3-07 Brown

advertisement
Infrasound Technology
Workshop
Bermuda, 2008
Enhancements to the CTBTO operational
automatic infrasound processing system
David Brown, Nicolas Brachet, and Ronan Le Bras
International Data Centre
Software Applications Section
Preparatory Commission for the Comprehensive
Nuclear-Test-Ban Treaty Organization
Provisional Technical Secretariat
Vienna International Centre
P.O. Box 1200, A-1400 Vienna, Austria
E-mail: David.Brown@ctbto.org
E-mail: Nicolas.Brachet@ctbto.org
Overview
Two enhancements to the IDC automatic infrasound processing
system are being tested:
 Station Noise characterization

purpose

requirements

method

testing

real data

still to be done
 Infrasound amplitude determination
Infrasound Technology Workshop, Bermuda

purpose

procedure

testing

still to be done
November 2008
Page 2
Station Noise Characterization
Purpose
 Provide a method for users of IDC external products to determine station ‘noise’
levels.
• will be useful in determining network detection capability
 Provide an additional utility for internal PTS station performance monitoring.
Requirements
 Noise data for each station would need to reveal:
• seasonal or monthly variations
• day and night time variations
 External/Internal Users need to access the binary data and plot files.
Method
 Employ Power Spectral Density methods
• along the lines of the analysis of Bowman et al.
• will run automatically on incoming data and update seasonal or monthly averages.
Infrasound Technology Workshop, Bermuda
November 2008
Page 3
Station Noise Characterization
Choose 4 hourly periods for each station:
 03:30-04:30, 09:30-10:30, 15:30-16:30, 21:30-22:30 local time
Choose a FFT window function:
 References:
• Harris 1978: On the use of Windows for Harmonic Analysis with the Discrete Fourier
Transform Proc. IEEE Vol 66, No. 1, 1978.
•
Heinzel 2002: https://www.lisa.unihannover.de/?page=publikationen&sub=publikationen&type=publi&publitype=3&lang=en
 Require moderate frequency resolution
 Require good amplitude resolution
Infrasound Technology Workshop, Bermuda
November 2008
Page 4
Station Noise Characterization
Windowing:
 a significant part of the PSD measurement process
no window
equivalent to
High Frequencies
Infrasound Technology Workshop, Bermuda
November 2008
Page 5
Station Noise Characterization
Windowing:
 non-periodic nature of the windowing process leads to spectral leakage
 frequency content and spectral leakage controlled by the transform of the window
function:
F ( f (t )  w(t )  D(t ))  F ( ) W ( )  D( )
W ( )
‘Dirac Comb’
‘Dirac Comb’
spectral leakage
Infrasound Technology Workshop, Bermuda
November 2008
Page 6
Station Noise Characterization
Windowing:
 Ensure the side lobes of the Window Transform are below the noise floor of the
sensor system (microbarometer + filter):
• MB2005:
approx. 10-7 Pa2/Hz → Nutall4a window (Heinzel et al, 2002)
• Chaparral : approx. 10-9 Pa2/Hz → Nutall4c window (Heinzel et al, 2002)
Infrasound Technology Workshop, Bermuda
November 2008
Page 7
Station Noise Characterization
‘Welch’ PSD method
 Divide hourly data into overlapping 3 minute windows
one hour
…
• without averaging, the variance of the PSD for a individual frequency picket is of the order
of the PSD at that frequency.
• 20 Hz sample-rate →3600 samples or 1800 frequency pickets:
• 0.0000000 Hz→ DC
• 0.0015625 Hz → 640 sec period
• 0.0031250 Hz → 320 sec period
•…
• 9.9984375 Hz
Infrasound Technology Workshop, Bermuda
November 2008
Page 8
Station Noise Characterization
 Number of overlapping windows is determined by the window ROV (Relative Overlap
Value, see Heinzel et al, 2002), indicated in the following Table:
window
ROV
%
3-minute
windows
per hour
ENBW
(Equivalent
Noise Band
Width, bins)
side
lobe
level
emax
%
3db
width
(bins)
(db)
high frequency resolution windows
Rectangular
0.0
20
1.0000
-13.3
-36.3
0.9
general purpose windows
Welch
29.3
28
1.2000
-21.3
-22.6
1.2
Hanning
50.0
39
1.5000
-31.5
-15.1
1.4
Hamming
50.0
39
1.3628
-42.7
-18.2
1.3
Nutall4a
68.0
60
2.1253
-82.6
-8.1
2.0
Nutall4c
65.6
56
1.9761
-98.1
-9.3
1.9
high amplitude resolution windows
Infrasound Technology Workshop, Bermuda
FTSRS
75.4
78
3.7702
-76.6
-0.2
3.7
HFT248D
84.1
120
5.6512
-284.4
+0.0
5.6
November 2008
Page 9
Station Noise Characterization
 Remove linear trend from each 3 minute data segment
• otherwise DC component may generate spectral leakage
 Multiply data segment with window function
• remember to divide by the processing gain, ie, multiply by
 Apply FFT
 Determine PSD at frequency picket i according to the rule:
2
N
  FFTi
PSDi  
2N
2

