I completed some trials with rate 1/8 K=25 and it works very
well indeed. I use 15 second symbols and a near-field source
set to the give same flux density at the rx as the signal from
DF6NM. The results are considerably better than 1/4 K=21 and
I look forward to trying the stronger code over a long distance.
The signal also decodes with 10 second symbols but is marginal,
ranking a few hundred down the Viterbi list and tolerant of
only about 20 degrees of phase error. Nevertheless the code
was able to correct 213 errors in 560 symbols.
I tried to find a better set of polynomials for K=25 but
without success. It becomes exponentially more difficult as
K increases. To illustrate, I set 128 CPU cores to search for
polynomials and let them run for 36 hours. They turned up a
couple of thousand possible codes and a further 16 cores tested
these. None produced a better distance spectrum than my existing
top performer although one came close.
Now I am motivated to test rate 1/16 and have started a search
for K=21 and K=25 codes. K=21 is a practical decode limit for
domestic computers and K=25 is the limit with AWS EC2
32 core VMs. Decode time is around 10 seconds but it takes
5 mins or so to search for the phase if the signal phase is
steady, and maybe 40 mins to an hour if there is significant
Doppler. I guess that's not unreasonable for a message that
takes a few hours to send.
Ready in a day or two for more far-field tests.
--
Paul Nicholson
--
|
