Sunteți pe pagina 1din 4

Computers ind. Engng Vol. 19, Nos 1-4, pp. 254-257, 1990 0360-8352/90 $3.00 + 0.

00
Printed in Great Britain. All rights reserved Copyright © 1990 Pergamon Press pie

M E T H O D O L O G Y FOR T O L E R A N C E DESIGN USING QUALITY LOSS FUNCTION

Kailash C. Kapur, Director & Professor


Shivakumar Raman, Assistant Professor
P. Simin Pulat, Assistant Professor

School of Industrial Engineering


The University of Oklahoma
202 West Boyd, Suite 124
Norman, OK 73019

ABSTRACT For example y can be an assembly dimension which is a function


of component dimensions x 1 . . . . . x n.
In this paper, we develop a methodology for tolerance design that
focuses on total variability from the target value. A general
optimization model is developed to allocate tolerances to assembly
parts. The constraint for the optimization model is the requirment
for assembly tolerances or variability and cost is related to the total
cost to reduce variance for the components and loss due to variation
from target for the assembly. Different quality loss functions are
@
also discussed. @ @ . ) y = quality characteristic
of the system
INTRODUCTION
@.
One of the most important tasks in any quality and manufacturing
program is to understand and evaluate the needs and/or expectations
-@
of the customer/users and then provide products and services that
meet and/or exceed these expectations. Quality of the manufactured
product is defined and evaluated by the customer. One of the first
steps in the design process is to translate the Voice of the
Customer into substitute quality characteristics, design
configurations, design parameters and technological characteristics Figure 1. System Configuration
that can be deployed through the whole organization. Products
have several characteristics and an "ideal" state or value of these
characteristics must be determined from the customer's viewpoint. Let Yo denote the target value for y. It is impossible to
This ideal state is called the target value. Due to manufacturing manufacture products at target due to process capabilities and
variation and other uncontrollable factors in the customer's variabilities, material imperfections, human error and other
environment, it is impossible to produce every product on target uncontrollable factors. Hence, an acceptable range for y placed
and have the products function in the ideal state. The variation of around Yo is specified as the tolerance zone for the product. Several
the characteristic from the target will result in degradation of innovative ideas have been implemented over the years in
function of product and result in some disutility from the view allocating and analyzing tolerances as evidenced by Chase and
point of the customer. This disutility may be measured in the Greenwood (1988). A common practice is that a product with y
form of losses to the producer and the user - thus the society. value outside the tolerance zone is characterized as "bad" where as a
Tolerances are a mechanism to control the variation. The purpose product with y value within the tolerance zone is characterized as
of this paper is to develop a rational basis to develop tolerances to "good", although the two y values may be very close to each other
minimize total losses to the society. as shown in Figure 2.
One of the concepts that is used to evaluate quality of the
manufactured product is conformance to spec~cations. Tolerances
are allowable variances defined by design specifications. Item t (Bid)
Everything that is within the specification limits is called good or
conforming and units outside the specification limits are bad or Item2 (Good)
nonconforming. Many systems to evaluate the quality of
manufactured product focus on the nonconforming units only.
Although in fact the difference in quality of an item that is just
outside the specification limit and an item that is just within the
specification limit may not be very much, we call one item "bad"
and the other item "good". A better way is to evaluate the quality
of all the items, both within as well as outside the specifications.
Thus, we evaluate the total population using a quality loss
function as we deviate from an ideal or target value. Lower
I
. - - ToleranceZone
for y
Upl~r Sj ¢cification
SlXCirtc~on Limit.
Limit (USL)
P R O B L E M DEFINITION (LSL)

Consider a system of n components (Xl, x 2 , . . . , Xn) as the one


shown in Figure 1. The system or assembly performance y is a
function of the characteristics of the components and is given by Figure 2. Binary Evaluation Using Specification
Limits
y = f(x 1, x2. . . . x,) (1)

254
K a p u t et al.: Methodology for Tolerance Design 255

A better evaluation of quality of a product is through the use of a Let us evaluate L for different quality characteristics. If y is
quality loss function (Chen and Kapur, 1989). We apply these "smaller the better" then a "good" approximation of L(y) is
concepts to tolerance allocation problem.
L (y) ffiky 2 y>0 (6)
The following discussion on quality loss function is based on Then
Kapur (1989). Further details are also given in Kapur & Wang
(1987), Kaput (1988) and Tagnchi (1986, 1987). L = E [L(Y)] = k [(E[Y]) 2 + V[Y]] (7)

