Professional Documents
Culture Documents
To cite this article: Qianmei Feng & Kailash C. Kapur (2006) Economic Design of Specifications
for 100% Inspection with Imperfect Measurement Systems, Quality Technology & Quantitative
Management, 3:2, 127-144, DOI: 10.1080/16843703.2006.11673105
Article views: 4
Keywords: 100% inspection, bivariate quality characteristics, inspection error, optimal specifications,
single quality characteristics.
______________________________________________________________________
1. Introduction
with respect to specification limits, which are decision variables in this model. Here, ETC
is the expected total cost per unit; EQL is the expected quality loss per unit; ESC is the
expected scrap cost per unit; and EIC denotes the expected inspection cost per unit.
Based on this general optimization model, different models have been formulated by
Kapur [6], Kapur and Wang [7], and Kapur and Cho [9] [10]. In terms of the inspection
error, Hong and Elsayed [4], Tang and Schneider [14] discussed the effects of inspection
error on a complete inspection plan for a single quality characteristic. They assumed that the
inspection is unbiased and the imprecision of measurement is independent of the true value
of the quality characteristic. Using the similar assumptions on measurement precision and
additional assumptions on the inspection costs associated with inspection error, we present
models to develop the optimal specifications to minimize the expected total cost per unit
shipped rather than per unit produced. Furthermore, the model for bivariate quality
characteristics is proposed with the consideration of inspection errors.
L ( y ) = k ( y − y0 ) 2 . (2)
Since the quality loss function is symmetric about y0 , the process mean should be
centered at the target, i.e., µ = y0 , and the specification limits are also symmetric about the
process mean, consisting of a lower limit µ − δ1 and an upper limit µ + δ1 . Hence, if the
observed value of Y is outside the interval ( µ − δ1 , µ + δ1 ) , the item is rejected and
scrapped. The expected quality loss ( EQL1 ) and expected scrap cost ( ESC1 ) are carefully
derived and given below:
⎡ ⎛δ ⎞ 2δ ⎛ δ ⎞ ⎤
EQL1 = kσ 2 ⎢ 2Φ ⎜ 1 ⎟ − 1 − 1 φ ⎜ 1 ⎟ ⎥ ,
⎣ ⎝σ ⎠ σ ⎝ σ ⎠⎦
Economic Design of Specifications for 100% Inspection 129
⎡ ⎛ δ ⎞⎤
ESC1 = SC ⎢ 2 − 2Φ ⎜ 1 ⎟ ⎥ ,
⎣ ⎝ σ ⎠⎦
where SC is the scrap cost per unit, Φ (⋅) and φ (⋅) are the cumulative distribution and
probability density functions respectively for a standard normal variable.
If we denote the fraction of shipped unit as q , we have
⎛δ ⎞
q = ∫µµ−+δδ11 f ( y )dy = 2Φ ⎜ 1 ⎟ − 1.
⎝σ ⎠
Thus the expected total cost per shipped unit is
1
ETC1 = ( EQL1 + ESC1 + IC )
q
1 ⎧ 2 ⎡ ⎛ δ1 ⎞ 2δ1 ⎛ δ1 ⎞ ⎤ ⎡ ⎛ δ ⎞⎤ ⎫
= ⎨kσ ⎢ 2Φ ⎜ ⎟ − 1 − φ ⎜ ⎟ ⎥ + 2SC ⎢1 − Φ ⎜ 1 ⎟ ⎥ + IC ⎬ , (3)
⎛ δ1 ⎞ ⎩ ⎣ ⎝ σ⎠ σ ⎝ σ ⎠⎦ ⎣ ⎝ σ ⎠⎦ ⎭
2Φ ⎜ ⎟ − 1
⎝σ ⎠
where IC is the inspection cost per unit. The minimum value of ETC1 with respect to δ1
can be obtained by optimization methods.
