Ulrich, +-3 sigma is 99% on a gaussian distribution.
If you want to look at other algorithms on this topic, here is one: http://www.control.isy.liu.se/books/adfilt/ br, Björn On Thu, December 14, 2006 12:55, Ulrich Bangert said: > Hello Said, > > from my own experience in detecting and removal of outliers i can asure > you that it is a challenging and ambitious task in statistical math. I > fear you expect too much from a receiver's TRAIM firmware to compute the > necessary statistic math on a pulse to pulse base. Nevertheless you can > do it on your own if the microcontroller that compares the LO to the > GPS's 1pps has enough RAM and processing power. > > It involves computing the MEDIAN over a number measurements (say 60) to > get a reliable estimate of the measurement's MEAN value in that time. > Note that the normal MEAN computation is susceptible to outliers while > the MEDIAN is not. On the other hand computing the MEDIAN involves > sorting the measurements in ascending order, much more stuff than simple > MEAN adding/dividing. Once you have the MEDIAN of your measurements you > can compute the distances between the single measurements and the > median. But in contrast to the simple STANDARD DEVIATION which is > susceptible to outliers itself, you now compute the MEDIAN over the > distances between the measurements and their MEDIAN to give you > something that is called the MAD (MEDIAN ABSOLUTE DEVIATION. As the > STANDARD DEVIATION the MAD is a measure of the witdh of the > measurement's distribution in the sense of a 'sigma' but not susceptible > to outliers. Once you have that number you can easily decide whether the > last measurement is within or without a range of say +/- 5 * sigma. I > guess +/- 5 * sigma resembles 99% of normal distributed values, so you > are on the safe side to throw away anything outside +/- 5 * sigma. Note > that you have to perform the above steps for every new measurement, i.e. > every second removing the oldest in your data buffer and adding the > latest measurement to you data buffer. It can be done and i do so in my > DIY GPSDO but actually it needs exactly this complexity. > > Best regards > Ulrich Bangert, DF6JB > >> -----Ursprüngliche Nachricht----- >> Von: [EMAIL PROTECTED] >> [mailto:[EMAIL PROTECTED] Im Auftrag von [EMAIL PROTECTED] >> Gesendet: Donnerstag, 14. Dezember 2006 01:56 >> An: time-nuts@febo.com >> Betreff: Re: [time-nuts] Best GPS 1PPS Accuracy >> >> >> Hi guys, >> >> slightly off-topic, but has anyone done stability tests with >> the m12+ or >> M12M when the antenna goes bad in TRAIM-enabled mode (e.g. >> less than 4 good sats >> being received by i.e. disconnecting the antenna)? >> >> I see that the 1PPS output goes away only after about 5 or so >> additional >> pulses have been generated, so these 5 pulses are "free-running". >> >> These pulses are somewhat inaccurate, and I am wondering if >> anyone has done >> a max/min analysis on the errand pulse's offsets? >> >> I wonder if the TRAIM limits can be set to say +-50ns >> effectively to kill >> these pulses and improve overall accuracy? >> >> I know lot's of folks run the M12 TRAIM limits at 1us, that >> seems excessive >> on a receiver that is supposed to do 10ns 1-Sigma, doesen't it? ... >> >> Thanks for your feedback, >> Said >> _______________________________________________ >> time-nuts mailing list >> time-nuts@febo.com >> https://www.febo.com/cgi-> bin/mailman/listinfo/time-nuts >> > > > _______________________________________________ > time-nuts mailing list > time-nuts@febo.com > https://www.febo.com/cgi-bin/mailman/listinfo/time-nuts > _______________________________________________ time-nuts mailing list time-nuts@febo.com https://www.febo.com/cgi-bin/mailman/listinfo/time-nuts