SEGMENTATION WITH PREDICTIVE ERROR AND RECURSIVE LIKELIHOOD
RATIO DEVIATION METHODS IN MATERIAL CHARACTERIZATION
S. Femmam and N. K. M'Sirdi
L.R.P.- U.V.S.Q. 10-12 Avenue de l'Europe V¶elizy 78140 France
e-mail : ffemmam, msirdig@robot.uvsq.fr
ABSTRACT
In this paper, abrupt changes detection techniques are
used for segmentation of ultrasound signals in order to
characterize materials. We describe the main motivations for the investigation of change detection problems.
We show that the use of the Recursive Likelihood Ratio Deviation (RLRD) algorithm provides better results
than the Linear Prediction Error (LPE) algorithm. The
choice of the ultrasound signal length problem is used in
the spectral ratio technique for determining the quality
factor of a material.
Keywords: Spectral ratio, quality factor, LPE, RLRD,
segmentation, faults diagnosis.
1
INTRODUCTION
Many practical problems arising in quality control and
non-destructive evaluation, can be processed parametric
models in which the parameters are subject to abrupt
changes at unknown time instants. Because a large part
of the information lies in the nonstationarities, and because most of the adaptive estimation algorithms basically can follow slow changes[2]. In this paper we implement and compare two di®erent algorithms (LPE and
RLRD) for automatic segmentation of ultrasound signal. The mostly applied technique for determining the
quality factor is the spectral ratio technique. The latter uses the entire ultrasound signals. But the main
information is localized in a piece of the signal, so we
need segmentation. After segmentation we apply the SR
method to the desired segments, the results are better
comparing the case when we use all the signal. Figure
1. a, b and c show respectively the ultrasound signal for
the aluminum, the granite and the limestone. In ¯gure
1. a [T 1; T 2] represents the main interval of information
and [T 2; T 3] is the reverberation signal.
2
PROBLEM FORMULATION
The spectral ratio method deals with the analysis of
the signals energy distribution. The method consists
of determining the spectral ratio magnitude of a reference signal and the signal of the material whose quality
factor is unknown. Afterward we calculate the slope
(¯gure 3. a and b) of the segment of the straight line
at the frequency of the excitation signal. The viscoelas2
2
tic model used in [1] is given by: r @@tU2 = M (!) @@xU2
where M(!) is the complex magnitude. A general solution for a °at wave is: U = U0 exp(¡®x) exp(i!(t ¡ xc ))
where c is a real constant, exp(¡®x) gives the wave attenuation, and ® is the attenuation factor. The quality
¼
factor Q is given by the relation: ® = QV
f: where f
and V are respectively the frequency and the velocity
of the ultrasound wave. The equation that gives the
spectral amplitude of a °at wave is given by: U(f) =
G(f; x) exp(¡®(f )x exp(i(2¼f t ¡ kx)) where G(f; x) is
the function which takes into account the geometrical
shape of the sample and the reference, x, is the length
of the material to be characterized, ®(f ), is the attenuation coe±cient. The spectral ratio can be written as:
·
¸
U(t)
1
1
¡
Log(
) = ¡¼
xf + cte
(1)
Ur (t)
QV
Qr Vr
The Aluminum is well known and considered as a reference material. Its quality factor Qr is very large ( Qr1Vr
¼
can be neglected), and (1) gives the slope a = ¡ QV
x
which allows us to determine Q.
2.1
Classical spectral analysis and Parametric
spectral estimation
By using the classical spectral estimation technique
(Fourier) the spectral ratio method does not give a good
spectrum and hence the determined slope is not smooth
so resulting on wrong Q's . The resulting curve is shown
in ¯gure 3 [3]. For this estimation the signal is supposed
to be stationary. However if the characteristics of the
process vary with time, the results are not able to follow
these variations. Due to these problems the parametric
models to characterize these processes have been chosen
to be used.
We have shown [3] that the adaptive parametric spectral estimation methods, for instance the Least Mean
Square, the Recursive Least Square, the Burg algorithm,
and the Marple algorithm in AR modeling provide a
very smooth spectrums and hence the spectral ratio.
This implies a constant slope estimation over the resonant frequency range and an e±cient quality factor.
