Sunteți pe pagina 1din 15

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/335056188

Optimal Design of Reliability Acceptance Sampling Plan Based on Data


Obtained from Partially Accelerated Life Test (PALT)

Article  in  International Journal of Reliability Quality and Safety Engineering · August 2019


DOI: 10.1142/S0218539320400021

CITATION READS

1 19

3 authors, including:

Mahesh Kumar Ramyamol P C


National Institute of Technology Calicut National Institute of Technology Calicut
14 PUBLICATIONS   21 CITATIONS    4 PUBLICATIONS   6 CITATIONS   

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Ramyamol P C on 13 January 2020.

The user has requested enhancement of the downloaded file.


Econ. Qual. Control 2016; 31 (1):23–36

Research Article

Mahesh Kumar* and Ramyamol P C

Design of Optimal Reliability Acceptance


Sampling Plans for Exponential Distribution
DOI: 10.1515/eqc-2015-0005
Received August 25, 2015; revised December 11, 2015; accepted March 26, 2016

Abstract: In this paper, an attempt is made to derive the most efficient economic reliability sampling plans
for accepting a lot containing identical units having exponentially distributed lifetime with parameter θ. We
consider two types of sampling plans, namely, (a) sequential sampling plan (t1 , t2 ) and (b) repetitive group
sampling plan (n, t1 , t2 ). Under plan (a), the lot is rejected when the time between successive failures (Y r )
is less than t1 , and accepted when Y r ≥ t2 . The testing will continue for t1 ≤ Y r < t2 . Also we formulate an
optimization problem that minimizes the total expected testing cost. Under plan (b), four different criteria
are used to derive the sampling plan. The optimization problem formulated under each criterion is solved
using a genetic algorithm to obtain the plan parameters (n, t1 , t2 ). Several numerical examples are discussed
to illustrate our plans. In addition, a real example is also considered to demonstrate our plan. Finally, we
compare the cost of our plan with that of an existing plan in the literature. Our plan has significant potential
to reduce the testing cost by about 50%.

Keywords: Acceptance Sampling Plan, Exponential Distribution, Type I Censoring, Type II Censoring,
Optimization

MSC 2010: 62N02, 62N05, 90C26

1 Introduction
In recent years, due to increasing demand for highly reliable products in the market, manufacturers in the
industry are stressing the need for developing innovative techniques to test the quality and reliability of the
products developed. In addition, they require that new methods developed have minimum cost of testing,
and hence reduce the burden of huge budget for quality testing. To asses the quality of the product, they
consider many procedures. They need to make sure that the reliability of units should meet the requirements
of the consumer. It is important for them to arrive at good decision in a small duration of time, since testing
involves cost.
Many authors have considered different methods of reliability sampling plans and reliability testing
plans. The reliability sampling plans are generally based on Type I, Type II, Progressive, Hybrid or Mixture
of these censorings. The above mentioned testing procedures for obtaining sampling plans can be seen in
[1–7, 15, 20]. These authors have found the minimum sample size for the test without considering any cost.
They have considered two points on the OC curve and the sample size is obtained by solving inequalities
involving Type I and Type II errors. Epstein and Sobel [12] have considered truncated life testing. They dis-
cussed the case in which observations are available in an ordered manner and calculated the number of
failed items using the ratio of two points on the OC curve. Here time is not a constraint. Later, Epstein [11]
extended this idea to Hybrid censoring, to obtain sample size and the number of failures in a predefined

*Corresponding author: Mahesh Kumar: Department of Mathematics, National Institute of Technology Calicut,
PIN 673601, Kerala, India, e-mail: mahesh@nitc.ac.in
Ramyamol P C: Department of Mathematics, National Institute of Technology Calicut, PIN 673601, Kerala, India,
e-mail: ramyaputhenpurackal@gmail.com

Brought to you by | De Gruyter / TCS


Authenticated
Download Date | 6/2/16 6:29 PM
24 | M. Kumar and Ramyamol P C, Optimal Reliability Acceptance Sampling Plans

time-bound. Srinivas Rao [18] described a group acceptance sampling plan based on truncated lifetimes,
when the lifetime of an item follows a generalized exponential distribution. For a given group size, the mini-
mum number of groups and the acceptance number required are determined for a specified consumer’s risk
and the test termination time. The values of the operating characteristic function for various quality levels
are calculated, and the minimum ratios of the true average life to the specified life at given producer’s risk
are obtained. A group acceptance sampling plan for truncated life tests based on the inverse Rayleigh and
Log-logistic distributions are considered by Aslam and Jun [4]. When the test facilities are limited, Jun and
Balamurali [13] proposed a variable sampling plan in which they observe only the first failure in a group.
Balamurali and Usha [8] considered the variable chain sampling plan to find the minimum sample size for
testing by formulating an optimization problem that minimizes the average sample number.
Sherman [19] introduced the concept of repetitive group sampling. He introduced a new acceptance sam-
pling plan which has a simple design and operation procedure. The operation of the repetitive group sampling
plan is as follows:
1. Take a random sample of size n.
2. Count the number of defectives in the sample. Let it be d.
3. If d ≤ c1 , accept the lot. If d > c2 , reject the lot. If c1 < d ≤ c2 , repeat Steps 1, 2 and 3.
Sherman obtained the long run acceptance and rejection probabilities, and the number of samples to get a
decision through mathematical formulation. Jun et al. [14] proposed a variable repetitive group sampling plan
under Type II censoring when the lifetime of a unit follows Weibull distribution with known shape parameter,
but unknown scale parameter. They introduced a new acceptance sampling policy based on the ratio of the
average life to the specified time, and formulated an optimization problem to find the minimum number of
failures.
Aslam et al. [6] used the concept of repetitive group sampling plans for Weibull and generalized exponen-
tial distributions under Type I censoring. To design the proposed plan, they used median life as the quality
parameter and found the minimum average number in a sample.
Aslam et al. [5] discussed the repetitive group sampling plan based on truncated tests for Weibull models.
They considered an attribute repetitive group sampling plan based on truncated life tests, when the lifetime
of the product under testing follows the Weibull distribution with known shape parameter. They considered
only the number of failures observed in the interval [0, T]. The time of failure of items is not incorporated in
the sampling plan.
In this paper, we consider acceptance sampling plans for units in a lot having exponential failure time.
The repetitive group sampling under different types of censoring schemes and sequential sampling plans are
considered. The usual repetitive group sampling plan considers the number of failures in a sample during
testing. In this work, we consider the average life of units in the sample to make a decision to accept or reject
the lot. Hence our plan is based on testing the lifetime of units in the lot, and thereby accepting the lot. This
plan is different from available plans in the literature, where the decision as to accept or reject is merely
based upon the number of units failed in a sample. Thus, these existing plans do not incorporate the lifetime
information, and which is a draw back as long as the lot to be accepted contains units having lifetime which
will change eventually.
We have considered exponential distribution for modeling the lifetime of units in the lot. Since we deal
with failure times of items such as fuses, transistors, bulbs, etc., the failure is caused due to sudden shocks
rather than wear and tear. In such circumstances, the assumption of exponentiality on lifetime is partic-
ularly justified. For example, Proschan [17] reported on the times between failures of the air-conditioning
equipment in ten Boeing 720 aircrafts and proved that the data follows exponential distribution. Despite the
inadequacy of the exponential distribution to accurately model the behavior of some products in the real
world, it is still widely used in today’s reliability practices, standards and methods in industries. Moreover,
the usefulness of exponential distribution lies in its simplicity. That is, the mean of the distribution com-
pletely characterizes the lifetime distribution of units in a sample and is a sufficient metric. However, if the
data are modeled by some other distribution, then the mean is not sufficient to describe data, and in many
cases, a poor reliability metric. Also by Drenick’s theorem [10], the distribution of the time to first failure of a
subassembly is asymptotically exponential as the number of components on the subassembly gets large (and