Q U A L I T Y LOSS FUNCTION Let E[Y] = It and V[Y] = o 2, then


Let Ll(Y ) be a measure of losses, disutility, failure rate,
Lfk~2+o 2] (8)
degradation, etc., of the quality associated with value y. Let us
assume that L(y) is a differentiable function (in the neighborhood
The advantage of the quadratic less function is that we can evaluate
of yo) and using Taylor's series expansion, we have
losses in terms of average and variance. If the function L(y) is
other than quadratic, then we will have higher order moments of
2 random variable Y in order to evaluate expected losses. Thus, the
L l (y) = Ll(Yo) + L~(yo)(y - Yo) + L~(yo)~ +... (2) above function (8), gives equal weight to bias (deviation of It from
the ideal target, which is zero) and standard deviation, o. Thus, in
order to reduce expected losses, we have to reduce bias as well as o.
The target Yo should be developed so that totalquality losses are It should be pointed out that the countermeasures to reduce bias
may be very different than countermeasures to reduce variance. The
minimum at Yo' and hence L~(yo)= 0. Thus, Ll(Yo) depends on above loss function combines the average and variance together.
We may be able to do separate optimization for both of these
the selection of the target value Yo and it is very important to fred measures of quality.
the best value of Yo to minimize quality losses. If we are
evaluating only losses due to variationfrom Yo' then let us define For "nominal the best" quality characteristic, we can use Eq. (4) as
2
an aPlxoximation for loss function and we have
L(y) = L l(y) - L l(yo) = L"(yo) ~ + ••• (3)
L(y) = k(y- yo)2 (9)

Hence, if we ignore the higher order terms, we may approximate in I'IgIIce


the neighborhood of Yo' L(y) by a quadratic function and hence we
L = E[L (Y)] = k E[(Y- yo)2] or (10)
have

L = k [It - yo)2 + 0"2]


L(y) "-"k(y - y,,)2 (4)
where Thus, to reduce losses, we must reduce bias (It - yo) and variance
L"(yo) 0 2 . Again we can develop other functions by giving different
2 weights to bias and standard deviation o.

Itshould be pointed out that we can use the actual function L(y) if Thus we focus on the total distribution rather than the items that
we wish (See Figure 3). Let f(y) be the probability density are outside the specifications in traditional quality system. In the
function for the random variable Y. Then, we can compute the traditional system, as soon as we have all the item within the
expected losses and these are given by limits, we have achieved "Zero Defects (ZD) "or perfection. In the
proposed system, we focus on reducing bias as well as variance.
Thus, this evaluation system improves quality on a never ending
L = ElL(Y)] = ~.,, L(y) f(y) dy (5) basis or promotes the philosophy of continual improvement.
Hence, we propose to develop the philosophy of tolerance design
Thus, the expected loss is based on all values of y and not just the based on reduction of variation from a target value rather than just
values of y that are outside the specifications. This is quite meeting specifications or tolerances.
different than evaluating only the cost or losses due to
nonconformance. Thus, in order to improve quality, we have to O P T I M I Z A T I O N MODEL
change the total distribution of y and not just look at items outside
a given set of specification limits. Quality loss function may be Consider an assembly between a shaft and a hole with dimensions
used as a means to evaluate improvements in the total distribution shown in Figure 4 One can define the functional relationship of
of Y, not just the tails of the distribution. the assembly as:

y = f(xl, x2) = x I - x2 (11)

I
L(y)
Let Yo be the targeted value for y. Variation of y due to several
factors result in a probability density function(pal0 f(y). Assuming
L(y)
normal distribution, + 30 limits on y will account for the 99.73%
f(Y) of the variability.

Yo

Figure 3. Quality Loss Function. Figure 4. Shaft Assembly


256 Proceedings of the 12th Annual Conference on Computers & Industrial Engineering

Let ty define the specified allowable deviation from the target Now, let us consider the following tolerance allocation model:
value, Yo' Then Yo + ty is the tolerance zone defined on y. The
n
problem is to allocate tolerances t 1 and t2 to x 1 and x2 such that
Min. TC = ~ c i (°i)
99.73% of the variability in the resultant y will be within the i=l
specified tolerance zone. Figure 5 describes the relationship of the
tolerance zone to the standard deviation of y, assuming that Ixy = S.t.
Yo"

fly) . ( 2 f (x 1. . . . . x,) /2

where x = (x 1. . . . . Xn) and Ix = (111. . . . . P-n)"


I I
I I The above model aims at determining optimal component
tolerances while observing assembly tolerance limits (determined

"% by o~) in a manner which minimizes total cost.

