Application of Independent Component Analysis (ICA) to
Download
Report
Transcript Application of Independent Component Analysis (ICA) to
Application of Independent Component
Analysis (ICA) to Beam Diagnosis
Xiaobiao Huang
Indiana University / Fermilab
5th MAP meeting at IU, Bloomington
3/18/2004
1
7/17/2015
Content
Review of MIA*
Principles of ICA
Comparisons (ICA vs. PCA**)
Brief Summary of Booster Results
*Model Independent Analysis (MIA), See J. Irvin, Chun-xi Wang, et al
**MIA is a Principal Component Analysis (PCA) method.
2
7/17/2015
Review of MIA
Each raw is
made zero
mean
1. Organize BPM
turn-by-turn data
2. Perform SVD
3. Identify modes
spatial pattern, m×1 vector
temporal pattern, 1×T vector
3
7/17/2015
Review of MIA
Features
1. The two leading modes are betatron modes
2. Noise reduction
3. Degree of freedom analysis to locate locale modes (e.g. bad BPM)
4. And more …
Comments: MIA is a Principal Component Analysis (PCA) method
4
7/17/2015
A Model of Turn-by-turn Data
BPM turn-by-turn data is considered as a linear*
mixture of source signals**
(1) Global sources
Betatron motion, synchrotron motion, higher order
resonance, coupling, etc.
(2) Local sources
Malfunctioning BPM.
Note: *Assume linear transfer function of BPM system.
** This is also the underlying model of MIA
5
7/17/2015
A Model of Turn-by-turn Data
Source signals are assumed to be independent,
meaning
where p{} is joint probability density function (pdf) and pi {si}
represents marginal pdf of si. This property is called statistical
independence.
Independence is a stronger condition than uncorrelatedness.
Independence
E{g ( x) f ( y)} E{g ( x)}E{ f ( y)}
Uncorrelatedness
E{xy} E{x}E{ y}
The source signals can be identified from measurements
under some assumptions with Independent Component
Analysis (ICA).
6
7/17/2015
An Introduction to ICA*
Three routes toward source signal separation, each
makes a certain assumption of source signals.
1. Non-gaussian: source signals are assumed to have non-gaussian
distribution.
Gaussian pdf
2. Non-stationary: source signals have slowly changing power spectra
3. Time correlated: source signals have distinct power spectra.
This is the one we are
going to explore
* Often also referred as Blind Source Separation (BSS).
7
7/17/2015
ICA with Second-order Statistics*
The model
with
Measured signals
Random noises
Source signals
Mixing matrix
Note:*See A. Belouchrani, et al, for Second Order Blind Identification (SOBI)
8
7/17/2015
ICA with Second-order Statistics
Assumptions
(1)
• Source signals are temporally correlated.
• No overlapping between power spectra of source signals.
As a convention, source signals are normalized, so
(2)
Noises are temporally white and spatially decorrelated.
9
7/17/2015
ICA with Second-order Statistics
Covariance matrix
So the mixing matrix A is the diagonalizer of the sample
covariance matrix Cx.
Although theoretically mixing matrix A can be found as an
approximate joint diagonalizer of Cx() with a selected set of , to
facilitate the joint diagonalization algorithm and for noise reduction, a
two-phase approach is taken.
10
7/17/2015
ICA with Second-order Statistics
Algorithm
D1,D2 are
diagonal
1. Data whitening
D1
C x (0) [U1 ,U 2 ]
1
[U1 ,U 2 ]T
D2
Set to remove
noise
with
0 max(D2 ) min(D1 )
Benefits of whitening:
1. Reduction of dimension
2. Noise reduction
2. Joint approximate diagonalization
3. Only rotation (unitary W) is
T
Cz ( ) W Cs ( )W for { i | i 1,2,, k} needed to diagonalize.
z D1 2U1T x Vx
E{zz T } In×n
The mixing matrix A and source signals s
s WVx
11
1
2
1
A (U D )W T
T
1
7/17/2015
Linear Optics Functions Measurements
The spatial and temporal pattern can be used to
measure beta function (), phase advance () and
dispersion (Dx)
1. Betatron function and phase advance
x Ab1s1 Ab 2 s2
Ab1
Ab 2
a( Ab21 Ab22 ) tan 1
2. Dispersion
x Al sl
Dx bAl
12
sl
b
7/17/2015
Betatron motion is decomposed to a
sine-like signal and a cosine-like signal
a, b are constants to be
determined
Orbit shift due to synchrotron oscillation
coupled through dispersion
Comparison between PCA and ICA
• Both take a global view of the BPM data and aim at re-interpreting
the data with a linear transform.
