From jimr@maia.usno.navy.mil Thu Jul 23 15:26:34 EDT 1998
Received: (from jimr@localhost) by maia.usno.navy.mil (8.7.5/8.7.3) id PAA20925 for gpst@maia; Thu, 23 Jul 1998 15:24:38 -0400 (EDT)
From: Jim Ray (USNO 202-762-1444) 
Message-Id: <199807231924.PAA20925@maia.usno.navy.mil>
Subject: Clock Resets at USNO
To: gpst@maia.usno.navy.mil
Date: Thu, 23 Jul 1998 15:24:38 EDT
X-Mailer: Elm [revision: 212.4]
Status: RO

   Clock Resets at USNO -- Comparison of 1 pps Data with Geodetic Analysis

                                Jim Ray
               U.S. Naval Observatory, Washington, DC, USA

Summary
=======
AOA TurboRogue receivers have a well known tendency to reset their internal
clocks occasionally, even when driven from a very stable external frequency
standard.  While this feature is necessary for the receiver to maintain
tracking when using a poor quality frequency standard, it is quite annoying
for high-precision timing applications.

The JPL group (Larry Young) has advised users that the TurboRogue 1 pps
output can be monitored to determine quantitatively the magnitude of such
clock jumps.  For the USNO IGS receiver installation, such a measurement
system was set up by the Time Service Dept. (Wendy King and Paul Wheeler) to
compare the 1 pps output from the TurboRogue against the 5 MHz signal input
(which comes from a steered H-maser aligned to UTC(USNO)).  Recently, we have
observed two resets of the USNO receiver clock (one spontaneous and the other
induced by a power outage) which provide data to compare the 1 pps monitor
performance with independent geodetic analyses.

An examination of these data show that: 1) the 1 pps measurement system
provides very stable monitor data; 2) the 1 pps data can be used to
calibrate spontaneous GPS receiver clock resets probably to the few ps
level; 3) the calibration value for one such event has been confirmed using
GPS geodetic analyis to within 0.1 to 0.2 ns.  The 1 pps measurement system
provides the means to monitor and calibrate GPS receiver resets, well
within the uncertainty of the geodetic results.


Spontaneous Reset on 1998 July 06 at ~12:22:35 UTC (MJD 51000.515683)
=====================================================================

Results from 1 pps measurements
-------------------------------
The 1 pps monitor data (at about 15-minute intervals) from the period
before the clock glitch are summarized below for the entire measurement
period and for the last day beforehand:

                  all data before        last day before
                       glitch                 glitch
             ----------------------   ----------------------
                  409.166 ns mean         409.166 ns mean
                    0.090 ns RMS            0.082 ns RMS
                     2626 #                   100 #
             50969.000243 MJD start   50999.479213 MJD start
              (1998 06 05 00:00:21)    (1998 07 05 11:30:04)
             51000.510475 MJD stop    51000.510475 MJD stop
              (1998 07 06 12:15:05)    (1998 07 06 12:15:05)

This shows that the measurement system is capable of monitoring the
1 pps signal with a stability of <100 ps per measurement and that there
is no evidence for a drift.  It also suggests that if the interval
between resets is sufficiently long, then the clock jumps can probably
be determined to the few-ps level (assuming Gaussian statistics).

Using the mid-point between measurements where the jump occurred, we can
estimate the epoch of the event as 1998 07 06 at ~12:22:35 UTC (MJD
51000.515683).

Similar 1 pps monitor data for the period after the clock glitch are
summarized below:

                  all data after         first day after
                       glitch                 glitch
             ----------------------   ----------------------
                  946.848 ns mean         946.864 ns mean
                    0.100 ns RMS            0.097 ns RMS
                      683 #                   100 #
             51000.520880 MJD start   51000.520880 MJD start
              (1998 07 06 12:30:04)    (1998 07 06 12:30:04)
             51007.625012 MJD start   51001.552141 MJD start
              (1998 07 13 15:00:01)    (1998 07 07 13:15:05)

The more recent 1 pps data are slightly less stable than the previous
period but insignificantly.

Thus, if we use the two longest periods and difference the mean values
then the following estimate for the clock jump is obtained:

       1 pps estimate for clock jump   +537.682 ns
                      standard error      0.004 ns
                               epoch   1998 07 06 at 12:22:35 UTC
                                                    +/- 07:30
                                       (MJD 51000.515683 +/- 0.005208)

Results from geodetic analysis
------------------------------
>From the USNO geodetic analysis performed for the IGS Rapid products, we
can examine the estimated clock time series for the same event.  The
geodetic analysis uses the USNO receiver as the clock solution reference
(i.e., all other receiver and satellite clocks are determined holding
USNO fixed at zero) and estimates are made at 7.5-minute intervals.  Thus
a clock jump at USNO shows up as a discontinuity for all the other clocks.
Since the actual epoch of the clock jump is not resolvable better than
some time between 1998 07 06 12:11:03 and 1998 07 06 12:26:03 GPS Time (the
intermediate data epoch was apparently affected by the reset and did not
produce usable results), stations equipped with very stable frequency
standards are preferred to estimate a value for the clock discontinuity.

