sean danaher z 1 z 1 z 1 z 1 z 1 z 1 [a] -k- [a] u1u1 u2u2 u3u3 u4u4 y2y2 y1y1 y3y3...
Post on 20-Dec-2015
224 views
TRANSCRIPT
Sean Danaher
z1
z1
z1
z1
z1
z1
[A]
-K-
-K-
-K-
-K-
-K-
-K-
-K-
[A]
u 1
u 2
u 3
u 4
y 2
y 1
y 3
MIM
O
SYS
TEM
+-
+-
+-
+-
Signal Processing For Acoustic Neutrino
Detection (A Tutorial)
Sean Danaher University of Northumbria,
Newcastle UK
Sean Danaher
z1
z1
z1
z1
z1
z1
[A]
-K-
-K-
-K-
-K-
-K-
-K-
-K-
[A]
u 1
u 2
u 3
u 4
y 2
y 1
y 3
MIM
O
SYS
TEM
+-
+-
+-
+-
What is Signal Processing?• Used to model signals and systems where
there is correlation between past and current inputs/outputs (in space or time)
• Two broad categories: Continuous and Sampled processes
• A host of techniques Fourier, Laplace, State space, Z-Transform, SVD, wavelets……
• Fast, accurate, robust and easy to implement• Sadly also a big area. Engineers spend years
learning the subject (Not long Enough!)• Smart as Physicists are I can’t cover a 3+ year
syllabus in 40 minutes!• Notation (not just i and j)!
Sean Danaher
z1
z1
z1
z1
z1
z1
[A]
-K-
-K-
-K-
-K-
-K-
-K-
-K-
[A]
u 1
u 2
u 3
u 4
y 2
y 1
y 3
MIM
O
SYS
TEM
+-
+-
+-
+-
At Virtually all stages of the process1. Accurate Parameterisation of distributions with no
amenable analytical form e.g. Shower Energy distribution profiles
2. Speed up computation e.g. Acoustic integrals by many orders of magnitude
3. Design, Understand, simulate Hydrophones, Microphones and other acoustic transducers (and amplifiers)
4. Design, Understand, simulate filters both analogue and digital (tailored amplitude, phase response, minimise computing time)
5. Estimate and recreate noise and background spectra6. Design Optimal Filters e.g. Matched Filters7. Design Optimal Classification Algorithms (Separating
Neutrino Pulses from Background)8. Improve Reconstruction accuracy……………Will cover 1-5 in this talk……
How Can Signal Processing be used in Acoustic Neutrino
Detection?
Sean Danaher
z1
z1
z1
z1
z1
z1
[A]
-K-
-K-
-K-
-K-
-K-
-K-
-K-
[A]
u 1
u 2
u 3
u 4
y 2
y 1
y 3
MIM
O
SYS
TEM
+-
+-
+-
+-
Singular Value Decomposition
W L VTW L VT
(((( )))) (((( ))))(((( ))))Decompose the Data matrix into 3 matrices.When multiplied we get back the original data.W and V are unitary. L contains the contribution from each of the eigenvectors in descending order along main diagonal.