Brought to you by | De Gruyter / TCS


Authenticated
Download Date | 6/2/16 6:29 PM
M. Kumar and Ramyamol P C, Optimal Reliability Acceptance Sampling Plans | 25

a condition that amounts to ‘no small number of failure rates dominates the group’ is satisfied). This is true
irrespective of the original time to failure distribution of the components. Also, if the lifetime distribution is
assumed as any two parameter family of distributions such as Gamma or Weibull, the distribution of the max-
imum likelihood estimator of the mean turns to be complicated and hence the corresponding optimization
problem becomes intractable. The use of simplest exponential distribution gives rise to acceptance sampling
plans which are based on formulating an optimization problem which minimizes the testing cost or maxi-
mizes the power of the test (1 − β). This is an innovative approach which is not addressed in the literature
(see [6, 14, 21, 22]).
The paper is organized as follows: In Section 2, we consider a sequential sampling plan, and define a pro-
cedure for accepting the lot based on the time between two successive failures. The plan parameters (t1 , t2 )
are obtained through an optimization problem which minimizes the total expected testing cost (ETC). In Sec-
tion 3, we discuss the repetitive group sampling plan under four different methods. The first method con-
siders a procedure to accept the lot based on the minimum of n observations in a random sample of size n.
The parameters involved in the plan are obtained through the optimization problem which minimizes the
ETC. Secondly, we consider the maximum of n observations as the criterion to obtain the acceptance sam-
pling plan. The plan parameters defined for accepting the lot are obtained through the optimization problem
which minimizes the ETC at a point on the OC curve. Next, Type I censoring is discussed, and the plan param-
eters defined for accepting the lot are obtained through the optimization problem which minimizes the ETC at
a point on the OC curve. Finally, under the repetitive group sampling plan, we introduce a notion of deriving
the acceptance sampling plan based upon the maximum power of the test (i.e., minimizing the probability
of Type II error). An optimization problem is formulated to obtain the plan parameters, by making use of the
repetitive group sampling plan under Type II censoring. In Section 4, some numerical results are discussed.
In addition, the results are compared with the existing sampling plan by Aslam et al. [5]. In Section 5, we
draw conclusions. It is observed that our plan has significant potential to reduce the testing cost by about
50%.

2 Sequential Sampling Plan and Model Development


Consider a lot of units having exponential failure time with probability density function given by

{ 1 exp( −x
θ ), x > 0, θ > 0,
f(x) = { θ (2.1)
0, otherwise.
{
Let θ0 denote the acceptable average life (AAL) and θ1 the unacceptable average life (UAL) of an item in the
lot. The decision on the lot as to accept or reject will be based upon the following probability requirements:

P(Reject the lot | θ ≥ θ0 ) ≤ α, (2.2)


P(Accept the lot | θ ≤ θ1 ) ≤ β, (2.3)

where α is the producer’s risk and β is the consumer’s risk. We use the concept of repetitive group sampling
described by Sherman [19] to obtain long-run acceptance and rejection probabilities. We test identical units
in the lot with replacement of failed units. Let X i denote the time of the i-th failure, i ∈ N. Our acceptance
sampling plan is based on the time between two successive failures, which is described as follows:
Let Y r+1 = X r+1 − X r , r = 1, 2, . . . and Y1 = X1 denote the time between successive failures. Note that Y r ,
r = 1, 2, 3, . . . follows exponential distribution with same parameter θ. Now define the acceptance sampling
plan (ASP) as follows: At any failure point,

continue the process if t1 ≤ Y r < t2 ,


accept the lot if Y r ≥ t2 ,
reject the lot if Y r < t1 .

Brought to you by | De Gruyter / TCS


Authenticated
Download Date | 6/2/16 6:29 PM
26 | M. Kumar and Ramyamol P C, Optimal Reliability Acceptance Sampling Plans

Our objective is to obtain the optimal pair (t1 , t2 ) such that the ETC is minimized subject to conditions (2.2)
and (2.3). The cost of testing is the cost obtained by multiplying the cost for testing an item for unit time with
the total time of testing. For each item, the time to failure is a random variable Y r with expectation θ. Then
the total time for testing is given by Y r times the number of items failed. Since Y r is a random quantity, the
total testing time is also a random quantity. Hence, we consider the ETC. Suppose the testing cost of an item
for unit time is C. Let p c , p a and p r denote the probabilities of continuing the process, accepting the lot and
rejecting the lot, respectively, while observing the failure of an item from the given lot. Then,
t1 t2
p c = P(t1 ≤ Y r < t2 ) = exp(− ) − exp(− ), (2.4)
θ θ
t2
p a = P(Y r ≥ t2 ) = exp(− ), (2.5)
θ
t1
p r = P(Y r < t1 ) = 1 − exp(− ). (2.6)
θ
pa pr
The long run acceptance and rejection probabilities are given by P A = 1−p c
and P R = 1−p c
, and the ex-
1
pected number of failed items is given by N = 1−p c (see [19]). Then the total expected testing cost is given
by

