reliability evaluation of high voltage insulation using weibull distribution
TRANSCRIPT
Science Academy
Publisher
International Journal of Research and Reviews in Electrical and Computer Engineering (IJRRECE)
Vol. 1, No. 2, June 2011
ISSN: 2046-5149
Copyright © Science Academy Publisher, United Kingdom
www.sciacademypublisher.com
Reliability Evaluation of High Voltage Insulation Using Weibull
Distribution
Sunitha Anup, A S. Deshpande, A N. Cheeran, and H. A. Mangalvedekar
Electrical Department, VJTI, Mumbai, India
Email: [email protected] , [email protected] , [email protected] , [email protected]
Abstract – Partial Discharge (PD) is the phenomenon observed in High Voltage (HV) Insulation Systems. During the event
of a Partial Discharge, there are significant statistical variability in properties such as partial discharge pulse charge
amplitude, pulse shape and time of occurrence. Due to the random nature of pulse charge amplitude and occurrence time that
describe the process of Partial Discharge, these can be described in terms of probability distributions. PD from different
sources (i.e., corona discharges, as well as surface and voids discharges) has different PD pulse height distribution (PDHD)
patterns. Modeling of these patterns can be performed using Weibull distribution. It is shown that Partial Discharge Height
Distributions fit the two-parameter Weibull Function when Partial Discharges having the same origin occur, while deviations
from the Weibull plot are detected when various PD sources occur simultaneously. In the presence of two simultaneous PD
phenomena, the tested specimen is ideally divided into two elements, each affected by a single type of PD source. This
problem is approached by additive mixing of Weibull Distribution function. The different approaches used for parameter
estimation of Weibull Distribution are Least Square Regression and Maximum Likelihood Estimation (MLE). The empirical
cumulative probability density functions are estimated using Mean rank, Median rank, Bernard‟s rank and Hazard plotting.
Comparison of these methods in estimation of parameters and in model validation is shown in this paper. The data of pulse
charge amplitude is corresponding to the failure data of insulation of the system. The characteristics of reliability function
are plotted with respect to pulse charge amplitude. The pulse charge amplitude appears to be inversely proportional to the
reliability function. Therefore, it is desirable to have minimum pulse charge amplitude so as to achieve better reliability
characteristics.
Keywords – Partial Discharge Height Distribution, Weibull Distribution, Parameter Estimation, Reliability.
1. Introduction
Partial Discharge is an electrical discharge that occurs in
gaseous inclusions within electrical insulation or over
contaminated or damp sections of the surface of such
insulations. PD can cause solid insulation to age or
deteriorate [1]. PD signals are of different nature depending
on the type and source of the defect. Analysis of PD enables
identification of source and type, leading to the assessment of
deterioration rate and the reliability of high voltage insulation
systems. PD phenomena being stochastic in nature [2],
statistical inferences become inevitable. Since PD
measurement generates enormous amount of data, usage of
various analysis methods to organize data within manageable
proportions is imperative to derive the desired inferences on
PD source/type identification. PDHD is the three dimensional
pattern, with the three parameters being φ-q-n i.e. phase
angle-charge amplitude-pulse repetition rate. PDHD data
resulting from a single and multiple PD sources are often
characterized using the Weibull distribution [3]. In this paper,
the parameters of Weibull function have been estimated on
two sets of PDHD data obtained from measurement of
rotating machines. The results of PDHD modelling using
Weibull distribution for different approaches like Mean rank,
Median rank, Bernard‟s rank, Hazard Function and MLE is
compared. Also, the quality of fitting of the standard Weibull
model has been confirmed by the comparison
with probability density function (pdf) plots of each method.
The pdf is the function that describes the relative likelihood
for the random variable to occur at a given point and is given
by the integral of the variable‟s density over the region of
study[4]. Quality of fitting of two parameter Weibull function
can be also performed with the help of estimated parameters
using Kolmogorov-Smirnov (K-S) test [5]. In reliability
analysis, a probability distribution is the key element which
enables the appropriate data handling and manipulation.
