Dear JCH!
Even if You gain no applause on this group, the target You aim to is
very good. And it's a dream for every control engineer to make it
working.
I fully disagree with You Peter, that we need transfer function to
control anything.Please answer yourself a question:
Does Formula 1 driver knows transfer function of anything?
This proves, that control without transfer function is possible.
AND THIS IS FUTURE - realtime system approximations!
Question to JCH?
How You estimate z1 and z2 disturbance. Is it a constant,time invariant
value?
robert
Get what working? JCH's examples ignored reality. The assumptions
weren't stated, units were not used, work wasn't shown. Doesn't this
bother you? Don't infinite gains bother you?
JCH has his transfer function, why do you single me out? I just am
trying to show a better way than tweaking gains and drinking coffee.
It takes a long time to train, tune, a Formula 1 driver and drivers
can't be duplicated. People can eventually be trained, tuned, to be
very good controllers but industry doesn't have the time. If I can
determine the system model in few minutes then I can save a lot of
time tuning a system by trial and error. Better yet, if I can write
an auto tuning program that saves my customers time and money.
Sure, but that isn't engineering that is guessing or trail and
error.
So must every end user learn how tune using trial and error? How much
time ( money ) with this take?
Isn't that a model or transfer function?
Peter Nachtwey
Ah, Peter! Thank you for answering "user". I was waiting until I could
muster the necessary patience and calm, and you saved me the effort. His
message was so full of misconceptions I didn't know where to begin.
Jerry
--
Engineering is the art of making what you want from things you can get.
¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯
Please take into account, that nowadays we don't have controller which
would compete with third rank driver.And that's because what You, Peter,
have mentioned: we have 5 minutes for tuning, and it should work for
decades.
5 minutes tuning is good to get money from Customer.They loop must tune
itself all the time.That's what I tempt to.
In one huge factory they purchased software for loop monitoring
(chemical process). And they found out, that even if they have control
engineer onsite, the 80% loops are badly tuned.
And installation was very fine, did fast, tuning took " 5minutes ".
Your point of view come from the ground, that because nobody counts
losses of mistuned loops, there are not losses.
All what I've said here could be rubbish for You, but I appreciate your
time.
This phrase didn't make sense.
And that's because what You, Peter,
I never said the tuning should last forever. Machinery changes slowly
as things wear. It shouldn't change like the weather. If it does then
there is something that needs repair.
How do you do that without a transfer function? One could use recursive
least squares. One could also make small changes in the gains and try
to find the gradient. This last method would be slow but it would work
without a model. We have used used that method in a more manual way.
So who's fault is that?
This phrase doesn't make sense.
that because nobody counts
Who's fault is that?
I don't see the point you are trying to make with this post.
Peter Nachtwey
It depends, what target we have. If we are short time runners You are
right. But if we consider long term performance You are wrong.
80% of loops I dealt with didn't have D term, because it makes problems
in the future.NOT DURING INSTALLATION (your 5 minutes).
Almost every of these loops performs better with D in mind.
That you must've experienced as well Peter.Everybody knows it.
But we still don't know good solution for this.
Maybe we talk about the same but different languages.
How am I wrong? What is wrong with tuning the system? It seems you
just want to argue without being specific
Why D term terms cause problems? You must be specific. What does this
have to do with me?
FOPDT, first order plus dead time systems do not require a derivative
gain. However, just about all of my motion systems do. First order
velocity systems don't. You have a limited view of the different types
of control systems.
NOT DURING INSTALLATION (your 5 minutes).
Are you suggesting that just because the plant changes one should use
the derivative gain? That is the wrong approach. One should try to
reduce the variations in the process if the gains change that much.
Again, you must be specific. It still seem like you just want to argue.
_____
Mr. User,
I just want to chime in on this thread for a moment. After studying
Control Engineering at Univ. of Manchester, and a 30 year career with a
large integrated oil refiner, i can say from solid experience that User
seems to be confused about the need for a process transfer function in
order to tune loops. Plant commissioning and loop tuning is a vast
subject, and much has been written about it. I also wrote a book on the
subject from a practical viewpoint for control engineers in the field.
In a nutshell, i see three broad categories of of tuning needs:
1) Commissioning phase of a process unit or units
2) Final tuning of recently commissioned loops
3) Ongoing monitoring and adjustments to tuning of loops as units age
For category 1) i recommend a table of conservative tuning constants to
get the plants up and running. Loops include flow, pressure,
temperature and so on. Category 2) requires that one obtains a process
transfer functions obtained from simple plant tests. For multi-variable
systems this can be quite a lot of work. After the transfer functions
have been obtained and verified, one can develop tuning constants using
a variety of well proven methods. I teach a graphical approach. The
majority of loops do not required derivative action. Category 3)
requires the same techniques for Category 2). It is a simple fact of
life that loops go "out of tune" when (a) the process ages due mainly to
fouling of equipment (heat exchangers), (b) throughput is permanently
increased or decreased (variable gain tuning comes to mind), (c)
debottle-necking which has combined effects of (a) and (b).
Trial and error tuning as used in the Ziegler-Nichols (and other)
methods is usually frowned upon by the process operator on duty in the
control room(s) because you will be causing the process to oscillate.
Similarly, plant step tests to get the transfer function can often be
difficult to do because one has to introduce a disturbance that can be
distinguished clearly from the normal noise signal.