ETC = .
1 − pc

2.1 Problem Formulation and Solution Procedure

Since ETC is a function of θ, the optimal plan that minimizes ETC will depend on θ. The minimization of ETC
is difficult, since θ is unknown, and no prior distribution has been assumed. Hence, we consider the total
expected testing cost at θ0 , which is the value of θ under H0 . Our problem is to minimize the total expected
testing cost satisfying the Type I and Type II error requirements, and thereby to get optimal (t1 , t2 ). Since θ0
is the AAL, by the definition of ASP, we have t2 ≤ θ0 . Hence the optimization problem becomes

Cθ0
Min ( ) such that (2.7)
(t1 ,t2 ) 1 − pc
P(Reject the lot | θ ≥ θ0 ) ≤ α,
P(Accept the lot | θ ≤ θ1 ) ≤ β,

where t1 , t2 > 0 and t2 > t1 . That is,


Cθ0
Min ( ) such that
(t1 ,t2 ) 1 − pc
pr
( | θ ≥ θ0 ) ≤ α,
1 − pc
pa
( | θ ≤ θ1 ) ≤ β,
1 − pc

where t1 , t2 > 0 and t2 > t1 . We substitute for p c , p a and p r the terms given in (2.4)–(2.6) to get

Cθ0
Min ( ) such that
(t1 ,t2 ) 1 − exp(− θ0 ) + exp(− θt20 )
t1

1 − exp(− tθ1 )
( | θ ≥ θ0 ) ≤ α, (2.8)
1− exp(− tθ1 ) + exp(− tθ2 )
exp(− tθ2 )
( | θ ≤ θ1 ) ≤ β, (2.9)
1− exp(− tθ1 ) + exp(− tθ2 )

where t1 , t2 > 0 and t2 > t1 .

Brought to you by | De Gruyter / TCS


Authenticated
Download Date | 6/2/16 6:29 PM
M. Kumar and Ramyamol P C, Optimal Reliability Acceptance Sampling Plans | 27

Also note that constraints (2.8) and (2.9) are functions of θ, which is unknown. Hence, we consider the
following optimization problem:

Cθ0
Min ( ) such that
(t1 ,t2 ) 1− exp(− θt10 ) + exp(− θt20 )
1 − exp(− tθ1 )
max( ) ≤ α, (2.10)
θ≥θ0 1 − exp(− tθ1 ) + exp(− tθ2 )
exp(− tθ2 )
max( )≤β (2.11)
θ≤θ1 1 − exp(− tθ1 ) + exp(− tθ2 )

Since the rejection probability P R in (2.10) is decreasing with respect to θ and the acceptance probability P A
in (2.11) is increasing with respect to θ, the maximum of P R in (2.10) will occur at θ = θ0 , and the maximum
of P A in (2.11) will occur at θ = θ1 . Thus, the problem defined in (2.7) becomes problem P1:

Cθ0
Min ( ) such that
(t1 ,t2 ) 1 − exp(− θt10 ) + exp(− θt20 )
1 − exp(− θt10 )
≤ α,
1 − exp(− θt10 ) + exp(− θt20 )
exp(− θt21 )
≤ β,
1 − exp(− θt11 ) + exp(− θt21 )

where t1 , t2 > 0 and t2 > t1 .


The above optimization problem is solved by the genetic algorithm solver in Matlab. Table 1 in Section 4
presents the solution of problem P1.

3 Repetitive Group Sampling Plan


In this section, we consider the repetitive group sampling plan based on the lifetime of items in the given lot.
We assume that the lifetime follows the exponential distribution with parameter θ and having density given
by equation (2.1). We also consider AAL and UAL given in Section 2. Four different methods and repetitive
group sampling plan (RGSP) criteria for these methods are discussed below. Sampling is carried out with
replacement of surviving items into the lot.

3.1 Sampling Plan Based on Minimum of n Observations

In this section, the acceptance sampling plan is developed based on the time of the first failure among the
items in a random sample of size n. Here also the decision on the lot as to accept or reject will be based upon
the conditions given in (2.2) and (2.3). In this case our acceptance sampling plan is as follows: Let C denote
the cost of testing an item for unit time.
1. Take a random sample of size n. Let X i , i = 1, 2, . . . , n be i.i.d. random variables which represent the
lifetime of the i-th item. Let M = min(X1 , X2 , . . . , X n ). Then M is an exponential random variable with
parameter nθ .
2. If M < t1 , reject the lot. If M ≥ t2 , accept the lot. Else go to Step 3.
3. If t1 ≤ M < t2 , repeat Steps 1 and 2.
Our problem is to determine (n, t1 , t2 ) such that the total expected cost of testing is minimized subject to
conditions (2.2) and (2.3). The cost of testing is the cost obtained by multiplying the cost for testing an item
for unit time with the total time of testing. The time to get a decision, while testing a sample, is the random

Brought to you by | De Gruyter / TCS


Authenticated
Download Date | 6/2/16 6:29 PM
28 | M. Kumar and Ramyamol P C, Optimal Reliability Acceptance Sampling Plans

variable M, with expectation nθ . Note that,


nt1 nt2
p c = P(t1 ≤ M < t2 ) = exp(− ) − exp(− ), (3.1)
θ θ
nt2
p a = P(M ≥ t2 ) = exp(− ), (3.2)
θ
nt1
p r = P(M < t1 ) = 1 − exp(− ). (3.3)
θ
Here M and the number of samples are random quantities, the total testing time is also a random quantity
θ
with expectation n(1−p c)
. Hence the total expected testing cost is considered and is given by

