Re: extending gr-trellis to perform Viterbi, MLSD on GMSK


#1

Nick,

there is no change required in the modules that perform viterbi decoding
in order to implement either MLSD in ISI or (coherent) GMSK demod.
The whole idea around gr-trellis was to disentangle the trellis aspect
of a modulation scheme from the details of the modulation/channel.

So, if you want to implement general (coherent) CPM demodulation, all
you have to do is to represent the CPM signal as a FSM followed by a
memoryless modulator. Look up the paper by
Bixio Rimoldi: A decomposition approach to CPM, or take a look at my
notes on this at
http://www.eecs.umich.edu/~anastas/docs/cpm.pdf

You will only need to write an additional
constructor for the fsm class that takes the CPM parameters and produces
the appropriate FSM. Similarly, you’ll need to write a piece of code
that takes the incoming waveform and does the metrics calculations
(eithar as a separate block as in trellis-metrics or inside the viterbi
block as in trellis-viterbi-combined).

If you want to work on this general problem for a generic CPM modulation
I can help you.

Adding ISI to this is a piece of cake: you need to combine the two
trellises into either a combined trellis or to use hard/soft decisions
from one to feed the other detector.

Achilleas


Date: Tue, 3 Feb 2009 02:43:01 +0000
From: Nick F. removed_email_address@domain.invalid
Subject: [Discuss-gnuradio] extending gr-trellis to perform Viterbi
MLSD on GMSK
To: gnuradio removed_email_address@domain.invalid
Message-ID: removed_email_address@domain.invalid
Content-Type: text/plain; charset=“windows-1252”

Hi all,

I’ve spent a few days familiarizing myself with gr-trellis as best as I
can, and I’m interested in extending the gr-trellis module to handle
Viterbi equalization of ISI channels for GMSK demodulation. I saw Toby
Oliver’s thread in Sept. '06
(http://www.mail-archive.com/removed_email_address@domain.invalid/msg05615.html)
discussing a possible modification with Achilleas A. but
never saw anything checked in as a result. I’m just looking to use the
trellis code to demodulate low-BT GMSK in a more optimal way than the
current Gnuradio implementation, and I have similar questions to Toby’s:

  • How should I go about modifying make_isi_lookup() to add support for
    two-dimensional modulations? What format is trellis expecting?

  • I see the test_viterbi_equalization1.py file, which appears to do MLSD
    on an ISI channel for 4-PAM (and other one-dimensional modulations). Am
    I correct that if make_isi_lookup() is modified to support quadrature
    modulations, simply changing the modulation type in this example would
    be enough to make it work? I guess I’m asking more specifically if
    trellis.viterbi_combined_X will support an ISI lookup table for PSK
    modulations without modification.

  • Is there a good reason I should avoid tackling this problem? I’d hate
    to be duplicating someone else’s work in this area, or barking up the
    wrong tree.

For further information, I’ve written a packet-based AIS decoder for
Gnuradio, and I’m disappointed at the quality of the data coming out of
the GMSK demodulator. It’s 9600 symbols per second @ BT=0.3, so there’s
enough ISI that I think MLSD would provide significant reduction in
observed BER. Besides, it seems like it would be a useful capability to
have added to Gnuradio. Any other tips anyone has that might help me in
doing this would certainly be welcome!

Nick


#2

Hi,

I am trying to decode CPM signal according to your email but I have
problem
with understanding some things. What I’d like to do is to decode some
kind
of I/Q samples according to following scenario:

  • decimation of symbol rate N-times
  • decoding I/Q symbols according to FSM file and provided constellation

For VA I am using:
dimensionality = 1
va =
trellis.viterbi_combined_cb(f,K,0,-1,dimensionality,constellation,trellis.TRELLIS_EUCLIDEAN)

However the results are not as expected to be. My guess is that
something
might be wrong with FSM file - I’ve looked through *.fsm files example
and
seems that table OS (output symbols) is generated somehow. What I know
about
my state machine is:

  • there is S unique states
  • each state can have transition to one of I states
  • each input symbol is equal to IB bits
  • transition from state S(i-1) to S(i) corresponds to O output symbols
  • each output symbol is written on OB bits

So in general - I have defined NS table. I thought that OS table in
*.fsm
file is just table that tells: transition from state S(a) to state S(b)
is
equal to output symbol O(ab) but it seems that I’m wrong. Is there any
easy
way to generate OS table in *.fsm file? I have seen fsm constructor that
generates these values according to (n,k) and G - is that only way to
generate OS table?

Regards,

On Tue, Feb 3, 2009 at 11:23 PM, Achilleas A.
<removed_email_address@domain.invalid