PID tuning. Lag = ~ integrating time

G

George Herold

Guest
Hi all, this is not so much a question, as me just spouting.
If I do say (spout) something wrong, or that you disagree with
or that you don't understand, then please do speak up.

I've been tuning loops for years. (By tuning I mean picking
some reasonable starting values, all loops IME need a little
real time tweaking from the starting values.) I've always
used the Zeigler-Nichols oscillation method, and looked no further.

Now my boss asked for some other technique and I started reading about
tuning from the step response. (What Z-N call the Process-reaction curve)
Oh I stuck the Z-N paper here, along with a few 'scope shots.

https://www.dropbox.com/sh/7lkqazuy6wvzknk/AADcZtTvFvaD3TEzeJiAjEy6a?dl=0

If only I'd read the paper ~20 years ago.

This is a more up to date rehash of the same things.
http://faculty.mercer.edu/jenkins_he/documents/TuningforPIDControllers.pdf

For completeness I'll stick in Tim W's excellent article.
(Though unfortunately he doesn't talk about Z-N directly.)
http://www.wescottdesign.com/articles/pid/pidWithoutAPhd.pdf
(Maybe he'll add an appendix for me. :^)

So if you look at the step response in my dropbox link, there is a lag
of 300-350 ms. And if you look at the oscillations at the "ultimate gain"
you'll see a period of about 2 seconds. And lo and behold,
2*pi*tau (the lag) = period! That is very satisfying.

Estimating the needed gain (proportional term) from the step response,
didn't work out that well in this case. But I'm running at gains
that are ~20% the ultimate gain, so oscillation method didn't work all
that well either in this regard.


Oh I do have one question for Tim.
You show a thermal control loop with just integrating control.
Does this really work? (I would have thought it would just oscillate.)
Or do you have a little bit of proportional term in there?

Cheers,

George H.
 
On Mon, 24 Apr 2017 12:29:24 -0700, George Herold wrote:

Hi all, this is not so much a question, as me just spouting.
If I do say (spout) something wrong, or that you disagree with or that
you don't understand, then please do speak up.

I've been tuning loops for years. (By tuning I mean picking some
reasonable starting values, all loops IME need a little real time
tweaking from the starting values.) I've always used the
Zeigler-Nichols oscillation method, and looked no further.

Now my boss asked for some other technique and I started reading about
tuning from the step response. (What Z-N call the Process-reaction
curve) Oh I stuck the Z-N paper here, along with a few 'scope shots.

https://www.dropbox.com/sh/7lkqazuy6wvzknk/AADcZtTvFvaD3TEzeJiAjEy6a?
dl=0

If only I'd read the paper ~20 years ago.

This is a more up to date rehash of the same things.
http://faculty.mercer.edu/jenkins_he/documents/
TuningforPIDControllers.pdf

For completeness I'll stick in Tim W's excellent article.
(Though unfortunately he doesn't talk about Z-N directly.)
http://www.wescottdesign.com/articles/pid/pidWithoutAPhd.pdf (Maybe
he'll add an appendix for me. :^)
Click to expand...

Nope. For two reasons:

First, because I've never used Z-N, never felt the need, and so never
built up practical experience with it.

Second, because all the industrial control guys I've talked to, both in
person and 'lectronically, tell me that Z-N gets you a good starting
point but then you need to either tweak things, do measurement + system
ID + design, or do frequency sweeps + design. There's enough material
out there on it, and too many people that think that it's the bees knees.

So if you look at the step response in my dropbox link, there is a lag
of 300-350 ms. And if you look at the oscillations at the "ultimate
gain"
you'll see a period of about 2 seconds. And lo and behold,
2*pi*tau (the lag) = period! That is very satisfying.

Estimating the needed gain (proportional term) from the step response,
didn't work out that well in this case. But I'm running at gains that
are ~20% the ultimate gain, so oscillation method didn't work all that
well either in this regard.


Oh I do have one question for Tim.
You show a thermal control loop with just integrating control.
Does this really work? (I would have thought it would just oscillate.)
Or do you have a little bit of proportional term in there?
Click to expand...

If you keep a firm lid on your ambitions, certainly. For most systems
you won't get nearly the speed you'd get with some proportional action or
even proportional + derivative -- but any system that has a flat gain
characteristic going down to zero frequency can be controlled with just
an integrator, if you're patient.

The one thermal loop that I've done had a simple first-order response
well beyond any reasonable loop bandwidth. So it got a simple PI
controller and worked just grand.

--

Tim Wescott
Wescott Design Services
http://www.wescottdesign.com

I'm looking for work -- see my website!
 
On Monday, April 24, 2017 at 8:52:04 PM UTC-4, Tim Wescott wrote:
On Mon, 24 Apr 2017 12:29:24 -0700, George Herold wrote:

Hi all, this is not so much a question, as me just spouting.
If I do say (spout) something wrong, or that you disagree with or that
you don't understand, then please do speak up.

I've been tuning loops for years. (By tuning I mean picking some
reasonable starting values, all loops IME need a little real time
tweaking from the starting values.) I've always used the
Zeigler-Nichols oscillation method, and looked no further.

Now my boss asked for some other technique and I started reading about
tuning from the step response. (What Z-N call the Process-reaction
curve) Oh I stuck the Z-N paper here, along with a few 'scope shots.

https://www.dropbox.com/sh/7lkqazuy6wvzknk/AADcZtTvFvaD3TEzeJiAjEy6a?
dl=0

If only I'd read the paper ~20 years ago.

This is a more up to date rehash of the same things.
http://faculty.mercer.edu/jenkins_he/documents/
TuningforPIDControllers.pdf

For completeness I'll stick in Tim W's excellent article.
(Though unfortunately he doesn't talk about Z-N directly.)
http://www.wescottdesign.com/articles/pid/pidWithoutAPhd.pdf (Maybe
he'll add an appendix for me. :^)

Nope. For two reasons:

First, because I've never used Z-N, never felt the need, and so never
built up practical experience with it.

Second, because all the industrial control guys I've talked to, both in
person and 'lectronically, tell me that Z-N gets you a good starting
point but then you need to either tweak things, do measurement + system
ID + design, or do frequency sweeps + design. There's enough material
out there on it, and too many people that think that it's the bees knees.
Click to expand...
Huh, Ok I only said it was my starting point. Then do some step responses
and see what it looks like. (I'm mostly an idiot when it comes to
thermal loops... I'm starting to think that 1/2 the time I got lucky
and put the sensor in a good place, and the other times, not so good.)
Well sometimes you just need to put the sensor at the point you care
about.... I could think more about where to put the heater.
Grumble....

What's your starting point?
So if you look at the step response in my dropbox link, there is a lag
of 300-350 ms. And if you look at the oscillations at the "ultimate
gain"
you'll see a period of about 2 seconds. And lo and behold,
2*pi*tau (the lag) = period! That is very satisfying.

Estimating the needed gain (proportional term) from the step response,
didn't work out that well in this case. But I'm running at gains that
are ~20% the ultimate gain, so oscillation method didn't work all that
well either in this regard.


Oh I do have one question for Tim.
You show a thermal control loop with just integrating control.
Does this really work? (I would have thought it would just oscillate.)
Or do you have a little bit of proportional term in there?

If you keep a firm lid on your ambitions, certainly. For most systems
you won't get nearly the speed you'd get with some proportional action or
even proportional + derivative -- but any system that has a flat gain
characteristic going down to zero frequency can be controlled with just
an integrator, if you're patient.

The one thermal loop that I've done had a simple first-order response
well beyond any reasonable loop bandwidth. So it got a simple PI
controller and worked just grand.
Click to expand...
Huh, OK I'll have to try, the loss is in the delay/lag?
This thermal loop is fast, (at least for me)
so ten seconds is no problem.

George H.


--

Tim Wescott
Wescott Design Services
http://www.wescottdesign.com

I'm looking for work -- see my website!
Click to expand...
 
On Mon, 24 Apr 2017 19:22:31 -0700, George Herold wrote:

On Monday, April 24, 2017 at 8:52:04 PM UTC-4, Tim Wescott wrote:
On Mon, 24 Apr 2017 12:29:24 -0700, George Herold wrote:

Hi all, this is not so much a question, as me just spouting.
If I do say (spout) something wrong, or that you disagree with or
that you don't understand, then please do speak up.

I've been tuning loops for years. (By tuning I mean picking some
reasonable starting values, all loops IME need a little real time
tweaking from the starting values.) I've always used the
Zeigler-Nichols oscillation method, and looked no further.

Now my boss asked for some other technique and I started reading
about tuning from the step response. (What Z-N call the
Process-reaction curve) Oh I stuck the Z-N paper here, along with a
few 'scope shots.

https://www.dropbox.com/sh/7lkqazuy6wvzknk/AADcZtTvFvaD3TEzeJiAjEy6a?
dl=0

If only I'd read the paper ~20 years ago.

This is a more up to date rehash of the same things.
http://faculty.mercer.edu/jenkins_he/documents/
TuningforPIDControllers.pdf

For completeness I'll stick in Tim W's excellent article.
(Though unfortunately he doesn't talk about Z-N directly.)
http://www.wescottdesign.com/articles/pid/pidWithoutAPhd.pdf (Maybe
he'll add an appendix for me. :^)

Nope. For two reasons:

First, because I've never used Z-N, never felt the need, and so never
built up practical experience with it.

Second, because all the industrial control guys I've talked to, both in
person and 'lectronically, tell me that Z-N gets you a good starting
point but then you need to either tweak things, do measurement + system
ID + design, or do frequency sweeps + design. There's enough material
out there on it, and too many people that think that it's the bees
knees.

Huh, Ok I only said it was my starting point. Then do some step
responses and see what it looks like. (I'm mostly an idiot when it
comes to thermal loops... I'm starting to think that 1/2 the time I got
lucky and put the sensor in a good place, and the other times, not so
good.) Well sometimes you just need to put the sensor at the point you
care about.... I could think more about where to put the heater.
Grumble....

What's your starting point?
Click to expand...

I usually have an idea of what the DC response of the loop is (flat,
integrating, double integrating), and some idea of the initial gain and
rolloffs. If not, then I take a step response and eyeball it (OK, kind
of your Z-N, only much more casual). Once I have the system stable and
staying within bounds, I put a frequency sweep on it and do Bode plot
design.

So if you look at the step response in my dropbox link, there is a
lag of 300-350 ms. And if you look at the oscillations at the
"ultimate gain"
you'll see a period of about 2 seconds. And lo and behold,
2*pi*tau (the lag) = period! That is very satisfying.

Estimating the needed gain (proportional term) from the step
response, didn't work out that well in this case. But I'm running at
gains that are ~20% the ultimate gain, so oscillation method didn't
work all that well either in this regard.


Oh I do have one question for Tim.
You show a thermal control loop with just integrating control.
Does this really work? (I would have thought it would just
oscillate.)
Or do you have a little bit of proportional term in there?

If you keep a firm lid on your ambitions, certainly. For most systems
you won't get nearly the speed you'd get with some proportional action
or even proportional + derivative -- but any system that has a flat
gain characteristic going down to zero frequency can be controlled with
just an integrator, if you're patient.

The one thermal loop that I've done had a simple first-order response
well beyond any reasonable loop bandwidth. So it got a simple PI
controller and worked just grand.
Huh, OK I'll have to try, the loss is in the delay/lag?
This thermal loop is fast, (at least for me)
so ten seconds is no problem.

George H.



--

Tim Wescott Wescott Design Services http://www.wescottdesign.com

I'm looking for work -- see my website!
Click to expand...




--
Tim Wescott
Control systems, embedded software and circuit design
I'm looking for work! See my website if you're interested
http://www.wescottdesign.com
 
On Mon, 24 Apr 2017 12:29:24 -0700 (PDT), George Herold
<gherold@teachspin.com> wrote:

Hi all, this is not so much a question, as me just spouting.
If I do say (spout) something wrong, or that you disagree with
or that you don't understand, then please do speak up.

I've been tuning loops for years. (By tuning I mean picking
some reasonable starting values, all loops IME need a little
real time tweaking from the starting values.) I've always
used the Zeigler-Nichols oscillation method, and looked no further.

Now my boss asked for some other technique and I started reading about
tuning from the step response. (What Z-N call the Process-reaction curve)
Oh I stuck the Z-N paper here, along with a few 'scope shots.

https://www.dropbox.com/sh/7lkqazuy6wvzknk/AADcZtTvFvaD3TEzeJiAjEy6a?dl=0

If only I'd read the paper ~20 years ago.

This is a more up to date rehash of the same things.
http://faculty.mercer.edu/jenkins_he/documents/TuningforPIDControllers.pdf

For completeness I'll stick in Tim W's excellent article.
(Though unfortunately he doesn't talk about Z-N directly.)
http://www.wescottdesign.com/articles/pid/pidWithoutAPhd.pdf
(Maybe he'll add an appendix for me. :^)

So if you look at the step response in my dropbox link, there is a lag
of 300-350 ms. And if you look at the oscillations at the "ultimate gain"
you'll see a period of about 2 seconds. And lo and behold,
2*pi*tau (the lag) = period! That is very satisfying.

Estimating the needed gain (proportional term) from the step response,
didn't work out that well in this case. But I'm running at gains
that are ~20% the ultimate gain, so oscillation method didn't work all
that well either in this regard.


Oh I do have one question for Tim.
You show a thermal control loop with just integrating control.
Does this really work? (I would have thought it would just oscillate.)
Or do you have a little bit of proportional term in there?

Cheers,

George H.
Click to expand...

If you do it in software, you may program a "rate predictor"
and know in advance how to feedback.


w.
 
On 04/24/2017 10:22 PM, George Herold wrote:
On Monday, April 24, 2017 at 8:52:04 PM UTC-4, Tim Wescott wrote:
On Mon, 24 Apr 2017 12:29:24 -0700, George Herold wrote:

Hi all, this is not so much a question, as me just spouting.
If I do say (spout) something wrong, or that you disagree with or that
you don't understand, then please do speak up.

I've been tuning loops for years. (By tuning I mean picking some
reasonable starting values, all loops IME need a little real time
tweaking from the starting values.) I've always used the
Zeigler-Nichols oscillation method, and looked no further.

Now my boss asked for some other technique and I started reading about
tuning from the step response. (What Z-N call the Process-reaction
curve) Oh I stuck the Z-N paper here, along with a few 'scope shots.

https://www.dropbox.com/sh/7lkqazuy6wvzknk/AADcZtTvFvaD3TEzeJiAjEy6a?
dl=0

If only I'd read the paper ~20 years ago.

This is a more up to date rehash of the same things.
http://faculty.mercer.edu/jenkins_he/documents/
TuningforPIDControllers.pdf

For completeness I'll stick in Tim W's excellent article.
(Though unfortunately he doesn't talk about Z-N directly.)
http://www.wescottdesign.com/articles/pid/pidWithoutAPhd.pdf (Maybe
he'll add an appendix for me. :^)

Nope. For two reasons:

First, because I've never used Z-N, never felt the need, and so never
built up practical experience with it.