ETC = .
n(1 − p c )
Problem Definition and Solution. Note that ETC is a function of θ (see Section 2.1), we consider ETC at θ0 .
Also by the above description of the ASP, we have t2 ≤ θ0 . So our optimization problem for minimizing the
ETC becomes
Cθ0
Min such that
(n,t1 ,t2 ) n(1 − p c )
P(Reject the lot | θ ≥ θ0 ) ≤ α,
P(Accept the lot | θ ≤ θ1 ) ≤ β,

where t1 , t2 > 0 and t2 > t1 . That is,


Cθ0
Min such that (3.4)
(n,t1 ,t2 ) n(1 − p c )
pr
( | θ ≥ θ0 ) ≤ α, (3.5)
1 − pc
pa
( | θ ≤ θ1 ) ≤ β, (3.6)
1 − pc
where t1 , t2 > 0 and t2 > t1 . We substitute for p c , p a and p r the terms given in (3.1)–(3.3). Also note that
the objective function in (3.4) and constraints (3.5), (3.6) are functions of θ, which is unknown. Hence, we
consider the following optimization problem P2:
Cθ0
Min ( ) such that
(n,t1 ,t2 ) n(1 − exp(− θ0 ) + exp(− nt
nt1 2
θ0 ))
1 − exp(− ntθ1 )
max( ) ≤ α, (3.7)
θ≥θ0 1 − exp(− ntθ1 ) + exp(− ntθ2 )
exp(− ntθ2 )
max( ) ≤ β, (3.8)
θ≤θ1 1 − exp(− ntθ1 ) + exp(− ntθ2 )
pr
where t1 , t2 > 0 and t2 > t1 . Since the rejection probability P R = 1−p c
in (3.7) is decreasing with respect to θ
pa
and the acceptance probability P A = 1−p c in (3.8) is increasing with respect to θ, the maximum of (3.7) will
occur at θ = θ0 , and the maximum of (3.8) will occur at θ = θ1 . Thus the optimization problem P2 reduces to
Cθ0
Min ( nt1
) such that
(n,t1 ,t2 )n(1 − exp(− θ0 ) + exp(− nt 2
θ0 ))
1 − exp(− nt 1
θ0 )
≤ α,
1− exp(− nt nt2
θ0 ) + exp(− θ0 )
1

exp(− nt 2
θ1 )
≤ β,
1− exp(− nt nt2
θ1 ) + exp(− θ1 )
1

where t1 , t2 > 0 and t2 > t1 . The optimization problem is solved by the genetic algorithm solver in Matlab.
Examples are given in Table 2 in Section 4.

Brought to you by | De Gruyter / TCS


Authenticated
Download Date | 6/2/16 6:29 PM
M. Kumar and Ramyamol P C, Optimal Reliability Acceptance Sampling Plans | 29

3.2 Sampling Plan Based on the Maximum of n Observations in a Sample


The maximum of n observations, while testing a random sample of size n, is taken into account for developing
the ASP. The acceptance or rejection of the lot will be based upon the following probability requirements:
P(Reject the lot | θ = θ0 ) ≤ α, (3.9)
P(Accept the lot | θ = θ1 ) ≤ β, (3.10)
The formulation of the ASP is as follows: Let C be the testing cost of a sample for unit time.
1. Take a random sample of size n. Let X i , i = 1, 2, . . . , n be i.i.d. random variables which represent the life-
time of the i-th item. Let M = max(X1 , X2 , . . . , X n ). Then P(M > t) = 1 − P(M ≤ t) = 1 − (1 − exp(− θt ))n .
M is an n-th order statistics.
2. If M < t1 , reject the lot. If M ≥ t2 , accept the lot. Else go to Step 3.
3. If t1 ≤ M < t2 , repeat Steps 1 and 2.
Then,
t2 n t1 n
p c = P(t1 ≤ M < t2 ) = (1 − exp(− )) − (1 − exp(− )) , (3.11)
θ θ
t2 n
p a = p(t2 ≤ M) = 1 − (1 − exp(− )) , (3.12)
θ
t1 n
p r = P(M < t1 ) = (1 − exp(− )) , (3.13)
θ
where (n, t1 ,t2 ) is determined so that the ETC at θ0 is minimum. The cost of testing is the cost obtained by
multiplying the cost for testing a sample for unit time with the total time of testing. Here M and the number
of samples are random quantities with expectations θ ∑ni=1 1i and 1−p 1
c
, respectively. The total testing time is
1 n 1
also a random quantity and its expected value equals 1−p c θ ∑i=1 i . Therefore the ETC is given by

Cθ ∑ni=1 1i
ETC = .
1 − pc
Modelling the Problem and Its Solution. The ETC is a function of θ (see Section 2.1). We consider the problem
of minimizing the ETC at θ0 . By the definition of the ASP given above, we have t2 ≤ θ0 . We get the following
optimization problem P3:
Cθ0 ∑ni=1 1
i
Min ( ) such that
(n,t1 ,t2 ) 1 − pc
P(Type I error | θ = θ0 ) ≤ α, (3.14)
P(Type II error | θ = θ1 ) ≤ β, (3.15)
where t1 , t2 > 0 and t2 > t1 . That is,
Cθ0 ∑ni=1 1
i
Min ( ) such that
(n,t1 ,t2 ) 1 − pc
pr
≤α at θ = θ0 ,
1 − pc
pa
≤β at θ = θ1 ,
1 − pc
where t1 , t2 > 0 and t2 > t1 . We substitute for p c , p a and p r the terms given in (3.11)–(3.13). Then we have
Cθ0 ∑ni=1 1
i
Min ( ) such that
(n,t1 ,t2 ) 1 − (1 − exp(− θt20 ))n − (1 − exp(− θt10 ))n
(1 − exp(− θt10 ))n
≤ α,
1 − (1 − exp(− θt20 ))n − (1 − exp(− θt10 ))n
1 − (1 − exp(− θt21 ))n
≤ β,
1 − (1 − exp(− θt21 ))n − (1 − exp(− θt11 ))n

Brought to you by | De Gruyter / TCS


Authenticated
Download Date | 6/2/16 6:29 PM
30 | M. Kumar and Ramyamol P C, Optimal Reliability Acceptance Sampling Plans

where t1 , t2 > 0 and t2 > t1 . This optimization can be solved using the genetic algorithm solver in Matlab.
Table 3 in Section 4 presents the solutions of the above defined optimization problem.