We can get an approximation of the original data by setting the L values to zero below a certain threshold
Similar techniques are used in statistics CVA, PCA and Factor AnalysisBased on Eigenvector Techniques
Good SVD algorithms exist in ROOT and MATLAB
Sean Danaher
z1
z1
z1
z1
z1
z1
[A]
-K-
-K-
-K-
-K-
-K-
-K-
-K-
[A]
u 1
u 2
u 3
u 4
y 2
y 1
y 3
MIM
O
SYS
TEM
+-
+-
+-
+-
SVD IISomewhere between vector five and seven we get the “Best” representation of the data
•Better than the observation as noise filtered out
•Highly compressed (only 1-2% of original size e.g. 6/500x1.5)
•Have basis vectors for data so can produce “similar” data+++
SVD done onNoisy data
Sean Danaher
z1
z1
z1
z1
z1
z1
[A]
-K-
-K-
-K-
-K-
-K-
-K-
-K-
[A]
u 1
u 2
u 3
u 4
y 2
y 1
y 3
MIM
O
SYS
TEM
+-
+-
+-
+-
Building the Radial Distribution
SVD works well with the radial distribution•Three vectors sufficient to fit the data•Can use a linear mixture of these three vectors as input to the Acoustic MC
We Reproduce both the shape and variation
4 examples chosen have maximum variation in shape
Hadronic Component using Geant IV
Sean Danaher
Acoustic Integrals
fast thermal energy deposition
2
2
d
dt
( ) ( / )0
4 p
E dP t r c t dv
C r dt
Band limited by •Water Properties•Attenuation
Slow decay
Sean Danaher
Method I
z1
z1
z1
z1
z1
z1
[A]
-K-
-K-
-K-
-K-
-K-
-K-
-K-
[A]
u 1
u 2
u 3
u 4
y 2
y 1
y 3
MIM
O
SYS
TEM
+-
+-
+-
+-
•Throw a number if MC points ~107
•Assume the energy at each point in the cascade deposited as a Gaussian Distribution. •Calculate distance to observer from each point •Propagate the Gaussian to the observer (an analytical solution is known for this)•Sum over all the points•Provided the width of the Gaussian is small in comparison to the shower radius we will get the pressure pulse
Expensive5 flops per time pointIf using 1024 pointson time axis5*1024*1e7=5.1e10 flops
-0.200.22
3
4
5
6
7
8
9
10
11
12
0 2 4 6 8 10 120
0.5
1
1.5
2
2.5
3 x 104
Longitudinal Distance (m)F
req
uen
cyRadial Distribution
-5 0 5-0.5
0
0.5
Time (ps)
P (
arb
itra
ry)
x 10-4-6 -4 -2 0 2 4 6-0.1
-0.05
0
0.05
0.1
Time (s)
Pre
ssu
re (
Pa)
Need to do this thousands of times for different distances, angles, distributions, etc.
Sean Danaher
u 1
u 2
u 3
u 4
y 2
y 1
y 3
MIM
O
SYS
TEM
+-
+-
+-
+-
The Convolution IntegralGiven a signal s(t) and a system with an
impulse response h(t) (response to (t)) then y(t) to an arbitrary input is given by
( ) ( ) ( )y t s t h t d
s(t) Bipolar Acoustic Pulse
h(t) Impulse response high pass filter (SAUND)
y(t) ElectricalPulse
We can use this to process all the MC points in the cascade simultaneously
Sean Danaher
z1
z1
z1
z1
z1
z1
[A]
-K-
-K-
-K-
-K-
-K-
-K-
-K-
[A]
u 1
u 2
u 3
u 4
y 2
y 1
y 3
MIM
O
SYS
TEM
+-
+-
+-
+-
* =
A Few examples
RC circuit
2d
Mixed signalThis situation is similar to our acoustic integral
Sean Danaher
z1
z1
z1
z1
z1
z1
[A]
-K-
-K-
-K-
-K-
-K-
-K-
-K-
[A]
u 1
u 2
u 3
u 4
y 2
y 1
y 3
MIM
O
SYS
TEM
+-
+-
+-
+-
Will speed the process still furtherWill speed the process still further• Convolution very expensive
computationally (o=n2)– Convert the signal and impulse response to
the frequency domain (Fourier Transform)– Provided the number of points n=2m very
efficient FFTs o=n log n– Multiply and take Inverse Transform
Convolution in the time domain is Multiplication in the frequency domain (and visa versa)
Convolution Theorem
Sean Danaher
z1
z1
z1
z1
z1
z1
[A]
-K-
-K-
-K-
-K-
-K-
-K-
-K-
[A]
u 1
u 2
u 3
u 4
y 2
y 1
y 3
MIM
O
SYS
TEM
+-
+-
+-
+-
Acoustic Integrals II•Assume each point produces a Delta Function•The sampled equivalent of a Delta function is 1 •The integral can be done using a histogram function(modern histogram algorithms are very efficient)•The derivative can be done in the frequency domain (d/dt i No overhead)•We then convolve the response with the entire signal
( ) ( / )
( / ) ( )
0
0 0
4
4 4
p
xyzp p
E dP t r c t dv
C r dt
E Ed dr c t dv E t
C dt r C dt
Sean Danaher
z1
z1
z1
z1
z1
z1
[A]
-K-
-K-
-K-
-K-
-K-
-K-
-K-
[A]
u 1
u 2
u 3
u 4
y 2
y 1
y 3
MIM
O
SYS
TEM
+-
+-
+-
+-
Only Histogram depends on the number of MC points
511
512
[ ] ( ) ( ) ( ) ( )n
i nTxyz w
n
P t E w D A e
FT of Exyz
Blackman window
Derivative j
Water Attenuation
Histogram: 103 flopsScale histogram by 1/d: 103 flops (only needed in near field)FFT nLogn c 104 flopsMultiplication c 4x103 flopsIFFT nLogn c 104 flops
6 orders of magnitude less than approach 1Limitation: Will not work if volume of the integral is so great that attenuation varies significantly (Large 100+m source in near field)
Sean Danaher
u1
u2
u3
u4 y
2
y1
y3
MI
MO
SYSTE
M
State Space AnalysisMIMO systemsMixed Mechanical/electrical models etcMatrix Based Method
X = AX +BU
Y = CX +DU
Pictured 20000 state simulation of a square acoustic membrane:(100x100 masses 2 states x and v)A Matrix 400x106 elements
All modern control algorithms use
SS methods.
