When not to use PID-controllers

[ecwai]

Hello. How do you know when you cannot apply PID-controllers to a system? Have you had an experience where you first thought that a PID-controller might work but eventually find it impossible? How would you decide?

 

[Skogsgurra]

My experience (drives and general electrical control) is that you seldom need a PID controller. PI is what you need in most cases. Even P controllers do their job if the plant is integrating (like a level control).

What are your applications? You cannot just say that PID is needed or not. It depends a lot.

Gunnar Englund

http://www.gke.org

--------------------------------------
100 % recycled posting: Electrons, ideas, finger-tips have been used over and over again...

 

[ecwai]

I was asking it as a general question. I don't have a specific application in mind. By "PID controllers" I meant the family of controllers with P and/or I and/or D controls (i.e. it includes P, PI, I-PD, etc).

I can only think of a couple cases:

1) when I need the controller to increase or decrease the dB/dec by more than 20 or -20, I know that I can't use a PID-controller;

2) if I need to extend the bandwidth, a PID controller doesn't give me the zeros to use.

But these two cases seem obvious that PID controllers cannot be used. I was wondering what case you might have where it is not as obvious.

 

[itsmoked]

For starters I don't really get your two examples...

1) dB or dec isn't an output I'm familiar with on PID controllers..

2) And again, Bandwidth isn't a common term to use in the same sentence with PID unless it's a discussion about sensor response, etc.

But a system like:
1) A batch-like process where input/s might be lumped.
2) Horribly non-linear systems like ph control.
3) A system dependant on noisy inputs.
4) A system with easy to wear out actuators.
5) A system that would benefit from predictive control.
6) A system with an easy to model response.



Keith Cress
Flamin Systems, Inc.-

http://www.flaminsystems.com

 

[ecwai]

Sorry, I was thinking of meeting gain margin and phase margin requirements but I didn't mention the context in my examples. In those two examples, one would use a lead-lag or a series of them to shape the response, because PID controllers don't give zeros to use. If I derive the frequency requirements from the time-domain requirements, and I see that the shape I want in the frequency domain is complicated then I know that I cannot use a PID-controller.

What would you use for (3) when the actuators are worn easily? What are the grey areas in terms of applications where PID-controllers work sometimes only?

- Thanks

 

[PNachtwey]

Hello. How do you know when you cannot apply PID-controllers to a system? Have you had an experience where you first thought that a PID-controller might work but eventually find it impossible? How would you decide?

I count the poles in the plant.

My experience (drives and general electrical control) is that you seldom need a PID controller. PI is what you need in most cases. Even P controllers do their job if the plant is integrating (like a level control).

Based on what?

2) And again, Bandwidth isn't a common term to use in the same sentence with PID unless it's a discussion about sensor response, etc.

Why not? I use it all the time.

Peter Nachtwey

 

[Skogsgurra]

Peter,

Your second question.

I base that on successful applications in paper machines and steel mills for more than forty years.

Gunnar Englund

http://www.gke.org

--------------------------------------
100 % recycled posting: Electrons, ideas, finger-tips have been used over and over again...

 

[JLSeagull]

On some small flow nasty applications like sour water boot I often apply on/off or PID with gap to avoid operating with a control valve near the seat.

 

[PNachtwey]

I base that on successful applications in paper machines and steel mills for more than forty years.

Good, but that is application specific and applies to the few applications in a paper or steel mill. How do you prove that a PI, PID is the best for the job?
How did you compensate for dead time in a paper mill. Are you saying you used only a simple PI? No model or Smith Predictor?

Even P controllers do their job if the plant is integrating (like a level control).

What about an integrating application like position control. Is a P gain good enough there? Why or why not?

I often come across systems that require a second derivative gain along with velocity, acceleration and jerk feed forwards. Obviously a standard simple PID alone will not do. It must be augmented in some way. Even knowing when to use a PID vs a I-PD can be a big help.

ecwal's question is very open ended and is good for a class room discussion. There are so many different types of systems out there that use different types of control. I am waiting for ecwal to explain why he needs bandwidth and why the PID didn't work. This makes me believe he had a motion system in mind but he hasn't made that clear yet.

