w4kek
/
WSJT-X
mirror of https://github.com/saitohirga/WSJT-X.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Use 3rd order polynomial fit to estimate the noise baseline. The 

polynomial fit is done over 400 Hz bandwidth for T/R periods longer
than 15s, and over approx. 600 Hz (10 times the signal bandwidth) for
T/R period of 15s.
This commit is contained in:
Steven Franke 2020-08-28 09:22:22 -05:00 committed by Bill Somerville
parent d82b9f5b0e
commit 5ca81a6507
No known key found for this signature in database
GPG Key ID: D864B06D1E81618F
3 changed files with 78 additions and 19 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
CMakeLists.txt
View File

@ -618,6 +618,7 @@ set (wsjt_FSRCS
lib/fst4/osd240_101.f90
lib/fst4/osd240_74.f90
lib/fst4/get_crc24.f90
lib/fst4/fst4_baseline.f90
)
# temporary workaround for a gfortran v7.3 ICE on Fedora 27 64-bit

48
lib/fst4/fst4_baseline.f90 Normal file
View File

@ -0,0 +1,48 @@
subroutine fst4_baseline(s,np,ia,ib,npct,sbase)
! Fit baseline to spectrum (for FST4)
! Input: s(npts) Linear scale in power
! Output: sbase(npts) Baseline
implicit real*8 (a-h,o-z)
real*4 s(np),sw(np)
real*4 sbase(np)
real*4 base
real*8 x(1000),y(1000),a(5)
data nseg/8/
do i=ia,ib
sw(i)=10.0*log10(s(i)) !Convert to dB scale
enddo
nterms=3
nlen=(ib-ia+1)/nseg !Length of test segment
i0=(ib-ia+1)/2 !Midpoint
k=0
do n=1,nseg !Loop over all segments
ja=ia + (n-1)*nlen
jb=ja+nlen-1
call pctile(sw(ja),nlen,npct,base) !Find lowest npct of points
do i=ja,jb
if(sw(i).le.base) then
if (k.lt.1000) k=k+1 !Save all "lower envelope" points
x(k)=i-i0
y(k)=sw(i)
endif
enddo
enddo
kz=k
a=0.
call polyfit(x,y,y,kz,nterms,0,a,chisqr) !Fit a low-order polynomial
sbase=0.0
do i=ia,ib
t=i-i0
sbase(i)=a(1)+t*(a(2)+t*(a(3))) + 0.2
! write(51,3051) i,sw(i),sbase(i)
!3051 format(i8,2f12.3)
enddo
sbase=10**(sbase/10.0)
return
end subroutine fst4_baseline