3.3 Sampling Plan under Type I Censoring


In this ASP, a random sample of size n from the given lot is taken and tested up to a predefined time. The
decision on the lot as to accept or reject will be based upon the conditions given in (3.9) and (3.10). In this
case our ASP is as follows: Suppose C is the cost of testing a sample for unit time.
1. Take a random sample of size n and test up to time T. Let r be the number of failures obtained with
corresponding lifetimes X1,n , X2,n , . . . , X r,n .
2. Calculate the maximum likelihood estimate of θ, which is given by
∑r X i,n + (n − r)T
θ̂ = i=i .
r
3. If t1 ≤ θ̂ < t2 , repeat Steps 1 and 2. Else go to Step 4.
4. If θ̂ < t1 , reject the lot. If θ̂ ≥ t2 , accept the lot.
But θ̂ is asymptotically normal with mean θ and standard deviation
θ
√ n(1 − exp( −T
θ ))

(see [9]). Then,

t1 − θ −T t −θ −T
p c = p(t1 ≤ θ̂ < t2 ) = p( √ n(1 − exp( )) ≤ Z < 2 √ n(1 − exp( ))), (3.16)
θ θ θ θ
t2 − θ −T
p a = p(t2 ≤ θ)̂ = p(Z ≥ √ n(1 − exp( ))), (3.17)
θ θ
t1 − θ −T
p r = p(θ̂ < t1 ) = p(Z < √ n(1 − exp( ))), (3.18)
θ θ
where Z is the standard normal random variable. Our problem is to determine (n, t1 ,t2 ) so that the ETC at θ0 is
minimum. Since the truncation time is T, there are two possibilities: (a) One can stop the testing at T. (b) One
can stop testing when all the items failed before time T. Let M be the time observed when all items have failed
for one particular sample. Then it is clear that M ≤ T. Note that the total number of samples to be tested is
1 T
1−p c . Therefore the total testing time is less than or equal to 1−p c . Since C is the cost of testing a sample for unit
CT
time, the total testing cost will always be less than or equal to 1−p c
, which is a random quantity. However, the
actual cost may still be less than this. Since it is difficult to minimize the random cost, we consider a problem
to minimize the expected value of the upper bound random testing cost (ETCub ). Hence the expected value of
the upper bound random testing cost becomes
CT
ETCub = .
1 − pc
Problem Description and Solution Procedure. Since ETCub is a function of θ, the optimal plan that minimizes
ETCub will depend on θ. But θ is unknown, and no prior distribution has been assumed. To overcome this
difficulty, we consider ETCub at θ = θ0 . Our aim is to minimize ETCub subject to conditions (3.9) and (3.10).
Since θ0 is the acceptable level of quality, we have t2 ≤ θ0 (using the definition of the ASP given above). Hence
we have the optimization problem P4 defined as follows:
CT
Min ( ) such that
(n,t1 ,t2 ) 1 − pc
p(Type I error | θ = θ0 ) ≤ α,
p(Type II error | θ = θ1 ) ≤ β,

Brought to you by | De Gruyter / TCS


Authenticated
Download Date | 6/2/16 6:29 PM
M. Kumar and Ramyamol P C, Optimal Reliability Acceptance Sampling Plans | 31

where t1 , t2 > 0 and t2 > t1 . That is,


CT
Min ( ) at θ = θ0 such that
(n,t1 ,t2 ) 1 − pc
pr
≤α at θ = θ0 ,
1 − pc
pa
≤β at θ = θ1 ,
1 − pc

where t1 , t2 > 0 and t2 > t1 . We substitute for p c , p a and p r the terms given in (3.16)–(3.18). Then we have

CT
Min ( ) such that
(n,t1 ,t2 )
(1 − p( t1θ−θ
0
0
√n(1 − exp( −T
θ0 )) <z< t2 −θ0
θ0
√n(1 − exp( −T
θ0 ))))

t1 −θ0
p(z < θ0
√n(1 − exp( −T
θ0 )))
≤ α,
1 − p( t1θ−θ
0
0
√n(1 − exp( −T
θ0 )) < z <
t2 −θ0
θ0
√n(1 − exp( −T
θ0 )))

t2 −θ1
p(z > θ1
√n(1 − exp( −T
θ1 )))
≤ β,
1 − p( t1θ−θ
1
1
√n(1 − exp( −T
θ1 )) < z <
t2 −θ1
θ1
√n(1 − exp( −T
θ1 )))

where t1 , t2 > 0 and t2 > t1 . This optimization problem P4 can be solved using the genetic algorithm solver
in Matlab. We give some examples in Table 4 in Section 4.

3.4 Sampling Plan under Type II Censoring

Here we consider a random sample of size n and test up to r failures. The decision on the lot as to accept or
reject will be based on RGSP criteria and the conditions given in (3.9) and (3.10). Here our ASP is as follows:
1. Take a random sample of size n and test up to r failures. Let X1,n , X2,n , . . . , X r,n be the lifetime of items
in increasing order.
2. Calculate the maximum likelihood estimate of θ, which is given by

∑r X i,n + (n − r)X r,n


θ̂ = i=i .
r

3. If t1 ≤ θ̂ < t2 , repeat Steps 1 and 2. Else go to Step 4.


4. If θ̂ < t1 , reject the lot. If θ̂ ≥ t2 , accept the lot.
The pair (t1 , t2 ) is determined so that the power of the test (1 − β) is maximum subject to the conditions given
̂
in (3.14) and (3.15). According to Epstein [11], 2rθθ follows the chi-square distribution with 2r degrees of
freedom. Therefore,
2rt1 2rt2
p c = p(t1 ≤ θ̂ < t2 ) = p( < χ2 (2r) < ),
θ θ
2rt2
p a = p(t2 ≤ θ)̂ = p( < χ2 (2r)),
θ
2rt1
p r = p(θ̂ < t1 ) = p r = p(χ2 (2r) < ).
θ
But
r−1
a ( a )i
p(χ2 (2r) > a) = ∑ exp(− ) 2 .
i=0
2 i!

Brought to you by | De Gruyter / TCS


Authenticated
Download Date | 6/2/16 6:29 PM
32 | M. Kumar and Ramyamol P C, Optimal Reliability Acceptance Sampling Plans

Hence the expressions for p a , p c and p r become


rt rt
r−1
rt1 ( θ1 )i r−1 rt2 ( θ2 )i
p c = ∑ exp(− ) − ∑ exp(− ) , (3.19)
i=0
θ i! i=0
θ i!
rt
r−1
rt2 ( θ2 )i
p a = ∑ exp(− ) (3.20)
i=0
θ i!
rt
r−1
rt1 ( θ1 )i
p r = 1 − ∑ exp(− ) . (3.21)
i=0
θ i!

