mock-data-challenges for ltp - aei.mpg.dehewitson/presentations/presentations2009/files/wed... ·...
TRANSCRIPT
AMALDI, NY, June 09
Outline
•Why MDCs•MDC1
•calibration of IFO outputs to out-of-loop acceleration•MDC2
•parameter estimation•MDC3 and beyond
2
Saturday, 20 June 2009
AMALDI, NY, June 09
Mock-Data-Challenges
3
Define MDCmodel(s),
assumptions, etc
1
Saturday, 20 June 2009
AMALDI, NY, June 09
Mock-Data-Challenges
3
Define MDCmodel(s),
assumptions, etc
1
Produce data sets based on 1)
2
Saturday, 20 June 2009
AMALDI, NY, June 09
Mock-Data-Challenges
3
Define MDCmodel(s),
assumptions, etc
1
Produce data sets based on 1)
2
Analyse data(based on some details from 1)
3
Saturday, 20 June 2009
AMALDI, NY, June 09
Compare results to expected
4
Mock-Data-Challenges
3
Define MDCmodel(s),
assumptions, etc
1
Produce data sets based on 1)
2
Analyse data(based on some details from 1)
3
Saturday, 20 June 2009
AMALDI, NY, June 09
MDC1
•Simple model of LTP (x-axis dynamics)•Data generation
•Model is based on 5 parameters of the system• stiffness of two test-masses• gains of two control servos• cross-coupling in IFO from X1 to X12
•Generate two IFO output time-series
4
Saturday, 20 June 2009
AMALDI, NY, June 09
MDC1
•Simple model of LTP (x-axis dynamics)•Data generation
•Model is based on 5 parameters of the system• stiffness of two test-masses• gains of two control servos• cross-coupling in IFO from X1 to X12
•Generate two IFO output time-series•Data analysis
•convert the two IFO outputs to out-of-loop acceleration• convert each to in-loop acceleration• account for control forces
4
Saturday, 20 June 2009
AMALDI, NY, June 09
Data generation
•Frequency-domain analytical model of transfer functions•Fit sets of digital filters to the transfer functions•Filter white-noise time-series to produce simulated IFO
outputs
5
o1
o12
Saturday, 20 June 2009
AMALDI, NY, June 09
Data generation
•Frequency-domain analytical model of transfer functions•Fit sets of digital filters to the transfer functions•Filter white-noise time-series to produce simulated IFO
outputs
5
o1
o12
Generation of simulated noise for LTP Mock Data ChallengesLuigi Ferraioli
Saturday, 20 June 2009
AMALDI, NY, June 09
Control loops
6
TM1
TM2
o12
o1
Cdf
Csus
x1
x2
ATAN
A
k1x1
k3x2
m
m
M
m1 = m2 = 1.96 kg
M = 475 kg
IFO/DMU
A1
A2 Asus
k1
k2
!A
!A
Saturday, 20 June 2009
AMALDI, NY, June 09
o1
o12
!12
interferometer
1s2 + !2
3
1s2 + !2
1Cdf
Asus
Adf
SC Force
Noise
TM
Differential
Force
Noise
Sensing
Noise
Sensing
Noise
Csus
Control loops
6
TM1
TM2
o12
o1
Cdf
Csus
x1
x2
ATAN
A
k1x1
k3x2
m
m
M
m1 = m2 = 1.96 kg
M = 475 kg
IFO/DMU
A1
A2 Asus
k1
k2
!A
!A
Saturday, 20 June 2009
AMALDI, NY, June 09
o1
o12
!12
interferometer
1s2 + !2
3
1s2 + !2
1Cdf
Asus
Adf
SC Force
Noise
TM
Differential
Force
Noise
Sensing
Noise
Sensing
Noise
Csus
Control loops
6
TM1
TM2
o12
o1
Cdf
Csus
x1
x2
ATAN
A
k1x1
k3x2
m
m
M
m1 = m2 = 1.96 kg
M = 475 kg
IFO/DMU
A1
A2 Asus
k1
k2
!A
!A
Saturday, 20 June 2009
AMALDI, NY, June 09
o1
o12
!12
interferometer
1s2 + !2
3
1s2 + !2
1Cdf
Asus
Adf
SC Force
Noise
TM
Differential
Force
Noise
Sensing
Noise
Sensing
Noise
Csus
Control loops
6
TM1
TM2
o12
o1
Cdf
Csus
x1
x2
ATAN
A
k1x1
k3x2
m
m
M
m1 = m2 = 1.96 kg
M = 475 kg
IFO/DMU
A1
A2 Asus
k1
k2
!A
!A
Saturday, 20 June 2009
AMALDI, NY, June 09
o1
o12
!12
interferometer
1s2 + !2
3
1s2 + !2
1Cdf
Asus
Adf
SC Force
Noise
TM
Differential
Force
Noise
Sensing
Noise
Sensing
Noise
Csus
Control loops
6
TM1
TM2
o12
o1
Cdf
Csus
x1
x2
ATAN
A
k1x1
k3x2
m
m
M
m1 = m2 = 1.96 kg
M = 475 kg
IFO/DMU
A1
A2 Asus
k1
k2
!A
!A
Saturday, 20 June 2009
