Journal of Applied Statistics

ISSN: 0266-4763 (Print) 1360-0532 (Online) Journal homepage: http://www.tandfonline.com/loi/cjas20

Parameter estimation of the flexible Weibull

distribution for type I censored samples

Arwa Nasr, Soufiane Gasmi & Fayçal Ben Hmida

To cite this article: Arwa Nasr, Soufiane Gasmi & Fayçal Ben Hmida (2016): Parameter
estimation of the flexible Weibull distribution for type I censored samples, Journal of Applied
Statistics, DOI: 10.1080/02664763.2016.1257588

To link to this article: http://dx.doi.org/10.1080/02664763.2016.1257588

Published online: 01 Dec 2016.

Submit your article to this journal

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=cjas20

Download by: [Fudan University] Date: 01 December 2016, At: 23:53

JOURNAL OF APPLIED STATISTICS, 2016
http://dx.doi.org/10.1080/02664763.2016.1257588

Parameter estimation of the flexible Weibull distribution for

type I censored samples
Arwa Nasra , Soufiane Gasmib and Fayçal Ben Hmidaa
a Department of Electrical engineering, University of Tunis, The National Higher Engineering School of Tunis,

Tunisia; b Department of Mathematics, University of Tunis, The National Higher Engineering School of Tunis,
Tunisia

ABSTRACT ARTICLE HISTORY

Often in practice one is interested in the situation where the lifetime Received 14 July 2015
data are censored. Censorship is a common phenomenon frequently Accepted 31 October 2016
encountered when analyzing lifetime data due to time constraints. KEYWORDS
In this paper, the flexible Weibull distribution proposed in Bebbing- Reliability; flexible Weibull
ton et al. [A flexible Weibull extension, Reliab. Eng. Syst. Safety 92 distribution; parameter
(2007), pp. 719–726] is studied using maximum likelihood technics estimation; statistical test;
based on three different algorithms: Newton Raphson, Levenberg censoreddata
Marquardt and Trust Region reflective. The proposed parameter esti-
mation method is introduced and proved to work from theoretical
and practical point of view. On one hand, we apply a maximum likeli-
hood estimation method using complete simulated and real data. On
the other hand, we study for the first time the model using simulated
and real data for type I censored samples. The estimation results are
approved by a statistical test.

1. Introduction
The Weibull distribution was proposed by Waloddi Weibull [15], this distribution was
widely used to resolve many reliability engineering problems and to model lifetime data.
Since 1970, this law was developed and modified to have several forms and extensions. We
can quote some of them: the four-parameter Weibull distribution proposed by Kies [7], the
exponentiated Weibull distribution studied by Mudholkar and Srivastave [11], then Lai et
al. [8] suggested the modified Weibull distribution. Famoye et al. [4] proposed the beta-
Weibull distribution by coupling the beta density and the Weibull distribution function.
Nikulin and Haghighi introduced the generalized power Weibull family [13].
Often, during a study for an industrial system the results are incomplete. In this case,
we obtain incomplete data called: censored data. These data come from the fact that we
have no access to all information. We will focus our work on the study of models based
on censored data which are not widely dealt by researchers. Notice that censorship is the
most frequently encountered phenomenon when collecting survival data. In the last few

CONTACT Arwa Nasr [email protected] Department of Electrical engineering, University of Tunis, The
National Higher Engineering School of Tunis, 5, Avenue Taha Hussein, B.P.56, Bab Menara 1008, Tunis, Tunisia