Problem Statement and Solution Methodology. Here we consider the optimization problem P5 of maximiz-
ing the power of the test (1 − β), subject to the constraints of Type I error and Type II error (β) at the acceptable
and rejectable level of average life, respectively. Hence we have the optimization problem P5:

Max(1 − β) such that


p(Type I error | θ = θ0 ) ≤ α,
p(Type II error | θ = θ1 ) ≤ β,

where t1 , t2 > 0 and t2 > t1 .


Since the probability of Type I error is the long run rejection probability and the probability of Type II
error is the long run acceptance probability, the above optimization problem becomes

Max(1 − β) such that


pa
≤ α at θ = θ0 ,
1 − pc
pr
≤ β at θ = θ1 ,
1 − pc
where t1 , t2 > 0, t2 > t1 and t2 < θ0 (by the definition of the ASP given above).
Note that β is a function of the unknown θ. We consider the following problem:

Max (1 − β(θ1 )) such that


(t1 ,t2 ,θ1 )
pr
≤ α, θ = θ0 ,
1 − pc
pa
≤ β, θ = θ1 ,
1 − pc
where t1 , t2 > 0 and t2 > t1 . Substituting for p c , p a and p r the terms given in (3.19)–(3.21), we arrive at
rt
( θ 1 )i
1 − ∑r−1 rt1
i=0 exp(− θ1 )
1
i!
Max( rt rt
) such that
( θ 1 )i ( θ 2 )i
1 − ∑r−1 rt1
i=0 exp(− θ1 )
1
i! + ∑r−1 rt2
i=0 exp(− θ1 )
1
i!
rt
( θ 1 )i
1 − ∑r−1 rt1
i=0 exp(− θ0 )
0
i!
rt1 i rt2 i
≤ α, (3.22)
rt1 ( θ0 ) rt2 ( θ0 )
1− ∑r−1
i=0 exp(− θ0 ) i! + ∑r−1
i=0 exp(− θ0 ) i!
rt
( θ 2 )i
∑r−1 rt2
i=0 exp(− θ1 )
1
i!
rt1 i rt2 i
≤ β, (3.23)
rt2 ( θ1 ) rt2 ( θ1 )
1− ∑r−1
i=0 exp(− θ1 ) i! + ∑r−1
i=0 exp(− θ1 ) i!

where t1 , t2 > 0 and t2 > t1 .


This optimization problem P5 can be solved for (t1 , t2 ) using the following steps.
1. Find the minimum value of r in the feasible region given by (3.22) and (3.23).
2. Using the value of r obtained in Step 1, find the optimal pair (t1 , t2 ).
This optimization problem is solved by the genetic algorithm solver in Matlab. Examples are given in Table 5
in Section 4.

Brought to you by | De Gruyter / TCS


Authenticated
Download Date | 6/2/16 6:29 PM
M. Kumar and Ramyamol P C, Optimal Reliability Acceptance Sampling Plans | 33

4 Numerical Results, Comparisons and Case Study


In this section, we present numerical results. Tables 1–5 present numerical results for various sampling
plans. Table 6 presents comparisons of testing costs associated with two plans, namely, sequential sam-
pling plan and RGSP (minimum of n observations). In this comparison, for example, for the set of param-
eters θ0 = 300, θ1 = 50, α = 0.05, β = 0.05, one can see 93% cost reduction in the sequential sampling plan
as compared to cost involved in RGSP based on the minimum of n observations. In Table 7, comparisons of
testing costs incurred in RGSP based on the maximum of n observations and RGSP under Type I censoring,
are presented. Among these comparisons, for the set of parameters θ0 = 300, θ1 = 100, α = 0.05, β = 0.05,
there is about 90% cost reduction in RGSP under Type I censoring. These comparisons show that the RGSP
under Type I censoring has the minimum cost of testing.
Finally, we compare our results for RGSP under Type I censoring with the work of Aslam et al. [5]. For
example, when θ0 = 200, θ1 = 100, T = 100, α = 0.05, β = 0.1, our sampling plan gives an average sample
number of 22.92 and an expected testing cost of 156.53 (C = 1). The sampling plan derived in [5] gives an
average sample number of 24.9 and expected testing cost of 274.44. One observes about 50% cost reduction
in our plan as compared to that in [5].

4.1 Case Study: A Real Example

Consider a real example discussed by Noughabi [16]. The data given here are from an experiment in which
new models of a small electrical appliance were being tested. The appliances were operated repeatedly by
an automatic testing machine; the lifetimes given here are the number of cycles of use completed until the
appliances are failed. This data follows exponential distribution. Data: 170, 2694, 3034, 3214, 2400, 2471,
329, 13403, 381, 708, 7846, 1062, 1594, 1925, 1990, 2223, 2327, 2451, 2551, 2565, 2568, 2702, 11, 35,
2761, 2831, 3059, 3112, 3478, 3504, 4329, 6367, 6976, 1167, 49.
Let our plan parameters be θ0 = 3000, θ1 = 600, α = 0.1 and β = 0.2. We can solve the optimization
problems P1, P2, P3, P4 and P5 for the plan parameters t1 and t2 . We note that t1 and t2 are real numbers.
To apply our plan to the above given data, it is necessary that t1 and t2 must be integers. In view of these, the
problems P1, P2, P3, P4 and P5 are solved first for t1 and t2 . If t1 and t2 are integers, then we can accept the
lot satisfying the appropriate conditions of acceptance. Suppose t1 and t2 are not integers, we restrict t1 and
t2 as integers and solve the corresponding non-linear integer programming problem P1, P2, P3, P4 and P5
for integer solutions.
For the above data, the various plans discussed in our paper give the following results:
1. The sequential sampling plan will give t1 = 204, t2 = 1579. From the data set, we reject the lot, since the
first observation is less than t1 = 204.
2. The RGSP based on the minimum of n observations will give t1 = 7, t2 = 115, n = 21.4868. Then, from
the above data, we consider the first 22 observations. Here the minimum of these 22 observations is
greater than 115. Hence we can accept the lot.
3. The RGSP based on the maximum of n observations will give t1 = 1333, t2 = 1692, n = 2.4. Then, from
the above data, we consider the first 3 observations and the maximum of these observations is greater
than 1692. Hence we accept the lot.
4. The RGSP based on Type I censoring will give t1 = 1206, t2 = 1208, n = 11. Then, from the above data,
we consider the first 11 items. Next, find the average lifetime when the number of cycles is truncated at
T = 2000. Then the average lifetime of the 11 observations from the data is 1426.45, which is greater
than t2 = 1208. Hence we accept the lot.
5. The RGSP based on Type II censoring will give t1 = 900, t2 = 1508, r = 8. Now consider the first 8 fail-
ure times from the data, then the average of the truncated sample is 4060.125, which is greater than
t2 = 1508. Hence we accept the lot.