2.2. Model 2: With Inspection Error and Constant Inspection Cost
When considering the inspection error, the observed value of a quality characteristic is
different from the true value of the same quality characteristic. We denote the observed
value of Y as X . The conditional distribution of X given Y is assumed as a normal
distribution, X |Y ∼ N ( y,σ e2 ) , which implies that the observed value is unbiased and the
variability of measurements is independent of Y as assumed by Hong and Elsayed [4],
Tang and Schneider [14]. The variance of the measurements, σ e2 , can be estimated using an
experiment where the potential sources of variation are changed systematically as described
by Czitrom and Spagon [2]. Given the conditional probability density function, g ( x | y ) ,
the joint distribution of X and Y is
h ( x, y ) = g ( x | y ) f ( y ) ,
where
( y − µ )2 ( x − y )2
−
1 − 1 2σ e2
f ( y) = e 2σ 2 and g ( x | y ) = e .
2πσ 2πσ e
h( x , y )
p( y|x ) = .
m( x )
Let α = 1/σ e2 and β = 1/σ 2 , we proved that both X and Y | X are normal variables, and
130 Feng and Kapur
⎛ α x + βµ −1 ⎞
X ∼ N ( µ ,1/α + 1/β ) and Y | X ∼ N ⎜ , (α + β ) ⎟ .
⎝ α +β ⎠
Because of the inspection error, the inspection result or the observed value X is used
as the criterion to determine whether or not an item is rejected, rather than the true value Y .
As a result, if the specification limits are ( µ − δ 2 , µ + δ 2 ) , the rejection area and acceptance
area can be shown in Figure 1.
Using the same quality loss function in Equation (1), the expected quality loss per item
is calculated for the acceptance area as
= ∫µµ−+δδ22 ∫−∞
∞
k ( y − y0 ) 2 p ( y | x )m ( x )dydx
⎡⎛ α x + βµ ⎞
2
1 ⎤
= ∫µµ−+δδ22 km ( x ) ⎢⎜ − y0 ⎟ + ⎥ dx
⎢⎣⎝ α + β ⎠ α + β ⎥⎦
kα 2 µ +δ 2 2 2 kα ⎛ βµ ⎞
= 2 ∫µ −δ 2
x m ( x )dx + ⎜ − y0 ⎟ ∫µµ−+δδ22 xm ( x )dx
(α + β ) α +β ⎝α + β ⎠
⎡⎛ βµ ⎞
2
1 ⎤ µ +δ 2
+ k ⎢⎜ − y0 ⎟ + ⎥ ∫µ −δ 2 m ( x )dx .
⎢⎣⎝ α + β ⎠ α + β ⎥⎦
For a normal random variable X ∼ N ( µ ,1/α + 1/β ) with p.d.f m ( x ) , the following
formulas are derived:
Economic Design of Specifications for 100% Inspection 131
µ +δ ⎡ ⎛ δ2 ⎞ ⎤
∫µ −δ 22 xm ( x )dx = µ ⎢ 2Φ ⎜ ⎟ − 1⎥ ,
⎢⎣ ⎝ σ m ⎠ ⎥⎦
µ +δ ⎡ ⎛ δ2 ⎞ ⎤ ⎛ δ2 ⎞
∫µ −δ 22 x m ( x )dx = ( µ + σ m ) ⎢ 2Φ ⎜ ⎟ − 1⎥ − 2δ 2σ mφ ⎜
2 2 2
⎟,
⎣⎢ ⎝ σ m ⎠ ⎥⎦ ⎝σm ⎠
where σ m = 1/α + 1/β is the standard deviation of X . Then the expected quality loss,
EQL2 is calculated as
3
k ⎤⎡ ⎛ δ ⎞ ⎤ 1
⎡ ⎛ α ⎞2 −2 ⎛ δ2 ⎞
EQL2 = ⎢ k ( µ − y0 ) 2 + ⎥ ⎢ 2Φ ⎜ 2 −
⎟ ⎥1 − 2 k δ 2⎜ ⎟ β φ⎜ ⎟.
⎣ β ⎦ ⎣⎢ ⎝ σ m ⎠ ⎦⎥ ⎝α + β ⎠ ⎝σm ⎠
If the process mean is centered at the target value, µ = y0 , then the expected quality
loss is simplified as
3
k ⎡ ⎛ δ2 ⎞ ⎤
1
⎛ α ⎞2 − 2 ⎛ δ2 ⎞
EQL2 = ⎢ 2Φ ⎜ ⎟ − 1⎥ − 2 kδ 2 ⎜ ⎟ β φ⎜ ⎟
β ⎢⎣ ⎝ σ m ⎠ ⎥⎦ ⎝α + β ⎠ ⎝ σm ⎠
If we denote the fraction shipped to the customer as q , we have
⎛δ ⎞
q = ∫µµ−+δδ22 m ( x )dx = 2Φ ⎜ 2 ⎟ − 1,
⎝ σm ⎠
and the expected scrap cost is ESC 2 = SC (1 − q ).
