[evlatests] Sub-band saturation and RFI

[Rick Perley]

    A wide-band L-band observation was made Thursday night which has 
demonstrated an effect not seen before -- but which will undoubtedly 
occur again. 

    The observation was of the strong source 3C84, using 8 subband 
pairs, each of 128 MHz, with full polarization.  Each cross-product had 
64 channels, so there is 2 MHz spectral resolution.  The observation was 
13 minutes in duration, with a scan (but not a mode) change every 40 
seconds. 

    In general, the data are of very good quality.  AIPS reported 7% 
integer zeros -- I have yet to find just where these 7% are -- they are 
not obvious using simple listings.  What was immediately obvious in 
simple column listings was the the RR correlations were showing notable 
changes in amplitude which were not matched by any changes in the LL 
correlations. 
   
    A simple CALIB solution was made to see how the antenna-based 
amplitudes were behaving -- this showed immediately that, beginning at 
19:46 IAT (about 9 minutes into the observation), the RCP gains changed 
every 40 seconds.  The changes are all of the '1 dB' variety -- so it's 
clear for each 40 second observation, the system was adjusting the T304 
attenuators in response to changes in external power.   Evidently, the 
'set and remember' mode of operation was not enabled for this test.  The 
overall range in gain change was quite large -- about a factor of 4 or 5 
in power.  Meanwhile, the LCP gains were dead steady, other than one or 
two antennas that changed only once, and at a time clearly unconnected 
to what was happening on the RCP side. 

    There are eight subbands ('spectral windows') in this dataset -- 
using the wonderful SPFLG program to view the data, it is clear that the 
problem driving the gains wild was isolated to subbands 2 and 5 -- these 
are the two which contain the GPS signals -- which just happen to be in 
RCP.  For short periods of time (typically a minute or two at most), all 
data within one or both of these two subbands were 'wiped out' -- 
meaning that all amplitudes (all 64 channels) jumped up by a factor of 
100 or more.  This effect is seen only in RCP, and is 
baseline-dependent.  Some baselines -- always the longest ones -- did 
not show the effect at all.  Most (including all the shortest ones) had 
significant saturation. 

    Use of POSSM to plot the spectra before, during, and after this 
phenomenon allows a better description. 

    A) A few minutes before the effect began, spectra look normal in all 
subbands.  'Normal' for subbands 2 and 5 (which include the GPS signals) 
means spectral flux density variations (in frequency) within each 
subband of up to a factor of 10 to 100 above the astronomical signal.  
RCP and LCP look similar, but are not identical.  This is as expected 
for RCP (or whatever polarization) signals entering through far 
sidelobes, as these are essentially randomly polarized.  The cross-hand 
amplitudes look like the parallel hand, but a factor of 10 to 20 lower 
for the 'sky' signal, and about equal to the parallel-hand amplitudes 
for the satellites.  All of this is normal. 
    B) Just prior to the saturation effect, the RCP amplitudes (only) 
rise dramatically, reaching a factor of 1000 (peak) above the 
astronomical signal.  It seems that at this point (when the interfering 
signal is about 30 to 35 dB above the noise), a saturation occurs. 
    C) When the satellite signal exceeds this threshold, all spectral 
channels *within the two affected subbands only* saturate, giving a 
completely false spectrum whose mean level is 100 to 1000 times too 
high.  This saturation effect can last as short as a few seconds, or as 
long as one minute. 
    D) All other subbands (i.e., those without the RFI) are not 
affected, other than the loss of gain caused by the T304 modules trying 
to adjust to the increased power.  Also, all other cross-power spectra 
(LL, RL and LR) are unaffected by the saturation seen in the affected 
(RR)  polarization product. 

    The basic cause of this first-seen phenomenon seems fairly clear:

    A strong source of RCP radiation (almost certainly a GPS satellite, 
but this is unproven) passed close to the boresight of the antennas.  
The GPS satellites have low gain -- they are designed to transmit to the 
entire planet, so all EVLA antennas should see about the same input 
power.  There are typically 10 to so of these satellite moving (fairly 
slowly -- I think it takes a couple hours for any one to pass over the 
visible hemisphere) along -- and we nearly always see these through the 
randomly polarized antenna sidelobes. 
    But in this case, one of them must have been close to our pointing 
direction, so that it entered our properly polarized RCP beam -- either 
the horn directly, or (more likely in my view, but less likely in 
Vivek's) the antenna beam itself.   This accounts for the rapid rise in 
RCP-only power. 
    At some point, when the maximum spectral power exceeds about 35 dB 
(a factor of a few thousand), something overflowed.  Since the effect is 
seen only in the subbands in which the incoming signal exists, we know 
it is not the antenna electronics.  And since the effect is limited to 
the particular polarization in which the emission occurs, we know it is 
not the requantization which follows the sub-band definition on the 
station boards.  Further, since the effect is absolutely baseline based, 
the cause must be at or after the baseline board. 
    Ken opines that the accumulator, which follows the 
cross-multiplication has overflowed.  This is a problem which will not 
be solved by going to 7-bit internal resampling.  It should be solved 
(up to a point) by utilizing a currently unused digital gain adjustment 
which is associated with the requantizer.  However, this can only take 
us so far, as in turning down the gain in response to greatly increased 
subband power (1 to 3 orders of magnitude) will eventually clip the 
noise, and lose us the desirable astronomical signal. 
    I suppose the ultimate solution is to employ 7-bit requantizer 
sampling, along with the digital gain adjustment to prevent overflows.  
(Faster dumps will help too I guess, but that will increase data volume 
...)  But the 7-bit requantization comes at a significant loss in 
correlator capability, so we would only want to use this if we think it 
needed -- when an sufficiently offensive satellite is expected to pass 
by.  I don't recall if this requantization adjustment can be done by 
subband -- being able to do so would help. 

    I emphasize that I believe this saturation/overflow effect is not 
common (no WIDAR0 data to my knowledge saw this), and only affects the 
subband/polarization/baseline in which the signals lie.  And -- it is of 
limited duration.  Post-quantizer gain adjustment should handle most 
cases.  That, and 7-bit requantization might be expected to handle all 
but the most pathological cases. 

    I have some illustrative plots to show at the next appropriate 
meeting.