© 2016 Informa UK Limited, trading as Taylor & Francis Group

2 A. NASR ET AL.

decades, the progressive censoring schemes have become commonly used in the literature
because of its flexibility in removing failed units from the test.
We cite some of these works: to estimate the model parameters, Lee et al. [9] extract
the derivatives of the log-likelihood function to obtain the maximum likelihood estimates
(MLEs) of the parameters in the case of censored simulated data. Also, Wang et al. [14]
used the MLE via the expectation maximization algorithm to estimate the BUR XII param-
eters using censored data. The MLEs of the inverse Weibull distribution parameters with
censored data were studied and obtained by Gusmao et al. [6]. The progressive type II cen-
soring schemes have been applied to exponential distributions in [10], also it is used by
Zhang et al. [17] for several popular lifetime models.
The beta-Weibull model was applied by Lee et al. [9] for two censored data sets of bus-
motor failures rate. In Nasr et al. [12], the reliability and the maintainability parameters
of the Weibull distribution were estimated based on type I censored data and by using
the virtual age model of Kijima type I and type II. Gasmi et al. [5] estimated the modi-
fied Weibull distribution parameters proposed by Zaindin and Sarhan [16] in the case of
type I censored data. Bebbington et al. [2] introduced a two-parameter aging distribution,
the flexible Weibull distribution (FWD), which represents a generalization of the Weibull
distribution.
The research objective of this paper is to develop an MLE method to obtain more accu-
rately estimated parameters based on complete and censored data and using the FWD
which is quite simple and yet very flexible to model reliability data.
Graphical resolution techniques have been frequently used in the literature but they are
not accurate enough. In this work, we present an analytical parameter estimation technique
– the maximum likelihood method – which is widely used by researches in different fields.
This technique is complicated for kinds of distributions because of the complexity of the
likelihood function but leads to less estimation errors.
The rest of the paper is organized as follows. Section 2 is devoted to the study of the flex-
ible Weibull model using simulated and real data sets in the case of complete samples. We
mention that our study is based on the maximum likelihood method of estimation using
three different algorithms – Newton Raphson (NR), Levenberg Marquardt (LM) and Trust
Region reflective (TR). After that, we performed the estimation of model parameters which
belongs to the family of Weibull. Our results were compared to those found by Bebbington
[2]. The K–S statistical test was established to check the compatibility between the model
and the real data sets.
Section 3 is dedicated to the parameter estimation of the FWD in the case of type I
censored samples. In this section, simulated and real censored data are used. The flexible
Weibull model has not been studied by researchers in the case of censored data hence we
obtained new estimation results. The estimation results were validated by the K–S statistical
test. The conclusion is presented in Section 4.

2. Estimation of the FWD in the case of complete data

We treat in this section the parameters estimation of the FWD in the maximum likelihood
sense, which is characterized by its asymptotic normality and consistency, and based on
complete simulated and real data.
 JOURNAL OF APPLIED STATISTICS 3

Figure 1. Reliability function.

2.1. Flexible Weibull distribution

The cumulative distribution function (CDF) and the survival function (SF) of the FWD
are given, respectively, by the following expression:

F(t, θ) = 1 − e(−e
at−b/t )
, t > 0, (1)

S(t, θ ) = e(−e
at−b/t )
, t > 0, (2)

where θ = (a, b) and a,b > 0.

The probability density function (PDF) and the hazard rate function (HRF) correspond-
ing to Equation (1) are, respectively, as follows:
 
b
f (t, θ) = a + 2 e(at−b/t) e(−e
at−b/t )
, (3)
t
 
b
h(t, θ) = a + 2 e(at−b/t) . (4)
t
In contrast to many other generalized Weibull distribution, we remark that the hazard
rate expression of this distribution is not complicated.
Figures 1, 2, 3 and 4 represent, respectively, the CDF, the reliability function R, the PDF
and the HRF for different selected values of a and b. We mention that when we decrease
b, the HRF becomes more bathtub shape and when a increases, the bathtub becomes less
deep. For different values of a and b, the HRF is considered as increasing failure rate (IFR),
increasing failure rate average (IFRA) and modified bathtub (MBT) shaped which explains
the flexibility of the studied model. Bebbington et al. [2] studied the behavior of the failure
rate function, so we can conclude that:

• If ab ≥ 27
64 , the distribution is IFR.
• If and only if ab ≥ 14 , the distribution is IFRA which includes the IFR case, as needed.
• If 14 < ab < 27 64 , the distribution is IFRA and not IFR.
• If ab < 27
64 , the hazard rate has an MBT shaped.
4 A. NASR ET AL.

Figure 2. Cumulative distribution function.

Figure 3. Probability density function.

Figure 4. Hazard rate function.

We remark that when b decreases, the failure rate function becomes more bathtub-like,
in consideration when a increases, the bathtub becomes shallower. So we can conclude that
the model has a great deal of flexibility.
 JOURNAL OF APPLIED STATISTICS 5

Table 1. Simulated data.