Brought to you by | De Gruyter / TCS


Authenticated
Download Date | 6/2/16 6:29 PM
34 | M. Kumar and Ramyamol P C, Optimal Reliability Acceptance Sampling Plans

(θ0 , θ1 ) (α, β) t1 t2 ETC at θ0

θ0 = 300, θ1 = 70 α = 0.1, β = 0.1 13.6317 275.0377 675.3388


θ0 = 300, θ1 = 50 α = 0.05, β = 0.05 6.9591 2756.1826 654.1617
α = 0.05, β = 0.1 8.0812 204.9714 564.3796
α = 0.05, β = 0.2 9.5 157.0392 481.0505
θ0 = 800, θ1 = 200 α = 0.1, β = 0.2 67.8962 526.3150 1523.4

Table 1. Sequential sampling plan (t1 , t2 ) for C = 1.

(θ0 , θ1 ) (α, β) n t1 t2 ETC at θ0

θ0 = 300, θ1 = 200 α = 0.05, β = 0.05 50 0 66.1840 1.1549 ⋅ 108


α = 0.05, β = 0.1 10 0.0001 298.8962 38958
θ0 = 500, θ1 = 300 α = 0.05, β = 0.05 30 0.0001 154.4318 160060
α = 0.05, β = 0.1 17 0.0012 217.7545 36232
θ0 = 1000, θ1 = 700 α = 0.05, β = 0.05 39 0 337.7583 1.0035 ⋅ 107
α = 0.05, β = 0.1 52 0 218.7054 1.3140 ⋅ 106
α = 0.1, β = 0.2 72 0.005 112.6939 4.4316 ⋅ 106

Table 2. RGSP using the minimum of n observations (n, t1 , t2 ) for C = 1.

(θ0 , θ1 ) (α, β) n t1 t2 ETC at θ0

θ0 = 300, θ1 = 50 α = 0.05, β = 0.05 4 196.9488 222.3383 637.3984


θ0 = 500, θ1 = 150 α = 0.05, β = 0.2 5 390.1137 495.1125 1098.7
α = 0.1, β = 0.15 3 222.9884 490.3024 974.6285
α = 0.1, β = 0.2 3 244.6331 429.3825 886.4062
θ0 = 10000, θ1 = 3000 α = 0.05, β = 0.2 5 8588 10000 23588

Table 3. RGSP using the maximum of n observations in a sample (n, t1 , t2 ) for C = 1.

(θ0 , θ1 ) (α, β) n t1 t2 ETCub at θ0

θ0 = 300, θ1 = 200, T = 50 α = 0.01, β = 0.01 33 237.1385 237.1385 50


α = 0.05, β = 0.05 30 246.5220 246.520 50
θ0 = 300, θ1 = 200, T = 100 α = 0.01, β = 0.01 28 234.6239 234.6239 100
α = 0.01, β = 0.05 46 227.4490 227.4490 100
θ0 = 3000, θ1 = 1500, T = 1500 α = 0.01, β = 0.01 36 1910.6 1924.4 1500

Table 4. RGSP under Type I censoring (n, t1 , t2 ) for C = 1.

(θ0 , θ1) (α, β) r t1 t2 1−β

θ0 = 300, θ1 = 200 α = 0.01, β = 0.01 37 187.7484 300 0.990715


α = 0.01, β = 0.05 26 168.2504 300 0.94999
α = 0.01, β = 0.05 27 171.5401 299.99 0.94999
θ0 = 9000, θ1 = 6000 α = 0.01, β = 0.01 37 5632.9 9000 0.998805
α = 0.01, β = 0.05 26 5073.0 9000 0.95153
α = 0.05, β = 0.05 21 5578.4 8952.1 0.950107
θ0 = 3000, θ1 = 1500 α = 0.01, β = 0.01 12 1228.8 2972.6 0.990051
α = 0.01, β = 0.05 9 1001.2 3000 0.9560
α = 0.01, β = 0.1 8 967.3 2864.6 0.90758

Table 5. RGSP under Type II censoring (r, t1 , t2 ).

Brought to you by | De Gruyter / TCS


Authenticated
Download Date | 6/2/16 6:29 PM
M. Kumar and Ramyamol P C, Optimal Reliability Acceptance Sampling Plans | 35

Sampling plan A B

θ0 = 300, θ1 = 50, α = 0.05, β = 0.05 654.162 41.0905


θ0 = 500, θ1 = 100, α = 0.05, β = 0.1 1224.4 50.0841
θ0 = 800, θ1 = 200, α = 0.1, β = 0.2 1537.3 60.6574

Table 6. Comparison of expected testing costs in the sequential sampling plan (column A) and
expected testing costs in the RGSP based on the minimum of n observations (column B). C = 1.

Sampling plan C D

θ0 = 300, θ1 = 50, α = 0.05, β = 0.05 2068.2 250


θ0 = 500, θ1 = 100, α = 0.05, β = 0.1 1224.4 300
θ0 = 800, θ1 = 200, α = 0.1, β = 0.2 1537.3 400

Table 7. Comparison of expected testing costs in the RGSP based on the maximum of n observations (column C)
and upper bound of expected testing costs involved in the RGSP under Type I censoring (column D). C = 1.

Sampling plan D E

θ0 = 200, θ1 = 100, T = 100, α = 0.05, β = 0.25 156.53 274.44


θ0 = 3000, θ1 = 1500, T = 1500, α = 0.05, β = 0.1 2290.35 3728.57

Table 8. Comparison of upper bound of expected testing costs involved in the RGSP under Type I censoring (column D)
and expected testing costs involved in the plan derived by Aslam et al. [5] (column E). C = 1.

