} where (
1
,
2
, ,
)
denote a collection of zero-mean random variables, the second,
third and fourth order cumulants are given, respectively, by
283
[12]:
(
1
,
2
) = E{
1
2
} (1)
(
1
,
2
,
3
) = E{
1
3
} (2)
(
1
,
2
,
3
,
4
) = E{
1
4
}
E{
1
2
}E{
3
4
}
E{
1
3
}E{
2
4
}
E{
1
4
}E{
2
3
}. (3)
The set of equations (1) to (3) can be rewritten as, [11]
2,
() = E{()( + )} (4)
3,
(
1
,
2
) = E{()( +
1
)( +
2
)} (5)
4,
(
1
,
2
,
3
) = E{()( +
1
)( +
2
)( +
3
)}
2,
(
1
)
2,
(
2
3
)
2,
(
2
)
2,
(
3
1
)
2,
(
3
)
2,
(
1
2
). (6)
where the second cumulant denote the autocorrelation of ().
For a random process symmetrically distributed, the third
cumulant is null; in these cases the fourth cumulant is used
[12].
Let {[]} be a random discrete signal where {[]} = 0.
The equations (4) to (6) can be expressed as:
2,
[] = {[][ + ]}, (7)
3,
[] = {[]
2
[ + ]}, (8)
4,
[] = {[]
3
[ + ]} 3
2,
[]
2,
[0]. (9)
According to stochastics aproximations, for a nite periodic
vector with length , if = 0, 1, , 1 , then equations
(7) to (9) become [14]:
2,
[] =
1
=0
[][( + )/], (10)
3,
[] =
1
=0
[]
2
[( + )/], (11)
4,
[] =
1
=0
[]
3
[( + )/]
2
1
=0
[][( + )/]
1
=0
2
[].(12)
These last set of equations, (10) to (12), can be used to
extract features from voltage signals for power quality analysis
purpose.
III. LINEAR DISCRIMINANTS
Features extracted from the signals are used as the input of
a classication system instead of the signal waveform itself,
as this usually leads to a much smaller system input. Selecting
a proper set of features is thus an important step toward
successful classication. It is desirable that the selected set
of features may characterize and distinguish different classes,
that the selected features be uncorrelated and that the total
number of features be small [9].
For multivariate normal densities, (x
) (
), the
discriminant function can be written as [9], [10]:
(x) =
1
2
(x
(x
2
ln 2
1
2
ln (
), (13)
where x is the data vector, is the dimension of x, is
the mean vector,
)
2
(
2
+
2
)
, (14)
where denotes the mean values and
2
represents the
variances.
This methodology was used to select the features for the
detection and estimation steps.
2) Mahalanobis distance: An approach can be made when
all classes have the same arbitrary covariance matrix. Previ-
ously it was diagonal and identical for all classes. In this case,
the discriminant function can be written as
(x) =
1
2
(x
1
(x
) + ln (
). (15)
The expansion of (x
1
(x
) results in a sum
in which the quadratic term, x
1
x, is independent of .
Considering this, (15) becomes
(x) = (W
x +
), (16)
where W
=
1
and
=
1
2
+ln (
). Again,
is equivalent to the
Mahalanobis distance:
1
(
) <
1
(
), = .
(17)
This method was used to dene the detection threshold.
IV. PROPOSED METHOD
The proposed algorithm is focused on the detection of sub-
harmonics disturbance and in the estimation of its frequency
and magnitudes. In order to achieve these goals a block
structure, shown in Figure 1, was dened.
In this structure the rst step is the segmentation of the
voltage input signal in frames. Each frame consists of one
signal cycle with 256 samples. HOS are used to extract
features from these signal frames and the most relevant is
selected. This feature is then compared to a threshold in order
284
Extraction and Selection
Detection
Input
Extraction and Selection
Decision
Notch filter
Estimation
Subharmonic detected
No subharmonic
Output
Figure 1. Block diagram
to determine if a sub-harmonic is present in that cycle or not.