1
ETC 2 = ( EQL2 + ESC 2 + IC )
q
⎡ 3
⎤
k 1 ⎢
1
⎛ α ⎞2 −2 ⎛ δ2 ⎞ ⎛ ⎛δ ⎞⎞ ⎥
= − 2 kδ 2 ⎜ ⎟ β φ⎜ ⎟ + 2SC ⎜ 1 − Φ ⎜ 2 ⎟ ⎟⎟ + IC ⎥ . (4)
β ⎢ ⎜
⎛δ ⎞
2Φ ⎜ 2 ⎟ − 1 ⎣⎢ ⎝α + β ⎠ ⎝σm ⎠ ⎝ ⎝σm ⎠⎠
⎦⎥
⎝σm ⎠
ETC 2 with respect to δ 2 is obtained by global search methods in the numerical example
illustrated later.
1
ETC 3 = [ EQL3 + ESC 3 + I c (σ e )]
q
132 Feng and Kapur
⎡ 3
⎤
k 1 ⎢
1
⎛ α ⎞2 −2 ⎛ δ3 ⎞ ⎛ ⎛ δ3 ⎞⎞ ⎥ . (5)
= − 2 kδ 3 ⎜ ⎟ β φ⎜ ⎟ + 2SC ⎜⎜ 1 − Φ ⎜ + I (
⎟ ⎟⎟ c e ⎥σ )
β ⎛ δ3 ⎞ ⎢ ⎝ α+β ⎠ ⎝σm ⎠ ⎝ ⎝σm ⎠⎠
2Φ ⎜ ⎟ − 1 ⎢⎣ ⎥⎦
⎝σm ⎠
1⎧ ⎡ ⎛δ ⎞ δ ⎛ δ ⎞⎤ ⎡ ⎛ δ ⎞⎤ ⎫
Min ETC1 = ⎨5(0.5) 2 ⎢ 2Φ ⎜ 1 ⎟ − 1 − 2 1 φ ⎜ 1 ⎟ ⎥ + 2 × 2.00 ⎢1 − Φ ⎜ 1 ⎟ ⎥ + 0.10 ⎬
q⎩ ⎣ ⎝σ ⎠ σ ⎝ σ ⎠⎦ ⎣ ⎝ σ ⎠⎦ ⎭
⎛δ ⎞
Subject to δ1 ≥ 0 , where q = 2Φ ⎜ 1 ⎟ − 1.
⎝σ ⎠
The optimal solution found by genetic algorithm [5] is: δ1∗ = 0.79, and
ETC1∗ = $1.110 . Since $1.110 < $1.250 , we should do the 100% inspection with the
optimal specification limits (9.21,10.79) , when there is no inspection error.
Model 2: 100% inspection with inspection error and constant inspection cost
It is reported that the measurement machine used for inspecting the fill volume of cans
is not perfect, and the standard deviation caused by inspection error is σ e = 0.05 . Thus we
have α = 1/σ e2 = 400 , β = 1/σ 2 = 4 and σ m = 1/α + 1/β = 1/ 400 + 1/ 4 . Using
Equation (4), the optimization model is given as below:
5 1⎡ ⎤
3
⎛ 400 ⎞ 2 − 12 ⎛ δ 2 ⎞ ⎛ ⎛ δ2 ⎞⎞
Min ETC 2 = − ⎢2 × 5δ 2 ⎜ ⎟ 4 φ ⎜ ⎟ + 2 × 2 . 00 ⎜ 1 − Φ ⎜ ⎟ ⎟ + 0. 10 ⎥
⎜ ⎟
4 q⎢
⎣ ⎝ 404 ⎠ ⎝σm ⎠ ⎝ ⎝ σm ⎠⎠ ⎥⎦
⎛δ ⎞
Subject to δ 2 ≥ 0 , where q = 2Φ ⎜ 2 ⎟ − 1.
