Just Curious, how many of you have changed the PID gains on-the-fly?

[kamenges]

Originally posted by rdrast:

I've found many times that since the transfer functions of the heating vs. cooling in a zone are so wildly different, it makes sense to have separate gains for each.

This would be similar to Peter having a separate gain set for hydraulic extend versus retract.

Did you also change the gain with changes in temperature setpoint or did you operate in the same setpoint range closely enough that it didn't matter? This would be like Peter changing gain based on cylinder position. I don't know if he does but I could see the need if the cylinder is very long.

Keith

 

[Peter Nachtwey]

I've found many times that since the transfer functions of the heating vs. cooling in a zone are so wildly different, it makes sense to have separate gains for each.

That would only make sense if you have an active cooling system and separate heaters. If cooling is done only by letting the heat decay to ambient you shouldn't need a separate PID gain.

Then there are the two features I love to see on ANY PID control, Integral HOLD and Integral RESET.

Yes but they need to be used with care. They shouldn't be used under normal conditions.

This would be similar to Peter having a separate gain set for hydraulic extend versus retract.

Only if the temperature system has a heater AND a cooler.

Gains do change as a function of temperature with heat exchangers. The heat exchange is governed by the LMTD or log mean temperature difference. A few years ago this topic came up on the www.controlguru.com website but Doug Cooper ducked this issue by tuning the system to work at only one temperature.

Keith, the hydraulic gains don't change as a function of position unless there is something external affecting the system. The natural frequency does change as a function of cylinder position. Natural frequency is lowest in the middle of the cylinder. We could change the gains as function of cylinder position but you would be surprised how many people don't know the specifications for there own cylinders. If necessary we have user programs that can change the gains on-the-fly for the advanced users.

 

[rdrast]

The heater controls that I have different gains are do have active cooling (normally water or oil loop cooling), and electric heaters. In some cases, we also control the temperature of the cooling loop supply, but that is rare.

I've not had to adjust gains based on setpoint, in any case I can remember.

@Peter - Integral reset is one I use often on large, position controlled accumulators, and sometimes on multi-pass dancers. If the line is stopped, proportional only error is enough to get close, and avoid integral creep. Once the line is moving, I re-enable the integral term and let it fly.

 

[kamenges]

Originally posted by Peter Nachtwey:

We could change the gains as function of cylinder position but you would be surprised how many people don't know the specifications for there own cylinders.

Wow. And you want people to give you plant transfer functions? I can see why you get frustrated sometimes.

Originally posted by rdrast:

If the line is stopped, proportional only error is enough to get close, and avoid integral creep. Once the line is moving, I re-enable the integral term and let it fly.

We do the same thing but with load cell based systems with velocity control. Given the amount of "feed forward" we do we often run dancer systems with velocity actuators without any integral gain at all. If your dancer is designed right the tension doesn't change much through the dancer travel so a small position error really doesn't mean anything.

Keith

 

[Peter Nachtwey]

Wow. And you want people to give you plant transfer functions? I can see why you get frustrated sometimes.

Yes, if the designers can't provide a transfer function or differential equation then they haven't really designed the system and have no f------ idea if or how their system will work. They expect the poor PLC guy in the field to find out and fix it in software. They don't care if the PLC guy wastes his time or that our time gets wasted. Then you get back to it is what it is, what ever that is.

We do the same thing but with load cell based systems with velocity control. Given the amount of "feed forward" we do we often run dancer systems with velocity actuators without any integral gain at all.

You can do that if your feed forward gains are precise and the position error doesn't make any difference. The feed forward gains are the inverse of the plant model so if you have good feed forward gains as a function of the radius you also have a good idea of what the plant model is as a function of the radius. You have be able to calculate the feed forwards as a function of radius and do it accurately since you don't use the integrator gain. However, I still recommend using the integrator gain and here is why. One trick I use is to plot the integrator's contribution to the control output. If the feed forwards are perfectly tuned then the integrator's contribution will be zero or near zero. If the integrators contribution to the control output is non-zero that means you have a problem with the feed forwards in that part of the motion profile.

I can calculate a the closed loop gains using the feed forward gains and a desired response. Tuning up the feed forward gains is relatively easy if you can compare the target and actual positions or velocities on the graph. It is easy IF the system is linear.

As I said I above I am writing an article about auto tuning. What I have found is that well designed systems are easy to tuning either manually or automatically. Poorly designed systems are difficult to tune either manually or automatically so too often when the auto tuning works you really didn't need it anyway and when it doesn't work is when you need it most. Sometimes the non-linear part can't be avoided but the designers should at least provide some information about this to the control guy.

Often auto tuning doesn't work because the system is non-linear so the tuning must be done in little segments that can be approximated as linear. It is apparent that few put in that much effort. My experience with our customers is similar to the result I have got from this thread.