AMALDI, NY, June 09
o1
o12
!12
interferometer
1s2 + !2
3
1s2 + !2
1Cdf
Asus
Adf
SC Force
Noise
TM
Differential
Force
Noise
Sensing
Noise
Sensing
Noise
Csus
Control loops
6
TM1
TM2
o12
o1
Cdf
Csus
x1
x2
ATAN
A
k1x1
k3x2
m
m
M
m1 = m2 = 1.96 kg
M = 475 kg
IFO/DMU
A1
A2 Asus
k1
k2
!A
!A
Saturday, 20 June 2009
AMALDI, NY, June 09
Calibration procedure
7
Convert to in-loopacceleration
[m]
[m]
!m s!2
"
!m s!2
"
Generate control signals
[m]
[m]
!m s!2
"
!m s!2
"
a1(t)
a2(t)
o1(t)
o12(t)
Saturday, 20 June 2009
AMALDI, NY, June 09
Calibration procedure
7
Convert to in-loopacceleration
[m]
[m]
!m s!2
"
!m s!2
"
Generate control signals
[m]
[m]
!m s!2
"
!m s!2
"
a1(t)
a2(t)
o1(t)
o12(t)
Double differentiation
Saturday, 20 June 2009
AMALDI, NY, June 09
Calibration procedure
7
Convert to in-loopacceleration
[m]
[m]
!m s!2
"
!m s!2
"
Generate control signals
[m]
[m]
!m s!2
"
!m s!2
"
a1(t)
a2(t)
o1(t)
o12(t)
Double differentiation
Filter with controllertransfer functions
Saturday, 20 June 2009
AMALDI, NY, June 09
Results
8
10−4 10−3 10−2 10−1 10010−14
10−13
10−12
10−11
10−10
10−9
10−8
10−7
Spectral Density [m][s−2][Hz−1/2]
Frequency [Hz]
DifferentialTM Force Noise
IFO Sensin
g Noise
Thruster Force Noise
a1
a1 ref
a2
a2 ref
Saturday, 20 June 2009
AMALDI, NY, June 09
MDC2
•Model same as MDC1•Analysis team does not know exact parameter values for the
model• Instead, they must be determined from a series of
experiments where the system is excited•3 different experiments•10 different ‘noise’ runs
9
Saturday, 20 June 2009
AMALDI, NY, June 09
Strategy
•Using ‘noise’ runs we determine the spectrum of the noise in the different channels•from these we construct whitening filters
10
Saturday, 20 June 2009
AMALDI, NY, June 09
Strategy
•Using ‘noise’ runs we determine the spectrum of the noise in the different channels•from these we construct whitening filters
•Outputs are whitened to produce statistically independent samples
10
Saturday, 20 June 2009
AMALDI, NY, June 09
Strategy
•Using ‘noise’ runs we determine the spectrum of the noise in the different channels•from these we construct whitening filters
•Outputs are whitened to produce statistically independent samples
•Estimate parameters from the whitened outputs
10
Saturday, 20 June 2009
AMALDI, NY, June 09
Strategy
•Using ‘noise’ runs we determine the spectrum of the noise in the different channels•from these we construct whitening filters
•Outputs are whitened to produce statistically independent samples
•Estimate parameters from the whitened outputs•Calibrate back to acceleration
• injected signals are removed (to some extent)
10
Saturday, 20 June 2009
AMALDI, NY, June 09
Strategy
•Using ‘noise’ runs we determine the spectrum of the noise in the different channels•from these we construct whitening filters
•Outputs are whitened to produce statistically independent samples
•Estimate parameters from the whitened outputs•Calibrate back to acceleration
• injected signals are removed (to some extent)•Look at residuals
•update whitening filter?
10
Saturday, 20 June 2009
AMALDI, NY, June 09
Experiments
11
i2
i1
Csus
CDFGDF
Gsus
D(!21)
o1
o12
!21
D(!23)
!12
interferometer
!2!
Saturday, 20 June 2009
AMALDI, NY, June 09
Experiments
• Experiment 1• inject signals into both control
loops and measure at the outputs• i1->o1 and i12->o12
• Gdf, Gsus (stiffnesses?)
11
i2
i1
Csus
CDFGDF
Gsus
D(!21)
o1
o12
!21
D(!23)
!12
interferometer
!2!
Saturday, 20 June 2009
AMALDI, NY, June 09
Experiments
• Experiment 1• inject signals into both control
loops and measure at the outputs• i1->o1 and i12->o12
• Gdf, Gsus (stiffnesses?)