BTW, did anybody look at thread just below this one about different PIDs?

 

[ScottyUK]

Yes, that was a good thread.


----------------------------------

Sometimes I only open my mouth to swap feet...

 

[PNachtwey]

The reason I asked is that one can see that a PID has one less zero than pole so the phase lag is only 90 degrees whereas a PI-D has a phase lag of 180 degrees and a I-PD has phase lag of 270 degrees. I don't see what point ecwai is trying to make.

If one wants more bandwidth one must just move the closed loop poles farther to the left or make them more negative without have the zeros mess things up. When zeros are complex they can cause the gain to drop sharply like a notch filter.

 

[ecwai]

I thought that PID controllers as a family can only give at most 90 degrees phase lag. I did not consider that PI-D or I-PD would give more.

In some cases of robust control where some parameters of the plant may vary, the boundary of the closed-loop gain takes the shape of a cylinder on the Nichols chart. In the design, I would use zeros and poles on the controller to make shape the open-loop response. When I don't have enough zeros and poles to use, I would have to increase the gain in the middle in order to make it fall later outside the cylindrical boundary.

So suppose I have requirements on the gain margin or phase margin, the open-loop plots can tell me whether a controller from the PID family can stabilize the system. However I assumed that the PID family can only give at most 90 degree phase lag.

 

[sreid]

It's a very good question and hard to answer. Perhaps it's better to ask what kinds of systems PID can control.

1) One and two pole systems.
2) Single and double integrator systems.
3) Both of the above with moderate delays
4) Both of the above with a slow Right Half Plane Zero.
5) All of the above with moderate non-linearities.
6) All of the above with moderate time dependent changes.

The above list is not meant to be all inclusive.

 

[jraef]

I count the poles in the plant.

What does the number of Polish workers have to do with PID control? Were you thinking PID means Polish Immigration Deterrence?

Sorry. Couldn't resist. I was really trying to figure out what your comment meant, but that version of the question was nagging me even after I figured it out. Had to get it off my mind by dumping it here!

 

[PNachtwey]

funny but I have heard these kinds of jokes before.

http://wilk4.com/humor/humore30.htm

See the joke about more poles.

 

[telmatrix]

For what it is worth... PID controllers remain the most widely used controller in industry today, primarily because they are simple, and usually easy to tune due to the reduced number of parameters that required consideration. There are several different classifications of PID algorithms which may have additional variations of theme dependent on the design engineer, the nature of the plant to be controlled, and the desired performance.

However someone mentioned above that a PI controller is more likely to be used than a PID. This is generally true because the "real" world is pretty noisy environment. The D term essentially behaves as a high pass filter on the error signal. Consequently, though D-control can be beneficial, it should be implemented judiciously as it easily introduces instability in a system and makes it more sensitive to noise.

Hope this helps...

 

[Skogsgurra]

Following up a few loose ends...

The OP wasn't very specific. The question could come from a university or a guy practising at the shop floor. I told her/him my opinion.

Peter asked on what I based that opinion. It is mostly drive systems. And some other applications. But I never do anything where a long dead-time gets into the loop (like g/m2 control or anything like that). Most of the loops I see are rather simple. A sum of short times plus a first order delay or an integrator. And that's where PI works well.

Feed forward is used a lot, but I do not see that as a controller characteristic. It, as its name says, a parallel path that takes care of large changes. Acceleration compensation in a winder is one such application where the feed-forward signal usually is totaly dominant. The web tension controller usually doesn't move much when the feed-forward is correctly adjusted and the inertia calculation is OK.

I have been thinking. How come that there is so much talk about different controller types when we do not seem to need them IRL? I guess that it is because most practical solutions are using cascaded loops where each loop controls its own variable. Take a drive for instance. There is the inner current loop (or torque loop), then the speed loop controlling the current set-point, and around that a web tension, or level control or air pressure or whatever, loop.

Dividing and conquering like this seems to be the rule today. And then PI works very well then.