⎝ σm ⎠
ETC2*
1.35
1.3
1.25 ETC0
1.2
1.15
1.1
ETC1* 0 0.2 0.4 0.6 0.8 1 1.2
σ e/σ
USL & LSL
12
11.5
USL
11
10.5
10
9.5
8.5
LSL
8
0 0.2 0.4 0.6 0.8 1 1.2
σe/σ
ETC2*
1.5
0.5
0
0 20 40 60 80 100 120
SC/IC
USL & LSL
12
USL
11
10
9
LSL
8
0 20 40 60 80 100 120
SC/IC
Figure 3. The effects of SC /IC on ETC2∗ , LSL and USL.
ETC2*
1.5
0.5
0
0 0.5 1 1.5 2 2.5
SC/k
USL & LSL
12
USL
11
10
9
LSL
8
0 0.5 1 1.5 2 2.5
SC/k
Model 3: 100% inspection with inspection error and variable inspection cost
In many cases, a company has choices as for the selection of the measurement system.
Suppose we have seven measurement systems to choose from for inspection, and each
system has different precision level with different inspection cost as given in Table 1. The
inspection cost is a non-increasing discrete function of σ e rather than a constant, which is
also shown in Figure 5. An optimization procedure using genetic algorithm is implemented
to minimize the expected total cost with this discrete function in Equation (5). The resulted
ETC 3∗ for each measurement system is shown in Figure 5. The optimal ETC 3∗ depends
heavily on the magnitude of I c (σ e ) . We can see that the minimum expected total cost is
obtained by using the fourth measurement system, whose imprecision is σ e = 0.040 . The
corresponding optimal specifications are (9.21, 10.79).
ETC3* Ic(σ e)
1.135 0.125
1.13
ETC3* 0.12
0.115
1.125 0.11
1.12 0.105
Ic(σe) 0.1
1.115
0.095
1.11 0.09
0 0.02 0.04 0.06 0.08 0.1 0.12
σe
Figure 5. ETC ∗
3 and I c (σ e ) for seven measurement systems.
1 ⎧⎪ 1 ⎡ ( y1 − µ1 ) 2 ( y1 − µ1 )( y2 − µ2 ) ( y2 − µ2 ) 2 ⎤ ⎫⎪
f ( y1 , y2 ) = exp ⎨− 2 ⎢
− 2 ρ + ⎥ ⎬,
2πσ 1σ 2 1− ρ2 ⎪⎩ 2(1 − ρ ) ⎣ σ 1
2
σ 1σ 2 σ 22 ⎦ ⎪⎭
σ 12
where ρ is the correlation of y1 and y2 , and ρ = .
σ 1σ 2
Unlike the case for a single quality characteristic, bivariate quality characteristics form
a specification region by two sets of specification limits, that is, ( µi − δ i , µi + δ i ) for i =1, 2.
The item is rejected if any value of y1 , y2 is outside the interval ( µi − δ i , µi + δ i ) . A
bivariate quality loss function is considered for evaluating quality for such systems [10]:
L ( y1 , y2 ) = k1 ( y1 − t1 ) 2 + k12 ( y1 − t1 )( y2 − t 2 ) + k2 ( y2 − t 2 ) 2 ,
where k1 , k12 , and k2 are positive constants, t1 and t 2 are the target values of the two
quality characteristics. The process means can be adjusted to the target values, and thus we
have µ1 = t1 and µ2 = t 2 . From now on, we use Y1 and Y2 to denote the quality
deviations between the quality characteristics and the target values to facilitate the derivation.
Without loss of generality, the bivariate quality loss function is rewritten as
L ( y1 , y2 ) = k1 y12 + k12 y1 y2 + k2 y22 .
The expected quality loss can be computed by EQL11 = ∫−+δδ22 ∫−+δδ11 L ( y1 , y2 ) f ( y1 , y2 )dy1dy2 ,
where y1 and y2 are the deviations between the quality characteristics and the target
values. The model to develop the specification region based on the expected total cost per
shipped unit is
1
Min ETC11 = [ EQL11 + SC (1 − q ) + IC ] , (6)
q
where EQL11 = ∫−+δδ22 ∫−+δδ11 ( k1 y12 + k12 y1 y2 + k2 y22 ) f ( y1 , y2 )dy1dy2 , q = ∫−+δδ22 ∫−+δδ11 f ( y1 , y2 )dy1dy2 .
