by looking at the thousands of offices and changing the destination of the return air from one "hot" office at a time (the one which is the warmest in the building) and simultaneously opening the supply damper of the "coldest office" to that same header.
I will not be surprised, to find that, if by the end of this century, we will all be using self-teaching computers. These devices would be watching the operation of a refinery or the landing and take-off of airplanes and eventually would possess as much knowledge as an experienced operator or pilot.
Some might argue that this would be a step forward, because machines do not forget, do not get tired, angry or sleepy. While this scenario might come true, I would still prefer to land in a human-piloted airplane because the knowledge of the past, which is the knowledge of the computer, is not enough to bring it safely to earth.
Bla Liptk, PE, process control consultant can be reached at Liptakbela@aol.com.