In ¯gure 3 the parametric spectral ratio estimation is
shown.
2.2 Segmentation
An other practical motivation for change detection. In
signal processing the segmentation of signals refers to
the automatic decomposition of a given signal into segments, the length of which is adapted to the local characteristics of the analyzed signal [2]. The homogeneity of
a segment can be in terms of the mean level or in terms
of the spectral characteristics. Due to the problem of
the estimation of the desired portion (portion containing the largest energy) of the ultrasound signal used in
the estimation of the Q's. For this task we have developed two algorithms which are the LPE and the RLRD
for automatic segmentation of the ultrasound P waves.
These results are validated by use of time-frequency and
time scale (wavelet) energy distribution, for more information see [4]. The use of time-frequency and wavelets
energy distribution can be a tool for faults discrimination in nondestructive evaluation, for more details refer
to [8].
3
CHANGE DETECTION ALGORITHMS
3.1 Linear Predictive Error algorithm
The objective of this algorithm is to detect parameters
changes of an AR model driven by a gaussian white
noise. The method consists of supervising the innovations (LPE) calculated from an AR model. The parameters are estimated by the algorithm of Morf [2]. The
model is constructed over a piece of the signal, after we
compare the LPE of variance ¾
b2 to a threshold value
¾ , ¸ is real. A change exists when "n exceeds S" ;
S" = ¸b
Nd times. Where Nd is called the alarm number. After
detection we update the model and in this case the new
window must be delayed by few samples after the last
detection. A discrete signal can be represented by (2):
p
X
xn =
ak xn¡k + En
(2)
k=1
and where En is a white noise of zero mean and a varip
P
ance ¾2 : Let x
bn =
b
ak xn¡k is the estimated of xn
k=1
constructed from the last p samples of the signal xn¡p ;
xn¡p+1; ... xn¡1 ; and the EPL can be written as in (3):
"(n) = xn ¡ x
bn = xn ¡
p
X
k=1
b
ak xn¡k
(3)
to En with a varithe estimation¡ is good ¢when£ "n tends
¤
ance ¾
b2 ' ¾2 ¾
b2 < ¾x2 :E "2 (n) , ¾"2 : . We introduce
a pointer variable as stated in (4).
P (n) = 1, f or j"(n)j  t and P (n) = 0, f or j"(n)j Á t
(4)
where t is a threshold value which is chosen as follows:
¯rst we de¯ne El , El = (P (n ¡ l + 1); :::; P (n)) where l
is an innovation sliding window.¤ Under the hypothesis
H0 (the probability of a ¯xed alarm PF A = ®), the
"threshold" value is given by:
½
µ ¶¾l
t
P [j"(n ¡ l + 1)j  t; :::; j"(n)j  t=H0 ] = ERF
¾"
(5)
¤ Under the hypothesis H1 , the probability of detection
PD is maximized by the parameter l.
3.2
Recursive Likelihood Ratio Deviation algorithm
3.2.1 Hypothesis test and Con¯dence interval
The hypothesis test can be stated by a con¯dence interb and for a value of ® which
val de¯ned for an estimator Á
its signi¯cant level is taken by its probability density
function[6]. We can consider:
* An accepted hypothesis region given by (H0 ):
Á®
Pr ob[Á1¡ ®2 < Áb < Á ®2 ] =
Z2
Á1¡ ®
2
b
p(Á)dÁ
=1¡®
(6)
* A critical region de¯ned by (H1 ) :
Pr ob[Áb > Á ®2 ] =
+1
Z
®
b
p(Á)dÁ
=
2
(7)
Á®
2
Â2 law, variance estimation and Likelihood ratio
For a gaussian random variables xi we can de¯ned a
nb
¾2
variance ¾2x = Â2n (xi ²N (0; 1); n = N ¡ 1)) as a new
x
variable Â2 with n degrees of freedom (number of lags
in processing) [5]. For an (1 ¡ ®)% required con¯dence
level. The estimated variance is compared to the accep2
¾x
nb
¾2
tance region Ânb
· ¾x2 (t) < Â2 x ® ; if this equation
2
®
n;
2
n;1¡
2
is satis¯ed The information signal is chosen to be the
innovation sequence "(t) then, there is no change (the
innovation is white x(t) = "(t)); else, there is a change
(the innovation is not white).