A similar model was developed for the expected total cost per produced unit by Kapur
and Cho [10], but they did not pursue the case with consideration of inspection error.
⎛ σ e21 σ e12 ⎞
X |Y = y ∼ N 2 ( y,Σ e ) , where y = ( y1 , y2 ), and Σ e = ⎜ ⎟.
⎜σ 2 ⎟
⎝ e12 σ e 2 ⎠
If we denote the probability density function of conditional distribution X |Y= y as
g ( x | y ) , then the joint distribution of X and Y , h( x, y ) , is calculated as
h( x , y ) = g ( x | y ) f ( y ).
h ( x, y ) g (x|y) f ( y)
p ( y|x ) = = .
m(x ) m(x )
One desirable property of normal multivariate distribution is that marginal
distributions and conditional distributions derived from multivariate normal distributions are
also normal distributions, see Anderson [1]. We proved in the Appendix A that
X ∼ N 2 (0, (Σ + Σ e )) , and Y | X ∼ N 2 ((Σ −1 + Σ e−1 ) −1 Σ e−1 x , (Σ −1 + Σ e−1 ) −1 ) . The observed values
X have bigger variability than the quality characteristics Y .
Because of the inspection error, the observed values or the inspection results X are
the criteria to determine whether or not an item is rejected, rather than Y . As a result, the
expected quality loss per unit using the bivariate quality loss function is
⎣ ⎦ ⎣ ⎦
+ k12 [ Σ12 + l1 ( x1 , x 2 )l2 ( x1 , x 2 ) ]}m ( x )dx1dx 2 .
1
ESC 22 = ( EQL22 + ESC 22 + IC ) (7)
q
=
1 δ 2 δ1
q
{
∫−δ 2 ∫−δ1 k1 ⎡⎣ (l1 ( x1 , x 2 )) + Σ11 ⎤⎦ + k2 ⎡⎣ (l2 ( x1 , x 2 )) + Σ 22 ⎤⎦
2 2
} ( ) 1 1
⎣ 12 + l1 ( x1 , x 2 ) l 2 ( x1 , x 2 ) ⎤⎦ m x dx1dx 2 + q SC (1 − q ) + q IC ,
+ k12 ⎡Σ
⎛ 0.5 0.3795 ⎞
Σ=⎜ ⎟.
⎝ 0.3795 0.8 ⎠
1
Min ETC11 = [ EQL11 + SC (1 − q ) + IC ] ,
q
where EQL11 = ∫−+δδ22 ∫−+δδ11 (30 y12 + 10 y1 y2 + 25 y22 ) f ( y1 , y2 )dy1dy2 , q = ∫−+δδ22 ∫−+δδ11 f ( y1 , y2 )dy1dy2 ,
Economic Design of Specifications for 100% Inspection 139
1 ⎧⎪ −1 ⎛ y12 y1 y2 y22 ⎞ ⎫⎪
f ( y1 , y2 ) = exp ⎨ 2 ⎜⎜ − 2(0 . 6) + ⎟⎬ ,
2π 0.5 0.8 1 − 0.62 ⎪⎩ 2(1 − 0.6 ) ⎝ 0.5 0.5(0.8) 0.8 ⎟⎠ ⎭⎪
0 ≤ δ1 ≤ 6.0σ 1 , 0 ≤ δ 2 ≤ 6.0σ 2 .
The function DBLQUAD in Matlab is used to evaluate the double integrals for the
above bivariate normal distribution. The optimal solutions we found is δ1∗ = 2.20σ 1 ,
δ 2∗ = 2.02σ 2 . Thus the optimum specification limits for Y1 and Y2 are (8.444,11.556) ,
(18.193, 21.807) . With this specification region, the minimum ETC ∗ is $37.479 , which is
smaller than $38.795 , the minimum total cost from Model 0. When there is no inspection
error, we should do the 100% inspection with the specification limits (8.444,11.556) and
(18.193, 21.807) for Y1 and Y2 , respectively.