Mode 1
Mode 14
Mode 100
didt
H
1
2
3
4
.
.
n
i H
Quantum Mechanics
Sean Danaher
u1
u2
u3
u4 y
2
y1
y3
MI
MO
SYSTE
M
A B
DC
SS Implementation
if
if
if position velocity
if velocity acceleration
c C
L L
dVx V I C Cx
dtdi
x I V L Lxdt
x x
x x
States Inputs
States
Outputs
States are degrees of freedom of the system. Things that store energy•Capacitor Voltages•Inductor currents•Positions and Velocities of masses
A is of size States StatesB is of size States InputsC is of size Outputs StatesD is of size Outputs Inputs
Sean Danaher
u1
u2
u3
u4 y
2
y1
y3
MI
MO
SYSTE
M
Simple Example LC ( )c L
c L
L c L C
dV dII C V L
dt dtV V I I
1 1 2 1 1
2 2 1 2 2
C
L
x V Cx x y x
x I Lx x y x
10
1 0
0 110
C
L
A C
-1.5
-1
-0.5
0
0.5
1
1.5
Vo
ltag
e
0 5 10 15 20 25 30 35 40-1.5
-1
-0.5
0
0.5
1
1.5
Cu
rren
t
Response to Initial Conditions
Time (sec)
Am
plit
ud
e
Maths simple to implementExample shows the behaviour with an initial 1V on the Capacitor (1F) and 1A flowing through the inductor (1H)
Sean Danaher
z1
z1
z1
z1
z1
z1
[A]
-K-
-K-
-K-
-K-
-K-
-K-
-K-
[A]
u 1
u 2
u 3
u 4
y 2
y 1
y 3
MIM
O
SYS
TEM
+-
+-
+-
+-
Speaker/MicrophoneConeFrame
Magnetassembly
Coil
S N
2
2e
d x dx xF Bi m r
dt dt s Mechanically Force produced by current
Loss to air term
dx diV B L Ri
dt dt Electrically coil moving in a magnetic field
Creates an EMF
x1=x, (position) x2=dx/dt, (velocity)x3=i, (current)
0 1 0 0 0
1 10
100
0 1 0 0 0( ) ( )
r B
sm m m mB R
LL L
X X U
Y X U
-160
-140
-120
-100
-80
-60
Mag
nitu
de (
dB)
102
103
104
105
106-180
-90
0
90
Pha
se (
deg)
Bode Diagram
Frequency (rad/sec)
0 0.5 1 1.5 2 2.5 3-0.5
0
0.5
1
1.5
2
2.5x
10
-4 Step Response
Time (ms)A
mp
litu
de
Magnetassembly
Sean Danaher
z1
z1
z1
z1
z1
z1
[A]
-K-
-K-
-K-
-K-
-K-
-K-
-K-
[A]
u 1
u 2
u 3
u 4
y 2
y 1
y 3
MIM
O
SYS
TEM
+-
+-
+-
+-
Simple Hydrophone Model
Massm
F=rv
F=ma
F=x/SF=KVc
VcKvIc
inV Vc
R
Vin
1 2 3, ,Cx V x v x x Heart of HydrophonePiezo electric crystalPiezo electric effect
Omni works in breathing mode
40cm
30cm
25 30 35 40 45 50-15
-10
-5
0
5
10
15
20
25
Time (eq. cm)
Am
plit
ud
e (A
rbit
rary
)
Simulated Hydrophone signal
0
x
I I e
3rd order simulation. Do we need higher?