As we have stated before the likelihood ratio method
is based on the statistical test of the probability density function P (xi (k)) of the gaussian random variables
xi (k). The PDF is maximum when xi (k) is a value of
change. The PDF of x(k) under H0 and H1 is:
L[x(k)] =
k p(x(t)=x(t ¡ 1); H1 )
p [x(k)=H1 ]
= ¦
(8)
p [x(k)=H0 ] t=k0 p(x(t)=x(t ¡ 1); H0 )
The likelihood ratio can be written in recursive form (9):
(x(k)) +
LnL[x(k)] = LLkk¡1
0
[p(x(t)=x(t ¡ 1); H1 )]
(9)
[p(x(t)=x(t ¡ 1); H0 )]
Dk = Dk¡1 + dlk ? »
(10)
The decision test becomes the Likelihood Ratio Deviation dlk ? ±:
3.2.2
Detection of abrupt changes of variance by
RLRD
The RLRD is an e®ective algorithm for abrupt changes
detection. Consider an independent random variables
x(k) with mean m0 and variance ¾02 for the hypothesis H0 and with mean m1 and variance ¾12 for the hypothesis H1 : We assume that it follows the normal distribution. µIts distribution¶ is given by: p(x(t)=Hi ) =
´2
³
1 x(t)¡m
1
p
exp
¡
, i=0,1, and ¾
b12 are
b02 and ¾
2
¾
bi
¾
b 2¼
i
where a is a reel constant. If we derive (11) with respect
to µb and assuming that x(t) and ¾x (t) are independent
T
b
of µ(t):
So, F (t) = E '(t)'(t)
: F (t) gives us information
2 (t)
¾x
about the parameters in its direction. For an e®ective
information we estimate F (t) on a sliding window with
tc
P
'(t)'(t)T
size nf ; F (tc ¡ nf + 1; tc ) =
in recur¾
b 2 (t)
sive form,
F (tc ) = F (tc ¡ nf + 1; tc ¡ 1) +
To avoid the
cision test is dl(k) =
complexity of the threshold choice we rearrange dl(k):
2
¡th
(x(k)¡m)2
where A¡1 = ¾b 1th
is a constant.
dl(k) = ¡ A
2
¾
b 12
A is the minimum and relative uncertainty of the estimated variance to be accepted. So we can consider
now the variable dl(k) as a Â2 variable with one degree
of freedom, normalized by the constant A of a mean m
and a variance ¾
b12 and the test becomes there a change
if the conditions jdl(k)j > Â21; ® and jdl(k)j < Â21;1¡ ®
2
2
are satis¯ed, and no change else where. Or we can say
2
¢b
¾
that there is a change of variance if ¾b 21 > A% and no
0
change else where.
3.2.3 Change compensation
Compensation by Fisher information matrix
The use of compensation technique provides a good estimation and allows an e®ective tracking performances of
parameters estimation. We use the compensation technique by computing the compensation term at the instant of change -0 (tc ) as a function of the Fisher information matrix F (t): From Cramer Rao inequality
e µ(t)
e T ) > F ¡1 where µ is the paramewe have E(µ(t)
ters vector and F is given by the expected value of the
probability function logarithm deviation with respect
2
b If we conto the parameters. F = ¡E @@µb2 Lnp(x=µ):
sider the signal x(t) is normally distributed
the PFD
µ
³
´2 ¶
x(t)¡m
1
1
x (t)
b
is : p(x(t)=µ(t)))
= ¾ (t)p2¼ exp ¡ 2
¾x (t)
x
If we assume that between two instants of changes
b
So,
that mx (t) may be replaced by the x
b(t=µ(t)).
Tb
b
mx (t) = x
b(t=µ(t)) = '(t) µ(t) substituting mx (t) in
b
p(x(t)=µ(t)))
and rearranging, we can write that:
b
Lnp(x(t)=µ(t))
= a ¡ Ln¾x (t) ¡
pi (tc ) =
b 2
(x(t) ¡ '(t)T µ(t))
2
¾x (t)
(11)
'(t)'(t)T
¾
bx2 (t)
fi F (tc ; tc ¡ nf + 1)fiT
p
P
fi F (tc ; tc ¡ nf + 1)fiT
(12)
(13)
i=1
1
2
.