Among the different distributions, Weibull distribution aptly
represents various physical phenomena like partial discharge
[6]. Analysis of PD enables identification of source and type,
leading to the assessment of deterioration rate and the
reliability of high voltage insulation systems. The reliability
of the system being a crucial factor in the engineering design,
the asset management tasks with respect to reliability and
maintainability has to be coordinated throughout its life
cycle. This coordination demands better understanding and
International Journal of Research and Reviews in Electrical and Computer Engineering (IJRRECE) 56
utilization of the partial discharge measurements of various
insulation specimens in the system. The reliability of a
product depends on technical decisions made during the
design and manufacturing of the product. Thus, Weibull
models assist in decision making process in the asset
management of insulation system design. Similarly, attention
is paid to the pulse charge height distribution of the sample.
As brought out above, since the pulse charge height appears
to be inversely proportional to the reliability function, it is
desirable to have minimum pulse charge height so as to
achieve better reliability characteristics.
The subject has been detailed in the succeeding sections
as follows. Section 2 explains PDHD modeling using
Weibull distribution with emphasis on the methodology for
standard Weibull model and mixed Weibull model.
Parameter estimation using graphical method and analytical
method are explained for the standard Weibull model and
mixed Weibull model. Section 3 gives the result obtained
when the proposed statistical procedure has been applied to
standard Weibull model and additive-mixed Weibull model.
Reliability characteristics are plotted with respect to the pulse
charge amplitude for both standard Weibull model and
additive-mixed Weibull model.
2. PDHD Modeling using Weibull Distribution
Assuming that different PD sources can produce different
PD-pulse height distributions, the problem of PD separation
and identification can be approached resorting to the
stochastic analysis of PDHD Shape [7]. Partial Discharges
taking place in the tested insulation system are detected at the
measuring impedance Zm and fed into signal conditioning
circuit having correct bandwidth and gain [3].At this stage
pulses are stretched (in shape and time) to match the
instrumentation requirements for the input signals. The
analog-to-digital converter(A/D-C) turns the peak magnitude
of PD electric signal (mV or mA) into discrete digital form
Si, where i ranges from 1 to the maximum number of
quantization levels or channels available on the measurement
system. Hence the interpretation of PD measurements can be
approached by means of the shape analysis of partial
discharge height distribution (PDHD). The PDHD is obtained
by plotting the ratio of ni to the total number of pulses NT as a
function of their amplitude Si. The results of the PD
measurements is stored as an array of integers [ni, Si] where
ni is the number of discharge signals in the channel i that
constitutes the histogram to be processed. Due to the high
degree of randomness in the Partial Discharge data, it can be
very well modeled by a suitable probability model such as a
distribution function so that the data can be viewed as
observed outcomes (values) of random variables from the
distribution. The Black-Box approach to modeling [9] is
performed with the help of data, which is preceded by
determination of type of mathematical formulation
appropriate to model the data. The statistics inference of the
black box modeling is executed in such a way that the
available data makes statements about the probability model,
either in terms of probability distribution itself or in terms of
its parameters or some other characteristics. Among the many
standard probability distribution functions, Weibull
distribution is significant to model the data sets of widely
differing characteristics. PDHD histograms, that is, number
of pulses, ni, as a function of their pulse charge amplitude, qi,
derived from 3D pattern, were processed according to two
different stochastic models, that is: the two-parameter
standard Weibull model and the five-parameter additive
mixed Weibull function. From the studies [8], it is clear that
the data fits a two-parameter Weibull function when partial
discharges having the same nature i.e. generated by the same
type of defect, occur while an additive mixed-Weibull
distribution holds when partial discharge phenomena of two
different natures are superimposed.
2.1. Methodology for Standard Weibull Models
In a standard Weibull probability distribution function is
defined as
(
) (1)
where β>0 and α>0 are the parameters referred to as shape
and scale parameters of the distribution. The strength of the
Weibull model lies in its flexible shape as a model for many
different kinds of data. The parameter β plays the role in
determining, how the Weibull will appear. The methodology
for the standard Weibull model can be explained as follows.
2.1.1. Collecting the Data
Figure 1 shows the components of PD Analyzer [10] are
Pulse Height detector, Phase Detector and Interface Unit to
PC. The Pulse Height Detector identifies the PD Pulse,
sample and digitize the peak of the pulse(height). The phase
detector gives the phase of detected pulse with the zero
crossing of the applied voltage as the reference. The interface
unit synchronizes the pulse height and the phase data transfer
to the PC where the data is stored as PD Record.