Second, because all the industrial control guys I've talked to, both in
person and 'lectronically, tell me that Z-N gets you a good starting
point but then you need to either tweak things, do measurement + system
ID + design, or do frequency sweeps + design. There's enough material
out there on it, and too many people that think that it's the bees knees.
Huh, Ok I only said it was my starting point. Then do some step responses
and see what it looks like. (I'm mostly an idiot when it comes to
thermal loops... I'm starting to think that 1/2 the time I got lucky
and put the sensor in a good place, and the other times, not so good.)
Well sometimes you just need to put the sensor at the point you care
about.... I could think more about where to put the heater.
Grumble....

What's your starting point?

So if you look at the step response in my dropbox link, there is a lag
of 300-350 ms. And if you look at the oscillations at the "ultimate
gain"
you'll see a period of about 2 seconds. And lo and behold,
2*pi*tau (the lag) = period! That is very satisfying.

Estimating the needed gain (proportional term) from the step response,
didn't work out that well in this case. But I'm running at gains that
are ~20% the ultimate gain, so oscillation method didn't work all that
well either in this regard.


Oh I do have one question for Tim.
You show a thermal control loop with just integrating control.
Does this really work? (I would have thought it would just oscillate.)
Or do you have a little bit of proportional term in there?

If you keep a firm lid on your ambitions, certainly. For most systems
you won't get nearly the speed you'd get with some proportional action or
even proportional + derivative -- but any system that has a flat gain
characteristic going down to zero frequency can be controlled with just
an integrator, if you're patient.

The one thermal loop that I've done had a simple first-order response
well beyond any reasonable loop bandwidth. So it got a simple PI
controller and worked just grand.
Huh, OK I'll have to try, the loss is in the delay/lag?
This thermal loop is fast, (at least for me)
so ten seconds is no problem.
Click to expand...

I get starting points for thermal loops by looking at the open-loop step
response. In small systems it's usually well approximated by a time
delay cascaded with an integrator. Thus the plant's transfer function
is approximately

H_plant(f) ~ (1/j 2 pi f tau_I) exp(-j 2 pi f tau_delay)

Then I do the usual frequency-compensation thing, aiming for about a
60-degree phase margin. If the system is reasonably well insulated, the
effect of the heat leak is usually small enough to ignore for frequency
compensation purposes, but if it isn't, you need to move the pole from
zero out to (j 2 pi f tau_leak).

For things like diode lasers and IR photodiodes, this is usually close
enough.

Cheers

Phil Hobbs



--
Dr Philip C D Hobbs
Principal Consultant
ElectroOptical Innovations LLC
Optics, Electro-optics, Photonics, Analog Electronics

160 North State Road #203
Briarcliff Manor NY 10510

hobbs at electrooptical dot net
http://electrooptical.net
 
On Tuesday, April 25, 2017 at 1:22:30 PM UTC-4, Phil Hobbs wrote:
On 04/24/2017 10:22 PM, George Herold wrote:
On Monday, April 24, 2017 at 8:52:04 PM UTC-4, Tim Wescott wrote:
On Mon, 24 Apr 2017 12:29:24 -0700, George Herold wrote:

Hi all, this is not so much a question, as me just spouting.
If I do say (spout) something wrong, or that you disagree with or that
you don't understand, then please do speak up.

I've been tuning loops for years. (By tuning I mean picking some
reasonable starting values, all loops IME need a little real time
tweaking from the starting values.) I've always used the
Zeigler-Nichols oscillation method, and looked no further.

Now my boss asked for some other technique and I started reading about
tuning from the step response. (What Z-N call the Process-reaction
curve) Oh I stuck the Z-N paper here, along with a few 'scope shots.

https://www.dropbox.com/sh/7lkqazuy6wvzknk/AADcZtTvFvaD3TEzeJiAjEy6a?
dl=0

If only I'd read the paper ~20 years ago.

This is a more up to date rehash of the same things.
http://faculty.mercer.edu/jenkins_he/documents/
TuningforPIDControllers.pdf

For completeness I'll stick in Tim W's excellent article.
(Though unfortunately he doesn't talk about Z-N directly.)
http://www.wescottdesign.com/articles/pid/pidWithoutAPhd.pdf (Maybe
he'll add an appendix for me. :^)

Nope. For two reasons:

First, because I've never used Z-N, never felt the need, and so never
built up practical experience with it.

Second, because all the industrial control guys I've talked to, both in
person and 'lectronically, tell me that Z-N gets you a good starting
point but then you need to either tweak things, do measurement + system
ID + design, or do frequency sweeps + design. There's enough material
out there on it, and too many people that think that it's the bees knees.
Huh, Ok I only said it was my starting point. Then do some step responses
and see what it looks like. (I'm mostly an idiot when it comes to
thermal loops... I'm starting to think that 1/2 the time I got lucky
and put the sensor in a good place, and the other times, not so good.)
Well sometimes you just need to put the sensor at the point you care
about.... I could think more about where to put the heater.
Grumble....

What's your starting point?

So if you look at the step response in my dropbox link, there is a lag
of 300-350 ms. And if you look at the oscillations at the "ultimate
gain"
you'll see a period of about 2 seconds. And lo and behold,
2*pi*tau (the lag) = period! That is very satisfying.

Estimating the needed gain (proportional term) from the step response,
didn't work out that well in this case. But I'm running at gains that
are ~20% the ultimate gain, so oscillation method didn't work all that
well either in this regard.


Oh I do have one question for Tim.
You show a thermal control loop with just integrating control.
Does this really work? (I would have thought it would just oscillate.)
Or do you have a little bit of proportional term in there?

If you keep a firm lid on your ambitions, certainly. For most systems
you won't get nearly the speed you'd get with some proportional action or
even proportional + derivative -- but any system that has a flat gain
characteristic going down to zero frequency can be controlled with just
an integrator, if you're patient.

The one thermal loop that I've done had a simple first-order response
well beyond any reasonable loop bandwidth. So it got a simple PI
controller and worked just grand.
Huh, OK I'll have to try, the loss is in the delay/lag?
This thermal loop is fast, (at least for me)
so ten seconds is no problem.

I get starting points for thermal loops by looking at the open-loop step
response. In small systems it's usually well approximated by a time
delay cascaded with an integrator. Thus the plant's transfer function
is approximately

H_plant(f) ~ (1/j 2 pi f tau_I) exp(-j 2 pi f tau_delay)
Click to expand...
Right, I went looking for a pulse response with "your"
delay*RC time constant, model in mind.
Seem's like Z-N did it before we were born... which I love.
They also give a guess at the gain, from the pulse response.

Then I do the usual frequency-compensation thing, aiming for about a
60-degree phase margin. If the system is reasonably well insulated, the
effect of the heat leak is usually small enough to ignore for frequency
compensation purposes, but if it isn't, you need to move the pole from
zero out to (j 2 pi f tau_leak).
Click to expand...
Huh, Phil, lots of the time you're talking over my head,
(or at least not where my head is at the moment.)

I'm not sure where 60 degree's is, but I look at the set point
step response and lower the gain if there's not too much ringing.
Say a Q of 1-2. (probably closer to one, and then in production
each unit is a bit different.)
For things like diode lasers and IR photodiodes, this is usually close
enough.
Click to expand...

I was thinking I'd like to make my Rb cell heaters faster.
A cylinder of glass (25mm daim X 25 mm length)
Oh here's some pics on drop box.
There's another one, similar but with an Al cylinder
holding the heater wire, rather than a glass cylinder.
https://www.dropbox.com/sh/9mv2b813latvvla/AAAEdQqEFEr9iw_6k9QONagHa?dl=0

I stick the thermal couple in next the the Rb cylinder.
Would I do better sticking it on the heater? (I, the students,
can monitor the light transmission to give a measure of when the
cell temperate stabilizes...)
(Hey I found that all my old dropbox links still work, which is nice.)


George H.

Cheers

Phil Hobbs



--
Dr Philip C D Hobbs
Principal Consultant
ElectroOptical Innovations LLC
Optics, Electro-optics, Photonics, Analog Electronics

160 North State Road #203
Briarcliff Manor NY 10510

hobbs at electrooptical dot net
http://electrooptical.net
Click to expand...
 
On Tuesday, April 25, 2017 at 8:02:54 PM UTC-4, George Herold wrote:
On Tuesday, April 25, 2017 at 1:22:30 PM UTC-4, Phil Hobbs wrote:
On 04/24/2017 10:22 PM, George Herold wrote:
On Monday, April 24, 2017 at 8:52:04 PM UTC-4, Tim Wescott wrote:
On Mon, 24 Apr 2017 12:29:24 -0700, George Herold wrote:

Hi all, this is not so much a question, as me just spouting.
If I do say (spout) something wrong, or that you disagree with or that
you don't understand, then please do speak up.

I've been tuning loops for years. (By tuning I mean picking some
reasonable starting values, all loops IME need a little real time
tweaking from the starting values.) I've always used the
Zeigler-Nichols oscillation method, and looked no further.

Now my boss asked for some other technique and I started reading about
tuning from the step response. (What Z-N call the Process-reaction
curve) Oh I stuck the Z-N paper here, along with a few 'scope shots.

https://www.dropbox.com/sh/7lkqazuy6wvzknk/AADcZtTvFvaD3TEzeJiAjEy6a?
dl=0

If only I'd read the paper ~20 years ago.

This is a more up to date rehash of the same things.
http://faculty.mercer.edu/jenkins_he/documents/
TuningforPIDControllers.pdf

For completeness I'll stick in Tim W's excellent article.
(Though unfortunately he doesn't talk about Z-N directly.)
http://www.wescottdesign.com/articles/pid/pidWithoutAPhd.pdf (Maybe
he'll add an appendix for me. :^)

Nope. For two reasons:

First, because I've never used Z-N, never felt the need, and so never
built up practical experience with it.

Second, because all the industrial control guys I've talked to, both in
person and 'lectronically, tell me that Z-N gets you a good starting
point but then you need to either tweak things, do measurement + system
ID + design, or do frequency sweeps + design. There's enough material
out there on it, and too many people that think that it's the bees knees.
Huh, Ok I only said it was my starting point. Then do some step responses
and see what it looks like. (I'm mostly an idiot when it comes to
thermal loops... I'm starting to think that 1/2 the time I got lucky
and put the sensor in a good place, and the other times, not so good.)
Well sometimes you just need to put the sensor at the point you care
about.... I could think more about where to put the heater.
Grumble....

What's your starting point?

So if you look at the step response in my dropbox link, there is a lag
of 300-350 ms. And if you look at the oscillations at the "ultimate
gain"
you'll see a period of about 2 seconds. And lo and behold,
2*pi*tau (the lag) = period! That is very satisfying.

Estimating the needed gain (proportional term) from the step response,
didn't work out that well in this case. But I'm running at gains that
are ~20% the ultimate gain, so oscillation method didn't work all that
well either in this regard.


Oh I do have one question for Tim.
You show a thermal control loop with just integrating control.
Does this really work? (I would have thought it would just oscillate.)
Or do you have a little bit of proportional term in there?

If you keep a firm lid on your ambitions, certainly. For most systems
you won't get nearly the speed you'd get with some proportional action or
even proportional + derivative -- but any system that has a flat gain
characteristic going down to zero frequency can be controlled with just
an integrator, if you're patient.

The one thermal loop that I've done had a simple first-order response
well beyond any reasonable loop bandwidth. So it got a simple PI
controller and worked just grand.
Huh, OK I'll have to try, the loss is in the delay/lag?
This thermal loop is fast, (at least for me)
so ten seconds is no problem.

I get starting points for thermal loops by looking at the open-loop step
response. In small systems it's usually well approximated by a time
delay cascaded with an integrator. Thus the plant's transfer function
is approximately

H_plant(f) ~ (1/j 2 pi f tau_I) exp(-j 2 pi f tau_delay)
Right, I went looking for a pulse response with "your"
delay*RC time constant, model in mind.
Seem's like Z-N did it before we were born... which I love.
They also give a guess at the gain, from the pulse response.


Then I do the usual frequency-compensation thing, aiming for about a
60-degree phase margin. If the system is reasonably well insulated, the
effect of the heat leak is usually small enough to ignore for frequency
compensation purposes, but if it isn't, you need to move the pole from
zero out to (j 2 pi f tau_leak).
Huh, Phil, lots of the time you're talking over my head,
(or at least not where my head is at the moment.)

I'm not sure where 60 degree's is, but I look at the set point
step response and lower the gain if there's not too much ringing.
Click to expand...
oops delete "not" ............................^^^
(please forgive all the other numerous mistakes...)
GH.
Say a Q of 1-2. (probably closer to one, and then in production
each unit is a bit different.)

For things like diode lasers and IR photodiodes, this is usually close
enough.

I was thinking I'd like to make my Rb cell heaters faster.
A cylinder of glass (25mm daim X 25 mm length)
Oh here's some pics on drop box.
There's another one, similar but with an Al cylinder
holding the heater wire, rather than a glass cylinder.
https://www.dropbox.com/sh/9mv2b813latvvla/AAAEdQqEFEr9iw_6k9QONagHa?dl=0

I stick the thermal couple in next the the Rb cylinder.
Would I do better sticking it on the heater? (I, the students,
can monitor the light transmission to give a measure of when the
cell temperate stabilizes...)
(Hey I found that all my old dropbox links still work, which is nice.)


George H.


Cheers

Phil Hobbs



--
Dr Philip C D Hobbs
Principal Consultant
ElectroOptical Innovations LLC
Optics, Electro-optics, Photonics, Analog Electronics

160 North State Road #203
Briarcliff Manor NY 10510

hobbs at electrooptical dot net
http://electrooptical.net
Click to expand...
 
On Tue, 25 Apr 2017 17:02:48 -0700, George Herold wrote:

On Tuesday, April 25, 2017 at 1:22:30 PM UTC-4, Phil Hobbs wrote:
On 04/24/2017 10:22 PM, George Herold wrote:
On Monday, April 24, 2017 at 8:52:04 PM UTC-4, Tim Wescott wrote:
On Mon, 24 Apr 2017 12:29:24 -0700, George Herold wrote:

Hi all, this is not so much a question, as me just spouting.
If I do say (spout) something wrong, or that you disagree with or
that you don't understand, then please do speak up.

I've been tuning loops for years. (By tuning I mean picking some
reasonable starting values, all loops IME need a little real time
tweaking from the starting values.) I've always used the
Zeigler-Nichols oscillation method, and looked no further.

Now my boss asked for some other technique and I started reading
about tuning from the step response. (What Z-N call the
Process-reaction curve) Oh I stuck the Z-N paper here, along with a
few 'scope shots.

https://www.dropbox.com/sh/7lkqazuy6wvzknk/
AADcZtTvFvaD3TEzeJiAjEy6a?
dl=0

If only I'd read the paper ~20 years ago.