¡ 12 ¾bth2 (x(k)¡m)
¾
b 12 ¡th
1
x
The change probability and the compensation term are
given by (13) and (14)
respectively the estimates of ¾02 and ¾12 : The minimum deviation of the variance is ¾
¾02 > th > 0
b12 -b
where th is the threshold value for the minimum variance deviation to be accepted. Dk becomes Dk =
2
and the deLLk1 (x(k)) = LL1k¡1 (x(k)) ¡ 12 ¾bth2 (x(k)¡m)
¾
b 2 ¡th
1
i=tc ¡nf +1
2
-0 (tc ) = s0 s (tc )
p
X
pi (tc )fi fiT
(14)
i=1
and ¯nally the compensated gain becomes:
P ((t = tc ) + 1) = P (t = tc ) + -0 (tc )
(15)
where fi is the change direction vector, s0 is a constant
for the compensation level, s2 (tc ) this function is chosen
as the information signal covariance and p is the number
of parameters to be estimated[7].
4
RESULTS
Fig. 1. a, b and c show respectively the ultrasound
signal for the aluminum, the granite and the limestone
with the instant of segmentation by use of RLRD algorithm. In ¯g.1. d the limestone with its instant of
detection by LPE algorithm is represented. Fig. 2.
a, b, and c show respectively the piece of the segment
containing the main information of the aluminum, the
granite and the limestone signals with segmentation by
LPE algorithm. We can remark that the delay of the
instant of detection is considerable, this means a loss
of information at the beginning of the signals. This is
due to the non implementation of the failure compensation technique in the LPE algorithm. Fig. 2. d, e and
f show respectively the desired segments by use of the
RLRD algorithm of the aluminum, the granite and the
limestone signals. The use of RLRD algorithm is much
better and more sensitive to the instant of detection, to
make clear the failure compensation technique the implementation of the RLRD algorithm without the failure
compensation technique is shown by ¯g. 2. f. Fig. 3. a
and b show respectively the estimated and the original
(Fourier) SR of the desired segments by LPE algorithm
of the aluminum (reference signal) over the granite signal and the aluminum over the limestone signal. In ¯g.
3. c and d we represent respectively the estimated and
the original SR of the desired segments by RLRD algorithm of the aluminum over the granite signal and the
aluminum over the limestone signal. The estimated SR
give a very smooth negative slope compared to the original SR where nonparametric method is used. this slope
is used for characterizing the materials.
(a)
15
10
10
(b)
Signals + Abrupt changes
2
Esti. & Orig. Spectral Ratio
Original
5
5
0
Estimated
0
Ampl dB
Estimated
-5
Ampl dB
-5
-10
Original
-10
-15
-15
-20
-25
0
0.5
(c)
(a)
(b)
Esti. & Orig. Spectral Ratio
15
1
1.5
frequency (Hz)
2
2.5
x 10
-20
0
0.5
1
1.5
frequency (Hz)
6
(d)
Esti. & Orig. Spectral Ratio
15
2
2.5
x 10
6
Esti. & Orig. Spectral Ratio
15
Signals + Abrupt changes
1
10
Original
10
5
1.5
1
0.5
Ampl mvolt
Ampl mvolt
Estimated
0
5
Ampl dB
-5
Ampl dB
0
Original
-10
0.5
Estimated
-15
-5
0
-20
0
T1
T2
-1
0
200
400
(c)
1
-10
T3
-0.5
600
Samples
800
1000
1200
-0.5
0
200
400
(d)
Signals + Abrupt changes
1
600
Samples
800
1000
1200
0
0.5
1
1.5
frequency (Hz)
2
2.5
3
x 10
6
-25
0
0.5
1
1.5
frequency (Hz)
2
2.5
3
x 10
6
Figure 3. Estimated & original SR of segments
Signals + Abrupt changes
References
0.5
0.5
[1] A. Kavetsky et al, A model of acoustic pulse propagation
Ampl mvolt
Ampl mvolt
0
-0.5
0
0
200
400
600
Samples
800
1000
1200
-0.5
0
200
400
600
Samples
800
1000
1200