Three such geodetic estimates are listed below based on differences of
clock estimates immediately before and after the reset:

                           FORT   +537.688 ns  (Sigma Tau H-maser)
                   formal error      0.204 ns

                           GRAZ   +537.383 ns  (Cesium)
                   formal error      0.197 ns

                           TSKB   +537.242 ns  (Cesium)
                   formal error      0.202 ns

The formal error is merely the RSS of the errors from the GIPSY Kalman
filter.  Because such errors are time-correlated, they can over-estimate
the uncertainty for a differential quantity such as this.  On the other
hand, the formal error does not include any contribution due to intrinsic
variations in each clock during the 15-minute gap (including all clock-like
effects, not only the underlying frequency standard).  Thus, the FORT
result, using an auto-tuned H-maser, is probably the most reliable.  This
conclusion is supported by the following statistical observations for the
3 hours before and after the clock jump:

                  FORT      0.073 ns RMS for 3 hr before jump
                            0.072 ns RMS for 3 hr after jump

                  GRAZ      0.208 ns RMS for 3 hr before jump
                            0.381 ns RMS for 3 hr after jump

                  TSKB      0.917 ns RMS for 3 hr before jump
                            1.074 ns RMS for 3 hr after jump

To attempt to verify the FORT result, we can examine results from other
H-masers with large drifts provided that rate corrections are applied.
This has been done for the H-maser stations below using first-differences
immediately before and after the glitch interpolated to the mid-point
(1998 07 06 12:18:33 GPS Time):

                           ALGO   +537.890 ns
                estimated error      0.304 ns

                           DRAO   +536.846 ns
                estimated error      0.503 ns

                           FAIR   +537.907 ns
                estimated error      0.339 ns

                           GOLD   +538.097 ns
                estimated error      0.223 ns

                           KOKB   +537.981 ns
                estimated error      0.258 ns

                           NLIB   +537.888 ns
                estimated error      0.215 ns

                           TIDB   +537.844 ns
                estimated error      0.216 ns

                           WTZR   +538.244 ns
                estimated error      0.348 ns

                           YELL   +537.573 ns
                estimated error      0.261 ns

The error estimates here include the formal errors from the Kalman filter
as well as the full effect of the clock drift over the 15-minute window.
The variation in error estimates is due to different drifts.

If we compute the weighted mean of all 12 estimates above for the clock
jump, the result below is obtained.  In weighting the cesium clock results,
the observed variance of the clocks before/after the glitch is also included,
although the weighted mean is not sensitive to whether the cesium results are
included or not.  Indeed, the overall mean itself is not sensitive to whether
weighting is used or not:

        weighted mean for clock jump   +537.827 ns
                        weighted RMS      0.270 ns

      unweighted mean for clock jump   +537.715 ns
                      unweighted RMS      0.395 ns

Based on all the geodetic results, it appears that the estimate derived from
the FORT data alone is the best overall.  Hence we conclude:

    geodetic estimate for clock jump   +537.688 ns
                     estimated error      0.204 ns
                               epoch   1998 07 06 at 12:18:33 GPS Time
                                                    +/- 07:30
                                       (MJD 51000.512882 +/- 0.005208)

Comparison of results
---------------------
The geodetic estimate agrees very well with the 1 pps measurement value,
better than expected from the error estimates which is probably fortuitous.
Using the very stable FORT data, the geodetic error estimate could well be
pessimistic although it is unlikely to be better than ~0.1 ns.  Much more
sophisticated analyses of the geodetic results could be performed but it
is unlikely that the conclusions will be greatly affected.  Clearly, the
1 pps measurement system provides the means to monitor and calibrate GPS
receiver resets, well within the uncertainty of the geodetic results.


Power Reset on 1998 July 13 shortly after 15:00:01 UTC (MJD 51007.625012)
=========================================================================
Some time within 15 minutes after 1998 07 13 15:00:01, the UPS power for
the GPS receiver, 1 pps measurement system, and other equipment was
interrupted.  Because manual intervention was required to restart the
various systems, about a day elapsed before everything was fully restored.
For this reason it is not feasible to perform a quantitative comparison
of clock offset determinations as with the 1998 July 06 event.

However, from the 1 pps measurement data available, before and after the
power interruption, a precise estimate for the clock offset can be 
determined.  For the period since the power outage, we have:

                  all data after   
                   power break     
             ----------------------
                  224.252 ns mean  
                    0.099 ns RMS   
                      589 #        
             51008.886794 MJD start
              (1998 07 14 21:16:59)
             51014.989618 MJD start
              (1998 07 20 23:45:03)

which, together with the previous offset for the period from 1998 July 06
to 13 gives the following estimate for the clock offset:

       1 pps estimate for clock jump   -722.596 ns
                      standard error      0.006 ns
                               epoch   1998 07 13 after 15:00:01 UTC
                                       (MJD 51007.625012)