This is a more up to date rehash of the same things.
http://faculty.mercer.edu/jenkins_he/documents/
TuningforPIDControllers.pdf

For completeness I'll stick in Tim W's excellent article.
(Though unfortunately he doesn't talk about Z-N directly.)
http://www.wescottdesign.com/articles/pid/pidWithoutAPhd.pdf (Maybe
he'll add an appendix for me. :^)

Nope. For two reasons:

First, because I've never used Z-N, never felt the need, and so
never built up practical experience with it.

Second, because all the industrial control guys I've talked to, both
in person and 'lectronically, tell me that Z-N gets you a good
starting point but then you need to either tweak things, do
measurement + system ID + design, or do frequency sweeps + design.
There's enough material out there on it, and too many people that
think that it's the bees knees.
Huh, Ok I only said it was my starting point. Then do some step
responses and see what it looks like. (I'm mostly an idiot when it
comes to thermal loops... I'm starting to think that 1/2 the time I
got lucky and put the sensor in a good place, and the other times,
not so good.) Well sometimes you just need to put the sensor at the
point you care about.... I could think more about where to put the
heater.
Grumble....

What's your starting point?

So if you look at the step response in my dropbox link, there is a
lag of 300-350 ms. And if you look at the oscillations at the
"ultimate gain"
you'll see a period of about 2 seconds. And lo and behold,
2*pi*tau (the lag) = period! That is very satisfying.

Estimating the needed gain (proportional term) from the step
response, didn't work out that well in this case. But I'm running
at gains that are ~20% the ultimate gain, so oscillation method
didn't work all that well either in this regard.


Oh I do have one question for Tim.
You show a thermal control loop with just integrating control.
Does this really work? (I would have thought it would just
oscillate.)
Or do you have a little bit of proportional term in there?

If you keep a firm lid on your ambitions, certainly. For most
systems you won't get nearly the speed you'd get with some
proportional action or even proportional + derivative -- but any
system that has a flat gain characteristic going down to zero
frequency can be controlled with just an integrator, if you're
patient.

The one thermal loop that I've done had a simple first-order
response well beyond any reasonable loop bandwidth. So it got a
simple PI controller and worked just grand.
Huh, OK I'll have to try, the loss is in the delay/lag?
This thermal loop is fast, (at least for me)
so ten seconds is no problem.

I get starting points for thermal loops by looking at the open-loop
step response. In small systems it's usually well approximated by a
time delay cascaded with an integrator. Thus the plant's transfer
function is approximately

H_plant(f) ~ (1/j 2 pi f tau_I) exp(-j 2 pi f tau_delay)
Right, I went looking for a pulse response with "your"
delay*RC time constant, model in mind.
Seem's like Z-N did it before we were born... which I love.
They also give a guess at the gain, from the pulse response.


Then I do the usual frequency-compensation thing, aiming for about a
60-degree phase margin. If the system is reasonably well insulated,
the effect of the heat leak is usually small enough to ignore for
frequency compensation purposes, but if it isn't, you need to move the
pole from zero out to (j 2 pi f tau_leak).
Huh, Phil, lots of the time you're talking over my head,
(or at least not where my head is at the moment.)

I'm not sure where 60 degree's is, but I look at the set point step
response and lower the gain if there's not too much ringing.
Say a Q of 1-2. (probably closer to one, and then in production each
unit is a bit different.)
Click to expand...

You need to learn frequency-domain tuning. Just looking at the time
domain response is not a reliable way to get the best response and a
known loop robustness.

--

Tim Wescott
Wescott Design Services
http://www.wescottdesign.com

I'm looking for work -- see my website!
 
On Tuesday, April 25, 2017 at 11:20:50 PM UTC-4, Tim Wescott wrote:
On Tue, 25 Apr 2017 17:02:48 -0700, George Herold wrote:

On Tuesday, April 25, 2017 at 1:22:30 PM UTC-4, Phil Hobbs wrote:
On 04/24/2017 10:22 PM, George Herold wrote:
On Monday, April 24, 2017 at 8:52:04 PM UTC-4, Tim Wescott wrote:
On Mon, 24 Apr 2017 12:29:24 -0700, George Herold wrote:

Hi all, this is not so much a question, as me just spouting.
If I do say (spout) something wrong, or that you disagree with or
that you don't understand, then please do speak up.

I've been tuning loops for years. (By tuning I mean picking some
reasonable starting values, all loops IME need a little real time
tweaking from the starting values.) I've always used the
Zeigler-Nichols oscillation method, and looked no further.

Now my boss asked for some other technique and I started reading
about tuning from the step response. (What Z-N call the
Process-reaction curve) Oh I stuck the Z-N paper here, along with a
few 'scope shots.

https://www.dropbox.com/sh/7lkqazuy6wvzknk/
AADcZtTvFvaD3TEzeJiAjEy6a?
dl=0

If only I'd read the paper ~20 years ago.

This is a more up to date rehash of the same things.
http://faculty.mercer.edu/jenkins_he/documents/
TuningforPIDControllers.pdf

For completeness I'll stick in Tim W's excellent article.
(Though unfortunately he doesn't talk about Z-N directly.)
http://www.wescottdesign.com/articles/pid/pidWithoutAPhd.pdf (Maybe
he'll add an appendix for me. :^)

Nope. For two reasons:

First, because I've never used Z-N, never felt the need, and so
never built up practical experience with it.

Second, because all the industrial control guys I've talked to, both
in person and 'lectronically, tell me that Z-N gets you a good
starting point but then you need to either tweak things, do
measurement + system ID + design, or do frequency sweeps + design.
There's enough material out there on it, and too many people that
think that it's the bees knees.
Huh, Ok I only said it was my starting point. Then do some step
responses and see what it looks like. (I'm mostly an idiot when it
comes to thermal loops... I'm starting to think that 1/2 the time I
got lucky and put the sensor in a good place, and the other times,
not so good.) Well sometimes you just need to put the sensor at the
point you care about.... I could think more about where to put the
heater.
Grumble....

What's your starting point?

So if you look at the step response in my dropbox link, there is a
lag of 300-350 ms. And if you look at the oscillations at the
"ultimate gain"
you'll see a period of about 2 seconds. And lo and behold,
2*pi*tau (the lag) = period! That is very satisfying.

Estimating the needed gain (proportional term) from the step
response, didn't work out that well in this case. But I'm running
at gains that are ~20% the ultimate gain, so oscillation method
didn't work all that well either in this regard.


Oh I do have one question for Tim.
You show a thermal control loop with just integrating control.
Does this really work? (I would have thought it would just
oscillate.)
Or do you have a little bit of proportional term in there?

If you keep a firm lid on your ambitions, certainly. For most
systems you won't get nearly the speed you'd get with some
proportional action or even proportional + derivative -- but any
system that has a flat gain characteristic going down to zero
frequency can be controlled with just an integrator, if you're
patient.

The one thermal loop that I've done had a simple first-order
response well beyond any reasonable loop bandwidth. So it got a
simple PI controller and worked just grand.
Huh, OK I'll have to try, the loss is in the delay/lag?
This thermal loop is fast, (at least for me)
so ten seconds is no problem.

I get starting points for thermal loops by looking at the open-loop
step response. In small systems it's usually well approximated by a
time delay cascaded with an integrator. Thus the plant's transfer
function is approximately

H_plant(f) ~ (1/j 2 pi f tau_I) exp(-j 2 pi f tau_delay)
Right, I went looking for a pulse response with "your"
delay*RC time constant, model in mind.
Seem's like Z-N did it before we were born... which I love.
They also give a guess at the gain, from the pulse response.


Then I do the usual frequency-compensation thing, aiming for about a
60-degree phase margin. If the system is reasonably well insulated,
the effect of the heat leak is usually small enough to ignore for
frequency compensation purposes, but if it isn't, you need to move the
pole from zero out to (j 2 pi f tau_leak).
Huh, Phil, lots of the time you're talking over my head,
(or at least not where my head is at the moment.)

I'm not sure where 60 degree's is, but I look at the set point step
response and lower the gain if there's not too much ringing.
Say a Q of 1-2. (probably closer to one, and then in production each
unit is a bit different.)

You need to learn frequency-domain tuning. Just looking at the time
domain response is not a reliable way to get the best response and a
known loop robustness.
Click to expand...

Grin, right. (There are bucket loads of things I need to learn.)
You do a frequency sweep with the loop closed?
Do you have a video showing that? :^)

Once you know what you are doing (I'm not claiming that *I* know.)
Doesn't the step response (I'm thinking closed loop) contain
the same information as a frequency sweep?

George H.
--

Tim Wescott
Wescott Design Services
http://www.wescottdesign.com

I'm looking for work -- see my website!
Click to expand...
 
On Wed, 26 Apr 2017 04:45:23 -0700, George Herold wrote:

On Tuesday, April 25, 2017 at 11:20:50 PM UTC-4, Tim Wescott wrote:
On Tue, 25 Apr 2017 17:02:48 -0700, George Herold wrote:

On Tuesday, April 25, 2017 at 1:22:30 PM UTC-4, Phil Hobbs wrote:
On 04/24/2017 10:22 PM, George Herold wrote:
On Monday, April 24, 2017 at 8:52:04 PM UTC-4, Tim Wescott wrote:
On Mon, 24 Apr 2017 12:29:24 -0700, George Herold wrote:

Hi all, this is not so much a question, as me just spouting.
If I do say (spout) something wrong, or that you disagree with
or that you don't understand, then please do speak up.

I've been tuning loops for years. (By tuning I mean picking
some reasonable starting values, all loops IME need a little
real time tweaking from the starting values.) I've always used
the Zeigler-Nichols oscillation method, and looked no further.

Now my boss asked for some other technique and I started reading
about tuning from the step response. (What Z-N call the
Process-reaction curve) Oh I stuck the Z-N paper here, along
with a few 'scope shots.

https://www.dropbox.com/sh/7lkqazuy6wvzknk/
AADcZtTvFvaD3TEzeJiAjEy6a?
dl=0

If only I'd read the paper ~20 years ago.

This is a more up to date rehash of the same things.
http://faculty.mercer.edu/jenkins_he/documents/
TuningforPIDControllers.pdf

For completeness I'll stick in Tim W's excellent article.
(Though unfortunately he doesn't talk about Z-N directly.)
http://www.wescottdesign.com/articles/pid/pidWithoutAPhd.pdf
(Maybe he'll add an appendix for me. :^)

Nope. For two reasons:

First, because I've never used Z-N, never felt the need, and so
never built up practical experience with it.

Second, because all the industrial control guys I've talked to,
both in person and 'lectronically, tell me that Z-N gets you a
good starting point but then you need to either tweak things, do
measurement + system ID + design, or do frequency sweeps +
design. There's enough material out there on it, and too many
people that think that it's the bees knees.
Huh, Ok I only said it was my starting point. Then do some step
responses and see what it looks like. (I'm mostly an idiot when
it comes to thermal loops... I'm starting to think that 1/2 the
time I got lucky and put the sensor in a good place, and the other
times, not so good.) Well sometimes you just need to put the
sensor at the point you care about.... I could think more about
where to put the heater.
Grumble....

What's your starting point?

So if you look at the step response in my dropbox link, there is
a lag of 300-350 ms. And if you look at the oscillations at the
"ultimate gain"
you'll see a period of about 2 seconds. And lo and behold,
2*pi*tau (the lag) = period! That is very satisfying.

Estimating the needed gain (proportional term) from the step
response, didn't work out that well in this case. But I'm
running at gains that are ~20% the ultimate gain, so oscillation
method didn't work all that well either in this regard.


Oh I do have one question for Tim.
You show a thermal control loop with just integrating control.
Does this really work? (I would have thought it would just
oscillate.)
Or do you have a little bit of proportional term in there?

If you keep a firm lid on your ambitions, certainly. For most
systems you won't get nearly the speed you'd get with some
proportional action or even proportional + derivative -- but any
system that has a flat gain characteristic going down to zero
frequency can be controlled with just an integrator, if you're
patient.

The one thermal loop that I've done had a simple first-order
response well beyond any reasonable loop bandwidth. So it got a
simple PI controller and worked just grand.
Huh, OK I'll have to try, the loss is in the delay/lag?
This thermal loop is fast, (at least for me)
so ten seconds is no problem.

I get starting points for thermal loops by looking at the open-loop
step response. In small systems it's usually well approximated by a
time delay cascaded with an integrator. Thus the plant's transfer
function is approximately

H_plant(f) ~ (1/j 2 pi f tau_I) exp(-j 2 pi f tau_delay)
Right, I went looking for a pulse response with "your"
delay*RC time constant, model in mind.
Seem's like Z-N did it before we were born... which I love.
They also give a guess at the gain, from the pulse response.


Then I do the usual frequency-compensation thing, aiming for about a
60-degree phase margin. If the system is reasonably well insulated,
the effect of the heat leak is usually small enough to ignore for
frequency compensation purposes, but if it isn't, you need to move
the pole from zero out to (j 2 pi f tau_leak).
Huh, Phil, lots of the time you're talking over my head,
(or at least not where my head is at the moment.)

I'm not sure where 60 degree's is, but I look at the set point step
response and lower the gain if there's not too much ringing.
Say a Q of 1-2. (probably closer to one, and then in production each
unit is a bit different.)

You need to learn frequency-domain tuning. Just looking at the time
domain response is not a reliable way to get the best response and a
known loop robustness.

Grin, right. (There are bucket loads of things I need to learn.) You do
a frequency sweep with the loop closed?
Do you have a video showing that? :^)
Click to expand...

It's on the short list -- which is about half a dozen titles long.

I describe it in my book. And in this article:

<http://wescottdesign.com/articles/FreqMeas/freq_meas.html>

The basic idea is that if you measure the phase and amplitude of a sine
wave going into a block, and the phase and amplitude of a sine wave
coming out of that block, then you can calculate the block's response at
that frequency by doing a complex division (amplitude out / amplitude in,
phase out - phase in). Then repeat for a whole bunch of frequencies.

It's easy to build this into code if you're working with a
microprocessor; if you're not there are instruments out there ("dynamic
system analyzers" or "transfer function analyzers") that do it. If
National Instruments isn't utterly crazy there's a pre-made VI to do this
in labview with one analog I/O brick.

I suspect that there's one out there for less than $500. And you can
always do it by hand with a signal generator and an O-scope.

Once you know what you are doing (I'm not claiming that *I* know.)
Doesn't the step response (I'm thinking closed loop) contain the same
information as a frequency sweep?
Click to expand...

In theory, yes. In practice the good stuff is often swamped out by
either noise or distortion. In the past I've done a pretty good job of
doing system identification from a customer-supplied step response -- but
I was limiting the performance of the loop based on what I didn't know
about the plant, rather than a sure idea of what the plant wasn't capable
of.

You could probably take a whole bunch of step responses and average them
together, and then fit a system model to the resulting average -- that
would at least knock down the effects of noise. But you'd want to do
real honest-to-god system ID, where you're fitting a model in the Laplace
domain to the step response, not just that Z-N stuff.