Figure 1. Schematic of PD Analyzer [10].
2.1.2. Analysis of Data
The analysis of data assists in determining whether a
particular model is appropriate or not to model a given data
set. The basic quantities of PD Analysis are the PD Pulse
Magnitude and Phase Data. In this paper, statistical approach
is used for the probabilistic model for PD Analysis and
distribution of charge magnitude and phase position is used
for representation of Analysis.
2.1.3. Model Selection
This step involves choosing an appropriate model
formulation i.e., distribution function. Due to the significant
variability of the data set (q1, q2….qn) it is viewed as random
International Journal of Research and Reviews in Electrical and Computer Engineering (IJRRECE) 57
variable (Q1,Q2,…..Qn), each Q is modelled by the
Probability Distribution function
F (q, θ) = P(Q ≤ q) (2)
where θ denotes the set of parameters for the distribution.
The reliability function of standard Weibull model is given
by
(
)
(3)
Weibull Distribution can be used to model failures due to
fatigue extremely well. The distribution function is given as
(
)
(4)
The parameters of the distribution are given by set θ = {α, β}
with α>0, β>0.
2.1.4. Parameter Estimation
The parameters in the Standard Weibull model i.e. 2-
parameter Weibull Distribution viz., shape (β) and scale (α)
parameters are estimated graphically and analytically. In this
paper, the graphical methods used are Least Square
Regression using the Rank Regression Techniques and
Hazard Plotting. The analytical method used is Maximum
likelihood estimation.
a) Graphical Methods
The data q (1) ≤q (2) ≤……… q (n) is reordered and the
empirical distribution F^q (i) is computed using the ranks. The
three different approaches are Mean rank, Median rank and
Bernard‟s rank. Hazard function is also a graphical method of
standard Weibull model parameter estimation. The ranking
method is exclusively associated with how to approximate
the probability associated with a particular observation. The
philosophy of rank methods is such that, it estimates the
probability of observing an event non-parametrically. In other
words, it is the probability of observing the ith
event out of N. This implies that, there is an associated probability without
any model assumption. However to interpolate and
sometimes extrapolate, it is essential to perform model
fitting. Using regression, this distribution is fitted to these
points. Weibull Probability Plot (WPP) verify the linearity of
a specific distribution. Figure 2 shows the linear regression of
cumulative distribution of WPP [11].
Figure 2: Linear regression of cumulative distribution in a WPP[11]
Thus, distribution line (cumulative distribution function)
is a straight line on a WPP. If the data follows the standard
Weibull distribution, then that data is linear on that
distribution‟s WPP.
i) Mean Rank Approach
A method of handling data which has the same observed
frequency occurring at two or more consecutive ranks; it
consists of assigning the average of the ranks as the rank for
the common frequency. The empirical distribution function is
given as
(5)
ii) Median Rank Approach
Median ranks are used to obtain an estimate of the
unreliability, F(q(i)), for each failure. The cumulative
binomial gives you the probability of occurrence of the ith
observation at a particular confidence level. The choice of the
confidence level for probability plotting is usually 50% CL.
The empirical distribution function is given as
(6)
iii) Bernard’s rank approach
For a sample size n, the ith
rank, as obtained from the
Bernard‟s Approximation is given as
(7)
iv) Hazard Plot:
Failure Rate/Hazard Function of Weibull distribution is given
as
(8)
Cumulative Hazard Function is
∫
(9)
The hazard function h (q) is estimated from the inverse of
the reverse rank of the ordered failures; and the cumulative
hazard function, H (q), is the cumulative of the values of h(q).
As seen in Figure 2, WPP is plotted for each empirical
distribution F^ q(i), and the best straight line fit using
regression is obtained. The slope of the line yields the shape
parameter β^, the estimate of β. From the y-intercept of the
fitted line, the scale parameter α^ the estimate of α is
obtained.
b) Analytical Methods
Maximum likelihood estimation (MLE) is a
popular analytical method used for fitting a statistical
model to data, and providing estimates for the model's
parameters. Maximum likelihood estimates depend on the
type of data available. Maximum likelihood estimation seeks
the parameter values that are most likely to have produced
the observed distribution The likelihood function is given by
α β ∏ β
β
αβ
β
αβ ] (10)
The maximum-likelihood estimates are obtained by
solving the equations resulting from setting the two partial
derivatives of L(α, β) to zero. The maximum likelihood
estimate of {α, β}(the unknown parameter in the model) is
International Journal of Research and Reviews in Electrical and Computer Engineering (IJRRECE) 58
that value that maximizes the log-likelihood, given the data.