• Both assume no knowledge of the transform matrix in advance.
• Both find un-correlated components.
1. However, the two methods have different criterion in defining the
goal of the linear transform.
For PCA: to express most variance of data in least possible
orthogonal components. (de-correlation + ordering)
For ICA: to find components with least mutual information.
(Independence)
2. ICA makes use of more information of data than just the covariance
matrix (here it uses the time-lagged covariance matrix).
13
7/17/2015
Comparison between PCA and ICA
So,
ICA modes are more likely of single physical origin, while PCA modes
(especially higher modes) could be mixtures.
ICA has extra benefits (potentially) while retaining that of PCA method :
1. More robust betatron motion measurements. (Less sensitive to
disturbing signals)
2. Facilitate study of other modes (synchrotron mode, higher order
resonance, etc.)
14
7/17/2015
A case study: PCA vs. ICA
Data taken with Fermilab Booster
DC mode, starting turn index 4235, length 1000 turns.
Horizontal and vertical data were put in the same data matrix (x,
z)^T. Similar results if only x or z are considered.
Only temporal pattern and its FFT spectrum are shown.
Only first 4 modes are compared due to limit of space.
The example supports the statement made in the previous slide.
15
7/17/2015
A case study: PCA vs. ICA
16
ICA Mode 1,4
7/17/2015
A case study: PCA vs. ICA
17
ICA Mode 2,3
7/17/2015
A case study: PCA vs. ICA
18
PCA Mode 1,4
7/17/2015
A case study: PCA vs. ICA
19
PCA Mode 2,3
7/17/2015
A case study: PCA vs. ICA
20
ICA Mode 8, 14
7/17/2015
A case study: PCA vs. ICA
21
PCA Mode 8, 14
7/17/2015
Another Case Study with APS data*
22
ICA Mode 1,3
7/17/2015
*Data supplied by Weiming Guo
Another Case Study with APS data*
23
PCA Mode 1,3
7/17/2015
*Data supplied by Weiming Guo
Booster Results (, )
(b)
(a)
(1915,1000)*, MODE 1: (a)
Spatial pattern; (b) temporal
pattern; (c) FFT spectrum of (b)
24
*(Starting turn index, number of turns)
7/17/2015
(c)
Booster Results (, )
(b)
(a)
(c)
(1915,1000)*, MODE 2: (a) Spatial
pattern; (b) temporal pattern;
(c) FFT spectrum of (b)
25
7/17/2015
Booster Results (, )
120
60
MAD
Measured
115
50
110
40
x
x
105
100
30
95
20
90
10
0
0
85
10
20
30
40
BPM index
(a) σ=7%
50
80
0
MAD
Measured
5
10
15
Period index
(b)
20
25
σ =3 deg
Comparison of (, ) between MAD model and measurements. (a)
Measured with error bars. (b) phase advance in a period (S-S).
26
Note: Horizontal beam size is about 20-30 mm at large ; Betatron
amplitude was about 0.6mm; BPM resolution 0.08mm.
7/17/2015
Booster Results (Dx)
(b)
(a)
1000 turns from turn index 1.
(a)Temporal pattern. (b) Spatial pattern.
(t=0)= -0.3×10-3
27
7/17/2015
Booster Results (Dx)
5.5
5
MAD
Measured
4.5
4
D
x
3.5
3
2.5
2
1.5
1
0.5
0
10
20
30
BPM index
40
50
(a) σD=0.11 m
Comparison of dispersion between MAD model and measurements.
28
7/17/2015
Summary
ICA provides a new perspective and technique for
BPM turn-by-turn data analysis.
ICA could be more useful to study coupling and
higher order modes than PCA method.
More work is needed to:
1. Explore new algorithms.
2. Refine the algorithms to suit BPM data.
3. More rigorous understanding of ICA and PCA.
29
7/17/2015