Figure 1. Signals with detection of abrupt changes
(a)
2
Desired segment (LPE)
(d)
2
1.5
1
Ampl mvolt
Ampl mvolt
0.5
0.5
0
[3] S. Femmam, N.K. M'Sirdi, S. Guillaume, AR process-
0
-0.5
-0.5
-1
0
50
100
150
Samples
200
250
300
-1
350
(b)Desired segment (LPE)
0.8
0
50
100
150
Samples
200
250
300
200
250
300
(e)Desired segment (RLRD)
0.8
0.6
0.6
0.4
0.4
0.2
0.2
Ampl mvolt
Ampl mvolt
0
0
-0.2
-0.2
-0.4
-0.4
-0.6
0
50
(c)
0.7
100
150
Samples
200
250
-0.6
300
0
50
Desired segment (LPE)
100
150
Samples
0.6
0.5
0.4
0.4
0.3
Ampl mvolt
0.2
0.2
Ampl mvolt
0.1
0
0
-0.1
-0.2
-0.2
-0.3
ing of ultrasonic signals and its application to determine
characteristics for a given material, Submitted 27 07
96 at Material Evaluation journal, 1711 Arlingate Lane,
POB 28518, OH, USA.
[4] S. Femmam, N.K. M'sirdi, Application of change de-
(f)Desired segment (RLRD)
0.8
0.6
0
50
100
150
200
Samples
250
300
350
400
-0.4
0
50
100
150
Samples
200
250
300
Figure 2. Desired segments used in SR method
5
[2] M. Basseville, P. Flandrin, N. Martin, Signal segmentation, Traitement du signal, 9, Suppl¶ement au N 1, pp.,
115-147, 1992.
Desired segment (RLRD)
1.5
1
and its application to determine Q for rock mass. Int. J.
Rock. Mech. Min. Sci. & Geomech. Abstr. Vol. 27, N± 1,
pp. 33-41, 1990.
CONCLUSION
The detection, with diagnosis, of abrupt changes in ultrasound signal characteristics requires, a ¯ne study of
the signal of the changes, and consists of the study
of the system using a convenient de¯nition and tools,
which allow to build the adequate "residuals" for testing the presence of a precise type of change. This paper
describes the performances of the segmentation by use
of the nonstationarity detection based on the threshold
analysis of the LPE and the RLRD algorithms, for the
ultrasound signals. The latter are used for characterization (determination of the quality factor in nondestructive evaluation), perception, and diagnosis of materials. Furthermore the RLRD method gives better results
compared to the LPE one, since the delay is less for the
same applied ultrasound signal and hence a precise segmentation. The RLRD is very sensitive either for the
positive or negative jumps because of the estimation of
the jump characteristics (amplitude and position), the
detection depends essentially on the instantaneous value
of the signal at the instant jump.
tection technique to ultrasound signal processing. Use
of spectral parametric estimation, time-frequency and
wavelets, 30th IEEE SSST, March 8-10, 1998, West Virginia Univ. Morgantown, USA.
[5] R. Ramdani, N.K. M'sirdi, Change detection statistical
tests in non stationary signal, European Journal of Diagnosis and Safety in Automation, Volume 4, N± 2, pp.
229-271, 1994.
[6] H. Tjokronegoro, Estimation adaptative des paramµetres
de systemes variant dans le temps et detection de ruptures, Thµese de doctorat de INPG 90.
[7] S. Femmam and N.K. M'sirdi, Application of signal segmentation and parametric estimation to determine material characteristics. IEEE, CESA98, IMACS Hammamet,
Tunisia, April 1-4 1998 .
[8] Femmam, N.K. M'sirdi, Comparison of the performances
of spectral parametric estimation, time frequency, time
scale, and segmentation with detection of abrupt changes
methods in non-destructive evaluation, 3rd IFAC workshop On Line fault detection and supervision in the
chemical process industries. 15-16 June 1998, Lyon,
France (Accepted).