If no sub-harmonic is detected, the output receives no alarm.
But, if there is a sub-harmonic in the cycle, then an alarm
is send to the output and the samples are submeted to a 60
Hz IIR notch lter, in order to generate an error signal ().
The lter consists in a second order Chebyshev and have two
cycles of transient. New cumulants are calculated and selected,
now for the error signal. These cumulants are used to estimate
the magnitude and frequency of the sub-harmonic which are
also send to output.
The input signals have 60 Hz of fundamental frequency
and were corrupted by sub-harmonics with frequencies ranging
from 0 Hz to 55 Hz and amplitude ranging from 0 p.u. to 0.9
p.u., both sorted randonly using uniform distribution. They
have 10 cycles, with 256 samples per cycle.
Three input signals, with 10 dB, 30 dB and 50 dB signal
to noise ratios can be seen in Figure 2. They were disturbed
by a 29 Hz/0.48 p.u. sub-harmonic.
To perform features extraction, the signal is segmentated in
frames which consists of one cycle, i.e. 256 signal samples.
Equations (10), (11) and (12) presented and discussed in
section II, was used to calculate the signal features.
The set of cumulant features calculated by HOS generate
256 parameters per sample, for each cumulant order. In order
to reduce the amount of data, Fishers discriminant ratio,
discussed above, was used for feature selection. The most
relevant cumulant value was shown to be the rst lag.
1.16 1.18 1.2 1.22 1.24 1.26 1.28 1.3 1.32 1.34
2
0
2
(a)
time in seconds
m
a
g
n
i
t
u
d
e
1.16 1.18 1.2 1.22 1.24 1.26 1.28 1.3 1.32 1.34
2
0
2
(b)
time in seconds
m
a
g
n
i
t
u
d
e
1.16 1.18 1.2 1.22 1.24 1.26 1.28 1.3 1.32 1.34
2
0
2
(c)
time in seconds
m
a
g
n
i
t
u
d
e
Figure 2. Input signal with sub-harmonic disturbance which has frequency
of 55 Hz and 0.9 p.u. of magnitude. (a) SNR = 10dB(b) SNR = 30dB. (c)
SNR = 50 dB.
A. The detector
In order to perform the detection of spurious components
in the main signal, after the extraction and selection of the
input signal cumulants, these features are disposed in maps.
The feature space shows the discrimination capability of the
second and fourth order cumulants in Figure 3 , in which
disturbances class, i.e. signals with sub-harmonis components,
is represented by black and the class without disturbance
is displayed in red. Figure 3 also emphasizes the threshold
of separation between the classes, which was obtained by
Mahalanobis distance.
0.1 0.2 0.3 0.4 0.5 0.6 0.7
0.05
0.1
0.15
0.2
0.25
0.3
2nd order cumulant
4
t
h
o
r
d
e
r
c
u
m
u
l
a
n
t
without disturbance
with disturbance
Mahalanobis distance
threshold
0.49 0.5 0.51
0.115
0.12
0.125
0.13
0.135
Figure 3. Feature map of a signal corrupeted by sub-harmonic component and
with SNR=30dB. The black circles denotes the features due to disturbance
signals and the red stars represent the clean signals features. The blue
region depicted the detection threshould, which is dened by the Mahalanobis
distance.
285
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
magnitude
2
n
d
o
r
d
e
r
c
u
m
u
l
a
n
t
real values
estimated values
Figure 4. Sub-harmonic magnitude estimative by exponencial curve tting.
Real data (second order cumulant) is represented by the blue stars and its
estimative is presented in red dots.
B. Identication
To nd the parameters of a mathematical model that de-
scribes the variation of second and fourth order cumulants
with respect to magnitude and frequency of sub-harmonic
distortion, were performed an approach by curve tting.
The result is an exponential model that is capable to estimate
the magnitude and the frequency of a sub-harmonic component
by high order cumulants. The model that ts the present data
is decribed by:
() =
, and (18)
() =
, (19)
where denotes the cumulant of second or fourth order,