The bilateral confidence limits [5] at probability 95% of the
mean values of the parameters were estimated according to
the expression
√ (11)
where M is the arithmetic mean of the parameter values
obtained from a series of N observations, s is the estimation
of the standard deviation of this series and t is the Student
variable (t=1.96).
2.1.5. Model Validation
Calculation of shape and scale parameters are proceeded
by control of the quality of fitting of the two-parameter
Weibull function .The good quality of fitting of the two
parameter Weibull function is confirmed [12] by the
probability density graphs which are relevant to the
respective data. The probability density function, f(x),
describes the behavior of the random variable „q‟ [6]. Figure
3 shows the bell shaped Probability Density curve [4]. The
area below the probability density function to the left of a
given value, q, is equal to the probability of the random
variable represented on the x-axis being less than the given
value q.
Figure 3. .Bell shaped Probability density curve following the data
histogram[4].
2.2. Methodology for mixed Weibull Models
Deviation from the standard two parameter Weibull
model is detected in the presence of simultaneous PD
phenomena, which superimposes each pattern with the
previous one [8,12]. This scenario prompts the investigation
of mixed distributions, based on the Weibull function, which
are able to fit the recorded PDH distribution. In the presence
of two simultaneous PD phenomena, it can be assumed that
the tested specimen is ideally divided into two elements, each
affected by different PD phenomena. The five parameter
Weibull function permits the separation of the two sub
distributions relevant to each PD source, so that each source
can be described by its characteristic statistics [13].
2.2.1. Model structure of mixed Weibull model
Each sub population of the mixed model are standard two
parameter Weibull distribution given by (2). The density
function is given by
g (q) = Σpi fi (q) (12)
where fi(q) is the density function associated with F(i)
When there are two subpopulations, the above mixture model
is a two-fold mixture model with cumulative distribution can
be given as
G(q)= pF1(q)+(1-p)F2(q) (13)
The model is characterized by five parameters –the shape
and scale parameters of two subpopulations and the mixing
parameter p(0<p<1).The density function of the two –fold
mixture model is given as
g (q) =pf1 (q) + (1-p) f2 (q) (14)
The reliability function of two-fold mixed Weibull
distribution is defined by
R(q) = pR1(q) + (1− p)R2 (q) (15)
2.3. Model Analysis
For small q (very close to zero) G (q) can be approximated as
G (q) ~cF1 (q) (16)
where c= Σ p (
)
β
For large q, G(q) can be approximated by
G(q) = 1- p1[1- F1(q)] (17)
Similarly the density function can be approximated as
g (q)~ cf1(q) for small q and by g (q)~ p1f1(q) for large q.
This implies that density function is increasing for small q if
β1 >1. For a mixed Weibull model, the WPP plot is smooth
curve in the x-y plane. Figure 4 gives the WPP plot of
mixture model [9].The shape of the curve is significantly
influenced by the two shape parameters of the
subpopulations.
Figure 4. WPP plot of Weibull mixture model (n=2,α1=2,α2=10,β1=
β2=2,p=0.4) [9]
2.3.1. Parameter Estimation
For the two-fold mixed Weibull distribution, there are
totally five unknown parameters ( p , α1, α2, β1, β2). Both
graphical and analytical approaches can been used for the
parameter estimation [15].