Model 2: 100% inspection with inspection error
It is reported that the measurement machine used for inspection is not perfect. We
assume that the observed values are unbiased and the variabilities of measurements are
independent of the true values. The variance matrix of the conditional distribution of
X1 , X 2 given Y1 ,Y2 is
⎛ 0.01 0 ⎞
Σe = ⎜ ⎟,
⎝ 0 0.04 ⎠
which implies that the two observed variables are independent. Thus, the marginal
distribution of the observed variables is
⎛ ⎛ 0.51 0.3795 ⎞ ⎞
X ∼ N 2 ⎜ 0, ⎜ ⎟ ⎟.
⎝ ⎝ 0.3795 0.84 ⎠ ⎠
Therefore, we have σ x21 = 0.51 , σ x22 = 0.84 and the correlation between X 1 and X 2
is ρ x = 0.5798 . Following the notations given in Appendix B, we can calculate the mean
vector and the variance matrix of the conditional distribution Y given X , which are
=
1 δ 2 δ1
q
{
∫−δ 2 ∫−δ1 30 ⎣⎡( l1 ( x1 , x 2 ) ) + 0.0097 ⎤⎦ + 25 ⎡⎣( l 2 ( x1 , x 2 ) ) + 0.0371⎤⎦
2 2
80 2
+10 [ l1 ( x1 , x 2 )l2 ( x1 , x 2 ) + 0.0005]} m ( x )dx1dx 2 + (1 − q ) + ,
q q
4. Conclusions
As a means of providing protection for both the producer and the customer, 100%
inspection plays an important role for industrial or other decision making processes where
the consequences of excessive deviations for target values are very high. When inspection
error exists due to imperfect measurement systems, the effect of passing a nonconforming
item or rejecting a conforming item can be enormous.
In this paper, we investigated the economic and statistical effects of inspection error on
the design of specifications for both single quality characteristic and bivariate quality
characteristics. For single quality characteristic, we reviewed the case without inspection
error as Model 1, and developed two additional models, Model 2 and Model 3, with
consideration of inspection error and associated inspection cost. Based on the practical
situation (with or without inspection error consideration; constant or variable inspection
cost), one of the three models can be used to make effective decisions for inspection.
For bivariate quality characteristics, we discussed the case without inspection error first,
and then proposed a model for the case with the consideration of inspection error. An
example is used to illustrate the decision-making process for bivariate quality characteristics.
The model for multiple quality characteristics can be developed in a similar manner.
Acknowledgement
The authors wish to acknowledge the helpful comments and suggestions of the referees
whose suggestions on earlier versions of this work have significantly improved this paper.
References
1. Anderson, T. W. (2003). An Introduction to Multivariate Statistical Analysis. 3rd edition.
John Wiley & Sons, New York.
2. Czitrom, V., Spagon, P. D. (1997). Statistical Case Studies for Industrial Process Improvement.
Philadelphia, PA: Society for Industrial and Applied Mathematics; Alexandria, VA:
American Statistical Association.
3. Golhar, D. Y. (1987). Determination of the Best Mean Contents for a Canning
Problem. Journal of Quality Technology, 19(1), 82-84.
4. Hong, S. H., Elsayed, E. A. (1999). The optimum mean for processes with normally
distributed measurement error. Journal of Quality Technology, 31(3), 338-344.
5. Houck, C. R., Joines, J. A., Kay, M. G. (1995). A genetic algorithm for function
optimization: a Matlab implementation. NCSU-IE Technical Report, 95-09.
6. Kapur, K. C. (1987). Quality loss function and inspection. Proceeding of TMI, Quality
Control, Innovations in Quality: Concepts and Applications, Detroit, MI.
7. Kapur, K. C. and Wang, D. J. (1987). Economic design of specifications based on
Taguchi’s concept of quality loss function. Proceeding of American Society of Mechanical
Engineers, Winter Annual Meeting, Boston, MA.
Economic Design of Specifications for 100% Inspection 141
8. Kapur, K. C. (1988-89). An approach for development of specifications for quality
improvement. Quality Engineering, 1(1), 63-78.
9. Kapur, K. C. and Cho, B. (1994). Economic design and development of specifications.
Quality Engineering, 6(3), 401-417.
