singular value decomposition-based modeling of time domain signals in broadband microwave...
TRANSCRIPT
Singular Value Decomposition-Based Modeling of Time Domain Signals in Broadband
Microwave Spectroscopy
A. J. MineiCollege of Mount St. Vincent
S. A. CookePurchase College SUNY
Motivation:
Summer 2012 recorded some strong free induction decays pertaining to the rotational spectrum of 1H,2H-perfluorocyclobutane.
800,000 points. 25 ps/point.
Could we do something with the time domain signals?
Comments- Very useful for resolving overlapping
lines- Implemented in FTMW++
But:
- Need to know how many lines in the FID
- Performed iteratively- Non-linear least squares fit, needs
starting parameters
Motivation:
This is reference 7:
K
kknkkn
kbnt
etvcty1
2cos)(
ck, vk, fk, and bk represent amplitude, frequency, phase and damping factor for the kth signal, tn = nDt, with Dt = 25 ps.
The Algorithm:
Use a function to model the transient emission:
x1 x2 x3 x4
x2 x3x4 . . . .
. . . . . . xn
x3 x4 . . . .
.
.
.
X =
The Hankel Matrix
TVUX Singular value decomposition:
S is a diagonal vector, with the singular values along the diagonal,U and V are matrices for which columns contain the left and right singular vectors.
Then find Z which satisfies:
ZUU
tvibz kk 2exp
do 77 k = 1,kfit damp(k) = dlog(cdabs(root(k))) freq(k) = datan2(dimag(root(k)),dble(root(k))) / (2.0d0*pi) 77 continue
Diagonalize Z to obtain K signal poles, or roots, zK
Then create the Vandermonde matrix:
12
13
233
12
222
11
211
1
1
1
1
Mkkk
M
M
M
zzz
zzz
zzz
zzz
Then perform linear least squares fit:
nk xc ' )exp(' kkk icc
do 88 k = 1,kfit ampl(k) = cdabs (x(k)) fase(k) = datan2(dimag(x(k)),dble(x(k))) 88 continue
Output:
Parameters of the model function for several transitions of 1H,2H-perfluorocyclobutane using the HSVD-method. Only the first 8192 data points (1% or 200ns) of the FID were used.
k JK-1K+1 – JK-1K+1 ck bk / ms
fk / deg. vk / MHz vk / MHz (1024k FFT)
1 321 – 211 0.00111 1.42 -29.2 8753.1008 8753.2290
2 312 – 202 0.00222 63.84 23.6 8756.5952 8756.6013
3 330 – 220 0.00171 2.39 -96.2 9265.7422 9265.7423
4 331 – 221 0.00372 13.73 63.6 9369.4766 9369.4739
5 422 – 312 0.00195 12.30 58.5 11595.9084 11565.8884
6 423 – 313 0.00146 5.35 -147.1 11903.3664 11903.3749
7 432 – 322 0.00121 8.21 -39.0 12075.4803 12075.4601
8 441 – 331 0.00103 12.34 -140.8 12549.3485 12549.3555
K
kknkkn
kbnt
etvcty1
2cos)(
Data pts Time ck bk / ms vk / MHz vkFFT – vk / kHz fk / deg.
512 1 0.00390 0.12 9369.4188 55.1 62.7
1024 2 0.00371 5.41 9369.4901 -16.2 62.5
2048 9 0.00369 6.42 9369.6226 -148.7 61.8
4096 61 0.00368 3.25 9369.5171 -43.2 62.8
8192 513 0.00372 13.72 9369.4766 -2.7 63.6
Parameters of the model function for the 331 ← 221 transition for 1H, 2H perfluorocyclobutane for different numbers of FID data points
K
kknkkn
kbnt
etvcty1
2cos)(
Performance as a function of transition amplitude
Performance as a function of transition frequency
Possible applications:
1. Signal reconstruction2. Identifying non-molecular signals3. Alternative to zero-filling4. Time domain signal filtering5. Peak picking6. De-noising (!)7. Experimental rate improvement?8. …?
Matrix dimension
Time / s
Problems
Intel Core i7-3610QM @ 2.30 GHz16 Gb RAMgfortran compiled code (LAPACK/BLAS)
Future work:
1. Make it go faster
Thank you!
Acknowledgements:Thanks to Dirk van Ormondt – Delft USouthern New England Microwave Spectroscopy Consortium
Different compilerMulti-threading (ScaLAPACK, MPI)