--
Tim Wescott
Control systems, embedded software and circuit design
I'm looking for work! See my website if you're interested
http://www.wescottdesign.com
 
On Wednesday, April 26, 2017 at 10:47:38 AM UTC-4, Tim Wescott wrote:
On Wed, 26 Apr 2017 04:45:23 -0700, George Herold wrote:

On Tuesday, April 25, 2017 at 11:20:50 PM UTC-4, Tim Wescott wrote:
On Tue, 25 Apr 2017 17:02:48 -0700, George Herold wrote:

On Tuesday, April 25, 2017 at 1:22:30 PM UTC-4, Phil Hobbs wrote:
On 04/24/2017 10:22 PM, George Herold wrote:
On Monday, April 24, 2017 at 8:52:04 PM UTC-4, Tim Wescott wrote:
On Mon, 24 Apr 2017 12:29:24 -0700, George Herold wrote:

Hi all, this is not so much a question, as me just spouting.
If I do say (spout) something wrong, or that you disagree with
or that you don't understand, then please do speak up.

I've been tuning loops for years. (By tuning I mean picking
some reasonable starting values, all loops IME need a little
real time tweaking from the starting values.) I've always used
the Zeigler-Nichols oscillation method, and looked no further.

Now my boss asked for some other technique and I started reading
about tuning from the step response. (What Z-N call the
Process-reaction curve) Oh I stuck the Z-N paper here, along
with a few 'scope shots.

https://www.dropbox.com/sh/7lkqazuy6wvzknk/
AADcZtTvFvaD3TEzeJiAjEy6a?
dl=0

If only I'd read the paper ~20 years ago.

This is a more up to date rehash of the same things.
http://faculty.mercer.edu/jenkins_he/documents/
TuningforPIDControllers.pdf

For completeness I'll stick in Tim W's excellent article.
(Though unfortunately he doesn't talk about Z-N directly.)
http://www.wescottdesign.com/articles/pid/pidWithoutAPhd.pdf
(Maybe he'll add an appendix for me. :^)

Nope. For two reasons:

First, because I've never used Z-N, never felt the need, and so
never built up practical experience with it.

Second, because all the industrial control guys I've talked to,
both in person and 'lectronically, tell me that Z-N gets you a
good starting point but then you need to either tweak things, do
measurement + system ID + design, or do frequency sweeps +
design. There's enough material out there on it, and too many
people that think that it's the bees knees.
Huh, Ok I only said it was my starting point. Then do some step
responses and see what it looks like. (I'm mostly an idiot when
it comes to thermal loops... I'm starting to think that 1/2 the
time I got lucky and put the sensor in a good place, and the other
times, not so good.) Well sometimes you just need to put the
sensor at the point you care about.... I could think more about
where to put the heater.
Grumble....

What's your starting point?

So if you look at the step response in my dropbox link, there is
a lag of 300-350 ms. And if you look at the oscillations at the
"ultimate gain"
you'll see a period of about 2 seconds. And lo and behold,
2*pi*tau (the lag) = period! That is very satisfying.

Estimating the needed gain (proportional term) from the step
response, didn't work out that well in this case. But I'm
running at gains that are ~20% the ultimate gain, so oscillation
method didn't work all that well either in this regard.


Oh I do have one question for Tim.
You show a thermal control loop with just integrating control.
Does this really work? (I would have thought it would just
oscillate.)
Or do you have a little bit of proportional term in there?

If you keep a firm lid on your ambitions, certainly. For most
systems you won't get nearly the speed you'd get with some
proportional action or even proportional + derivative -- but any
system that has a flat gain characteristic going down to zero
frequency can be controlled with just an integrator, if you're
patient.

The one thermal loop that I've done had a simple first-order
response well beyond any reasonable loop bandwidth. So it got a
simple PI controller and worked just grand.
Huh, OK I'll have to try, the loss is in the delay/lag?
This thermal loop is fast, (at least for me)
so ten seconds is no problem.

I get starting points for thermal loops by looking at the open-loop
step response. In small systems it's usually well approximated by a
time delay cascaded with an integrator. Thus the plant's transfer
function is approximately

H_plant(f) ~ (1/j 2 pi f tau_I) exp(-j 2 pi f tau_delay)
Right, I went looking for a pulse response with "your"
delay*RC time constant, model in mind.
Seem's like Z-N did it before we were born... which I love.
They also give a guess at the gain, from the pulse response.


Then I do the usual frequency-compensation thing, aiming for about a
60-degree phase margin. If the system is reasonably well insulated,
the effect of the heat leak is usually small enough to ignore for
frequency compensation purposes, but if it isn't, you need to move
the pole from zero out to (j 2 pi f tau_leak).
Huh, Phil, lots of the time you're talking over my head,
(or at least not where my head is at the moment.)

I'm not sure where 60 degree's is, but I look at the set point step
response and lower the gain if there's not too much ringing.
Say a Q of 1-2. (probably closer to one, and then in production each
unit is a bit different.)

You need to learn frequency-domain tuning. Just looking at the time
domain response is not a reliable way to get the best response and a
known loop robustness.

Grin, right. (There are bucket loads of things I need to learn.) You do
a frequency sweep with the loop closed?
Do you have a video showing that? :^)

It's on the short list -- which is about half a dozen titles long.

I describe it in my book. And in this article:

http://wescottdesign.com/articles/FreqMeas/freq_meas.html

The basic idea is that if you measure the phase and amplitude of a sine
wave going into a block, and the phase and amplitude of a sine wave
coming out of that block, then you can calculate the block's response at
that frequency by doing a complex division (amplitude out / amplitude in,
phase out - phase in). Then repeat for a whole bunch of frequencies.
Click to expand...
Right. I've mostly just kept track of the amplitude response,
though you can always see the phase bounce around too.
(It must be Krammers and Kronig, or someone earlier, who observed
that you can see the phase response before the amplitude.)

Hmm I've got a Z-N oscillation question, moving to SED,
in hopes of more experience.


It's easy to build this into code if you're working with a
microprocessor; if you're not there are instruments out there ("dynamic
system analyzers" or "transfer function analyzers") that do it. If
National Instruments isn't utterly crazy there's a pre-made VI to do this
in labview with one analog I/O brick.
Click to expand...

My Rigol sig gen goes down to ~uHz. I can do a frequency sweep,
It's got two outputs which I can phase lock and set at 90 degrees.
In principle, I can measure everything myself.

I hate long time sweeps... you don't know till hours later
if you f'ed something up. It's much faster to understand the
step response, or oscillation method.

George H.
I suspect that there's one out there for less than $500. And you can
always do it by hand with a signal generator and an O-scope.

Once you know what you are doing (I'm not claiming that *I* know.)
Doesn't the step response (I'm thinking closed loop) contain the same
information as a frequency sweep?

In theory, yes. In practice the good stuff is often swamped out by
either noise or distortion. In the past I've done a pretty good job of
doing system identification from a customer-supplied step response -- but
I was limiting the performance of the loop based on what I didn't know
about the plant, rather than a sure idea of what the plant wasn't capable
of.

You could probably take a whole bunch of step responses and average them
together, and then fit a system model to the resulting average -- that
would at least knock down the effects of noise. But you'd want to do
real honest-to-god system ID, where you're fitting a model in the Laplace
domain to the step response, not just that Z-N stuff.

--
Tim Wescott
Control systems, embedded software and circuit design
I'm looking for work! See my website if you're interested
http://www.wescottdesign.com
Click to expand...
 
On Thursday, April 27, 2017 at 6:50:06 PM UTC-4, George Herold wrote:
On Wednesday, April 26, 2017 at 10:47:38 AM UTC-4, Tim Wescott wrote:
On Wed, 26 Apr 2017 04:45:23 -0700, George Herold wrote:

On Tuesday, April 25, 2017 at 11:20:50 PM UTC-4, Tim Wescott wrote:
On Tue, 25 Apr 2017 17:02:48 -0700, George Herold wrote:

On Tuesday, April 25, 2017 at 1:22:30 PM UTC-4, Phil Hobbs wrote:
On 04/24/2017 10:22 PM, George Herold wrote:
On Monday, April 24, 2017 at 8:52:04 PM UTC-4, Tim Wescott wrote:
On Mon, 24 Apr 2017 12:29:24 -0700, George Herold wrote:

Hi all, this is not so much a question, as me just spouting.
If I do say (spout) something wrong, or that you disagree with
or that you don't understand, then please do speak up.

I've been tuning loops for years. (By tuning I mean picking
some reasonable starting values, all loops IME need a little
real time tweaking from the starting values.) I've always used
the Zeigler-Nichols oscillation method, and looked no further.

Now my boss asked for some other technique and I started reading
about tuning from the step response. (What Z-N call the
Process-reaction curve) Oh I stuck the Z-N paper here, along
with a few 'scope shots.

https://www.dropbox.com/sh/7lkqazuy6wvzknk/
AADcZtTvFvaD3TEzeJiAjEy6a?
dl=0

If only I'd read the paper ~20 years ago.

This is a more up to date rehash of the same things.
http://faculty.mercer.edu/jenkins_he/documents/
TuningforPIDControllers.pdf

For completeness I'll stick in Tim W's excellent article.
(Though unfortunately he doesn't talk about Z-N directly.)
http://www.wescottdesign.com/articles/pid/pidWithoutAPhd.pdf
(Maybe he'll add an appendix for me. :^)

Nope. For two reasons:

First, because I've never used Z-N, never felt the need, and so
never built up practical experience with it.

Second, because all the industrial control guys I've talked to,
both in person and 'lectronically, tell me that Z-N gets you a
good starting point but then you need to either tweak things, do
measurement + system ID + design, or do frequency sweeps +
design. There's enough material out there on it, and too many
people that think that it's the bees knees.
Huh, Ok I only said it was my starting point. Then do some step
responses and see what it looks like. (I'm mostly an idiot when
it comes to thermal loops... I'm starting to think that 1/2 the
time I got lucky and put the sensor in a good place, and the other
times, not so good.) Well sometimes you just need to put the
sensor at the point you care about.... I could think more about
where to put the heater.
Grumble....

What's your starting point?

So if you look at the step response in my dropbox link, there is
a lag of 300-350 ms. And if you look at the oscillations at the
"ultimate gain"
you'll see a period of about 2 seconds. And lo and behold,
2*pi*tau (the lag) = period! That is very satisfying.

Estimating the needed gain (proportional term) from the step
response, didn't work out that well in this case. But I'm
running at gains that are ~20% the ultimate gain, so oscillation
method didn't work all that well either in this regard.


Oh I do have one question for Tim.
You show a thermal control loop with just integrating control.
Does this really work? (I would have thought it would just
oscillate.)
Or do you have a little bit of proportional term in there?

If you keep a firm lid on your ambitions, certainly. For most
systems you won't get nearly the speed you'd get with some
proportional action or even proportional + derivative -- but any
system that has a flat gain characteristic going down to zero
frequency can be controlled with just an integrator, if you're
patient.

The one thermal loop that I've done had a simple first-order
response well beyond any reasonable loop bandwidth. So it got a
simple PI controller and worked just grand.
Huh, OK I'll have to try, the loss is in the delay/lag?
This thermal loop is fast, (at least for me)
so ten seconds is no problem.

I get starting points for thermal loops by looking at the open-loop
step response. In small systems it's usually well approximated by a
time delay cascaded with an integrator. Thus the plant's transfer
function is approximately

H_plant(f) ~ (1/j 2 pi f tau_I) exp(-j 2 pi f tau_delay)
Right, I went looking for a pulse response with "your"
delay*RC time constant, model in mind.
Seem's like Z-N did it before we were born... which I love.
They also give a guess at the gain, from the pulse response.


Then I do the usual frequency-compensation thing, aiming for about a
60-degree phase margin. If the system is reasonably well insulated,
the effect of the heat leak is usually small enough to ignore for
frequency compensation purposes, but if it isn't, you need to move
the pole from zero out to (j 2 pi f tau_leak).
Huh, Phil, lots of the time you're talking over my head,
(or at least not where my head is at the moment.)

I'm not sure where 60 degree's is, but I look at the set point step
response and lower the gain if there's not too much ringing.
Say a Q of 1-2. (probably closer to one, and then in production each
unit is a bit different.)

You need to learn frequency-domain tuning. Just looking at the time
domain response is not a reliable way to get the best response and a
known loop robustness.

Grin, right. (There are bucket loads of things I need to learn.) You do
a frequency sweep with the loop closed?
Do you have a video showing that? :^)

It's on the short list -- which is about half a dozen titles long.

I describe it in my book. And in this article:

http://wescottdesign.com/articles/FreqMeas/freq_meas.html

The basic idea is that if you measure the phase and amplitude of a sine
wave going into a block, and the phase and amplitude of a sine wave
coming out of that block, then you can calculate the block's response at
that frequency by doing a complex division (amplitude out / amplitude in,
phase out - phase in). Then repeat for a whole bunch of frequencies.

Right. I've mostly just kept track of the amplitude response,
though you can always see the phase bounce around too.
(It must be Krammers and Kronig, or someone earlier, who observed
that you can see the phase response before the amplitude.)

Hmm I've got a Z-N oscillation question, moving to SED,
in hopes of more experience.


It's easy to build this into code if you're working with a
microprocessor; if you're not there are instruments out there ("dynamic
system analyzers" or "transfer function analyzers") that do it. If
National Instruments isn't utterly crazy there's a pre-made VI to do this
in labview with one analog I/O brick.

My Rigol sig gen goes down to ~uHz. I can do a frequency sweep,
It's got two outputs which I can phase lock and set at 90 degrees.
In principle, I can measure everything myself.

I hate long time sweeps... you don't know till hours later
if you f'ed something up. It's much faster to understand the
step response, or oscillation method.

George H.

I suspect that there's one out there for less than $500. And you can
always do it by hand with a signal generator and an O-scope.

Once you know what you are doing (I'm not claiming that *I* know.)
Doesn't the step response (I'm thinking closed loop) contain the same
information as a frequency sweep?

In theory, yes. In practice the good stuff is often swamped out by
either noise or distortion. In the past I've done a pretty good job of
doing system identification from a customer-supplied step response -- but
I was limiting the performance of the loop based on what I didn't know
about the plant, rather than a sure idea of what the plant wasn't capable
of.

You could probably take a whole bunch of step responses and average them
together, and then fit a system model to the resulting average -- that
would at least knock down the effects of noise. But you'd want to do
real honest-to-god system ID, where you're fitting a model in the Laplace
domain to the step response, not just that Z-N stuff.

--
Tim Wescott
Control systems, embedded software and circuit design
I'm looking for work! See my website if you're interested
http://www.wescottdesign.com
Click to expand...

Hey what about doing the FFT of the pulse response?
Right, I know about FFT's (Fir spectrometers as a post doc.)
There's low frequency information in the long tail.. noisy
part of the response... In this case I'm more interested in the
high frequency, I think?

George H.
 
