[time-nuts] LPRO-101 with Brooks Shera's GPS locking circuit

[Magnus Danielson]

From: "Poul-Henning Kamp" <phk at phk.freebsd.dk>
Subject: Re: [time-nuts] LPRO-101 with Brooks Shera's GPS locking circuit
Date: Sat, 16 Dec 2006 15:56:32 +0000
Message-ID: <29148.1166284592 at critter.freebsd.dk>

> In message <000001c72126$ecc89270$0202fea9 at athlon>, "Ulrich Bangert" writes:
> >Poul,
> >
> >i appreciate your comments always a lot! But dynamical methods are
> >especially usefull when the input parameters are subject of change,
> >aren't they?
> 
> They are also very useful for amateur projects where the users do
> not have the necessary measurement facilities and likely use random
> components bought on ebay :-)

Indeed. If you buy a particular GPS diciplination kit, only the TIC performance
is assumed to be known. If you include serial communication (to allow for
sawtooth corrections) to the GPS, for some receivers we could have included
apriori knowledge (since we now may ask the GPS what type it is), but not for
all. Many GPS receiver do not emit sawtooth corrections unfortunatly.
Also, the big range of oscillators being trained is not under control. Also,
the actual condition of one such oscillator may deviate from another measured,
so it is really only in the experienced time-nut (type of person rather than
list member) hands that the actual values could be used as a priori knowledge.

Also, if an oscillator has been knocked around etc, it will have worse ADEV
than when it has healed itself. So not being overly dependent on a priori
knowledge might actually be a key for optimum performance not on average but
rather in the momemnt. This is kind of the point of using GPS steering IMHO.

> >But how would you solve the system of equations with TWO
> >unknowns (LO and GPS jitter) if you have only ONE information?
> 
> The way you do this is by measuring the ADEV between your two sources
> and how it changes with changes in your timeconstant.

I.e. out of your TIC. The trouble is that you do not get one result but
several. Either you just drive the time-constant to minimize ADEC(100) or
something or you are a little more creative and make a (possibly weighted)
sum of a few ADEVs (say 10, 30, 100, 300 and 1000) and then let that be the
parameter to be minimized. Maybe an integral over ADEV to integrate the
energy would be appropriate, maybe not.

However, there have already been work done that shows painstakingly well that
this is a task for Kalman filters and these days we should also have a look at
particle filters. This is computingwise more complex than the above solution,
but may be of interest if one wants go further.

I think that something like the above may be helpfull, althought not being done
as scientifically as Ulrich explanation.

> In my experience, the better way is to start with a short timeconstant
> and increase it, until the ADEV shows signs of detoriation.

That should be simple enought to put into a PIC or AVR for a dirt-cheap
solution. To keep tracking the time constant to actual performance it can then
wiggle it up and down and decide if a step up or down is to be done.

Cheers,
Magnus