Parameter Estimations and Optimal Design of Simple Step-Stress Model for Gamma Dual Weibull Distribution

This paper considers life-testing experiments and how it is effected by stress factors: namely temperature, electricity loads, cycling rate and pressure. A major type of accelerated life tests is a step-stress model that allows the experimenter to increase stress levels more than normal use during the experiment to see the failure items. The test items are assumed to follow Gamma Dual Weibull distribution. Different methods for estimating the parameters are discussed. These include Maximum Likelihood Estimations and Confidence Interval Estimations which is based on asymptotic normality generate narrow intervals to the unknown distribution parameters with high probability. MathCAD (2001) program is used to illustrate the optimal time procedure through numerical examples.


Introduction
The Accelerated Life Tests ALTs play pivotal role in reliability analysis of products or systems.During an experiment, the stress levels increase under normal operating conditions and at pre-fixed time points on test units to obtain quickly information on the parameters of the lifetime distributions in a short time.In addition, failure resulting from units of the experiment is probably due to several stress factors, including temperature, cycling, pressure, vibration and voltage, etc.This procedure is called a Step-Stress test SSt.
Many authors mentioned different choices for the SSt.(Drop et al 1996) provided a Bayes estimation for step-stress model and they assumed the failure times have exponential distribution while the specification of strict compliance to the transformation function of time is not desired.(Dharmadhikari & Monsur 2003) assumed that the scale parameter of Weibull and Lognormal models depend on the age of the present levels.They provided a parametric model to the failure time distribution for the SSt and estimated the parameters involved in it.(Chen et al 2006) considered the exponential population as a failure time to SSt.They applied the tampered failure rate model to obtain some measures such evidence for the existence, strong consistency, uniqueness and the asymptotic normality of the MLE of mean.Under a cumulative exposure model, (Al-Ghamdi & Hassan 2009) considered the problem using the Lomax distribution.By minimizing with respect to the asymptotic variance of the MLE of a given 50th percentile and the change time scale parameter, they obtained the optimum test plan.
Where     , the exponential distribution holds.This paper consists of four sections corresponding to 2, 3, 4 and 5, respectively.section 2 discusses the model description and assumptions.section 3, provides the likelihood function estimators and the asymptotic Fisher information matrix which contains the second partial derivatives of logarithm of likelihood function.Confidence intervals for the parameters are examined based on asymptotic normality in section 4. The last section considers a numerical investigation to study the properties of the new estimators.

Model Description and Assumptions
Based on the CDF of the iii.
At the time In SSt which is the one of ALTs, the cumulative exposure model is: and the pdf of the failure time is: In next section, the MLEs and asymptotic Fisher information matrix are provided.

Likelihood Function and Asymptotic Fisher Information Matrix
Let the failure times , 1, 2, , , 1, nn are the number of items which are fail at the first stress level 1 Z and high stress 2 Z respectively.Then, their likelihood and log likelihood functions respectively are: and The first derivatives of   12 ; , , ,    and 2  are, respectively: ln ln ln ln where is the digamma function . in ( 8), ( 9), ( 10) and (11).So, Newton-Raphson method is used to solve these equations in numerical analysis.It is an iterative method for solving equations ( ) 0 fb where ()  are obtained to construct the asymptotic Fisher information matrix as follows: . where . ln .ln 1 1 . ln ln where is the digamma function and .

Approximate Confidence Interval
Let 12 , , , n X X X L be two independent random samples drawn from 2.W  distribution as Eq. ( 5) with parameters . Now, using following equations and confidence level 1 0.95   , the lower and upper confidence intervals can be constructed as:

Simulation Study and Data Analysis
The computer programs MathCAD ( 2001) is used to obtain numerical illustration for the last theoretical results.1000 samples generated from gamma dual Weibull distribution with parameters   Also, the lower and upper confidence intervals of the acceleration factor and both two shape and scale parameters are obtained.The steps of Simulation are: 1.
The random samples are generate from cumulative exposure for gamma dual Weibull distribution as Eq. ( 2) at various values of 1 Z , 2 Z with various size samples .

3.
The (R Bias), (MSE) and (RE) of the four parameter estimators are obtained.Also, for different sample sizes, asymptotic variance and covariance matrix of the estimators are calculated.

4.
The lower and upper confidence intervals at significant level 0.05

 
of the two shape and scale parameters are construct.Now, from table (1) and ( 2), note that, the (R Bias), (MSE) and (RE) of the four parameters   12 , , ,     are decreasing by increasing the sample size n .Also, in table (3), the lower and upper confidence intervals of the acceleration factor and two shape and scale parameters are asymptotic, so the length of interval is decreasing by increasing the sample size n .

(
Donghoon, H. & Balakrishnan2010) considered the simple SSt under time constant when the failure times of the different risk factors are distributed as exponential distribution.Under the assumption of a cumulative exposure model, they derived the MLEs of the unknown parameters.(Abd-Elfattah and Al-Harbey2010) studied the estimation problem when the lifetime distribution of the test items is following Burr type III distribution.They derived the MLEs for the acceleration factor and distribution parameters in case type II censored samples.Based on type II right censored data from the SSt, (Lee et al 2013) studied the problem for exponential products.They performed tests of hypotheses about model parameters based on the likelihood method.New family of distributions namely G  due to (Zagrafos and Balakrishnan 2009).They added a shape parameter to the baseline G distribution.In (2015), (Castellares and Lemonte) studied the concept of Zagrafos and Balakrishnan and they proposed a very useful explanation for its pdf.Another family of distributions was introduced by (Ristic and Balakrishnan2012) based on random variables taken from gamma distribution.The new family say 2.G  may be regarded to G  family.(Castellares and Lemonte 2016) introduced three interesting propositions to study the relation between G  and 2.G  families and used the last family to generate the new three parameters gamma dual Weibull 2.W  distribution which has CDF and pdf respectively as:

1 
 in (2), then the pdf of the Weibull distribution comes.If 1   , then (2) represents the pdf of the gamma dual exponential distribution.When 2   , the gamma dual Rayleigh distribution arises.And if 1

Fig. 1 .
Fig. 1.Different forms for pdf of the 2.W  with various values of the parameters   ,  at

12 ;
is assumed to have a continuous derivative () fb  .Given a function () fband its derivative () fb  , a first guess 0 b is initialed .Then, an approximation of 1 b is estimate. (seeKotz et al (2003) [11].Now, the second derivatives of   Z .Then, the maximum likelihood estimation  of  is asymptotically normally distributed with mean  and variance 2 ()   under regularity conditions and large sample.It can be constructed as:  are used, respectively at different values of 1

Table 1 : R BIAS, MSE and RE Values for parameters
Pak.j.stat.oper.res.Vol.XIV No.1

Table 3 : Upper and lower bounds of Confidence intervals at significant level 0.05
This paper studied a step stress which is a major type of accelerate life tests by assuming that data follow Gamma Dual Weibull distribution.Different methods for estimating the parameters are discussed including Maximum Likelihood Estimations, Fisher Information Matrix and Confidence Interval Estimations.In simulation study, three criteria measures have been calculated for the four parameters   12, , ,     .These measures are decreasing by increasing the sample size n .Also, the lower and upper