FFTi
 
N
w
i
if frequency is DC or Nyquist
otherwise
N = samples;  = sample frequency;  = ENBW
N 1
N
w
i 0
2
i
 N 1 
  wi 
 i 0 
2
Accommodates the accumulation ability
of the main lobe for white noise
 average the PSD for all sub windows and take the log of the average
• could consider taking the average of the log’s, but Bowman et al’s work suggests not
log-normal distribution.
Infrasound Technology Workshop, Bermuda
November 2008
Page 10
Station Noise Characterization
Testing
 use artificial digitizer noise
• Digitizing process generates (white) noise power spectral density:
(see Lyons, 1997; Heinzel, 2002)
2
U LSB
~
U dig 
6
U LSBis the Pa corresponding to one least-significant bit

is the sampling frequency, Hz.
 Example:
f1  0.3123456789 Hz
u  A1 sin( 2f1t )  A2 sin( 2f 2t )
 u

y  int 
 0.5 U LSB
 U LSB

f 2  2.0Hz
with
A1  2.123456789 Pa
A2  1.0 Pa
U LSB  0.001 Pa
2
 U LSB

~
  8.079
log 10 U dig  log 10 
 6 
Infrasound Technology Workshop, Bermuda
November 2008
  20
Page 11
Station Noise Characterization
Testing
 use 4 different windows:
A
2  PSD
N
 Rectangular (no window)
 Hanning
 Nutall4a (Heinzel 2002)
 HFT248D (Heinzel 2002)
-8.079
-8.079
Infrasound Technology Workshop, Bermuda
November 2008
Page 12
Station Noise Characterization
Testing
 have initiated a series of ‘blind-tests’ with Lars Ceranna at BGR to assess the
validity of the method.
• using randomly chosen IMS array data
• IS07, IS27, IS32, IS55
• 2007-2008, 1st day of the month 4 times per day.
• will ‘manipulate’ the PSD methods until we get the same answers.
 compare with known published results: Bowman et al.
Infrasound Technology Workshop, Bermuda
November 2008
Page 13
Station Noise Characterization
Real Data
 PSD noise data is being determined automatically for every IMS infrasound station.
 Two outputs are generated:
• Output 1: log-average PSD data for the given hours
• binary data file + plot (.ps)
• should be free of contamination from spectral leakage
• have complete hourly and monthly-averaged PSD information for each station for September
2008 , October 2008 and the current part of November 2008.
• for each 8-element station the data growth rate is :
4 x (80000 + 130000 ) bytes = 840Kb
(.jpg
Infrasound Technology Workshop, Bermuda
binary)
November 2008
Page 14
Station Noise Characterization
Real Data
 PSD noise data is being determined automatically for every IMS infrasound station.
 Two outputs are generated:
• Output 1: log-average PSD data for the given hours
• binary data file + plot (.ps)
• should be free of contamination from spectral leakage
• have complete hourly and monthly-averaged PSD information for each station for September
2008 , October 2008 and the current part of November 2008.
• for each 8-element station the data growth rate is:
4 x (80000 + 130000 ) bytes = 520Kb / day
(approx. 5 Gb per year for 60-stations)
(generate graphic on the fly)
Infrasound Technology Workshop, Bermuda
November 2008
Page 15
Station Noise Characterization
Real Data
 PSD noise data is being determined automatically for every IMS infrasound station.
 Two outputs are generated:
• Output 2: monthly-average PSD data + variance smoothed using an 11-point 6-order SavitzkyGolay Filter for the specified times.
• Savitzky-Golay filter
o polynomial regression that preserves features of the distribution such as relative maxima,
minima and width
• have complete hourly and monthly-averaged PSD information for each station for September
2008 , October 2008 and the current part of November 2008.
Infrasound Technology Workshop, Bermuda
November 2008
Page 16
Station Noise Characterization
Still to be done
 incorporate as part of mainstream IDC processing
 make data accessible to external users via a web-service interface
 incorporate into Station Performance Monitoring tool for internal PTS use