On Thu, 27 Apr 2017 16:10:26 -0700, George Herold wrote:

On Thursday, April 27, 2017 at 6:50:06 PM UTC-4, George Herold wrote:
On Wednesday, April 26, 2017 at 10:47:38 AM UTC-4, Tim Wescott wrote:
On Wed, 26 Apr 2017 04:45:23 -0700, George Herold wrote:

On Tuesday, April 25, 2017 at 11:20:50 PM UTC-4, Tim Wescott wrote:
On Tue, 25 Apr 2017 17:02:48 -0700, George Herold wrote:

On Tuesday, April 25, 2017 at 1:22:30 PM UTC-4, Phil Hobbs
wrote:
On 04/24/2017 10:22 PM, George Herold wrote:
On Monday, April 24, 2017 at 8:52:04 PM UTC-4, Tim Wescott
wrote:
On Mon, 24 Apr 2017 12:29:24 -0700, George Herold wrote:

Hi all, this is not so much a question, as me just
spouting.
If I do say (spout) something wrong, or that you disagree
with or that you don't understand, then please do speak up.

I've been tuning loops for years. (By tuning I mean
picking some reasonable starting values, all loops IME need
a little real time tweaking from the starting values.)
I've always used the Zeigler-Nichols oscillation method,
and looked no further.

Now my boss asked for some other technique and I started
reading about tuning from the step response. (What Z-N call
the Process-reaction curve) Oh I stuck the Z-N paper here,
along with a few 'scope shots.

https://www.dropbox.com/sh/7lkqazuy6wvzknk/
AADcZtTvFvaD3TEzeJiAjEy6a?
dl=0

If only I'd read the paper ~20 years ago.

This is a more up to date rehash of the same things.
http://faculty.mercer.edu/jenkins_he/documents/
TuningforPIDControllers.pdf

For completeness I'll stick in Tim W's excellent article.
(Though unfortunately he doesn't talk about Z-N directly.)
http://www.wescottdesign.com/articles/pid/
pidWithoutAPhd.pdf
(Maybe he'll add an appendix for me. :^)

Nope. For two reasons:

First, because I've never used Z-N, never felt the need, and
so never built up practical experience with it.

Second, because all the industrial control guys I've talked
to, both in person and 'lectronically, tell me that Z-N gets
you a good starting point but then you need to either tweak
things, do measurement + system ID + design, or do frequency
sweeps + design. There's enough material out there on it,
and too many people that think that it's the bees knees.
Huh, Ok I only said it was my starting point. Then do some
step responses and see what it looks like. (I'm mostly an
idiot when it comes to thermal loops... I'm starting to think
that 1/2 the time I got lucky and put the sensor in a good
place, and the other times, not so good.) Well sometimes you
just need to put the sensor at the point you care about.... I
could think more about where to put the heater.
Grumble....

What's your starting point?

So if you look at the step response in my dropbox link,
there is a lag of 300-350 ms. And if you look at the
oscillations at the "ultimate gain"
you'll see a period of about 2 seconds. And lo and behold,
2*pi*tau (the lag) = period! That is very satisfying.

Estimating the needed gain (proportional term) from the
step response, didn't work out that well in this case. But
I'm running at gains that are ~20% the ultimate gain, so
oscillation method didn't work all that well either in this
regard.


Oh I do have one question for Tim.
You show a thermal control loop with just integrating
control. Does this really work? (I would have thought it
would just oscillate.)
Or do you have a little bit of proportional term in there?

If you keep a firm lid on your ambitions, certainly. For
most systems you won't get nearly the speed you'd get with
some proportional action or even proportional + derivative
-- but any system that has a flat gain characteristic going
down to zero frequency can be controlled with just an
integrator, if you're patient.

The one thermal loop that I've done had a simple first-order
response well beyond any reasonable loop bandwidth. So it
got a simple PI controller and worked just grand.
Huh, OK I'll have to try, the loss is in the delay/lag?
This thermal loop is fast, (at least for me)
so ten seconds is no problem.

I get starting points for thermal loops by looking at the
open-loop step response. In small systems it's usually well
approximated by a time delay cascaded with an integrator. Thus
the plant's transfer function is approximately

H_plant(f) ~ (1/j 2 pi f tau_I) exp(-j 2 pi f tau_delay)
Right, I went looking for a pulse response with "your"
delay*RC time constant, model in mind.
Seem's like Z-N did it before we were born... which I love.
They also give a guess at the gain, from the pulse response.


Then I do the usual frequency-compensation thing, aiming for
about a 60-degree phase margin. If the system is reasonably
well insulated,
the effect of the heat leak is usually small enough to ignore
for frequency compensation purposes, but if it isn't, you need
to move the pole from zero out to (j 2 pi f tau_leak).
Huh, Phil, lots of the time you're talking over my head,
(or at least not where my head is at the moment.)

I'm not sure where 60 degree's is, but I look at the set point
step response and lower the gain if there's not too much
ringing.
Say a Q of 1-2. (probably closer to one, and then in production
each unit is a bit different.)

You need to learn frequency-domain tuning. Just looking at the
time domain response is not a reliable way to get the best
response and a known loop robustness.

Grin, right. (There are bucket loads of things I need to learn.)
You do a frequency sweep with the loop closed?
Do you have a video showing that? :^)

It's on the short list -- which is about half a dozen titles long.

I describe it in my book. And in this article:

http://wescottdesign.com/articles/FreqMeas/freq_meas.html

The basic idea is that if you measure the phase and amplitude of a
sine wave going into a block, and the phase and amplitude of a sine
wave coming out of that block, then you can calculate the block's
response at that frequency by doing a complex division (amplitude out
/ amplitude in,
phase out - phase in). Then repeat for a whole bunch of frequencies.

Right. I've mostly just kept track of the amplitude response, though
you can always see the phase bounce around too.
(It must be Krammers and Kronig, or someone earlier, who observed that
you can see the phase response before the amplitude.)

Hmm I've got a Z-N oscillation question, moving to SED,
in hopes of more experience.


It's easy to build this into code if you're working with a
microprocessor; if you're not there are instruments out there
("dynamic system analyzers" or "transfer function analyzers") that do
it. If National Instruments isn't utterly crazy there's a pre-made
VI to do this in labview with one analog I/O brick.

My Rigol sig gen goes down to ~uHz. I can do a frequency sweep, It's
got two outputs which I can phase lock and set at 90 degrees. In
principle, I can measure everything myself.

I hate long time sweeps... you don't know till hours later if you f'ed
something up. It's much faster to understand the step response, or
oscillation method.

George H.

I suspect that there's one out there for less than $500. And you can
always do it by hand with a signal generator and an O-scope.

Once you know what you are doing (I'm not claiming that *I* know.)
Doesn't the step response (I'm thinking closed loop) contain the
same information as a frequency sweep?

In theory, yes. In practice the good stuff is often swamped out by
either noise or distortion. In the past I've done a pretty good job
of doing system identification from a customer-supplied step response
-- but I was limiting the performance of the loop based on what I
didn't know about the plant, rather than a sure idea of what the
plant wasn't capable of.

You could probably take a whole bunch of step responses and average
them together, and then fit a system model to the resulting average
-- that would at least knock down the effects of noise. But you'd
want to do real honest-to-god system ID, where you're fitting a model
in the Laplace domain to the step response, not just that Z-N stuff.

--
Tim Wescott Control systems, embedded software and circuit design I'm
looking for work! See my website if you're interested
http://www.wescottdesign.com

Hey what about doing the FFT of the pulse response?
Right, I know about FFT's (Fir spectrometers as a post doc.)
There's low frequency information in the long tail.. noisy part of the
response... In this case I'm more interested in the high frequency, I
think?
Click to expand...

You're most interested in the frequencies around the loop closing
frequency. You can take the FFT of the response to a step or a square
wave (I'd get it w.r.t a square wave, with lots of reps).

* Take the Fourier transform of the excitation
* Take the Fourier transform of the response (get the delay right!)
* Zero out the response's FFT in those bins where the excitation's FFT is
zero
* Divide the response's FFT by the excitation's FFT for the non-zero bins

You'll have gain and phase, at least at those points.

I haven't done this so I can't guarantee it -- but what the heck, maybe
it'll work out just fine.

--

Tim Wescott
Wescott Design Services
http://www.wescottdesign.com

I'm looking for work -- see my website!
 
On Thursday, April 27, 2017 at 7:45:30 PM UTC-4, Tim Wescott wrote:
On Thu, 27 Apr 2017 16:10:26 -0700, George Herold wrote:

On Thursday, April 27, 2017 at 6:50:06 PM UTC-4, George Herold wrote:
On Wednesday, April 26, 2017 at 10:47:38 AM UTC-4, Tim Wescott wrote:
On Wed, 26 Apr 2017 04:45:23 -0700, George Herold wrote:

On Tuesday, April 25, 2017 at 11:20:50 PM UTC-4, Tim Wescott wrote:
On Tue, 25 Apr 2017 17:02:48 -0700, George Herold wrote:

On Tuesday, April 25, 2017 at 1:22:30 PM UTC-4, Phil Hobbs
wrote:
On 04/24/2017 10:22 PM, George Herold wrote:
On Monday, April 24, 2017 at 8:52:04 PM UTC-4, Tim Wescott
wrote:
On Mon, 24 Apr 2017 12:29:24 -0700, George Herold wrote:

Hi all, this is not so much a question, as me just
spouting.
If I do say (spout) something wrong, or that you disagree
with or that you don't understand, then please do speak up.

I've been tuning loops for years. (By tuning I mean
picking some reasonable starting values, all loops IME need
a little real time tweaking from the starting values.)
I've always used the Zeigler-Nichols oscillation method,
and looked no further.

Now my boss asked for some other technique and I started
reading about tuning from the step response. (What Z-N call
the Process-reaction curve) Oh I stuck the Z-N paper here,
along with a few 'scope shots.

https://www.dropbox.com/sh/7lkqazuy6wvzknk/
AADcZtTvFvaD3TEzeJiAjEy6a?
dl=0

If only I'd read the paper ~20 years ago.

This is a more up to date rehash of the same things.
http://faculty.mercer.edu/jenkins_he/documents/
TuningforPIDControllers.pdf

For completeness I'll stick in Tim W's excellent article.
(Though unfortunately he doesn't talk about Z-N directly.)
http://www.wescottdesign.com/articles/pid/
pidWithoutAPhd.pdf
(Maybe he'll add an appendix for me. :^)

Nope. For two reasons:

First, because I've never used Z-N, never felt the need, and
so never built up practical experience with it.

Second, because all the industrial control guys I've talked
to, both in person and 'lectronically, tell me that Z-N gets
you a good starting point but then you need to either tweak
things, do measurement + system ID + design, or do frequency
sweeps + design. There's enough material out there on it,
and too many people that think that it's the bees knees.
Huh, Ok I only said it was my starting point. Then do some
step responses and see what it looks like. (I'm mostly an
idiot when it comes to thermal loops... I'm starting to think
that 1/2 the time I got lucky and put the sensor in a good
place, and the other times, not so good.) Well sometimes you
just need to put the sensor at the point you care about.... I
could think more about where to put the heater.
Grumble....

What's your starting point?

So if you look at the step response in my dropbox link,
there is a lag of 300-350 ms. And if you look at the
oscillations at the "ultimate gain"
you'll see a period of about 2 seconds. And lo and behold,
2*pi*tau (the lag) = period! That is very satisfying.

Estimating the needed gain (proportional term) from the
step response, didn't work out that well in this case. But
I'm running at gains that are ~20% the ultimate gain, so
oscillation method didn't work all that well either in this
regard.


Oh I do have one question for Tim.
You show a thermal control loop with just integrating
control. Does this really work? (I would have thought it
would just oscillate.)
Or do you have a little bit of proportional term in there?

If you keep a firm lid on your ambitions, certainly. For
most systems you won't get nearly the speed you'd get with
some proportional action or even proportional + derivative
-- but any system that has a flat gain characteristic going
down to zero frequency can be controlled with just an
integrator, if you're patient.

The one thermal loop that I've done had a simple first-order
response well beyond any reasonable loop bandwidth. So it
got a simple PI controller and worked just grand.
Huh, OK I'll have to try, the loss is in the delay/lag?
This thermal loop is fast, (at least for me)
so ten seconds is no problem.

I get starting points for thermal loops by looking at the
open-loop step response. In small systems it's usually well
approximated by a time delay cascaded with an integrator. Thus
the plant's transfer function is approximately

H_plant(f) ~ (1/j 2 pi f tau_I) exp(-j 2 pi f tau_delay)
Right, I went looking for a pulse response with "your"
delay*RC time constant, model in mind.
Seem's like Z-N did it before we were born... which I love.
They also give a guess at the gain, from the pulse response.


Then I do the usual frequency-compensation thing, aiming for
about a 60-degree phase margin. If the system is reasonably
well insulated,
the effect of the heat leak is usually small enough to ignore
for frequency compensation purposes, but if it isn't, you need
to move the pole from zero out to (j 2 pi f tau_leak).
Huh, Phil, lots of the time you're talking over my head,
(or at least not where my head is at the moment.)

I'm not sure where 60 degree's is, but I look at the set point
step response and lower the gain if there's not too much
ringing.
Say a Q of 1-2. (probably closer to one, and then in production
each unit is a bit different.)

You need to learn frequency-domain tuning. Just looking at the
time domain response is not a reliable way to get the best
response and a known loop robustness.

Grin, right. (There are bucket loads of things I need to learn.)
You do a frequency sweep with the loop closed?
Do you have a video showing that? :^)

It's on the short list -- which is about half a dozen titles long.

I describe it in my book. And in this article:

http://wescottdesign.com/articles/FreqMeas/freq_meas.html

The basic idea is that if you measure the phase and amplitude of a
sine wave going into a block, and the phase and amplitude of a sine
wave coming out of that block, then you can calculate the block's
response at that frequency by doing a complex division (amplitude out
/ amplitude in,
phase out - phase in). Then repeat for a whole bunch of frequencies.

Right. I've mostly just kept track of the amplitude response, though
you can always see the phase bounce around too.
(It must be Krammers and Kronig, or someone earlier, who observed that
you can see the phase response before the amplitude.)

Hmm I've got a Z-N oscillation question, moving to SED,
in hopes of more experience.


It's easy to build this into code if you're working with a
microprocessor; if you're not there are instruments out there
("dynamic system analyzers" or "transfer function analyzers") that do
it. If National Instruments isn't utterly crazy there's a pre-made
VI to do this in labview with one analog I/O brick.

My Rigol sig gen goes down to ~uHz. I can do a frequency sweep, It's
got two outputs which I can phase lock and set at 90 degrees. In
principle, I can measure everything myself.

I hate long time sweeps... you don't know till hours later if you f'ed
something up. It's much faster to understand the step response, or
oscillation method.

George H.

I suspect that there's one out there for less than $500. And you can
always do it by hand with a signal generator and an O-scope.

Once you know what you are doing (I'm not claiming that *I* know.)
Doesn't the step response (I'm thinking closed loop) contain the
same information as a frequency sweep?