IX -t ix=y. If we are given the cost functions c i (oi), i = 1. . . . . n and o~yfor


the quality characteristic, we can solve the above problem using
optimization techniques. The authors are working on the
Figure 5. Variability in y and the tolerance zone. development of special algorithms which exploit the special
structure of the problem.
Let fl (Xl) and f2 (x2) be the pdf's of x I and x 2 and IX1'P'2 and o 1,
0 2 be the associated means and standard deviations, respectively. It should also be pointed out that we have to figure out how to
Since for the hole and shaft assembly problem, we have
develop target for o~y. If bias is equal to zero, then using equation
y=x 1 -x 2 (10), we have
L = k 02 (16)
then IZy = t.t1 - I.t2 (12)
Thus, in addition to considering cost due to variance of the
2 2 components, we should also consider the cost or losses due to o-~y.
O~Y= 01 + 02
Then, the general optimization model is as follows:
A simple model is
Min TC = ~ c (Gi) + k o~
th th ty (13) i=l
o, = T °2 = T o, = 5-
subject to
Thus the optimization problem is to find the values of cr 1 and G2
given a specified value for Oy. Given cost functions associated
~f~f(x I . . . . . Xn) }2
with 01 and 02 , we can optimize the total cost to achieve a given (17)
requirement for Oy. In real life there is a cost (or loss) function
associated with Oy and hence we have an iterative optimization
oi>0 foralli, a y > 0
problem. We now present the general optimization problem.

Using, eq. (1), we can approximate the variance of the random For discussion on development of the constant k, see Kapur
variable y given variance of x 1. . . . . x n using Taylor's series (1989). The authors plan to apply the above model to simple
approximation (Kaput and Lamberson, 1977) and this is given by engineering design problems and demonstrate its applicability.

2 CONCLUSIONS

(14) The quality of a product should not be defined using the concept of
i= I ~ ~Xi x=p. conformance to specifications. This binary system of evaluation is
very simplistic. This is also a barrier to continual improvement.
The tolerance design approach presented in this paper is based on
the total distribution of the quality characteristic. If the disutility
or loss function is quadratic, then the optimization can be done
where using the first two moments of the distribution. In this paper, we
x = (x 1. . . . . Xn) and IX = (It I . . . . . t.tn). presented a general optimization model in terms of costs associated
with variances of the components and losses (costs) associated with
The total cost is the sum of the costs associated with the the variability from the target for the quality characteristic.
components. Total cost is made up of a fixed part and a variable
part which depends on the tolerance associated with that
component. Several cost-tolerance relationships appear in literature
with the exponential cost curve being the most widely used
function. Let e 1 (or1) and c 2 (02) denote the costs associated with
x 1 and x 2, respectively. Then, the total cost TC for our example
is given by
TC = c 1 (o'l) + c 2 (02)
Kapur et al.: Methodology for Tolerance Design 257

REFERENCES

1. Chase, K.W. and Greenwood, W.H. (1988). "Design Issues in


Mechanical Tolerance Analysis," Manufacturing Review,
Vol. 1, No. 1, pp. 50-59.
2. Chen, G. and Kapur, K.C. (1989). "Quality evaluation
systems using loss function", The Proceedings of 1989
International Industrial Engineering Conference. Toronto,
Canada, May 1989.
3. Kapur, K.C. and Lamberson, L.R. (1977), Reliability in
Engineering Design, John Wiley & Sons, N.Y.
4. Kapur, K.C. and Wang, C.1. (1987). "Economic Design of
Specifications Based on Taguchi's Concept of Quality Loss
Function," Proceedings of ASME Symposium on
Quality: Design, Planning & Control, Boston, MA,
Dec. 1987.
5. Kapur, K.C. (1988). "Product and Process Design
Optimization by Design of Experiments using Taguchi
Method," Proceedings of Society of Automotive
Engineers, SAE Earth Moving Conference, Technical
Paper No. 880821, April 1988.
6. Kapur, K.C. (1989). "An Approach for Development of
Specifications for Quality improvement," Q u a l i t y
Engineering, Vol. 1, No. 1, pp. 63-77.
7. Taguchi, G. (1986). Introduction to Quality Engineering.
UNIPUB/Kraus International Publications, White Plains, NY
and American Supplier Institute Inc., Dearborn, MI.
8. Taguchi, G. (1987). System of Experimental Design, English
Translation. Vol. 1 and Vol. 2. UNIPUB/Kraus International
Publications, White Plains, NY and American Supplier
Institute Inc., Dearborn, MI.
ACKNOWLEDGEMENTS

This research was partially supported by NASA, George C.


Marshall Space Flight Center.

S-ar putea să vă placă și