10. Kapur, K. C. and Cho, B. R. (1996). Economic design of the specification region for
multiple quality characteristics. IIE Transactions, 28(3), 237-248.
11. Papadakis, E. P. (1985). The deming inspection criterion for choosing zero or 100
percent inspection. Journal of Quality Technology, 17(2), 121-127.
12. Taguchi, G. (1986). Introduction to Quality Engineering: Designing Quality into Products and
Processes. Asian Productivity Organization, Tokyo.
13. Taguchi, G. (1987). System of Experimental Design, Volume I and II. Quality Resources,
NY and American Supplier Institute, MI.
14. Tang, K. and Schneider H. (1987). The effects of inspection error on a complete
inspection Plan. IIE Transactions 19(4), 421-428.
15. Tang, K. and Tang, J. (1994). Design of screening procedures: a review. Journal of
Quality Technology, 26, 209-226.
⎛ σ 2 σ 12 ⎞
Σ=⎜ 1 ⎟.
⎜σ 2 ⎟
⎝ 12 σ 2 ⎠
The conditional distribution of X 1 , X 2 given Y1 , Y2 is also a bivariate normal variate,
X |Y = y ∼ N 2 ( y,Σ e ) , where y = ( y1 , y2 ), and
⎛ σ 2 σ e12 ⎞
Σe = ⎜ e1 ⎟.
⎜σ 2 ⎟
⎝ e12 σ e 2 ⎠
1 ⎡ 1 ⎤ 1 ⎡ 1 ⎤
where f ( y ) = 1/ 2
exp ⎢ − y tΣ −1 y ⎥ , and g ( x | y ) = 1/ 2
exp ⎢ − ( x − y ) t Σ e−1 ( x − y ) ⎥ .
2π | Σ | ⎣ 2 ⎦ 2π | Σ e | ⎣ 2 ⎦
( ) ( )
t t
y Σ −1 y + x − y Σ e−1 x − y
( )
t t t t
= y Σ −1 + Σ e−1 y − x Σ e−1 y − y Σ e−1 x + x Σ e−1 x
t
( )
= ⎡ y − Σ −1 + Σ e−1 Σ e−1 x ⎤ Σ −1 + Σ e−1 ⎡ y − Σ −1 + Σ e−1 ( ) ( ) Σ e−1 x ⎤
−1 −1
⎢⎣ ⎥⎦ ⎢⎣ ⎥⎦
t ⎡
+ x ⎢ I − Σ e Σ −1 + I
⎣
((
−1 t ⎤ −1
⎥ Σe x
⎦
) )
t
(
= ⎡ y − Σ −1 + Σ e−1 ) (
Σ e−1 x ⎤ Σ −1 + Σ e−1 ⎡ y − Σ −1 + Σ e−1 ) ( ) Σ e−1 x ⎤
−1 −1
⎣⎢ ⎦⎥ ⎢⎣ ⎦⎥
t
+ x ( Σ + Σe ) x.
−1
The last step is valid due to the fact that Σ and Σ e are symmetric matrixes. Thus the joint
distribution of X and Y is rewritten as
h( x , y ) = g ( x | y ) f ( y )
⎧ 1 t
⎫
exp ⎨− ⎡ y − Σ −1 + Σ e−1 ( ) (
Σ e−1 x ⎤ Σ −1 + Σ e−1 ⎡ y − Σ −1 + Σ e−1 ) ( ) Σ e−1 x ⎤ ⎬
−1 −1
⎢ ⎥⎦ ⎢⎣ ⎥⎦
= ⎩ 2⎣ ⎭
−1 1 / 2
(
2π Σ −1 + Σ e−1 )
×
{
exp − 12 x t ( Σ + Σ e ) x
−1
}.
1/ 2
2π Σ + Σ e
1 ⎧ 1 −1 ⎫
∞
m ( x ) = ∫−∞ h ( x , y )dy = 1/ 2
exp ⎨− x t ( Σ + Σ e ) x ⎬ ,
2π Σ + Σ e ⎩ 2 ⎭
because the first term in h( x, y ) is a probability density function of a bivariate normal
variable N 2 ((Σ −1 + Σ e−1 ) −1 Σ e−1 x , (Σ −1 + Σ e−1 ) −1 ) .