In theory, yes. In practice the good stuff is often swamped out by
either noise or distortion. In the past I've done a pretty good job
of doing system identification from a customer-supplied step response
-- but I was limiting the performance of the loop based on what I
didn't know about the plant, rather than a sure idea of what the
plant wasn't capable of.

You could probably take a whole bunch of step responses and average
them together, and then fit a system model to the resulting average
-- that would at least knock down the effects of noise. But you'd
want to do real honest-to-god system ID, where you're fitting a model
in the Laplace domain to the step response, not just that Z-N stuff.

--
Tim Wescott Control systems, embedded software and circuit design I'm
looking for work! See my website if you're interested
http://www.wescottdesign.com

Hey what about doing the FFT of the pulse response?
Right, I know about FFT's (Fir spectrometers as a post doc.)
There's low frequency information in the long tail.. noisy part of the
response... In this case I'm more interested in the high frequency, I
think?

You're most interested in the frequencies around the loop closing
frequency. You can take the FFT of the response to a step or a square
wave (I'd get it w.r.t a square wave, with lots of reps).

* Take the Fourier transform of the excitation
* Take the Fourier transform of the response (get the delay right!)
* Zero out the response's FFT in those bins where the excitation's FFT is
zero
* Divide the response's FFT by the excitation's FFT for the non-zero bins

You'll have gain and phase, at least at those points.

I haven't done this so I can't guarantee it -- but what the heck, maybe
it'll work out just fine.
Click to expand...

Hmm this was only a half baked idea. But what do I pulse?
I mostly step the setpoint, and look at the error signal. But this
gives me a sharp edge... So do I have to change the load?
put a pulse of heat into my plate?

I'm not sure about all your subtractions. I think that if
I just do a pulse that is much shorter than all the times
in the system, then I can ignore the pulse FFT... it's a delta function.

(I'll give that a try today... it's Friday I get to have some fun. :^)

George H.
--

Tim Wescott
Wescott Design Services
http://www.wescottdesign.com

I'm looking for work -- see my website!
Click to expand...
 
On 04/27/2017 07:45 PM, Tim Wescott wrote:
On Thu, 27 Apr 2017 16:10:26 -0700, George Herold wrote:

On Thursday, April 27, 2017 at 6:50:06 PM UTC-4, George Herold wrote:
On Wednesday, April 26, 2017 at 10:47:38 AM UTC-4, Tim Wescott wrote:
On Wed, 26 Apr 2017 04:45:23 -0700, George Herold wrote:

On Tuesday, April 25, 2017 at 11:20:50 PM UTC-4, Tim Wescott wrote:
On Tue, 25 Apr 2017 17:02:48 -0700, George Herold wrote:

On Tuesday, April 25, 2017 at 1:22:30 PM UTC-4, Phil Hobbs
wrote:
On 04/24/2017 10:22 PM, George Herold wrote:
On Monday, April 24, 2017 at 8:52:04 PM UTC-4, Tim Wescott
wrote:
On Mon, 24 Apr 2017 12:29:24 -0700, George Herold wrote:

Hi all, this is not so much a question, as me just
spouting.
If I do say (spout) something wrong, or that you disagree
with or that you don't understand, then please do speak up.

I've been tuning loops for years. (By tuning I mean
picking some reasonable starting values, all loops IME need
a little real time tweaking from the starting values.)
I've always used the Zeigler-Nichols oscillation method,
and looked no further.

Now my boss asked for some other technique and I started
reading about tuning from the step response. (What Z-N call
the Process-reaction curve) Oh I stuck the Z-N paper here,
along with a few 'scope shots.

https://www.dropbox.com/sh/7lkqazuy6wvzknk/
AADcZtTvFvaD3TEzeJiAjEy6a?
dl=0

If only I'd read the paper ~20 years ago.

This is a more up to date rehash of the same things.
http://faculty.mercer.edu/jenkins_he/documents/
TuningforPIDControllers.pdf

For completeness I'll stick in Tim W's excellent article.
(Though unfortunately he doesn't talk about Z-N directly.)
http://www.wescottdesign.com/articles/pid/
pidWithoutAPhd.pdf
(Maybe he'll add an appendix for me. :^)

Nope. For two reasons:

First, because I've never used Z-N, never felt the need, and
so never built up practical experience with it.

Second, because all the industrial control guys I've talked
to, both in person and 'lectronically, tell me that Z-N gets
you a good starting point but then you need to either tweak
things, do measurement + system ID + design, or do frequency
sweeps + design. There's enough material out there on it,
and too many people that think that it's the bees knees.
Huh, Ok I only said it was my starting point. Then do some
step responses and see what it looks like. (I'm mostly an
idiot when it comes to thermal loops... I'm starting to think
that 1/2 the time I got lucky and put the sensor in a good
place, and the other times, not so good.) Well sometimes you
just need to put the sensor at the point you care about.... I
could think more about where to put the heater.
Grumble....

What's your starting point?

So if you look at the step response in my dropbox link,
there is a lag of 300-350 ms. And if you look at the
oscillations at the "ultimate gain"
you'll see a period of about 2 seconds. And lo and behold,
2*pi*tau (the lag) = period! That is very satisfying.

Estimating the needed gain (proportional term) from the
step response, didn't work out that well in this case. But
I'm running at gains that are ~20% the ultimate gain, so
oscillation method didn't work all that well either in this
regard.


Oh I do have one question for Tim.
You show a thermal control loop with just integrating
control. Does this really work? (I would have thought it
would just oscillate.)
Or do you have a little bit of proportional term in there?

If you keep a firm lid on your ambitions, certainly. For
most systems you won't get nearly the speed you'd get with
some proportional action or even proportional + derivative
-- but any system that has a flat gain characteristic going
down to zero frequency can be controlled with just an
integrator, if you're patient.

The one thermal loop that I've done had a simple first-order
response well beyond any reasonable loop bandwidth. So it
got a simple PI controller and worked just grand.
Huh, OK I'll have to try, the loss is in the delay/lag?
This thermal loop is fast, (at least for me)
so ten seconds is no problem.

I get starting points for thermal loops by looking at the
open-loop step response. In small systems it's usually well
approximated by a time delay cascaded with an integrator. Thus
the plant's transfer function is approximately

H_plant(f) ~ (1/j 2 pi f tau_I) exp(-j 2 pi f tau_delay)
Right, I went looking for a pulse response with "your"
delay*RC time constant, model in mind.
Seem's like Z-N did it before we were born... which I love.
They also give a guess at the gain, from the pulse response.


Then I do the usual frequency-compensation thing, aiming for
about a 60-degree phase margin. If the system is reasonably
well insulated,
the effect of the heat leak is usually small enough to ignore
for frequency compensation purposes, but if it isn't, you need
to move the pole from zero out to (j 2 pi f tau_leak).
Huh, Phil, lots of the time you're talking over my head,
(or at least not where my head is at the moment.)

I'm not sure where 60 degree's is, but I look at the set point
step response and lower the gain if there's not too much
ringing.
Say a Q of 1-2. (probably closer to one, and then in production
each unit is a bit different.)

You need to learn frequency-domain tuning. Just looking at the
time domain response is not a reliable way to get the best
response and a known loop robustness.

Grin, right. (There are bucket loads of things I need to learn.)
You do a frequency sweep with the loop closed?
Do you have a video showing that? :^)

It's on the short list -- which is about half a dozen titles long.

I describe it in my book. And in this article:

http://wescottdesign.com/articles/FreqMeas/freq_meas.html

The basic idea is that if you measure the phase and amplitude of a
sine wave going into a block, and the phase and amplitude of a sine
wave coming out of that block, then you can calculate the block's
response at that frequency by doing a complex division (amplitude out
/ amplitude in,
phase out - phase in). Then repeat for a whole bunch of frequencies.

Right. I've mostly just kept track of the amplitude response, though
you can always see the phase bounce around too.
(It must be Krammers and Kronig, or someone earlier, who observed that
you can see the phase response before the amplitude.)

Hmm I've got a Z-N oscillation question, moving to SED,
in hopes of more experience.


It's easy to build this into code if you're working with a
microprocessor; if you're not there are instruments out there
("dynamic system analyzers" or "transfer function analyzers") that do
it. If National Instruments isn't utterly crazy there's a pre-made
VI to do this in labview with one analog I/O brick.

My Rigol sig gen goes down to ~uHz. I can do a frequency sweep, It's
got two outputs which I can phase lock and set at 90 degrees. In
principle, I can measure everything myself.

I hate long time sweeps... you don't know till hours later if you f'ed
something up. It's much faster to understand the step response, or
oscillation method.

George H.

I suspect that there's one out there for less than $500. And you can
always do it by hand with a signal generator and an O-scope.

Once you know what you are doing (I'm not claiming that *I* know.)
Doesn't the step response (I'm thinking closed loop) contain the
same information as a frequency sweep?

In theory, yes. In practice the good stuff is often swamped out by
either noise or distortion. In the past I've done a pretty good job
of doing system identification from a customer-supplied step response
-- but I was limiting the performance of the loop based on what I
didn't know about the plant, rather than a sure idea of what the
plant wasn't capable of.

You could probably take a whole bunch of step responses and average
them together, and then fit a system model to the resulting average
-- that would at least knock down the effects of noise. But you'd
want to do real honest-to-god system ID, where you're fitting a model
in the Laplace domain to the step response, not just that Z-N stuff.

--
Tim Wescott Control systems, embedded software and circuit design I'm
looking for work! See my website if you're interested
http://www.wescottdesign.com

Hey what about doing the FFT of the pulse response?
Right, I know about FFT's (Fir spectrometers as a post doc.)
There's low frequency information in the long tail.. noisy part of the
response... In this case I'm more interested in the high frequency, I
think?

You're most interested in the frequencies around the loop closing
frequency. You can take the FFT of the response to a step or a square
wave (I'd get it w.r.t a square wave, with lots of reps).

* Take the Fourier transform of the excitation
* Take the Fourier transform of the response (get the delay right!)
* Zero out the response's FFT in those bins where the excitation's FFT is
zero
* Divide the response's FFT by the excitation's FFT for the non-zero bins

You'll have gain and phase, at least at those points.

I haven't done this so I can't guarantee it -- but what the heck, maybe
it'll work out just fine.
Click to expand...

Or just get an HP 35665A, 35660A, or 3562A DSA from eBay. I got one for
$300 back in 2009, and got an all-options ROM (customized with the
unit's serial number) for $65 from glkinst.com. That gets you the swept
sine, AWG and noise waveform sources, plus HP Instrument Basic, extra
memory, curve synthesis, and some other nice things. It's worth getting
the pc-style keyboard for it (those are cheap).

FFTs are famous for generating reasonable-looking wrong answers when one
is careless.

Cheers

Phil Hobbs

--
Dr Philip C D Hobbs
Principal Consultant
ElectroOptical Innovations LLC
Optics, Electro-optics, Photonics, Analog Electronics

160 North State Road #203
Briarcliff Manor NY 10510

hobbs at electrooptical dot net
http://electrooptical.net
 
On 04/25/2017 08:02 PM, George Herold wrote:
On Tuesday, April 25, 2017 at 1:22:30 PM UTC-4, Phil Hobbs wrote:
On 04/24/2017 10:22 PM, George Herold wrote:
On Monday, April 24, 2017 at 8:52:04 PM UTC-4, Tim Wescott wrote:
On Mon, 24 Apr 2017 12:29:24 -0700, George Herold wrote:

Hi all, this is not so much a question, as me just spouting.
If I do say (spout) something wrong, or that you disagree with or that
you don't understand, then please do speak up.

I've been tuning loops for years. (By tuning I mean picking some
reasonable starting values, all loops IME need a little real time
tweaking from the starting values.) I've always used the
Zeigler-Nichols oscillation method, and looked no further.

Now my boss asked for some other technique and I started reading about
tuning from the step response. (What Z-N call the Process-reaction
curve) Oh I stuck the Z-N paper here, along with a few 'scope shots.

https://www.dropbox.com/sh/7lkqazuy6wvzknk/AADcZtTvFvaD3TEzeJiAjEy6a?
dl=0

If only I'd read the paper ~20 years ago.

This is a more up to date rehash of the same things.
http://faculty.mercer.edu/jenkins_he/documents/
TuningforPIDControllers.pdf

For completeness I'll stick in Tim W's excellent article.
(Though unfortunately he doesn't talk about Z-N directly.)
http://www.wescottdesign.com/articles/pid/pidWithoutAPhd.pdf (Maybe
he'll add an appendix for me. :^)

Nope. For two reasons:

First, because I've never used Z-N, never felt the need, and so never
built up practical experience with it.

Second, because all the industrial control guys I've talked to, both in
person and 'lectronically, tell me that Z-N gets you a good starting
point but then you need to either tweak things, do measurement + system
ID + design, or do frequency sweeps + design. There's enough material
out there on it, and too many people that think that it's the bees knees.
Huh, Ok I only said it was my starting point. Then do some step responses
and see what it looks like. (I'm mostly an idiot when it comes to
thermal loops... I'm starting to think that 1/2 the time I got lucky
and put the sensor in a good place, and the other times, not so good.)
Well sometimes you just need to put the sensor at the point you care
about.... I could think more about where to put the heater.
Grumble....

What's your starting point?

So if you look at the step response in my dropbox link, there is a lag
of 300-350 ms. And if you look at the oscillations at the "ultimate
gain"
you'll see a period of about 2 seconds. And lo and behold,
2*pi*tau (the lag) = period! That is very satisfying.

Estimating the needed gain (proportional term) from the step response,
didn't work out that well in this case. But I'm running at gains that
are ~20% the ultimate gain, so oscillation method didn't work all that
well either in this regard.


Oh I do have one question for Tim.
You show a thermal control loop with just integrating control.
Does this really work? (I would have thought it would just oscillate.)
Or do you have a little bit of proportional term in there?

If you keep a firm lid on your ambitions, certainly. For most systems
you won't get nearly the speed you'd get with some proportional action or
even proportional + derivative -- but any system that has a flat gain
characteristic going down to zero frequency can be controlled with just
an integrator, if you're patient.

The one thermal loop that I've done had a simple first-order response
well beyond any reasonable loop bandwidth. So it got a simple PI
controller and worked just grand.
Huh, OK I'll have to try, the loss is in the delay/lag?
This thermal loop is fast, (at least for me)
so ten seconds is no problem.

I get starting points for thermal loops by looking at the open-loop step
response. In small systems it's usually well approximated by a time
delay cascaded with an integrator. Thus the plant's transfer function
is approximately

H_plant(f) ~ (1/j 2 pi f tau_I) exp(-j 2 pi f tau_delay)
Right, I went looking for a pulse response with "your"
delay*RC time constant, model in mind.
Seem's like Z-N did it before we were born... which I love.
They also give a guess at the gain, from the pulse response.