Graphical Method: Jensen Peterson method
In the graphical method, the estimates are obtained from
plotting the data; it is popular due to its simplicity and
visibility [14] .A shown in fig 5,the sample data is plotted on
the WPP. The cumulative distribution function is empirically
estimated using Bernard‟s rank. The plot follows a smooth
curve. The coordinates (x1, y1) with smallest slope is
obtained. The ordinate value corresponds to ln(-ln(1-F))
International Journal of Research and Reviews in Electrical and Computer Engineering (IJRRECE) 59
=y1.The value of F is equivalent to p, the mixing weight of
the sub population. Thereafter, α1 and α2 are obtained from
the abscissa which corresponds to (0.632 p) and (p+0.632p)
level horizontally. The slopes of the tangent lines at the two
ends of the cumulative distribution graph estimates the shape
parameters β1and β2. The approximation is valid only when
the scale parameters differ significantly, and the shape
parameter for subpopulation 1 is bigger than that for
subpopulation 2. These are the basic features of the graphical
method of parameter estimation of a 2-Weibull mixture,
when two subpopulations are well separated.
Figure 5. WPP plot of a 2-Weibull Mixture, estimating the Weibull
parameters graphically.[14]
2.3.2. Modeling the Data Set
The WPP plot serves as a good starting point to determine
if a Weibull mixture model is appropriate to model a given
data set. If a smooth fit to the plotted data has a shape similar
to that shown in Figure 4, one can tentatively assume that the
two-fold Weibull mixture model can be an appropriate
model.
3. Results and Discussion The proposed statistical procedure has been applied to the
shape analysis of PDHD obtained from measurements
performed on rotating machine‟s stator bars.
3.1. Two-Parameter Weibull Function
Using the method of standard Weibull modelling
mentioned in section 2.1.4, cumulative function F^ has been
estimated. This F^ has been used to get the following PDHD
plots (using ln(-ln(1-F)) versus. ln(q) for two types of
discharges i.e. surface discharge and delamination. Figure 6
shows this PDHD plot for different approaches viz. Mean
rank, Median rank, Bernard‟s rank, Hazard plotting and
MLE.
The linearity observed in these plots shows that the given
data fits fairly well to the two parameter approximation of
Weibull function which corresponds to single type of
discharge. This implies that empirical calculation of
cumulative distribution function using rank regression
approaches are matching with the analytical estimation using
MLE, though the MLE curve stands out from the rank
regression WPPs. Table 1 shows the comparison of these
approaches for the parameter estimation of the standard
model Weibull distribution. The scale parameter α follows
the unit of pulse charge amplitude, while the shape parameter
β is a dimensionless quantity. For the delamination discharge
in the sample, numerical value of β is a positive quantity that
lies between 1 and 2 (1<β<2).
Figure 6. PDHD plotted in Weibull plot for delamination at 3.78kV
Table 1. Standard model Weibull distribution parameters
of delamination discharge at 3.78 kV
Method α(μC) β
Mean Rank 37.8 1.269
Median Rank 37.7 1.296
Bernard‟s
Rank
37.7 1.284
Hazard Plot 37.68 1.288
MLE 35.7 1.635
Figure 7. Comparison of probability density functions of each mentioned
methods with the pulse rate Vs. charge curve.
As explained in section 2.1.5, the standard Weibull model
validation of the considered PDHD data is performed
graphically using the probability density function graphs.
Pulse Charge amplitude is the random variable and it is seen
that the pdf plot follows the pulse rate vs charge curve which
implies that the probability of occurrence of discharge (pdf)
is a bell shaped curve with a single mode and it has the same
International Journal of Research and Reviews in Electrical and Computer Engineering (IJRRECE) 60
behaviour of the pulse rate variation with pulse charge
amplitude.
All the rank regression pdfs are overlapping, while the pdf
of MLE stands out. This implies that maximum likelihood
parameter estimation determines the distribution parameters
that maximize the probability (likelihood) of the sample data.
From a statistical point of view, the method of maximum
likelihood is considered to be more robust and yields
estimators with good statistical properties.
3.2. Reliability function Characteristics
The most frequently used function in life data analysis
and reliability engineering is the reliability function or the
survival function. This function gives the probability of an
item operating for a certain amount of time without failure
[5]. It is dependent on the distribution parameters and the
failure data. Once the Weibull distribution (with a particular
set of parameters) has been fit to the data, computation of
cumulative distributive function (F(q) given in equation (4))
is possible. The reliability function is also sometimes referred
to as the survivorship or survival function (since it describes
the probability of not failing or of surviving until a certain
time q. Reliability theory and reliability engineering make
extensive use of Weibull distribution. The reliability function
is expressed in terms of the random variable, „q‟(pulse charge
height). The distribution parameters enable to estimate the
reliability function (equation 3) and it is calculated as
compliment of cumulative distribution function. Since the
result in Figure 6 shows that linearity of MLE is more than
that in regression approaches, the parameters of MLE is used
in estimating the reliability characteristics using equation 3.