0.0372 0.0451 0.0479 0.0581 0.0595 0.0601 0.0638 0.0653 0.0669 0.0688 0.0714 0.0783 0.0783 0.0792
0.0805 0.0816 0.0818 0.0867 0.0870 0.0872 0.0909 0.0927 0.0959 0.0977 0.0992 0.1004 0.1012 0.1029
0.1040 0.1054 0.1067 0.1069 0.1085 0.1098 0.1120 0.1140 0.1142 0.1155 0.1170 0.1179 0.1227 0.1234
0.1236 0.1249 0.1259 0.1294 0.1427 0.1454 0.1484 0.1511 0.1511 0.1521 0.1530 0.1572 0.1590 0.1594
0.1608 0.1622 0.1640 0.1644 0.1718 0.1723 0.1733 0.1779 0.1847 0.1916 0.1950 0.1993 0.2147 0.2151
0.2173 0.2254 0.2262 0.2269 0.2282 0.2422 0.2429 0.2540 0.2552 0.2638 0.2665 0.2749 0.2773 0.2797
0.2961 0.3020 0.3151 0.3215 0.3257 0.3316 0.3361 0.3379 0.3439 0.3445 0.3764 0.3837 0.4077 0.4107
0.4159 0.4170 0.4194 0.4362 0.4543 0.4606 0.4664 0.4728 0.5152 0.5220 0.5455 0.5739 0.5744 0.5751
0.5839 0.5843 0.6531 0.6572 0.6845 0.7046 0.7305 0.7305 0.8158 0.8967 0.9339 0.9568 1.0016 1.0863
1.1011 1.1118 1.1222 1.1604 1.1684 1.1714 1.1734 1.3669 1.3756 1.4401 1.4629 1.5162 1.5574 1.6267
1.6604 1.8987 2.0276 2.0457 2.0859 2.1226 2.1518 2.1632 2.1934 2.1999 2.2949 2.4126 2.4228 2.4671
2.4807 2.4810 2.6383 2.6662 2.6736 2.6785 2.7448 2.7477 2.9442 2.9876 2.9923 3.0443 3.0776 3.1126
3.1879 3.3403 3.4742 3.5379 3.8037 3.8141 4.2147 4.3987 4.4541 4.5193 4.6251 4.6852 4.6860 4.7397
4.7734 5.0678 5.0710 5.1921 5.1925 5.4180 5.4995 5.6697 5.8701 5.9070 5.9193 5.9849 5.9924 6.0019
6.0845 6.2048 6.4564 8.6892

Another advantage of this distribution is the fact that: limt→0 h(t) = 0 and
limt→+∞ h(t) = +∞. We can conclude that the HRF is always increasing.

2.2. MLEs using simulated data

This subsection is devoted to the lifetime data simulation and the parameter estimation
in the case of simulated data. We mention that the maximum likelihood technique is
a common statistical method used to infer the probability distribution parameters for a
given sample. The parameter estimation of the model by the likelihood method requires
two steps: Step 1: get the likelihood expression function from the model formulation and
assumptions. Step 2: find the parameter values maximizing this function. This method con-
sists in constructing of the likelihood function (constructed from the PDF of the period of
system operation) and maximizes its logarithm with respect to the unknown parameters.
The simulation has been made by writing some computer programs using Matlab. To
achieve this purpose, we need to set the FWD parameters as follows: a = 0.2 and b = 0.2,
and generate samples from the FWD population by solving the following equation:

eati −b/ti + log(1 − U) = 0, (5)

where i = 1, . . . , n, n indicates the sample size and U is a random variable uniformly dis-
tributed on the interval (0,1). Table 1 represents 200 simulated data generated from the
FWD.
Let x = (x1 , . . . , xn ) be a random sample of the FWD with unknown parameter vector
θ = (a, b).
The likelihood function L(t, θ) is given by


n
L(t, θ) = f (ti , θ)
i=1
n 
 
b ati −b/ti )
= a+ 2 e(at−b/ti ) e(−e . (6)
i=1
ti
6 A. NASR ET AL.

Figure 5. Log-likelihood function.

