|
|
|
Control Overview
The process industries, ranging from chemical plants to paper manufacturing, from plastics to refining and all stops in between, typically use tried and trusted algorithms in order to control the plant. Before looking in detail, let us look at how most plants are controlled. This basically boils down to reactive control using PID (Proportional, Integral and Derivative) action controllers, 3-term control. Often only 1 or 2 of these actions are used, depending on the requirements of the process. A number of tools and methodologies are required to ensure plant stability, as a PID controlled process has the possibility of going into oscillation. Such a controller has 3 signals, and only 3; these are the set-point [SP] (where the operator wants the particular variable to be, such as pressure), the measured value [MV] (where the actual process really is) and output [CO] (the output from the controller which is required to balance the MV and SP). Controllers can be set up in different ways, and cascaded together to provide more complex control. However they are setup, they experience common features.
Here the operator sets the desired flow from a control dial. The flow transmitter sends a flow related signal to the controller. To balance the flow (the difference in MV and SP should be zero) the control output adjusts a valve, which opens just enough to give the desired flowrate.
It can be seen that, in order to cause the valve to change position, an error must first appear at the measuring element. This raises some important points:
We are launched into the world of tolerance and Process Capability. If there is a large time delay, the controller will first output the desired new output. But, for a while, nothing happens and so the controller outputs a stronger and stronger signal. Eventually the change starts to filter through at the sensors, but too late as now the controller has altered the valve too much, and it must move in the opposite direction. This is called overshoot and is a feature of this type of control. The degree of overshoot may affect product quality, hence strict tolerances must be defined and warning signals given to the process operators. It is possible to set the controller parameters to reduce or remove the overshoot but at the expense of response speed so, as ever, a compromise is made as to the optimal settings.
It is clear that, during the disturbance, control is poor
In the second scenario, although local control is not a problem, the changes in the rest of the plant cause the local situation to vary. It can be very difficult to stabilise such an interactive system using this type of control.
There are ways to reduce the transient errors, but we will look at those later. In the meantime, let's look at some data: (units are arbitrary)
Anything funny about this? Well, the SP hasn't been changed and the MV is remarkably stable (because its an illustration!), which is as it should be. The controller is doing a wonderful job and there is no reason for alarm. Or is there? And what is any of this to do with the environment? Actually, there is a problem and it does affect the environment in two main areas: energy and waste. That's forgetting the cost of the plant being shut down and all that overtime putting it right! If you look carefully at time period 10 you will see a small shift in CO. Not much, but it's there. In fact the mean value of periods 1 to 9 is 39.6 and that of 10 to 18 is 42.8. Still not worried? Suppose I say that the standard deviation of periods 1 to 9 is 1.02? Ah! You're looking now - the later average is over 3 standard deviations away from the earlier results (3.23 actually). This means that there is, all else being equal, only a 0.06% likelihood that something is not wrong, or at least needs investigation. Your controllers would never have picked that up. But, again, what's this to do with the environment? Well, it's like this - if you had known that an anomaly had occurred, before your alarm system detected a major limit exceeded, you would have had time to investigate and probably correct it before everything went wrong. If a manufacturing limit is exceeded, you create waste, which requires disposal or rework. In either case, it costs you money and cuts into the bottom line.
What can be done? Process improvement! Solutions range from inexpensive to highly costly and you should select a solution with appropriate benefits. For example, in the instance above where traditional methods deliver poor control, then some form of model based control could be highly beneficial. There are a number of solutions around with differing degrees of applicability. Techniques range from detailed mathematical modelling through to empirical neural network applications. Each has their own strengths and weaknesses which need consideration prior to selection. For the statistical anomaly above, a relatively simple Statistical Process Control (SPC) package would have picked that up. These range from packages designed to work with a specific platform to generic Excel based software with wide applicability. There are all sorts of improvements possible - discuss them with your system vendor or an independent engineer. Remember - If its not measured, its not controlled. SPC software is not typically looking at the same things your controllers are looking at. SPC is aimed at the control and reduction of process variability and is a key element in Quality Control. The software analyses current behaviour and compares this with an historic database. Where known poor results were experienced in the past they should be excluded. Thus the relationships between signals (inputs and outputs), using techniques like sampling, averaging, comparison and even Principle Component Analysis, may be learnt and significant deviations from the norm detected. Packages typically come with average (X-bar), Range (R), standard deviation (sigma), CUSUM and EWMA plots. SPC and Process Capability are very much intertwined. X-Bar The average of a sample. SPC packages should allow the sample period or number to be defined. Over a number of samples the average of the X-Bars is often taken, called the X-bar-bar value. Range The range of a sample - the difference between the upper and lower readings. Like X-bar, an average over several sample is often useful, and called R-bar. (Note: The word "bar" comes from the mathematical notation of putting a bar, like a minus sign, above the letter. It is written longhand here as not all computers have the correct fonts installed for mathematical notation) Standard Deviation (SD) A widely used and extremely valuable measure of the variability of a set of data. For instance: (A) 2,2,2,2,2,2,2,2 (B) 2,3,2,1,2,3,2,1, The average of each set is 2. It can be seen that the second set has much more variation in though. (SD (A) = 0 and SD(B) = 0.7) CUSUM (Cumulative Sum) is simple in principle and typically compares the sum, over a period of time, of deviation of the process against the desired value. In a neutral situation there should be as many negative errors as positive ones and the sum should be zero on average, within limits. If there isn't, something is changing and may need looking at. EWMA (Exponentially Weighted Moving Average) Charts are more complex, but can rapidly detect out-of-control situations. Basically it creates values based on weighted past and current values and from that creates control limits. If a control limit is exceeded a warning notification can be made.
| |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Controller Overshoot
Waste costs money twice - creating it and disposing of it