I can understand that there are many exotic plants that need unusual controllers. But I have to admit that I have been lucky to not have seen them. Or I have been able to handle them using simple techniques. Time varying and highly non-linear plants are something I either try to avoid or use an observer/model for. But, I have to admit again, that is also something I mostly leave for the guys interested in that kind of problems.

Gunnar Englund

http://www.gke.org

--------------------------------------
100 % recycled posting: Electrons, ideas, finger-tips have been used over and over again...

 

[PNachtwey]

"Peter asked on what I based that opinion. It is mostly drive systems."
OK, but you must realize that drive systems are just one of many types of systems that are controlled by PIDs or PIs. What bothers me is that people make statements about how to tune the PID often without knowing anything about the system that needs to be tuned. I have worked out the formulas for tuning many different types of systems and there is a general theme but the details about whether to use or not use a derivative gain or even a second derivative gain is dependent on the model of the system.

Noise should not prevent one from using the derivative gain. When I am confronted with this I need to know the source of the 'noise'.
1. Is it electrical?
2. Is it poor feedback resolution?
3. Is the process itself noisy?

There is no excuse for the first two reasons. The third reason can be over come by using a filter, an observer or even a Kalman type of filter.

About cascaded loops.

One should only use a cascade loop when there is a different and more accurate feedback device available for the inner loops. It makes no sense to differentiate an encoder to get feedback for an inner velocity loop. This is because the inner and outer loops can be combined to make a simpler controller.

Think about this"
The proportional gain for the outer loop is doing the same thing as the integrator gain for the inner velocity loop. The proportional gain for the inner velocity loop is doing the same thing as the derivative gain in a outer position loop.

 

[Skogsgurra]

Yes. I understand that there are many different systems. But, remember that the OP question was very general (" How do you know when you cannot apply PID-controllers to a system?") and that it was asked about experience.

That said, I must comment on what you say about making controllers simple. "This is because the inner and outer loops can be combined to make a simpler controller."

That is, in my view, a reasoning that leads to complex solutions. It is then that you need other controller structures than the simple PI, sometimes PID, controller with the problems it brings with it.

The need to make the controller "simple" was perhaps real back in the forties when every controller did contain a set of thermionic valves that had low realiability. Combining loops to make the controller "simple" (less valves) was a way of keeping MTBF up (and cost and heat down).

That trend was still strong during the sixties but abandoned in the seventies when the integrated opamps became available. And, today, all these things are done in algorithms, so the "cost", reliability-wise, for an extra loop is not high.

The cost for complex combined loops, that are poorly designed and that do not correspond to what is happening in the plant, is high though.

It is a fascinating area for research and for scientific papers, but it should be applied to real world systems with some carefulness and restriction. I live in Sweden and it is close to impossible not to meet guys like C-J Åström, P Ljung, Gunnar Bengtsson etc and I have had discussions about these topics with them. I have also been involved in the rebuild of a few autotuning projects where the "combined loop economy" made a paper machine's runnability close to zero.

There are many ways of looking at these things. And all are more or less valid. I base my view on what I have lived through - and continue to live through - during forty+ active years in industry.

Gunnar Englund

http://www.gke.org

--------------------------------------
100 % recycled posting: Electrons, ideas, finger-tips have been used over and over again...

 

[telmatrix]

PNachtwey... You said:
"Noise should not prevent one from using the derivative gain."

Noise can have lots of origins, and it may be cost-prohibitive to address it completely via "filters". For example, mechanical noise, which leads to torque oscillations, which reverbrate into the electrical system (or vice versa). Depending on the type of system, addition of mechananical damping could be expensive or may not really even be feasible. That said, certainly a "filter" on the electrical side can help, and prudent to implement, as will reduce some harmonic content, but will not remove all. If your physical system is relatively simple, "noise" may not be terribly significant. However if you are dealing with huge loads turning on/off such as in a shipboard power distribution system, electro-mechanical-hydraulic automotive cam phaser, etc. the "noise" is inevitable and inescapable, and will definitely may limit implementation of the "D" term of a PID controller.