The corresponding log-likelihood function is given by

   

n
b 
n 
n
b 
n
log(L) = log a + 2 + a ti − − e(ati −b/ti ) . (7)
ti ti
i=1 i=1 i=1 i=1

Therefore, we obtained a nonlinear system consisting of two equations in two param-

eters as in [2]. To obtain the FWD estimates, we have to solve the first partial derivatives
equations by equating it to zero. Since it is not possible to solve analytically the system, we
have chosen the numerical resolution using iterative techniques such as TR and NR [3].
To estimate the FWD parameters, we use the following three algorithms: NR, LM and TR
algorithms. To find the optimal values that maximizes the log-likelihood function, we solve
the obtained nonlinear system using the simulated data mentioned in Table 1. The initial
values introduced at the beginning of the program should belong to the domain conver-
gence of the distribution. The estimation results using the MLE technique for complete
simulated data are â = 0.2095 and b̂ = 0.1995.
Figure 5 illustrates the empirical CDF, the CDF and the 95% lower and upper confidence
bounds, noted, respectively, LCD and UCB, for the CDF of the 200 simulated data by setting
a = 0.2 and b = 0.2. We remark that the CDF follows properly the empirical CDF.

2.3. Application
To demonstrate the performance of the FWD in practice, we consider two real data sets
which represent the data of primary and secondary pumps installed inthe RSG gas reactor
[1] mentioned, respectively, in Tables 2 and 3. The primary and secondary pumps are the
mechanical components in the RSG-GAS reactor cooling system which are used for routine
operation.
The data are taken from the operation log-book of the RSG-GAS reactor during 10 years
of reactor operation since 1987.
For the two data sets, we compare the results of the FWD fits and its sub-model: the
exponential distribution ED.
 JOURNAL OF APPLIED STATISTICS 7

Figure 6. CDF and empirical CDF of the FWD(a,b) for complete simulated data.

Table 2. Time between failures (thousands of hours) of primary

reactor pumps.
Pumps TBF
JEOI-AP01 2.767 0.114 2.319 9.026 0.003 0.258
JEOI-AP02 2.805 0.008 1.132 10.136
JEOI-AP03 0.054 0.595 0.018 0.009

Table 3. Time between failures (thousands of hours) of secondary reactor pumps.

Pumps TBF
PAOI-APO1 2.160 0.746 0.402 0.954 0.491 6.560 4.992
PAO2-APO2 3.474 0.150 0.358 0.101 1.359 3.465 1.060 0.614 1.921 4.082 0.199
PAO3-APO3 0.605 0.273 0.070 0.062 5.320

The SF of the Weibull distribution:

β
S(t) = e(−t/η) , t  0, (8)

with β, η > 0.

S(t) = e−αt , t  0, (9)

with α > 0.
Tables 2 and 3 represent the time between failure (TBF) of the primary pump and the
secondary pump. We mention that JEOI-AP01, JEOI-AP02 and JEOI-AP03 are the three
primary pumps and PAOI-APO1, PA02-AP02 and PA03-AP03 of the three secondary
pumps of the RSG-GAS reactor, they are identically in type and duty [1].
We start modeling of the FWD in the case of complete data. For this purpose, we use
the time between failure data set of the primary pumps mentioned in Table 2.
In this case, contrary to the authors of [2], we have to use Xij in the likelihood function
to represent the time between successive failures of pumps instead of ti which materially
represents the time to failure (TTF) and it is why authors have not found a good estimation
results.
8 A. NASR ET AL.

Table 4. Parameter estimation using TBF of the primary pumps data.

â b̂ p-Values K–S L
FWD 0.1342 0.0125 .7009 0.2476 −13.7660
ED 0.4787 – .1846 0.3956 −24.3126

Table 5. Parameter estimation using TBF of the secondary pumps data.

â b̂ p-Values K–S L
FWD 0.2084 0.2519 .9150 0.1558 −32.7878
ED 0.5835 – .7639 0.1866 −35.3908

So the log-likelihood function becomes

   
n  k
b b
Log(L) = log a + + aXij − − e(aXij −b/Xij ) . (10)
i=1 j=1
(Xij )2 Xij

The first partial derivatives of the log-likelihood function with respect to the two
parameters of the FWD are
⎛⎛ ⎞ ⎞
∂Log(L)   ⎝⎝ 1 ⎠
n k
= + Xij − Xij e(aXij −b/Xij ) ⎠, (11)
∂a a + b2
i=1 j=1 Xij
 