48
lib/fst4_decode.f90
View File

@ -49,7 +49,7 @@ contains
complex, allocatable :: cframe(:)
complex, allocatable :: c_bigfft(:) !Complex waveform
real llr(240),llrs(240,4)
real candidates(200,4)
real candidates(200,5)
real bitmetrics(320,4)
real s4(0:3,NN)
real minsync
@ -254,14 +254,19 @@ contains
nhicoh=1
nsyncoh=8
fa=max(100,nint(nfqso+1.5*hmod*baud-ntol))
fb=min(4800,nint(nfqso+1.5*hmod*baud+ntol))
minsync=1.2
if(iwspr.eq.1) then
fa=1400.0
fb=1600.0
else
fa=max(100,nint(nfqso+1.5*hmod*baud-ntol))
fb=min(4800,nint(nfqso+1.5*hmod*baud+ntol))
endif
minsync=1.20
if(ntrperiod.eq.15) minsync=1.15
! Get first approximation of candidate frequencies
call get_candidates_fst4(c_bigfft,nfft1,nsps,hmod,fs,fa,fb, &
minsync,ncand,candidates,base)
minsync,ncand,candidates)
ndecodes=0
decodes=' '
@ -317,7 +322,7 @@ contains
enddo
ncand=ic
xsnr=0.
!write(*,*) 'ncand ',ncand
do icand=1,ncand
sync=candidates(icand,2)
fc_synced=candidates(icand,3)
@ -465,6 +470,7 @@ contains
do i=1,NN
xsig=xsig+s4(itone(i),i)
enddo
base=candidates(icand,5)
arg=600.0*(xsig/base)-1.0
if(arg.gt.0.0) then
xsnr=10*log10(arg)-35.5-12.5*log10(nsps/8200.0)
@ -645,14 +651,15 @@ contains
end subroutine fst4_downsample
subroutine get_candidates_fst4(c_bigfft,nfft1,nsps,hmod,fs,fa,fb, &
minsync,ncand,candidates,base)
minsync,ncand,candidates)
complex c_bigfft(0:nfft1/2) !Full length FFT of raw data
integer hmod !Modulation index (submode)
integer im(1) !For maxloc
real candidates(200,4) !Candidate list
real candidates(200,5) !Candidate list
real, allocatable :: s(:) !Low resolution power spectrum
real, allocatable :: s2(:) !CCF of s() with 4 tones
real, allocatable :: sbase(:) !noise baseline estimate
real xdb(-3:3) !Model 4-tone CCF peaks
real minsync
data xdb/0.25,0.50,0.75,1.0,0.75,0.50,0.25/
@ -668,17 +675,17 @@ contains
signal_bw=4*(12000.0/nsps)*hmod
analysis_bw=min(4800.0,fb)-max(100.0,fa)
xnoise_bw=10.0*signal_bw !Is this a good compromise?
if(analysis_bw.gt.xnoise_bw) then
ina=ia
inb=ib
else
fcenter=(fa+fb)/2.0 !If noise_bw > analysis_bw,
fl = max(100.0,fcenter-xnoise_bw/2.)/df2 !we'll search over noise_bw
if(xnoise_bw .lt. 400.0) xnoise_bw=400.0
if(analysis_bw.gt.xnoise_bw) then !Estimate noise baseline over analysis bw
ina=0.9*ia
inb=min(int(1.1*ib),nfft1/2)
else !Estimate noise baseline over noise bw
fcenter=(fa+fb)/2.0
fl = max(100.0,fcenter-xnoise_bw/2.)/df2
fh = min(4800.0,fcenter+xnoise_bw/2.)/df2
ina=nint(fl)
inb=nint(fh)
endif
nnw=nint(48000.*nsps*2./fs)
allocate (s(nnw))
s=0. !Compute low-resolution power spectrum
@ -692,12 +699,16 @@ contains
ina=max(ina,1+3*hmod) !Don't run off the ends
inb=min(inb,nnw-3*hmod)
allocate (s2(nnw))
allocate (sbase(nnw))
s2=0.
do i=ina,inb !Compute CCF of s() and 4 tones
s2(i)=s(i-hmod*3) + s(i-hmod) +s(i+hmod) +s(i+hmod*3)
enddo
call pctile(s2(ina+hmod*3:inb-hmod*3),inb-ina+1-hmod*6,30,base)
s2=s2/base !Normalize wrt noise level
npct=30
call fst4_baseline(s2,nnw,ina+hmod*3,inb-hmod*3,npct,sbase)
if(any(sbase(ina:inb).le.0.0)) return
s2(ina:inb)=s2(ina:inb)/sbase(ina:inb) !Normalize wrt noise level
ncand=0
candidates=0
if(ia.lt.3) ia=3
@ -717,12 +728,11 @@ contains
s2(k)=max(0.,s2(k)-0.9*pval*xdb(i))
endif
enddo
! s2(max(1,iploc-2*hmod*3):min(nnw,iploc+2*hmod*3))=0.0
ncand=ncand+1
candidates(ncand,1)=df2*iploc !Candidate frequency
candidates(ncand,2)=pval !Rough estimate of SNR
candidates(ncand,5)=sbase(iploc)
enddo
return
end subroutine get_candidates_fst4