Then I do the usual frequency-compensation thing, aiming for about a
60-degree phase margin. If the system is reasonably well insulated, the
effect of the heat leak is usually small enough to ignore for frequency
compensation purposes, but if it isn't, you need to move the pole from
zero out to (j 2 pi f tau_leak).
Huh, Phil, lots of the time you're talking over my head,
(or at least not where my head is at the moment.)
Click to expand...

Not difficult, just too telegraphic. You replace the factor of
1/j omega f tau_I

with

1/(tau_leak/tau_i + j omega f tau_i),

so the gain maxes out and Delta-T decays exponentially when the heater
is turned off, instead of just sitting there the way it would if it were
really an integrator.

The heat leak isn't usually too relevant out at the unity gain cross,
because if the ratio isn't pretty big you won't have good thermal
control anyway.

I'm not sure where 60 degree's is, but I look at the set point
step response and lower the gain if there's not too much ringing.
Say a Q of 1-2. (probably closer to one, and then in production
each unit is a bit different.)
Click to expand...

I usually like thermal loops to be a bit better damped than that, to
avoid surprises if something in the plant drifts a bit.

Cheers

Phil Hobbs

--
Dr Philip C D Hobbs
Principal Consultant
ElectroOptical Innovations LLC
Optics, Electro-optics, Photonics, Analog Electronics

160 North State Road #203
Briarcliff Manor NY 10510

hobbs at electrooptical dot net
http://electrooptical.net
 
On Fri, 28 Apr 2017 10:58:58 -0400, Phil Hobbs wrote:

On 04/27/2017 07:45 PM, Tim Wescott wrote:
On Thu, 27 Apr 2017 16:10:26 -0700, George Herold wrote:

On Thursday, April 27, 2017 at 6:50:06 PM UTC-4, George Herold wrote:
On Wednesday, April 26, 2017 at 10:47:38 AM UTC-4, Tim Wescott wrote:
On Wed, 26 Apr 2017 04:45:23 -0700, George Herold wrote:

On Tuesday, April 25, 2017 at 11:20:50 PM UTC-4, Tim Wescott wrote:
On Tue, 25 Apr 2017 17:02:48 -0700, George Herold wrote:

On Tuesday, April 25, 2017 at 1:22:30 PM UTC-4, Phil Hobbs wrote:
On 04/24/2017 10:22 PM, George Herold wrote:
On Monday, April 24, 2017 at 8:52:04 PM UTC-4, Tim Wescott
wrote:
On Mon, 24 Apr 2017 12:29:24 -0700, George Herold wrote:

Hi all, this is not so much a question, as me just spouting.
If I do say (spout) something wrong, or that you disagree
with or that you don't understand, then please do speak up.

I've been tuning loops for years. (By tuning I mean picking
some reasonable starting values, all loops IME need a little
real time tweaking from the starting values.) I've always
used the Zeigler-Nichols oscillation method, and looked no
further.

Now my boss asked for some other technique and I started
reading about tuning from the step response. (What Z-N call
the Process-reaction curve) Oh I stuck the Z-N paper here,
along with a few 'scope shots.

https://www.dropbox.com/sh/7lkqazuy6wvzknk/
AADcZtTvFvaD3TEzeJiAjEy6a?
dl=0

If only I'd read the paper ~20 years ago.

This is a more up to date rehash of the same things.
http://faculty.mercer.edu/jenkins_he/documents/
TuningforPIDControllers.pdf

For completeness I'll stick in Tim W's excellent article.
(Though unfortunately he doesn't talk about Z-N directly.)
http://www.wescottdesign.com/articles/pid/
pidWithoutAPhd.pdf
(Maybe he'll add an appendix for me. :^)

Nope. For two reasons:

First, because I've never used Z-N, never felt the need, and
so never built up practical experience with it.

Second, because all the industrial control guys I've talked
to, both in person and 'lectronically, tell me that Z-N gets
you a good starting point but then you need to either tweak
things, do measurement + system ID + design, or do frequency
sweeps + design. There's enough material out there on it, and
too many people that think that it's the bees knees.
Huh, Ok I only said it was my starting point. Then do some step
responses and see what it looks like. (I'm mostly an idiot
when it comes to thermal loops... I'm starting to think that
1/2 the time I got lucky and put the sensor in a good place,
and the other times, not so good.) Well sometimes you just need
to put the sensor at the point you care about.... I could think
more about where to put the heater.
Grumble....

What's your starting point?

So if you look at the step response in my dropbox link, there
is a lag of 300-350 ms. And if you look at the oscillations
at the "ultimate gain"
you'll see a period of about 2 seconds. And lo and behold,
2*pi*tau (the lag) = period! That is very satisfying.

Estimating the needed gain (proportional term) from the step
response, didn't work out that well in this case. But I'm
running at gains that are ~20% the ultimate gain, so
oscillation method didn't work all that well either in this
regard.


Oh I do have one question for Tim.
You show a thermal control loop with just integrating
control. Does this really work? (I would have thought it
would just oscillate.)
Or do you have a little bit of proportional term in there?

If you keep a firm lid on your ambitions, certainly. For most
systems you won't get nearly the speed you'd get with some
proportional action or even proportional + derivative -- but
any system that has a flat gain characteristic going down to
zero frequency can be controlled with just an integrator, if
you're patient.

The one thermal loop that I've done had a simple first-order
response well beyond any reasonable loop bandwidth. So it got
a simple PI controller and worked just grand.
Huh, OK I'll have to try, the loss is in the delay/lag?
This thermal loop is fast, (at least for me)
so ten seconds is no problem.

I get starting points for thermal loops by looking at the
open-loop step response. In small systems it's usually well
approximated by a time delay cascaded with an integrator. Thus
the plant's transfer function is approximately

H_plant(f) ~ (1/j 2 pi f tau_I) exp(-j 2 pi f tau_delay)
Right, I went looking for a pulse response with "your" delay*RC
time constant, model in mind.
Seem's like Z-N did it before we were born... which I love. They
also give a guess at the gain, from the pulse response.


Then I do the usual frequency-compensation thing, aiming for
about a 60-degree phase margin. If the system is reasonably
well insulated,
the effect of the heat leak is usually small enough to ignore
for frequency compensation purposes, but if it isn't, you need
to move the pole from zero out to (j 2 pi f tau_leak).
Huh, Phil, lots of the time you're talking over my head,
(or at least not where my head is at the moment.)

I'm not sure where 60 degree's is, but I look at the set point
step response and lower the gain if there's not too much ringing.
Say a Q of 1-2. (probably closer to one, and then in production
each unit is a bit different.)

You need to learn frequency-domain tuning. Just looking at the
time domain response is not a reliable way to get the best
response and a known loop robustness.

Grin, right. (There are bucket loads of things I need to learn.)
You do a frequency sweep with the loop closed?
Do you have a video showing that? :^)

It's on the short list -- which is about half a dozen titles long.

I describe it in my book. And in this article:

http://wescottdesign.com/articles/FreqMeas/freq_meas.html

The basic idea is that if you measure the phase and amplitude of a
sine wave going into a block, and the phase and amplitude of a sine
wave coming out of that block, then you can calculate the block's
response at that frequency by doing a complex division (amplitude
out / amplitude in,
phase out - phase in). Then repeat for a whole bunch of
frequencies.

Right. I've mostly just kept track of the amplitude response, though
you can always see the phase bounce around too.
(It must be Krammers and Kronig, or someone earlier, who observed
that you can see the phase response before the amplitude.)

Hmm I've got a Z-N oscillation question, moving to SED,
in hopes of more experience.


It's easy to build this into code if you're working with a
microprocessor; if you're not there are instruments out there
("dynamic system analyzers" or "transfer function analyzers") that
do it. If National Instruments isn't utterly crazy there's a
pre-made VI to do this in labview with one analog I/O brick.

My Rigol sig gen goes down to ~uHz. I can do a frequency sweep, It's
got two outputs which I can phase lock and set at 90 degrees. In
principle, I can measure everything myself.

I hate long time sweeps... you don't know till hours later if you
f'ed something up. It's much faster to understand the step response,
or oscillation method.

George H.

I suspect that there's one out there for less than $500. And you
can always do it by hand with a signal generator and an O-scope.

Once you know what you are doing (I'm not claiming that *I* know.)
Doesn't the step response (I'm thinking closed loop) contain the
same information as a frequency sweep?

In theory, yes. In practice the good stuff is often swamped out by
either noise or distortion. In the past I've done a pretty good job
of doing system identification from a customer-supplied step
response -- but I was limiting the performance of the loop based on
what I didn't know about the plant, rather than a sure idea of what
the plant wasn't capable of.

You could probably take a whole bunch of step responses and average
them together, and then fit a system model to the resulting average
-- that would at least knock down the effects of noise. But you'd
want to do real honest-to-god system ID, where you're fitting a
model in the Laplace domain to the step response, not just that Z-N
stuff.

--
Tim Wescott Control systems, embedded software and circuit design
I'm looking for work! See my website if you're interested
http://www.wescottdesign.com

Hey what about doing the FFT of the pulse response?
Right, I know about FFT's (Fir spectrometers as a post doc.)
There's low frequency information in the long tail.. noisy part of the
response... In this case I'm more interested in the high frequency, I
think?

You're most interested in the frequencies around the loop closing
frequency. You can take the FFT of the response to a step or a square
wave (I'd get it w.r.t a square wave, with lots of reps).

* Take the Fourier transform of the excitation * Take the Fourier
transform of the response (get the delay right!)
* Zero out the response's FFT in those bins where the excitation's FFT
is zero * Divide the response's FFT by the excitation's FFT for the
non-zero bins

You'll have gain and phase, at least at those points.

I haven't done this so I can't guarantee it -- but what the heck, maybe
it'll work out just fine.


Or just get an HP 35665A, 35660A, or 3562A DSA from eBay. I got one for
$300 back in 2009, and got an all-options ROM (customized with the
unit's serial number) for $65 from glkinst.com. That gets you the swept
sine, AWG and noise waveform sources, plus HP Instrument Basic, extra
memory, curve synthesis, and some other nice things. It's worth getting
the pc-style keyboard for it (those are cheap).
Click to expand...

Oooh yea. Useful instrument, if you're doing control loops in analog.
Super-easy to build the pertinent bits into code if you're doing control
loops in software, and reasonably easy to do it so that any processing
that's incurred doesn't impact the calculation delay for the loop --
which lets you get exact sweeps.

Kinda big compared to the itty bitty things that you can get these days,
but essential if you're doing serious work sans software.

FFTs are famous for generating reasonable-looking wrong answers when one
is careless.
Click to expand...

Yes -- George, in case it wasn't apparent, I feel a sweep is MUCH BETTER,
for just this reason.

--
Tim Wescott
Control systems, embedded software and circuit design
I'm looking for work! See my website if you're interested
http://www.wescottdesign.com
 
On Fri, 28 Apr 2017 06:23:45 -0700, George Herold wrote:

On Thursday, April 27, 2017 at 7:45:30 PM UTC-4, Tim Wescott wrote:
On Thu, 27 Apr 2017 16:10:26 -0700, George Herold wrote:

On Thursday, April 27, 2017 at 6:50:06 PM UTC-4, George Herold wrote:
On Wednesday, April 26, 2017 at 10:47:38 AM UTC-4, Tim Wescott
wrote:
On Wed, 26 Apr 2017 04:45:23 -0700, George Herold wrote:

On Tuesday, April 25, 2017 at 11:20:50 PM UTC-4, Tim Wescott
wrote:
On Tue, 25 Apr 2017 17:02:48 -0700, George Herold wrote:

On Tuesday, April 25, 2017 at 1:22:30 PM UTC-4, Phil Hobbs
wrote:
On 04/24/2017 10:22 PM, George Herold wrote:
On Monday, April 24, 2017 at 8:52:04 PM UTC-4, Tim Wescott
wrote:
On Mon, 24 Apr 2017 12:29:24 -0700, George Herold wrote:

Hi all, this is not so much a question, as me just
spouting.
If I do say (spout) something wrong, or that you
disagree with or that you don't understand, then please
do speak up.

I've been tuning loops for years. (By tuning I mean
picking some reasonable starting values, all loops IME
need a little real time tweaking from the starting
values.) I've always used the Zeigler-Nichols
oscillation method, and looked no further.

Now my boss asked for some other technique and I started
reading about tuning from the step response. (What Z-N
call the Process-reaction curve) Oh I stuck the Z-N
paper here, along with a few 'scope shots.

https://www.dropbox.com/sh/7lkqazuy6wvzknk/
AADcZtTvFvaD3TEzeJiAjEy6a?
dl=0

If only I'd read the paper ~20 years ago.

This is a more up to date rehash of the same things.
http://faculty.mercer.edu/jenkins_he/documents/
TuningforPIDControllers.pdf

For completeness I'll stick in Tim W's excellent
article. (Though unfortunately he doesn't talk about Z-N
directly.) http://www.wescottdesign.com/articles/pid/
pidWithoutAPhd.pdf
(Maybe he'll add an appendix for me. :^)

Nope. For two reasons:

First, because I've never used Z-N, never felt the need,
and so never built up practical experience with it.

Second, because all the industrial control guys I've
talked to, both in person and 'lectronically, tell me
that Z-N gets you a good starting point but then you need
to either tweak things, do measurement + system ID +
design, or do frequency sweeps + design. There's enough
material out there on it, and too many people that think
that it's the bees knees.
Huh, Ok I only said it was my starting point. Then do some
step responses and see what it looks like. (I'm mostly an
idiot when it comes to thermal loops... I'm starting to
think that 1/2 the time I got lucky and put the sensor in
a good place, and the other times, not so good.) Well
sometimes you just need to put the sensor at the point you
care about.... I could think more about where to put the
heater.
Grumble....

What's your starting point?

So if you look at the step response in my dropbox link,
there is a lag of 300-350 ms. And if you look at the
oscillations at the "ultimate gain"
you'll see a period of about 2 seconds. And lo and
behold,
2*pi*tau (the lag) = period! That is very satisfying.

Estimating the needed gain (proportional term) from the
step response, didn't work out that well in this case.
But I'm running at gains that are ~20% the ultimate
gain, so oscillation method didn't work all that well
either in this regard.


Oh I do have one question for Tim.
You show a thermal control loop with just integrating
control. Does this really work? (I would have thought
it would just oscillate.)
Or do you have a little bit of proportional term in
there?

If you keep a firm lid on your ambitions, certainly. For
most systems you won't get nearly the speed you'd get
with some proportional action or even proportional +
derivative -- but any system that has a flat gain
characteristic going down to zero frequency can be
controlled with just an integrator, if you're patient.

The one thermal loop that I've done had a simple
first-order response well beyond any reasonable loop
bandwidth. So it got a simple PI controller and worked
just grand.
Huh, OK I'll have to try, the loss is in the delay/lag?
This thermal loop is fast, (at least for me)
so ten seconds is no problem.

I get starting points for thermal loops by looking at the
open-loop step response. In small systems it's usually well
approximated by a time delay cascaded with an integrator.
Thus the plant's transfer function is approximately

H_plant(f) ~ (1/j 2 pi f tau_I) exp(-j 2 pi f tau_delay)
Right, I went looking for a pulse response with "your"
delay*RC time constant, model in mind.
Seem's like Z-N did it before we were born... which I love.
They also give a guess at the gain, from the pulse response.