Figure 8 shows the reliability plot of 2 parameter model
and Figure 11 shows the reliability plot of mixed Weibull
model. Reliability function is a decreasing curve, which
implies that as the pulse charge amplitude increases, the
survival or the reliability of the insulation is decreased. This
implies that the pulse charge amplitude is a vital quantity to
be monitored in order to maintain the reliability Table 2 and
table 4 shows the pulse charge amplitudes at various critical
values of reliability percentages for standard Weibull model
and mixed model respectively. Various percentiles of
reliability function at 1%, 50%, 90% and 99 % can be
estimated and corresponding charges can be obtained. There
has been drastic decrease in reliability from 90% to 50%
when pulse charge increased from 0.09µC to 0.2843µC. The
reliability function of a surface discharge at 5.75kV is also
plotted and it is observed that the increase of pulse charge
amplitude has resulted in the dip of reliability function. The
reliability characteristics at 3.78kV have steeper descent than
that at 5.75kV. These estimates are particularly valuable for
determining the percentage of insulation that can be expected
to have failed at particular points of time corresponding to
respective pulse charge amplitude.
3.3. 5-parameter Weibull Function
For PDHD plot in case of simultaneous PD phenomena,
where one type of discharge superimposes with the other, two
parameter Weibull function shows a non linear behaviour.
Each individual distributions(F1 and F2 corresponding to
slot and end winding) follows standard Weibull model; hence
each has two distribution parameters(one shape and one scale
parameter). The mixing proportion parameter(p) along with
two pairs of shape and scale parameters corresponds to five
parameters in the mixed Weibull model. Hence the number of
parameters are increased and the estimation is complicated.
Hence Graphical method is adopted for simplicity of
estimation of parameters. The additively mixed Weibull
mixture modeling is performed as explained in section 2.2.3.
Fig 9 shows the PDHD plotted in Weibull graph of mixture
discharges such as end winding and slot discharges at 6 kV.
As seen, F represents the Weibull graph of mixture
dishcharges. It has a deviation with some vestiges of
linearity. F1 and F2 represent the standard Weibull graph
corresponding to individual discharge (slot winding and end
winding discharges). It is seen that F lies between the F1 and
F2.
Figure 8. Plot of reliability functions for delamination at 3.78 kV and for
surface discharges at 5.75 kV.
Table 2. Critical values of reliability function
of Weibull distribution at 3.78 kV
Critical values of
Reliability
α(μC)
R (0.1) 0.5793
R (0.5) 0.2843
R (0.9) 0.09
R (0.99) 0.0018
Figure 9. PDHD plotted in Weibull graph for mixture discharges.
International Journal of Research and Reviews in Electrical and Computer Engineering (IJRRECE) 61
Table 3. Weibull Parameter estimation for mixture discharges at 6 kV
Method p α1(µC) β1 α2(µC) β2
Graphical
method
0.607 16.39 0.4714 42.038 2.1761
Standard Weibull
23.52 1.6359
Table 3 gives the comparison of distribution parameters
of mixture discharges with standard Weibull modeling of the
same data. As seen, the shape parameter of standard Weibull
estimated using MLE, for slot and end winding discharge (α
=23.52µC) lies in between the scale parameters of mixed
model(α1=16.39µC and α2= 42.038 µC). Similarly shape
parameter of standard Weibull for slot and end winding
discharge (β=1.6359). The mixing parameter value (p=0.607)
corresponds to the fact that there is 60.7% of one of the
discharges(slot or end winding dishcharge) and 39.3 % of the
other.Once the parameters are estimated, the probability
density function is plotted with pulse charge height as
random variable using the equation 14.
Figure 10 shows the model validation using the
probability density function graph. It follows the same shape
as pulse rate data. This implies that the probability of
occurrence of discharge (pdf) is a bell shaped curve with a
single mode and it has the same behaviour of the pulse rate
variation with pulse charge amplitude.