• Experiment 2• Match stiffness of two TMs• Inject in drag-free loop, measure in
X12 loop• i1->o12
• IFO cross-coupling
11
i2
i1
Csus
CDFGDF
Gsus
D(!21)
o1
o12
!21
D(!23)
!12
interferometer
!2!
Saturday, 20 June 2009
AMALDI, NY, June 09
Experiments
• Experiment 1• inject signals into both control
loops and measure at the outputs• i1->o1 and i12->o12
• Gdf, Gsus (stiffnesses?)
• Experiment 2• Match stiffness of two TMs• Inject in drag-free loop, measure in
X12 loop• i1->o12
• IFO cross-coupling
• Experiment 3• Un-matched stiffness• Same injection
• i1->o12• difference of stiffness
11
i2
i1
Csus
CDFGDF
Gsus
D(!21)
o1
o12
!21
D(!23)
!12
interferometer
!2!
Saturday, 20 June 2009
AMALDI, NY, June 09
Different approaches
•Non-linear least-squares•time-domain:
• (fast) time-domain simulation•freq-domain:
• fit to measured transfer function
12
Saturday, 20 June 2009
AMALDI, NY, June 09
Different approaches
•Non-linear least-squares•time-domain:
• (fast) time-domain simulation•freq-domain:
• fit to measured transfer function•Linear least-squares
• linearize transfer functions in the parameters•do in either frequency-domain or time-domain
12
Saturday, 20 June 2009
AMALDI, NY, June 09
Different approaches
•Non-linear least-squares•time-domain:
• (fast) time-domain simulation•freq-domain:
• fit to measured transfer function•Linear least-squares
• linearize transfer functions in the parameters•do in either frequency-domain or time-domain
•MCMC•‘explore’ the full parameter space to determine posterior PDFs
for each parameter
12
Saturday, 20 June 2009
AMALDI, NY, June 09
Time-domain fitting
•Use models of the transfer functions with fft filtering to produce time-domain templates
•Look at experiment 1•fit 2 gains and 2 stiffness values
•Minimise χ2 to find best parameter estimates
13
Saturday, 20 June 2009
AMALDI, NY, June 09
Time-domain fitting
•Use models of the transfer functions with fft filtering to produce time-domain templates
•Look at experiment 1•fit 2 gains and 2 stiffness values
•Minimise χ2 to find best parameter estimates
13
Saturday, 20 June 2009
AMALDI, NY, June 09
Time-domain fitting
•Use models of the transfer functions with fft filtering to produce time-domain templates
•Look at experiment 1•fit 2 gains and 2 stiffness values
•Minimise χ2 to find best parameter estimates
13
Saturday, 20 June 2009
AMALDI, NY, June 09
Linearised model
•Taylor expand model of Transfer function around the nominal parameter values
•Solve with linear least-squares routine
14
Ti1o1 ! Ti1o1 linear
= T0 +!T0
!G0(Gdf "G0) +
!T0
!"0("1 " "0)
Saturday, 20 June 2009
AMALDI, NY, June 09
Linearised model
•Taylor expand model of Transfer function around the nominal parameter values
•Solve with linear least-squares routine
14
Ti1o1 ! Ti1o1 linear
= T0 +!T0
!G0(Gdf "G0) +
!T0
!"0("1 " "0)
Saturday, 20 June 2009
AMALDI, NY, June 09
Linearised model
•Taylor expand model of Transfer function around the nominal parameter values
•Solve with linear least-squares routine
14
Ti1o1 ! Ti1o1 linear
= T0 +!T0
!G0(Gdf "G0) +
!T0
!"0("1 " "0)
Saturday, 20 June 2009
AMALDI, NY, June 09
Linearised model
•Taylor expand model of Transfer function around the nominal parameter values
•Solve with linear least-squares routine
14
Ti1o1 ! Ti1o1 linear
= T0 +!T0
!G0(Gdf "G0) +
!T0
!"0("1 " "0)
Linear Analysis of the Second Mock Data Challenge for LTPAnneke Monsky
Saturday, 20 June 2009
AMALDI, NY, June 09
MCMC approach
•Find maximum of the likelihood surface using a standard simplex method
•Use Metropolis sampling to map-out the likelihood surface around this maximum
15
Saturday, 20 June 2009
AMALDI, NY, June 09
MCMC approach
•Find maximum of the likelihood surface using a standard simplex method
•Use Metropolis sampling to map-out the likelihood surface around this maximum
15
Bayesian parameter estimation in LISA Pathfinder MDCsMiquel Nofrarias
Saturday, 20 June 2009
AMALDI, NY, June 09
MDC3 and beyond
•Aim to verify the analysis in S2-UTN-TN-3045•Do full experiment campaign•Confirm parameters can be measured with the expected
accuracy•Work through other major experiments
•move in to 2D and 3D• x-y cross-talk, angular couplings, etc
16
Saturday, 20 June 2009