Then I do the usual frequency-compensation thing, aiming for
about a 60-degree phase margin. If the system is reasonably
well insulated,
the effect of the heat leak is usually small enough to
ignore for frequency compensation purposes, but if it isn't,
you need to move the pole from zero out to (j 2 pi f
tau_leak).
Huh, Phil, lots of the time you're talking over my head,
(or at least not where my head is at the moment.)

I'm not sure where 60 degree's is, but I look at the set
point step response and lower the gain if there's not too
much ringing.
Say a Q of 1-2. (probably closer to one, and then in
production each unit is a bit different.)

You need to learn frequency-domain tuning. Just looking at the
time domain response is not a reliable way to get the best
response and a known loop robustness.

Grin, right. (There are bucket loads of things I need to
learn.) You do a frequency sweep with the loop closed?
Do you have a video showing that? :^)

It's on the short list -- which is about half a dozen titles long.

I describe it in my book. And in this article:

http://wescottdesign.com/articles/FreqMeas/freq_meas.html

The basic idea is that if you measure the phase and amplitude of a
sine wave going into a block, and the phase and amplitude of a
sine wave coming out of that block, then you can calculate the
block's response at that frequency by doing a complex division
(amplitude out / amplitude in,
phase out - phase in). Then repeat for a whole bunch of
frequencies.

Right. I've mostly just kept track of the amplitude response,
though you can always see the phase bounce around too.
(It must be Krammers and Kronig, or someone earlier, who observed
that you can see the phase response before the amplitude.)

Hmm I've got a Z-N oscillation question, moving to SED,
in hopes of more experience.


It's easy to build this into code if you're working with a
microprocessor; if you're not there are instruments out there
("dynamic system analyzers" or "transfer function analyzers") that
do it. If National Instruments isn't utterly crazy there's a
pre-made VI to do this in labview with one analog I/O brick.

My Rigol sig gen goes down to ~uHz. I can do a frequency sweep,
It's got two outputs which I can phase lock and set at 90 degrees.
In principle, I can measure everything myself.

I hate long time sweeps... you don't know till hours later if you
f'ed something up. It's much faster to understand the step
response, or oscillation method.

George H.

I suspect that there's one out there for less than $500. And you
can always do it by hand with a signal generator and an O-scope.

Once you know what you are doing (I'm not claiming that *I*
know.) Doesn't the step response (I'm thinking closed loop)
contain the same information as a frequency sweep?

In theory, yes. In practice the good stuff is often swamped out
by either noise or distortion. In the past I've done a pretty
good job of doing system identification from a customer-supplied
step response -- but I was limiting the performance of the loop
based on what I didn't know about the plant, rather than a sure
idea of what the plant wasn't capable of.

You could probably take a whole bunch of step responses and
average them together, and then fit a system model to the
resulting average -- that would at least knock down the effects of
noise. But you'd want to do real honest-to-god system ID, where
you're fitting a model in the Laplace domain to the step response,
not just that Z-N stuff.

--
Tim Wescott Control systems, embedded software and circuit design
I'm looking for work! See my website if you're interested
http://www.wescottdesign.com

Hey what about doing the FFT of the pulse response?
Right, I know about FFT's (Fir spectrometers as a post doc.)
There's low frequency information in the long tail.. noisy part of
the response... In this case I'm more interested in the high
frequency, I think?

You're most interested in the frequencies around the loop closing
frequency. You can take the FFT of the response to a step or a square
wave (I'd get it w.r.t a square wave, with lots of reps).

* Take the Fourier transform of the excitation * Take the Fourier
transform of the response (get the delay right!)
* Zero out the response's FFT in those bins where the excitation's FFT
is zero * Divide the response's FFT by the excitation's FFT for the
non-zero bins

You'll have gain and phase, at least at those points.

I haven't done this so I can't guarantee it -- but what the heck, maybe
it'll work out just fine.

Hmm this was only a half baked idea. But what do I pulse?
I mostly step the setpoint, and look at the error signal. But this gives
me a sharp edge... So do I have to change the load?
put a pulse of heat into my plate?

I'm not sure about all your subtractions. I think that if I just do a
pulse that is much shorter than all the times in the system, then I can
ignore the pulse FFT... it's a delta function.

(I'll give that a try today... it's Friday I get to have some fun. :^)
Click to expand...

If you can keep things at a reasonable temperature, do it open loop.

The problem with a narrow pulse is that you'll get a small signal, and
hence a low SNR on your measurement. You can fix that by averaging --
but by the time you average a bunch of narrow pulses, do all the math,
etc., you've done _more_ work than doing a swept-sine measurement, and
have more questionable results.

Listen to Dr. Phil on this one (since he's saying the same thing as
me...).

If you're doing this for students, you should equip the thing for the
various measurements needed for formal system ID -- you should NOT just
be teaching Z-N plus seat-of-the-pants tuning!

--
Tim Wescott
Control systems, embedded software and circuit design
I'm looking for work! See my website if you're interested
http://www.wescottdesign.com
 
On Friday, April 28, 2017 at 10:59:02 AM UTC-4, Phil Hobbs wrote:
On 04/27/2017 07:45 PM, Tim Wescott wrote:
On Thu, 27 Apr 2017 16:10:26 -0700, George Herold wrote:

On Thursday, April 27, 2017 at 6:50:06 PM UTC-4, George Herold wrote:
On Wednesday, April 26, 2017 at 10:47:38 AM UTC-4, Tim Wescott wrote:
On Wed, 26 Apr 2017 04:45:23 -0700, George Herold wrote:

On Tuesday, April 25, 2017 at 11:20:50 PM UTC-4, Tim Wescott wrote:
On Tue, 25 Apr 2017 17:02:48 -0700, George Herold wrote:

On Tuesday, April 25, 2017 at 1:22:30 PM UTC-4, Phil Hobbs
wrote:
On 04/24/2017 10:22 PM, George Herold wrote:
On Monday, April 24, 2017 at 8:52:04 PM UTC-4, Tim Wescott
wrote:
On Mon, 24 Apr 2017 12:29:24 -0700, George Herold wrote:

Hi all, this is not so much a question, as me just
spouting.
If I do say (spout) something wrong, or that you disagree
with or that you don't understand, then please do speak up.

I've been tuning loops for years. (By tuning I mean
picking some reasonable starting values, all loops IME need
a little real time tweaking from the starting values.)
I've always used the Zeigler-Nichols oscillation method,
and looked no further.

Now my boss asked for some other technique and I started
reading about tuning from the step response. (What Z-N call
the Process-reaction curve) Oh I stuck the Z-N paper here,
along with a few 'scope shots.

https://www.dropbox.com/sh/7lkqazuy6wvzknk/
AADcZtTvFvaD3TEzeJiAjEy6a?
dl=0

If only I'd read the paper ~20 years ago.

This is a more up to date rehash of the same things.
http://faculty.mercer.edu/jenkins_he/documents/
TuningforPIDControllers.pdf

For completeness I'll stick in Tim W's excellent article.
(Though unfortunately he doesn't talk about Z-N directly.)
http://www.wescottdesign.com/articles/pid/
pidWithoutAPhd.pdf
(Maybe he'll add an appendix for me. :^)

Nope. For two reasons:

First, because I've never used Z-N, never felt the need, and
so never built up practical experience with it.

Second, because all the industrial control guys I've talked
to, both in person and 'lectronically, tell me that Z-N gets
you a good starting point but then you need to either tweak
things, do measurement + system ID + design, or do frequency
sweeps + design. There's enough material out there on it,
and too many people that think that it's the bees knees.
Huh, Ok I only said it was my starting point. Then do some
step responses and see what it looks like. (I'm mostly an
idiot when it comes to thermal loops... I'm starting to think
that 1/2 the time I got lucky and put the sensor in a good
place, and the other times, not so good.) Well sometimes you
just need to put the sensor at the point you care about.... I
could think more about where to put the heater.
Grumble....

What's your starting point?

So if you look at the step response in my dropbox link,
there is a lag of 300-350 ms. And if you look at the
oscillations at the "ultimate gain"
you'll see a period of about 2 seconds. And lo and behold,
2*pi*tau (the lag) = period! That is very satisfying.

Estimating the needed gain (proportional term) from the
step response, didn't work out that well in this case. But
I'm running at gains that are ~20% the ultimate gain, so
oscillation method didn't work all that well either in this
regard.


Oh I do have one question for Tim.
You show a thermal control loop with just integrating
control. Does this really work? (I would have thought it
would just oscillate.)
Or do you have a little bit of proportional term in there?

If you keep a firm lid on your ambitions, certainly. For
most systems you won't get nearly the speed you'd get with
some proportional action or even proportional + derivative
-- but any system that has a flat gain characteristic going
down to zero frequency can be controlled with just an
integrator, if you're patient.

The one thermal loop that I've done had a simple first-order
response well beyond any reasonable loop bandwidth. So it
got a simple PI controller and worked just grand.
Huh, OK I'll have to try, the loss is in the delay/lag?
This thermal loop is fast, (at least for me)
so ten seconds is no problem.

I get starting points for thermal loops by looking at the
open-loop step response. In small systems it's usually well
approximated by a time delay cascaded with an integrator. Thus
the plant's transfer function is approximately

H_plant(f) ~ (1/j 2 pi f tau_I) exp(-j 2 pi f tau_delay)
Right, I went looking for a pulse response with "your"
delay*RC time constant, model in mind.
Seem's like Z-N did it before we were born... which I love.
They also give a guess at the gain, from the pulse response.


Then I do the usual frequency-compensation thing, aiming for
about a 60-degree phase margin. If the system is reasonably
well insulated,
the effect of the heat leak is usually small enough to ignore
for frequency compensation purposes, but if it isn't, you need
to move the pole from zero out to (j 2 pi f tau_leak).
Huh, Phil, lots of the time you're talking over my head,
(or at least not where my head is at the moment.)

I'm not sure where 60 degree's is, but I look at the set point
step response and lower the gain if there's not too much
ringing.
Say a Q of 1-2. (probably closer to one, and then in production
each unit is a bit different.)

You need to learn frequency-domain tuning. Just looking at the
time domain response is not a reliable way to get the best
response and a known loop robustness.

Grin, right. (There are bucket loads of things I need to learn.)
You do a frequency sweep with the loop closed?
Do you have a video showing that? :^)

It's on the short list -- which is about half a dozen titles long.

I describe it in my book. And in this article:

http://wescottdesign.com/articles/FreqMeas/freq_meas.html

The basic idea is that if you measure the phase and amplitude of a
sine wave going into a block, and the phase and amplitude of a sine
wave coming out of that block, then you can calculate the block's
response at that frequency by doing a complex division (amplitude out
/ amplitude in,
phase out - phase in). Then repeat for a whole bunch of frequencies.

Right. I've mostly just kept track of the amplitude response, though
you can always see the phase bounce around too.
(It must be Krammers and Kronig, or someone earlier, who observed that
you can see the phase response before the amplitude.)

Hmm I've got a Z-N oscillation question, moving to SED,
in hopes of more experience.


It's easy to build this into code if you're working with a
microprocessor; if you're not there are instruments out there
("dynamic system analyzers" or "transfer function analyzers") that do
it. If National Instruments isn't utterly crazy there's a pre-made
VI to do this in labview with one analog I/O brick.

My Rigol sig gen goes down to ~uHz. I can do a frequency sweep, It's
got two outputs which I can phase lock and set at 90 degrees. In
principle, I can measure everything myself.

I hate long time sweeps... you don't know till hours later if you f'ed
something up. It's much faster to understand the step response, or
oscillation method.

George H.

I suspect that there's one out there for less than $500. And you can
always do it by hand with a signal generator and an O-scope.

Once you know what you are doing (I'm not claiming that *I* know.)
Doesn't the step response (I'm thinking closed loop) contain the
same information as a frequency sweep?

In theory, yes. In practice the good stuff is often swamped out by
either noise or distortion. In the past I've done a pretty good job
of doing system identification from a customer-supplied step response
-- but I was limiting the performance of the loop based on what I
didn't know about the plant, rather than a sure idea of what the
plant wasn't capable of.

You could probably take a whole bunch of step responses and average
them together, and then fit a system model to the resulting average
-- that would at least knock down the effects of noise. But you'd
want to do real honest-to-god system ID, where you're fitting a model
in the Laplace domain to the step response, not just that Z-N stuff.

--
Tim Wescott Control systems, embedded software and circuit design I'm
looking for work! See my website if you're interested
http://www.wescottdesign.com

Hey what about doing the FFT of the pulse response?
Right, I know about FFT's (Fir spectrometers as a post doc.)
There's low frequency information in the long tail.. noisy part of the
response... In this case I'm more interested in the high frequency, I
think?

You're most interested in the frequencies around the loop closing
frequency. You can take the FFT of the response to a step or a square
wave (I'd get it w.r.t a square wave, with lots of reps).

* Take the Fourier transform of the excitation
* Take the Fourier transform of the response (get the delay right!)
* Zero out the response's FFT in those bins where the excitation's FFT is
zero
* Divide the response's FFT by the excitation's FFT for the non-zero bins

You'll have gain and phase, at least at those points.

I haven't done this so I can't guarantee it -- but what the heck, maybe
it'll work out just fine.


Or just get an HP 35665A, 35660A, or 3562A DSA from eBay. I got one for
$300 back in 2009, and got an all-options ROM (customized with the
unit's serial number) for $65 from glkinst.com. That gets you the swept
sine, AWG and noise waveform sources, plus HP Instrument Basic, extra
memory, curve synthesis, and some other nice things. It's worth getting
the pc-style keyboard for it (those are cheap).

FFTs are famous for generating reasonable-looking wrong answers when one
is careless.

Cheers

Phil Hobbs

--
Dr Philip C D Hobbs
Principal Consultant
ElectroOptical Innovations LLC
Optics, Electro-optics, Photonics, Analog Electronics

160 North State Road #203
Briarcliff Manor NY 10510

hobbs at electrooptical dot net
http://electrooptical.net
Click to expand...

Hmm well that's fine for me.. but I gotta tell someone else how to tune it too.
(which one of those will be cheaper... I looked for the 35660A (assuming
a lower part number would be earlier.) All were ~$1k and above.
(you bought sometime during/ near the recession.. probably a good time to
buy.)

We've got an SR770 spectrum analyzer. That I think goes down to ~0.5 mHz
(10^-3 Hz) or something like that. It's got a noise source, chirped source.
(sine's only at the center frequency of each bin) and maybe does a swept sine.

I'm still not real sure what I'd drive and what I'd measure?
With a closed loop, I'd put the source into a modulation input for the
temperature? And then look at the error? Or would you drive some
external load (heater) with the source?

George H.
 
You must log in or register to reply here.

Welcome to EDABoard.com

Sponsor

Back
Top