Infrasound Technology Workshop, Bermuda
November 2008
Page 17
Infrasound Amplitude Determination
Purpose
 Provide a method for determining infrasound amplitudes in Pascals
• useful information to write in the Bulletin
• may be useful in Network processing
Infrasound Technology Workshop, Bermuda
November 2008
Page 18
Infrasound Amplitude Determination
Procedure
 Apply Window Function in the time domain
• Use FTSRS ‘Flat Top’window
• process sufficient data such that the window function will not drop by more than 50% from its
peak value over the duration of the signal segment.
• otherwise signal will get mangled by the window division processes
window function
signal
0.5
0.0
extra data
P  103.36 0.019V D 1.36
P  105.31 0.0116V D 1.76
signal
start time
Infrasound Technology Workshop, Bermuda
signal
end time
November 2008
Page 19
Infrasound Amplitude Determination
Procedure
 Apply FIR filter
• two filter bands chosen:
• 0.5 to 3.0 Hz
• minfreq and maxfreq determined by pmcc detection algorithm
 Remove window function from signal segment by dividing
 Three amplitude measures applied to a time-aligned beam :
• beam aligned according to azimuth + slowness in CSS detection table
• Peak-to-Peak: used in several attenuation laws
• LANL P  103.360.019V D 1.36
• DASE P  105.310.0116V D 1.76
• RMS (user specifiable window length)
• Amplitude of the Analytic Trace (see Olson, 2000)
A(t )  x 2 (t ) H 2 (t )
 For each detection:
• 6 amplitude measures are being determined
• the Peak-to-Peak 0.5-3.0 Hz amplitude is being written to the CSS arrival table, all others are
being written to the amplitude table.
Infrasound Technology Workshop, Bermuda
November 2008
Page 20
Infrasound Amplitude Determination
Testing
 A sinc function is used to test beamformer, filtering procedure and amplitude
measurement
A sin 2f (t  t0 ) 
y (t ) 
A  2.0 f  0.3Hz   20 samples/se cond
2f (t  t0 )
• Peak-to-Peak (P-P) amplitude = 2.434 ; Analytic amplitude = 2.0
Nutall 4a window
0.001 - 9.999 Hz
0.15 - 0.60 Hz
Infrasound Technology Workshop, Bermuda
Overlapping traces:
• beamformer is working correctly
• filter process is not generating a phase-shift
• window division is not mangling the data
non-overlapping traces:
• filter process is not generating a phase-shift
• window division is not mangling the data
November 2008
Page 21
Infrasound Amplitude Determination
Testing
window
high
frequency
resolution
rectangle
welch
general
purpose
hanning
Results:
• for all windows the P-P
amplitude measure 0.001 to 9.999 Hz 0.15 to 0.6 Hz
measure provides the
P-P
2.434
2.104
correct amplitude in the
‘unfiltered’ case
RMS
1.623
0.900
• the HFT248D window
Analytic
1.987
1.215
provides the best analytic
P-P
2.434
2.104
amplitude in the unfiltered
RMS
1.618
0.900
case.
• all windows provided
Analytic
1.981
1.215
reduced amplitudes for a
P-P
2.434
2.104
2-octave band-pass filter
RMS
1.623
0.900
centred on the dominant
Analytic
1.986
1.215
frequency.
P-P
2.434
2.104
RMS
1.628
0.900
Analytic
1.991
1.215
P-P
2.434
2.104
RMS
1.632
0.900
Analytic
1.996
1.215
nutall4a
high
amplitude
resolution
HFT248D
Infrasound Technology Workshop, Bermuda
November 2008
Page 22
Infrasound Amplitude Determination
Still to be done
 evaluate results
 compare with analyst measured amplitudes
 investigate utility of different amplitude measures
 Analytic amplitude as a microbarom classifier ?
 the RMS amplitude as a signal significance measure ?
 incorporate as part of mainstream IDC processing
Infrasound Technology Workshop, Bermuda
November 2008
Page 23
Conclusions
Summary
 Two enhancements to the automatic infrasound processing system are
currently under development:
 station noise characterization
 amplitude determination
 Should be operational in around 3-4 months.
Infrasound Technology Workshop, Bermuda
November 2008
Page 24
Download