Best Book on PID ??

Aha!! The perfect pointer! Thanks!

And what do you know, old "Nichols Chart" himself... last used one about 40 years ago ;-)

...Jim Thompson

Honeywell used to (and may still) have in some of their controllers a PID algorithm that handled setpoint changes via integral action only. I think that was called PID B, as distinguished from PID A which was the standard ISA PID with independent gains.

Not exactly what you describe, but a step >>

Damn, topped again!

I did do some interesting things on a supertanker that earned me the nickname "Lights-Out Larkin."

John

I am frustrated with industrial PID controllers that force me to choose between having the P & D terms based entirely on error or entirely on the process measurement only. What I often need is a separate gain and derivative term for the process measurement and setpoint inputs to the controller. This is almost always superior the gain and/or derivative based entirely on error (setpoint - process measurement) or base entirely just on the process measurement and lets me optimize (after I define that word for the particular loop) the process disturbance response and setpoint change response, individually. For critical tuning, I have to use the feed forward connection and external math to to get all the factors I need. Why must I choose chocolate or vanilla when I really want a swirl.

I never could get derivative to work in this particular application. The thermocouple is in a dewar maybe a foot downstream of the heater, and the gas flow isn't fully mixed when it hits the t/c. The heater emits four little jets of hot air, sort of like Stripe toothpaste. So the temperature is literally noisy, and derivative goes sort of crazy if set to any useful-to-the-dynamics level. That's a common problem in flow systems. I designed all sorts of cute flow stirring devices (a simple spring isn't bad) but couldn't get anybody interested in poking them into the dewar.

Customers are the other big control problem.

John

Yes, especially if you're doing setpoint profile vs. time.

Best regards, Spehro Pefhany

Hmmm, guys, very interesting point. In most of the loops I've done, the setpoint change just shoots through the error amp and the pid stuff and sort of comes out OK, but I can see how that ought to be optimizable.

John

Part of the problem is that numbskull implementations don't properly LPF the signal before taking the derivative so that ordinary process and quantization noise makes the D term useless at the proper setting.

You didn't see that with analog controllers, because analog engineers understand stuff like that. And it doesn't generally appear in the textbooks.

Best regards, Spehro Pefhany

Their current incarnation of DCS system that I work with has forms A,B,C and D.

A bases P&D on error (most aggressive response to setpoint changes).

B bases P on error and D on process. Slightly less aggressive setpoint response, but practically the same since D is used so seldom. This is my preferred choice for inner loops in cascade systems (where one loop provides a live setpoint for another loop, so the setpoint is smoothly moving all the time).

C bases both P&D on process only. Uses only integral action to approach new setpoint. This is my default for loops set strictly by operators. The smooth ramp response to a setpoint step change allows an operator to type in a wrong value, see the output heading off and do an Oh shit! and make a correction before things get seriously wrong.

D is an integral only algorithm.

The system I want replaces A, B and C with s single choice that includes all those extremes and everything in between with the choice of two proportional and two derivative values.

Z-N tuning is an experimental controller tuning method. It tends to be a bit aggressive for real systems (so, increase the P. band). I've got a copy of the original 1942 paper around somewhere if you can't find it online (shouldn't be that hard).

Best regards, Spehro Pefhany

Walter Driedger has posted a PDF copy of the world famous Z-N paper on his web site:

Dead time and noise definitely are limiting factors when applying derivative. Usually the derivative effect has to include a low pass roll off pole so that the phase lead of the derivative applies to a narrow band of frequencies between the closed loop dominant pole and the 1/(dead time) and noise frequencies.

Often you can do better by adding a second control loop that measures the temperature very close to the heater that is not affected by that dead time and noise, and control its setpoint (between judicious limits) with the controller that watches the actual process that is downstream.

Has anyone here made use of fractional calculus (differentiation & integration) in control loops?

But perhaps the math is too new for the applications to have been tested...? There was also a time when roots and powers were always integral, as differentiation/integration is even now to most folk :-).

I'm interested to know what kinds of systems benefit from it, and how you implemented it - pure curiosity of course.

Clifford Heath.

Click the "Download" button... and download the 757KB PDF version... scanned, then OCRd into Word, touched up and then PDFd ;-)

...Jim Thompson

Okay, that's a nice HTML version. I've got a 5M PDF scan version of the ASME paper here:

Best regards, Spehro Pefhany

Ah, that's better!

Best regards, Spehro Pefhany

You could probably write a good one. A PID controller is just a follower amplifier (that forces a process measurement to follow a setpoint). The PID controller tuning is just a lead lag network that stabilizes that unity gain amplifier.

view.

If you want to learn the twiddle method of tuning, I have written a tutorial for non engineers that has been posted at:

If this is for that Devil's Transducer that you were asking about earlier I don't think Z-N is going to get you there. You'll be much better off getting a good Bode plot (like, with about 50 times more points) of the thing and designing your controller from there. Watch for that resonance at 10kHz -- it'll be a bitch.

As mentioned elsewhere if you treat it like a mystery circuit element you'll do fine -- you could probably even model the thing as some bizarre circuit in SPICE and do all the analysis in that context.