∂Log(L)  
n k
1 1 1 (aXij −b/Xij )
= − + e , (12)
∂b i=1 j=1
a(Xij )2 + b Xij Xij

with n being the number of pumps and k representing the TBFs data for each pumps.
We obtain the MLEs θ̂ = (â, b̂) by maximizing the log-likelihood function. The estima-
tion results of the FWD and the ED using the TBF of the primary pumps in the case of
complete data are summarized in Table 4.
Second, we proceed to the FWD reliability parameter estimation using the secondary
pumps data sets. So we try to solve the system consisting the derivatives with respect to
flexible Weibull parameters and we set them equal to zero.
The estimation results of the FWD and ED using the TBF for the secondary pumps data
are summarized in Table 5.
From Table 5, we remark that the estimates of the FWD using TBF of the secondary
pumps in the case of complete data are equal to (0.2084, 0.2519). We rectify the estimation
results using the maximum likelihood estimator presented by Bebbington in [2] as θ̂ =
(0.0207, 2.5875). To confirm our results, we check graphically the values of â and b̂ that
maximize the log-likelihood function. We mention that the parameter estimation results,
in both cases, are equal using NR, LM and TR algorithms.
Figure 6 illustrates the log-likelihood function in 3D, it varies depending on the shape
and the scale parameters. The optimum values of a and b that maximize the log-likelihood
function are checked by curves in Figures 7 and 8. We remark that the log-likelihood func-
tion have a maximum for â = 0.19 and b̂ = 0.26, which proves the results obtained using
maximum likelihood method, we notice that these values verifies the numerical results
found previously.
 JOURNAL OF APPLIED STATISTICS 9

Figure 7. Log-likelihood function with respect to a.

Figure 8. Log-likelihood function with respect to b.

As a second step, we perform the following null hypothesis test; H0 : b = 0 and we set
a = log(λ), the data follow the exponential distribution (ED). In favor of the alternative
hypothesis Ha : the data follow the FWD.
We use a nonparametric test statistics, the K–S test with a level of significance equal
to 0.05, to test the null hypothesis mentioned below against Ha . We accept H0 with the
p-value under the condition p−value > 0.05. In Tables 4 and 5, the K–S test values of the
studied models are summarized also the p-values and the log-likelihood function of each
models. We can conclude first from Table 4 that none of H0 is rejected at level ϑ ≤ 0.19 and
second from Table 5 that none of H0 is rejected for the studied models at level ϑ ≤ 0.77.
We deduce that the FWD represents the best model to fit the current data sets because it
has the biggest p-value and the lowest K–S value.
Figures 9 and 10 illustrate the plots of the empirical and fitted scaled TTT-transforms,
the empirical and parametric cumulative density functions, the empirical and fitted hazard
and PDFs for the flexible Weibull model and its correspondent sub-models for the primary
and the secondary pumps data.
We remark from Figures 9 and 10 that the FWD fits the data sets better than all other
distributions used here, because its fitted curve is closer to the empirical curve.
10 A. NASR ET AL.

Figure 9. (a) The empirical and estimated scaled TTT-transform plots of the ED, WD and FWD models; (b)
the empirical and estimated cumulative density function of the ED, WD and FWD models; (c) empirical
and estimated HRFs of the ED, WD and FWD models; (d) empirical and estimated PDF of the ED, WD and
FWD models, for TBF primary pumps data.

Figure 10. (a) The empirical and estimated scaled TTT-transform plots of the ED, WD and FWD models;
(b) the empirical and estimated cumulative density function of the ED, WD and FWD models; (c) empirical
and estimated HRFs of the ED, WD and FWD models; (d) empirical and estimated PDF of the ED, WD and
FWD models, for TBF secondary pumps data.
 JOURNAL OF APPLIED STATISTICS 11

3. Parameter estimation in the case of censored data

The phenomenon of censoring is encountered in the reliability data analysis which makes
estimating parameters more difficult. The modeling problem of the FWD in the presence
of censorship phenomenon was not treated in the literature. In this section, we discuss the
reliability parameter estimation using the maximum likelihood method in the case of type
I censored data. We have to use the reliability function in place of the PDF.
Consider n independent and identical components put on test. X = (X1 , . . . , Xn ) rep-
resent the random sample of the FWD distribution with unknown parameters vector
θ = (a, b). Here, we consider maximum likelihood estimation for censored data with-
out replacement. N items are independently observed and the observation of the ith item
(i = 1, . . . , N) is censored at time Ti .
The likelihood function becomes


n 
n
L(t, θ) = f (ti , θ) R(Ti , θ ), (13)
i=1 i=1

Li (mi , δi ) = (fX (ti , θ))δi (RX (Ti , θ ))(1−δi ) ,

mi = min(ti , Ti ) and
1 if a failure was observed,
δi =
0 if the observation was censored at time Ti .