5 Conclusions
Most of the existing literature have addressed the development of sampling schemes based on the number
of failures. The optimization problems were formulated for obtaining the average sample number. However,
no sampling scheme addresses the need of incorporating lifetime information of units in the lot. In our work,
we have considered the sampling plan based on the lifetime of items, and the optimization problem is formu-
lated for obtaining the total expected cost of testing, which has not been addressed so far in the literature. As
compared to the sampling plan derived by Aslam et al. [5], the sample size and the number of samples to be
taken has been reduced in our plan by considering various optimization problems. Since the cost involved in
our plans is random, except for RGSP under Type II censoring, the ETC has been considered and minimized.
For example, when θ0 = 200, θ1 = 100, T = 100, α = 0.05, β = 0.1, our sampling plan gives an average sam-
ple number of 22.92 and an expected testing cost of 156.53 (C = 1). The sampling plan derived by Aslam
et al. [5] gives an average sample number of 24.9 and an expected testing cost of 274.44. The cost reduction
in our plan is 50% compared to that in [5]. Table 6 presents comparisons of testing costs associated with
two plans, namely, sequential sampling plan and RGSP (minimum of n observations). In this comparison,
for example, for the set of parameters θ0 = 300, θ1 = 50, α = 0.05, β = 0.05, 93% cost reduction can be seen
in the sequential sampling plan as compared to the cost involved in the RGSP based on the minimum of n
observations. Table 7 compares the testing costs incurred in the RGSP based on the maximum of n observa-
tions and the RGSP under Type I censoring. Among these comparisons, for the set of parameters θ0 = 300,
θ1 = 100, α = 0.05, β = 0.05, there is about 90% cost reduction in the RGSP under Type I censoring. These
comparisons show that the RGSP under Type I censoring has the minimum cost of testing. A real example is
also added to demonstrate the practical application of our acceptance sampling plans. As a future scope of
work, one may investigate the sampling scheme having finite number of samples to be tested, in order to take
a decision on acceptance of lot.

Acknowledgment: The authors wish to express their thanks to the editor and the reviewer for their sugges-
tions to improve the presentation and quality of the paper.

Brought to you by | De Gruyter / TCS


Authenticated
Download Date | 6/2/16 6:29 PM
36 | M. Kumar and Ramyamol P C, Optimal Reliability Acceptance Sampling Plans

References
[1] A. I. Al-Omari, Acceptance sampling plan based on truncated life tests for three parameter kappa distribution, Econ. Qual.
Contr. 29 (2014), no. 1, 53–62.
[2] M. Aslam, Double acceptance sampling based on truncated life-tests in Rayleigh distribution, Eur. J. Sci. Res. 17 (2007),
605–611.
[3] M. Aslam and C.-H. Jun, A group acceptance sampling plan for truncated life test having Weibull distribution, J. Appl. Stat.
36 (2009), 1021–1027.
[4] M. Aslam and C.-H. Jun, A group acceptance sampling plans for truncated life tests based on the inverse Rayleigh and
log-logistic distributions, Pakistan J. Statist. 25 (2009), 1–13.
[5] M. Aslam, C.-H. Jun, A. J. Fernandez, M. Ahmad and M. Rasool, Repetitive group sampling plan based on truncated tests
for Weibull models, Res. J. Appl. Sci. Engrg. Technol. 7 (2014), 1917–1924.
[6] M. Aslam, S. T. A. Niaki, M. Rasool and M. S. Fallahnezhad, Decision rule of repetitive acceptance sampling plans assuring
percentile life, Sci. Iranica 19 (2012), 879–884.
[7] M. Azam, M. Aslam, S. Balamurali and A. Javaid, Two stage group acceptance sampling plan for half normal percentiles,
J. King Saud Univ. Sci. 27 (2015), 239–243.
[8] S. Balamurali and M. Usha, Optimal design of variable chain sampling plan by minimizing the average sample number,
Int. J. Manufac. Engrg. 3 (2013), 1–10.
[9] D. J. Bartholomew, The sampling distribution of an estimate arising in life testing, Technometrics 5 (1963), 361–374.
[10] R. F. Drenick, The failure law of complex equipment, J. Soc. Indust. Appl. Math. 8 (1960), 680–689.
[11] B. Epstein, Truncated life tests in the exponential case, Ann. Math. Statist. 25 (1954), no. 3, 555–564.
[12] B. Epstein and M. Sobel, Life testing, J. Amer. Statist. Assoc. 48 (1953), 486–502.
[13] C.-H. Jun, S. Balamurali and S.-H. Lee, Variable sampling plans for Weibull distribution lifetimes under sudden death test-
ing, IEEE Trans. Reliab. 55 (2006), 53–58.
[14] C.-H. Jun, L. Hyeseon, S.-H. Lee and S. Balamurali, A variables repetitive goup sampling plan under failure-censored relia-
bility tests for Weibull distribution, J. Appl. Stati. 37 (2010), 453–460.
[15] R. R. Kantam, K. Rosaiah and G. Srinivas Rao, Acceptance sampling plans based on life tests: Log-logistic model, J. Appl.
Stat. 28 (2001), 121–128.
[16] H. A. Noughabi, Testing exponentiality based on the likelihood ratio and power comparison, Ann. Data Sci. 2 (2015),
195–204.
[17] F. Proschan, Theoretical explanation of observed decreasing failure rate, Technometrics 5 (1963), 315–383.
[18] G. S. Rao, A group acceptance sampling plans for lifetimes following a generalized exponential distribution, Econ. Qual.
Contr. 24 (2009), 75–85.
[19] R. E. Sherman, Design and evaluation of a repetitive group sampling plan, Technometrics 7 (1965), 11–21.
[20] T. R. Tsai and S. J. Wu, Acceptance sampling based on truncated life test for generalized rayleigh distribution, J. Appl. Stat.
33 (2006), 595–600.
[21] C.-W. Wu, M. Aslam, J. C. Chen and C.-H. Jun, A repetitive group sampling plan by variables inspection for product accep-
tance determination, Eur. J. Indust. Engrg. 9 (2015), 308–326.
[22] C.-H. Yen, C.-H. Chang and M. Aslam, Repetitive variable acceptance sampling plan for one-sided specification, J. Stat.
Comput. Simul. 85 (2015), 1102–1116.

Brought to you by | De Gruyter / TCS


Authenticated
Download Date | 6/2/16 6:29 PM
View publication stats

S-ar putea să vă placă și