Figure 10. Comparison of probability density functions of mixed Weibull
function with the pulse rate Vs. charge amplitude curve.
Figure 11 shows the reliability function characteristics of
the mixture discharge and the standard Weibull model for
end winding and slot discharges at 6 kV.
Figure 11 Comparison of reliability plot of a mixture model with standard
Weibull model for slot and mixture discharges at 6kV.
As seen, both the curves are decreasing with increase in
pulse charge height. If the data is processed using the
standard Weibull model, the parameters can be estimated as
explained in section 2.1.4 (one shape parameter and one scale
parameter). From the fig 11, it is seen that up to a certain
pulse charge amplitude, the reliability characteristics of
standard Weibull model is above the characteristics of
mixture model, and beyond that reliability characteristics of
mixture model is above the standard model. This implies that
at higher pulse charge amplitudes, for more precision in the
analysis it is advisable to follow the additive–mixed Weibull
model, in case of simultaneous occurrence of partial
discharges. Thus, mixture model ensures better verification
than the standard model though standard Weibull modelling
is also possible with the same data.
Table 4 (a)
Reliability (Using
Mixture Model)
α(μC)
R (0.5) 21.92
R (0.9) 0.353
R (0.99) 0.17
Table 4 (b)
Reliability (Using
Mixture Model)
α(μC)
R (0.5) 18.75
R (0.9) 5.8
R (0.99) 0.353
Table 4(a) and 4(b) gives the comparison of critical
values of reliability of a standard Weibull model and mixture
model for same data of slot winding and end winding
discharges at 6 kV.
4. Conclusion
PD measurement is based on measurement of charge
transfer taking place at the discharge site. Thus, the measured
charge is proportional to the PD event. Study of statistical
behaviour of PD phenomena by considering pulse charge
amplitude as a random variable is presented in this paper.
This charge is responsible for degradation of the insulating
material in high voltage systems. Thus data of pulse charge
amplitude is corresponding to the failure data of insulation of
the system. Therefore failure analysis modeling is essential.
Failure data can be best modeled using Weibull Distribution.
Weibull distribution parameters are used to proceed towards
the reliability analysis. Rank regression (Mean rank, Median
rank, and Bernard‟s rank), hazard plotting and MLE method
are used for parameter estimation of two parameter Weibull
function while graphical method is used in case of five
parameter Weibull function. The PDHD plotted in the
Weibull graph (WPP) follows the linear relation in case of 2-
parameter standard Weibull model, while there is deviation in
the additively mixed Weibull model with some vestiges of
linearity. It was observed that the probability density graphs
of the Rank Regression method and MLE method matches
the experimental plot, with MLE having a better closeness.
International Journal of Research and Reviews in Electrical and Computer Engineering (IJRRECE) 62
The WPP of the additively mixed Weibull model is compared
with the standard 2-parameter Weibull model and it is seen
that the non-linear Weibull curve of mixed model lies within
the linear curves of standard Weibull model. This implies that
partial discharge analysis of the given data can be proceeded
to do the standard Weibull model analysis in case of single
discharge while in case of simultaneous discharges, PD
analysis of additively mixed Weibull model gives more
accurate result. The estimated parameters of the Weibull
distribution give the overview of the discharges. Evaluation
of insulation can be performed using Reliability function by
considering PD pulse charge height (q) as a random variable.
As the pulse amplitude increases, the reliability follows a
decreasing path. Thus the measured charge, which is
proportional to the PD event, determines the
reliability/survival of the life of insulation. The pulse charge
amplitude can be taken as an index to evaluate the percentiles
of reliability. The work can be extended to find the life
expectancy characteristics of an insulation taking into
account the rate of change of applied voltage and behavior /
physical parameters which are obtained from the pulse
sequence analysis.
References
[1] F, H, Krueger, Partial Discharge Detection in HV equipment, pp, 19-
21, London Butterworth, 1989.
[2] R. J. Van Brunt, “Stochastic Properties of Partial-discharge Phenomena”, IEEE Transactions on Electrical Insulation Vol. 28 No.
5, pg.902-948, Oct. 1991.