It means the marginal distribution of X follows a bivariate normal distribution
X ∼ N 2 (0, (Σ + Σ e )) , and the conditional distribution Y given X also follows a bivariate
normal distribution Y | X ∼ N 2 ((Σ −1 + Σ e−1 ) −1 Σ e−1 x , (Σ −1 + Σ e−1 ) −1 ).
⎛Σ Σ12 ⎞
Σ = ⎜ 11 ⎟,
⎝ Σ12 Σ 22 ⎠
⎛ Σ11 Σ12 ⎞
⎜ ⎟.
⎝ Σ12 Σ 22 ⎠
E [Y1 | X1 , X 2 ] = l1 ( x1 , x 2 ) ,
V [Y1 | X 1 , X 2 ] = Σ11 ,
E [Y2 | X1 , X 2 ] = l 2 ( x1 , x 2 ) ,
V [Y1 | X 1 , X 2 ] = Σ 22 ,
cov[Y1 ,Y2 | X1 , X 2 ] = Σ12 .
The above information is used to simplify the expected quality loss EQL22 , which is
The double integration over y1 and y2 can be separated into three parts and we first
have
∞ ∞ 2
∫−∞ ∫−∞k1 y1 p ( y1 , y2 | x1 , x 2 )dy1dy2
∞ ∞
= ∫−∞ k1 y12 ∫−∞ p ( y1 , y2 | x1 , x 2 )dy2 dy1
∞
= ∫−∞ k1 y12 q ( y1 | x1 , x 2 )dy1
= k1 E [Y12 | X1 , X 2 ]
= k1 ( E [Y1 | X1 , X 2 ]2 + V [Y1 | X1 , X 2 ])
= k1[(l1 ( x1 , x 2 )) 2 + Σ11 ],
= k2 E ⎡⎣Y22 | X 1 , X 2 ⎤⎦ = k2 ⎡⎢( l 2 ( x1 , x 2 ) ) + Σ 22 ⎤⎥ ,
2
⎣ ⎦
144 Feng and Kapur
and
{
EQL22 = ∫−δδ2 2 ∫−δδ1 1 k1 ⎡⎢( l1 ( x1 , x 2 ) ) + Σ11 ⎤⎥ + k2 ⎡⎢( l2 ( x1 , x 2 ) ) + Σ 22 ⎤⎥
⎣
2
⎦ ⎣
2
}
⎣ 12 + l1 ( x1 , x 2 ) l 2 ( x1 , x 2 ) ⎤⎦ m ( x ) dx1dx 2 .
+ k12 ⎡Σ
Authors' Biographies:
Qianmei Feng is an Assistant Professor in the Department of Industrial Engineering at the
University of Houston, TX. Her research interests are quality and reliability engineering,
and applied statistics. She received the Ph.D. degree in Industrial Engineering from the
University of Washington, Seattle, WA in 2005. She is a member of IIE, INFORMS and
Alpha Pi Mu.
Kailash C. Kapur is a Professor, Industrial Engineering, University of Washington, Seattle,
Washington. He was the Director of Industrial Engineering at the University of Washington
from January 1993 to September 1999. He was a Professor and the Director of the School of
Industrial Engineering, University of Oklahoma (Norman, Oklahoma) from 1989-1992 and
a professor in the Department of Industrial and Manufacturing Engineering at Wayne State
University, Detroit, Michigan from 1970-1989. Dr. Kapur has worked with General Motors
Research Laboratories as a senior research engineer and with Ford Motor Company as a
visiting scholar and the U.S. Army, Tank-Automotive Command as a reliability engineer. Dr.
Kapur received the Ph. D. degree (1969) in Industrial Engineering from the University of
California, Berkeley. He has co-authored the book Reliability in Engineering Design, John
Wiley & Sons, 1977. He has written chapters on reliability and quality engineering for
several handbooks such as Industrial Engineering and Mechanical Design. He has published
over sixty papers in technical, research, and professional journals. He received the Allan
Chop Technical Advancement Award from the Reliability Division and the Craig Award
from the Automotive Division of American Society for Quality. He is a Fellow of American
Society for Quality, a Fellow of the Institute of Industrial Engineers, and a registered
professional engineer.