If all Ti are equal: Ti = T, i = 1, . . . , N, the likelihood function has the following form:


n
Li (mi , δi ) = fX (ti , θ ) (RX (T, θ ))(N−n) (14)
i=1

n = n(T). In this case, we are using real data which represent the TBF of the primary pump
and the secondary pump in the place of the TTF as in the previous section, so we have to
introduce Xi,j in the likelihood function instead of ti .
Using Equations (2) and (3), we obtain the following likelihood function:
⎧ ⎫
⎨n 
k   ⎬
b aX −b/Xij
e(aXij −b/Xij ) e (−e ij )
(e(−e
aT−b/T )
L(Xij , θ) = Xij a+ 2 )(N−n) .
⎩ Xij ⎭
i=1 j=1

The log-likelihood function takes the following form:


n 
k    
b b
Log(L) = log a + 2 + aXij − − e(aXij −b/Xij ) − (N − n) e(aT−b/T) .
i=1 j=1
Xij Xij

Now, we must extract the log-likelihood function derivatives with respect to the FWD
parameters and equating it to zero.
12 A. NASR ET AL.

Table 6. Parameter estimation using censored simulated data.

Level of censoring â b̂
5% of censoring 0.2268 0.1982
10% of censoring 0.2343 0.1976
15% of censoring 0.2646 0.1954

Figure 11. CDF and empiricalCDF of the FWD(a,b) for censored simulated data.

The first partial derivative of the log-likelihood function with respect to a and b is given
by the following expressions:
⎛ ⎞
∂Log(L)   ⎝ 1
n k
= + Xij − Xij e(aXij −b/Xij ) ⎠ − (N − n)T e(aT−b/T) , (15)
∂a a+ b
i=1 j=1 Xij 2
  
∂Log(L)  
n k
1 1 1 (aXij −b/Xij ) 1
= − + e + (N − n) e(aT−b/T) .
∂b i=1 j=1
aXij 2 + b Xij Xij T
(16)

After obtaining the log-likelihood derivatives with respect to a and b, we solve the
system of nonlinear equations numerically in a and b to obtain MLEs for censored data.
We perform three type I censored samples from the simulated data mentioned in Table 1
and we look for the MLE’s of the FWD. We apply to our system the same steps to achieve
optimum values that maximize the log-likelihood function. To find the FWD estimates,
we need to introduce a0 and b0 at the beginning of the program. The choice of the initial
values is arbitrary, it just requires that they belong to the domain of convergence.
In the first place, we consider the parameter estimation of the FWD based on the three
performed censored samples. The system converges to the global maximum and gives the
results mentioned in Table 6.
Figure 11 illustrates the empirical CDF, the CDF and the 95% LCB and UCB for the CDF
of the censored simulated data. We remark that the CDF follows properly the empirical
CDF using censored simulated data.
As a first step, we consider parameter estimation of the FWD using censored real data.
 JOURNAL OF APPLIED STATISTICS 13

Table 7. Parameter estimation using TBF of the primary pumps censored data.
Model â b̂ K–S p-Value
Level 1 FWD 0.0320 0.0141 0.2857 .5635
ED 0.1548 – 0.6154 .0930
Level 2 FWD 0.0319 0.0143 0.3333 .4595
ED 0.1115 – 0.7500 .0120

Table 8. Parameters estimation using TBF of the secondary pumps censored data.
Model â b̂ K–S p-Value
Level 1 FWD 0.0325 0.3276 0.3478 .1021
ED 0.1415 – 0.5119 .0320
Level 2 FWD 0.0020 0.3884 0.4091 .0389
ED 0.0884 – 0.6364 .0140

We perform two artificial levels of censoring for each real data used in Section 2. In all
the cases, the parameters estimation was carried out using the NR, LM and TR algorithms.
We observe that the MLEs in the three cases turned out to be equals. Finally, our system
converges to the global maximum and we get the results mentioned in Table 7 and 8.
We used the nonparametric statistical test, K–S, to test the null hypothesis against Ha as
in Section 2. The FWD and ED estimates, the K–S statistic with p-value are cited in Tables 7
and 8 for two different levels of censoring. We can conclude first from Table 7 that none of
H0 is rejected at level ϑ ≤ 0.013 and second from Table 8 that none of H0 is rejected for
the studied models at level ϑ ≤ 0.015. The FWD is the best model compared to ED, to fit
the censored real data sets because it has the biggest p-value and the lowest K–S value for
all levels of censorship and for both used real data.