[3] M. Cacciari, A. Contin, G. Rabach and G. C. Montanari, „„An Approach to Partial Discharge Investigation by Height Distribution
Analysis‟‟, IEE Proc.: Sci., Meas. Technol., Vol. 142, pp. 102-108.
[4] Athanasious Papoulis, S. Unnikrishna Pillai,” Probability, Random variable and Stochastic Processes, Tata-McGraw Hill 2002.
[5] B.S.Dhillon, “Design reliability, Fundamentals and Applications”,
CRC Press, 1999. [6] L. A. Dissado, J. C. Fothergill and S. V. Wolfe, R. M. Hill,” Weibull
statistics in dielectric breakdown; theoretical basis, applications and
implications”, IEEE Transactions in Electrical Insulation Vol .EI-19, No. 3, June 1984.
[7] A. Contin, G. Contessotto, G. C. Montanari, M. Cacciari, “Comparing
Different Stochastic Models for the Identification and Separation Of concurrent partial discharge phenomena,” IEEE conf. Dielectric
Materials, Measurements and Applications, No. 473, pp.374-379,
2000. [8] A.Contin, G.C.Montanari, C. Ferraro, “PD source recognition by
Weibull processing of Pulse Height Distributions”, IEEE transactions
on dielectrics and electrical insulation, Vol 7 No. 1, Feb 2000. [9] D. N. Prabhakar Murthy, Min Xie, Renyan Jiang-“Weibull Models;” A
John Wiley & Sons, INC., Publication, 2004.
[10] S Senthil Kumar,” Enhancing Partial Discharge Data with Extended Resolution Analyzer,” XI International Symposium on High Voltage
Engineering, 1999, London. [11] Robert B Abernethy, „The New Weibull Handbook‟, Reliability and
Statistical Analysis for Predicting Life, Safety, Failures, and Test
Substantiation, ISBN 0965306224, 9780965306225, 2005. [12] M. Cacciari, A.Contin, G.Rabach, G.C.Montanari.” Use of Mixed-
Weibull Distributions for the Identification of Partial Discharges
Phenomena”, IEEE Trans. On D.E.I, Vol.2, N.4, August 1995. [13] A. Contin, E. Gulski, M. Cacciari, G. C. Montanari, “Applications of
the Weibull Function to Partial Discharge Data Coming from Different
Sources Typologies”, IEEE-CEIDP 95 Annual Report, pp335-338, 1995.
[14] Siyuan Jiang, Dmitri Kececioglu,” Graphical Representation of Two
Mixed-Weibull Distributions‟IEEE Transactions on reliability, vol. 41, no. 2, 1992 June .
[15] Dan Ling, Hong-Zhong Huang, Yu Liu,” A Method for Parameter
Estimation of Mixed Weibull Distribution” IEEE Trans on reliability 1-4244-2509-9/09,2009.
Sunitha Anup received B Tech degree from National
Institute of Technology(NIT), Calicut, Kerala, India in 2002 and currently pursuing
M Tech degree in Power Systems from Veermata
Jeejabhai Technological Institute (VJTI), Mumbai, India. Her fields of interest are high voltage
engineering and partial discharge phenomenon.
Amol Shripad Deshpande received the B.E. degree in 2006 and the M. Tech degree in 2008 from Dr.
Babasaheb Ambedkar University, Maharashtra, India
and now pursuing his Doctoral Programme at VJTI, Mumbai, India. He is currently working in the area of
Partial Discharge.
Dr. Alice Cheeran has completed B. E in Electrical
Engineering in 1984 from Kerala University. She joined
V J T I, Mumbai as a lecturer in 1987. She completed Masters in Electrical Engg with specialization in
Control Systems and in Electronics respectively in 1994
and 1996 from Mumbai University. Further she did Ph. D. from IIT Bombay in 2005 in the topic of “Speech
Processing for the Hearing Impaired. Currently she is
working as an Associate Professor in the Electrical Engineering department, V J T I, Mumbai.
Harivittal A. Mangalvedekar received the BE degree
in 1979, the ME degree in 1984, and the Ph.D degree in 1995 in electrical engineering from Bombay University,
India. He has been with V.J. technological Institute for
the last 22 years and is currently a Professor in the Elec. Engg. Dept. He has developed the high voltage
laboratory at VJTI and presently conducting research in the areas of power systems and high voltage.