I firmly recommend that one obtain a process transfer function
(First-Order-plus-Dead-Time [majority of processes], or
Second-Order-plus-Dead Time [few processes]) in order to derive suitable
tuning constants for the standard PID control algorithm. There are many
forms of PID algorithms, and one needs to know which form is used in the
DCS (sometimes PLC), or computer control system one is employing to
calculate the correct tuning parameters.
--
Regards / JCH (jch in Canada, not jch in Germany)
Hi Peter again!
I've just studied your paper (pdf) with crane's load example. Maybe I've
missed it, but where is the load weight factor in your PID loop.Does
this crane have fixed load?
The load isn't important as the damping factor and natural frequency
don't depend on it. What I did neglect to show is the load must not be
so large relative to the crane that the 'tail wags the dog'. I admit
that I made an assumption that the crane and move the load without the
swinging of the load affecting the crane too much. You are the first to
pin me down on that. At least you are paying attention. A better
simulation would take into account that it is force that moves the load,
not position. I didn't say this in one of my previous post. I was just
trying to make a point not make a work of art.
BTW, can you figure out how the load moves into position without swinging?
Peter Nachtwey
From the simple experiment (key swinging on the thread) You may see that
damping depends on load wieght.If the crane would cary 3 pensils, it
didn't care it's speed.
I don't know if your experiment is for lab purposes or industry.
I would start from observing all measurable system data, like for
example motor currents. motor current change (on trolley) is related to
rotational acceleration of the load.There is a full bucket of
information about the system.
all the best
Yes, the damping does vary due to air resistance and the length of the
cable. From other peoples research a damping factor of 0.04 is
typical but it can be as high as 0.1. There are a lot of .pdf files
on this.
I was just trying to show how one can control JCH's cart by a much
easier method. It wasn't complete but at least it didn't assume the
cart moved instantly as JCH's example did.
Yes, rotational and just the mass of the cart or trolly and the weight
and angle of the load. There are .pdf files on this that show how the
load affects the trolley. I took a short cut and assumed the cart
would not be affected too much by the load.
What about using accelerometer at the end of rope?
It simplifies everything when comparing load and trolley accelerations.
That way You can make the whole system more rigid to external
disturbances like wind, trolley track angle change,load swing direction
(these which are unpredictable) etc.
I would suggest self configurable state space model, then just a PID
based on transfer function estimation.
I real world implementations accelerometer can be built into the hook,
supplied with battery, and comunicating via radio freq. with controller
using rope as antenna(depending on wave ofcourse) .The cost would be
less then 10$
How does it simplify things more than what I have shown?
Yes, those are problems. I don't think the technique I used would
work unless the crane moved in simple x, y and z directions. Overhead
cranes to that.
OK, how do you do that? How would you improve my state space model?
I am sure there are systems like that. There are systems that use
cameras too.
So what would you do with it?
JCH assumed he had perfect position feed back and that the crane
position would go instantly to a position proportional to the control
voltage. How would you improve on what JCH did?
Peter Nachtwey
I think You realize, that swinging on the real crane is a 3-dimensional
process !!
Modern cranes don't go step by step, but move all directions all
together!!!
You need 3 PID loops with decoupling.
And not only cranes.
All depends on load range versus local frictions.
Have You tried to make control based on PV from motor loads?
Try to imaging how 5 year old kid would control your crane with a
joystick.Making i-pv method takes him 5 minutes.
All the best Peter
If you look at the graphs you can see I plot the actual position
( cart position ) and the load position.
I don't have a motor big enough but I have a hydraulic system that can
move a swinging load. The problem is the hydraulic system is very
powerful relative to any loads I can suspend from it so I don't think
I would learn much. It may make a good video though.
I-PV is just the form of PID that I used. The integrator is in the
forward path and the proportional and velocity feedback gains are in
the feed bacl path. The trick is how I modified the simple target
position = velocity times time motion profile. If you look at the
green line you can see the control output does not look normal for
following such a simple motion profile.
Notice also the difference in the cart position and the load
positions. The cart position is ahead of the load during the first
half of the move and behind the second half of the move.
Yes, the goal is to make it so a 5 year old can move the cart and the
load safely and quickly.
Peter Nachtwey
How do you design control without transfer function? Actually you
don't have to estimate anything, you cn measure them if they are
observable. Now, how do you know if something is observable without
put them into an observable canonical form. The question also applies
to controllability, stability, sensitivity, robustness and so on.....
I can't imagine what kind of engineering work this is. Especially
about robustness: will it work. Please show me if it is robust.
Boen S. Liong
Thank you. Why trial and error? Maybe something like adaptive,
linearization, or stochastic control maybe better. Trial and error? Is
it a method? A method call trial and error? Does it give a final
state? Final value?
I still don't understand the method call trial and error. Seems to me
very confusing.
Seems to me that if transfer function is not known, then we don't know
about the characteristics of the process. Anyway, it's very, very
confusing. And I don't think anybody will trust trial and error
method. Because it an error you make and causes unstable to the
system.
Recall that An Unstable System is Useless. So my conclusion is trial
and error is useless, because the system can be unstable.
- Boen S. Liong.
Polytechforum.com is a website by engineers for engineers. It is not affiliated with any of manufacturers or vendors discussed here.
All logos and trade names are the property of their respective owners.