Gaussian cross-section
“Erlangener” water tank (Niess)
10 1
0
10 1
0 00 1 0
1 10 0 0
0 1 0 0 0
K
RC CRC
K rA B
m m Sm
C DR R
= 0.09 =5e-3
Sean Danaher
u1
u2
u3
u4 y
2
y1
y3
MI
MO
SYSTE
M
Acoustic Simulation•Bipolar Pulse •1000 element simulation•Two types of medium; Heavy and light
•pulse reflects against walls •mismatch between sections•velocity on each section•amplitude change •dispersion
0 1 0 0 0 0 0 0 . . . . . 0
-8 0 4 0 0 0 0 0 . . . . . 0
0 0 0 1 0 0 0 0 . . . . . 0
4 0 -8 0 4 0 0 0 . . . . . 0A =
0 0 0 0 0 1 0 0 . . . . . 0
. . . . . . . . . . . . . .
. . . . . . . . . . . . . .
0 0 0 0 0 0 0 0 0 0 1 0 -2 0
Enhance to 2D?Embed Hydrophone models etc.?
Sean Danaher
+-
+-
+-
+-
Butterworth Filter
StableGood Frequency Response
1930s
0 10 20 30 40 50-0.2
0
0.2
0.4
0.6
0.8
1
1.2
Time(s)
Ou
tpu
t
Buterworth Impulse Response
1st
order
2nd
order
4th
order
8th
order
16th
order
32nd
orderNote Impulse response •Output Delay•Oscillation
2
1
1| ( ) |
nH
Sean Danaher
+-
+-
+-
+-
Implementation
+
-
+
-
Cascade 2nd Order sections to make Butterworth high order filters
1st Order Familiar RC
2nd Order LRC
3rd Order “”
Passive Filters
At acoustic frequencies easiest to use op-amps. Expensive to make high precision lossless inductors
1
1 1
4 3 2
2 2
s +2.6131s +3.4142s +2.6131s+1
s +0.7654s+1 s +1.8478s+1
Ra
Rb
=
2| ( ) | , 1
(3 ) 1DC a
DCDC b
A Rh s A
s A s R
=
Sean Danaher
+-
+-
+-
+-
Recovering Phase information
0 0.2 0.4 0.6 0.8 1 1.2
x 10-3
-0.1
-0.05
0
0.05
0.1
0.15
time (s)
Am
pli
tud
e (
Arb
itra
ry)
S+Ncausalnon causal
It is trivial to design digital filters which have a constant group delay d/d and hence no phase distortionIf however we know the filter response e.g. Butterworth we can run the data through the filter backwards.
This increases the order of the filter by a factor of 2
e.g. SAUND Data
Sean Danaher
z
1
z
1
z
1
z
1
z
1
z
1
[A]
-K-
-K-
-K-
-K-
-K-
-K-
-K-
[A]
The Digital Filter
0 1 2 0 1 21 2 1 2[ ] [ ] [ ] .... [ ] [ ] [ ] ....a y n a y n a y n b u n bu n b u n
The digital filter is simple! Based upon sampled sequences
10 52 66 45 5.59 78 8.55 6.50 2.58 44 6
11
2[ ] [ ] [ ]y n x n x n
Simple moving average Filter
Crude low-passCalled FIR orMA
Negative coefficients =>Causal
6 68 149 237 302 324 369 459 544 589 67
Perfect Integrator
Called IIR or AR
[ ] [ 1] [ ]y n y n x n
Sean Danaher
z
1
z
1
z
1
z
1
z
1
z
1
[A]
-K-
-K-
-K-
-K-
-K-
-K-
-K-
[A]
The Z Transform and Sampled Signals
( ) [ ]n
n
n
X z x n z
1 2
2 11
0
1 3 4 1 3 4
11
1
, , ( )
, , ,... ( )n n
n
if x X z z z
if x a a X z azaz
Geometric sequences are of prime interest because
( ) [ ]i t i ntu t e u t e is a geometric sequence
Sean Danaher
z
1
z
1
z
1
z
