[time-nuts] D term (was no subject)

[WarrenS]

I second  Poul-Henning Kamp's comments concerning D-terms,
(mostly) as done in the TBolt and likely other GPSDOs.

A 'D-term' helps fast loops like a TPLL where you
want a high bandwidth with the P gain as high as possible.
For slow noisy loops like a cleanup osc or a GPSDO,
what helps is a pre-filter.
A D-term and a pre-filter do the opposite of each other
and are therefore generally not used together because they
tend to cancel each other.

The Tbolt control loop has neither!
What it does is:
The P-gain in the Tbolt is set by the time-constant (& EFC-Gain)
which sets the freq error recovery time constant.
The Integrator gain is set with the Damping, which controls the
Phase error recovery time-constant.
If the damping is set high >> 10, the TBolt's loop becomes a P only
controller, AKA a Freq-Lock-Loop (FLL)  instead of a PLL.
That is with a very high damping setting, The Tbolt then corrects
for any freq error offsets but does not correct for any phase errors.

The Tbolt is indeed very flexible, allowing usable time-constant
settings from 1 sec to days in either a PLL or FLL mode.
What it is missing is a pre-filter, and to get the best performance that 
must be added externally.

ws
*********************************

>I have seen no evidence that the Thunderbolt, in particular,
> uses a D term.  ...
>
>Charles
************************************************

>>Before anybody gets any ideas that causes them to waste a lot of time:
>>D terms are themselves very temperamental because they, by definition,
>>amplify measurement jitter noise.

>>In the precision time/frequency domain, D-terms are almost never
>>realistic.
>>
>>Poul-Henning Kamp
*********************************

>>>Without a D term, PI loops can be unstable when the gain (P) is
>>>increased. If you will, with a large error, the correction will itself
>>> be large and as the system corrects itself,
>>>it may overshoot the target value, going into a low (or high if you
>>>really blew it) level oscillation around the target value.
>>>The D term slows it down just enough and minimizes that overshoot
>>>while maintaining a high gain (low steady state error) and a fast 
>>>response.
>>>
>>>Didier KO4BB
************************************

On January 24, 2015 8:05:38 PM CST, Bob Camp <kb8tq at n1k.org> wrote:
>Hi
>
>A classic control loop in it's simplest form has only one term. That is
>often referred to as a proportional term. When the control signal (or
>error) changes by A the output changes by A times that term. Often in
>shorthand notation this term is refereed to as a P term.
>
>The next thing that some people add to a control loop is an integrator.
>It looks at the control signal (or error) has a constant offset of A,
>the integrator sums up the A's. The output of an integrator would
>eventually go to infinity with a constant control input (or error) into
>it. This term is often referred to as an I term.
>
>Lastly people add a term to the control loop that responds to the rate
>of change in the control signal (or error). The faster the change, the
>bigger this signal gets. This is commonly refereed to as a Derivative
>term. In shorthand it's talked about as the D term.
>
>The net result is a three element control loop running what's called a
>PID algorithm .
>
>The P and I can also be described by a time constant and a damping.
>That's what the Trimble software lets you do. The implication is that
>it's just a PI loop. In fact it appears to be a PID loop and you can't
>get at the D term.
>
>For a much more clearly worded explanation of all this, there's
>
>http://en.wikipedia.org/wiki/PID_controller
> >...
>>There does appear to be a D in the TBolt loop. For what ever reason,
>>that's not a changeable value. The D does scale with the time constant. .

>>Bob
>