4. Conclusions
In this paper, we show the performance of the FWD. The maximum likelihood estimations
of the FWD are discussed based on NR, LM and TR algorithms. The parameter estimation
was established using first complete real and simulated data and second censored real and
simulated data. Two real data sets are analyzed using the FWD and it is compared with sub-
models. The results of the comparisons showed that the FWD provides a better fit for the
two data sets using complete and censored data. Thus, in the presence of the phenomenon
of censoring we recommend the use of MLE via NR, LM and TR algorithms to estimate
the FWD parameters.

Disclosure statement
No potential conflict of interest was reported by the authors.

References
[1] R.E. Barlow and D. Bernard, Analysis of time between failures for repairable components, Inter-
national Conference on Nuclear System Reliability Engineering and Risk Assessment (4), 1977,
pp. 543–561.
[2] M. Bebbington, C.-D. Lai, and R. Zilitikis, A flexible Weibull extension, Reliab. Eng. Syst. Safety
92 (2007), pp. 719–726.
14 A. NASR ET AL.

[3] R.L. Burden and J.D. Faires, Numerical Analysis, Vol. 12, Cengage Learing, 2011.
[4] F. Famoye, C. Lee, and O. Olumolade, The Beta-Weibull distribution, J. Stat. Theory Appl.
4 (2005), pp. 121–136.
[5] S. Gasmi and M. Berzig, Parameters estimation of the modified Weibull distribution based on
type 1 censored samples, Appl. Math. Sci. 5 (2011), pp. 2899–2917.
[6] F. Gusmão, E. Ortega, and G. Cordeiro, The generalized inverse Weibull distribution, Stat. Papers
52 (2009), pp. 591–619.
[7] J.A. Kies, The Strength of Glass, Vol. 98, Naval Research Laboratory, 1958.
[8] C.D. Lai, M. Xie, and D.N.P. Murthy, A modified Weibull distribution, IEEE Trans. Reliab.
52 (2003), pp. 33–37.
[9] C. Lee, F. Famoye, and O. Olumolade, Beta-Weibull distribution: Some properties and applica-
tions to censored data, J. Mod. Appl. Stat. Methods 6 (2007), pp. 6173–6186.
[10] J. Lee and R. Pan, Bayesian analysis of step-stress accelerated life test with exponential distribu-
tion, Quality Reliab. Eng. Int. 28 (2012), pp. 353–361.
[11] G.S. Mudholkar and D.K. Srivastava, Exponentiated Weibull family for analyzing bathtub
failure-rate data, IEEE Trans. Reliab. 42 (1993), pp. 299–302.
[12] A. Nasr, S. Gasmi, and M. Sayadi, Estimation of the parameters for a complex repairable sys-
tem with preventive and corrective maintenance, IEEE International Conference on Electrical
Engineering and Software Applications (ICEESA) (3), 2013, pp. 1–6.
[13] M. Nikulin and F. Haghighi, A chi-squared test for the generalized power Weibull family for the
head-and-neck cancer censored data, J. Math. Sci. 133 (2006), pp. 1333–1341.
[14] F.K. Wang and Y.F. Cheng, Em algorithm for estimating the burr xii parameters with multiple
censored data, Quality Reliab. Eng. Int. 26 (2009), pp. 615–630.
[15] W. Weibull, A statistical distribution function of wide applicability, J. Appl. Mech. 18 (1951),
pp. 293–297.
[16] M. Zaindin and A.M. Sarhan, Parameters estimation of the modified Weibull distribution, Appl.
Math. Sci. 11 (2009), pp. 541–550.
[17] M. Zhang, Z. Ye, and M. Xie, A stochastic em algorithm for progressively censored data analysis,
Quality Reliab. Eng. Int. 30 (2014), pp. 711–722.