1
z
1
z
1
[A]
-K-
-K-
-K-
-K-
-K-
-K-
-K-
[A]
-1
0
1
-4-3-2-10123
-1 0 1
-1
0
12 1
1
2 110
2 1 210
( )cos
[ ] [ ] cos [ ] [ ]
h zz z
y n x n y n y n
0 5 10 15 20 25 30
-15
-10
-5
0
5
10
0 5 10 15 20 25 30
-1.5-1
-0.50
0.51
1.52
0 5 10 15 20 25 30-1
0
1
0 0.2 0.4 0.6 0.8 10
5
10
15
20
Fraction of Nyquist frequency
Frequency Response
Poles must be inside the unit circle for stability
/10, 1
/ 5, 0.95
/14, 1.05
A few z transforms
Sean Danaher
z
1
z
1
z
1
z
1
z
1
z
1
[A]
-K-
-K-
-K-
-K-
-K-
-K-
-K-
[A]
Frequency response simply determined by running around the unit circle Corresponds to the Nyquist Frequency (fs/2)
0 0.2 0.4 0.6 0.8 10
0.5
1
1.5
2
2.5
3
frequency (Nyquist=1)
|H(
)|
0.5( )
0.5[ ] [ ] 0.5 [ 1]
0.5 [ 1]
zh z
zy n x n x n
y n
-1
0
1
12
0 0.2 0.4 0.6 0.8 10
10
20
30
40
50
60
frequency (Nyquist=1)
H(
) (d
egre
es)
1
2
1 2
| ( ) |
( )
H z
H z
Simple Transfer Function
Sean Danaher
z
1
z
1
z
1
z
1
z
1
z
1
[A]
-K-
-K-
-K-
-K-
-K-
-K-
-K-
[A]Notch Filter
5 5
5 5
( )( )( )
( 0.95 )( 0.95 )
y[n]=x[n]-1.6180x[n-1]+x[n-2]
+1.5371y[n-1]-0.9025y[n-2]
i i
i i
z e z eh z
z e z e
-1 0 1
-1
0
1
/5
0 100 200 300 400 5000
0.2
0.4
0.6
0.8
1
1.2
1.4
frequency
|H(
)|
Fs=1000Hz
0 0.05 0.1 0.15 0.2 0.25-1.5
-1
-0.5
0
0.5
1
1.5
Time
Signal buried in 100Hz
Recovered signal
Sean Danaher
z
1
z
1
z
1
z
1
z
1
z
1
[A]
-K-
-K-
-K-
-K-
-K-
-K-
-K-
[A]Using the Fourier Transform is seldom the best way to get a spectrum. Normally methods based around• Autocorrelation (AC)•Linear Prediction are used
0 0.2 0.4 0.6 0.8 10
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
Frequency (relative)
Am
pli
tud
e(R
elat
ive)
fftaclpc
-150 -100 -50 0 50 100 150-0.5
0
0.5
1
Lag
Am
pli
tud
e (r
elat
ive)
Autocorrelation
Linear Prediction0 1 21 2[ ] [ ] [ ].... [ ]a y n a y n a y n e n
0 100 200 300 400 500-4
-3
-2
-1
0
1
2
3
4
Sample number
Am
pli
tud
e
First 500 of 2048 data points
AC methodUse FT of ACTo get PSD
All pole filter driven by white noise. Need to choose the order with care. But can now reproduce the spectrum
Spectral Analysis
Sean Danaher
z
1
z
1
z
1
z
1
z
1
z
1
[A]
-K-
-K-
-K-
-K-
-K-
-K-
-K-
[A]
Spectral estimation example
64k
933
2.4k
gain vocal tractfilter
pitch
noise
voiced/unvoicedswitch
synthesizedspeech
Frequency (Hz)0 1000 2000 3000 40000
1
2
3
4
5
6
7
8spectrum of frame 48
|h(
)|
0 10 20 30 40 50-1
-0.5
0
0.5
1
1.5
2
Sample [n]h
(t)
Impulse Response
•8 bit 8kHz
•Split speech into frame 240 samples long
•Use LPC10 to estimate spectrum
•Reconstruct
Sean Danaher
z1
z1
z1
z1
z1
z1
[A]
-K-
-K-
-K-
-K-
-K-
-K-
-K-
[A]
u 1
u 2
u 3
u 4
y 2
y 1
y 3
MIM
O
SYS
TEM
+-
+-
+-
+-
Conclusions
• SP Techniques Useful• Thanks for listening• Questions