Loss Functions-Lecture 5

THE QUALITY LOSS FUNCTION
Quality ultimate system performance measure Variability relates to quality Variability increases quality lost Can this loss be measured?
The concept % defective has been widely used as a measure of quality level When defective product units are not shipped it should not be considered a quality problem but a cost problem How to evaluate the quality level of products shipped to customers is the problem of concern.
10/08/2007
ENGN8101 Modelling and Optimization
Loss Function and Quality Level
In the past, % defective, process capability index and warranty cost have been used as measures of quality level for shipped products. One major weakness of the process capability index is that there is no apparent immediate basis for specifying the optimal value of Cp. % defective or warranty costs are understandable because they are monetary related measures
2
10/08/2007
How far from target can a system be before it should be rejected and changed? Different to customer tolerance o point at which customers have to take economic action because of off-target performance Often o incurs greater costs than manufacturing limit
10/08/2007
Quality Loss Concept
Deviation from target results in loss.
Lower than target Greater than target Both lose
10/08/2007
How can we use Quality Loss Concept in real life?

Lets assume that you are an Operations Manager at a company that produces custom made doors and door frames. During the winter, due to cold weather, doors tend to shrink which lets in cold air through the cracks between the door and the door frame. During the summer doors tend to expand beyond its normal size due to hot weather, which makes to door hard to open because it rubs against the door frame. Your job is to produce a door, where dimensions (length and width of the door) are set to a specific target level. In other words in the Winter the door cannot let any cold air in the house and it should open properly in the Summer.
10/08/2007
How can we use Quali concept in rea/life

Average Temp Winter Temp
.s:
0
0 Q)
.s:
0
0 Q)
.
M
.
LD
co
co
Traditional Quality Metric

All products within specifications equality good. All products outside specifications equally bad.
Equally unacceptable
All products equally good
LSL
10/08/2007
USL
7
Loss Function (Cont.)

Unfortunately, this definition has led to a mindset which becomes a barrier to improvement in our industry. We have come to view all products which fall within specification limits as being of equal quality. Consider the following:
10/08/2007
loss Function (Con
Or. Taguchf defines and q antfffes quality loss via Functio"'. He un"ftes the financial loss with the specification through a quadratic relationship that come Taylor Series Expans f on (wh fch f s an approxi Tatfon method} have a more sensitive/accurate analytical loss function. t Taylor Expansion ill be unnecessary. However. we ay quadratic expansion if there fs no other function available.
To apply the concept of loss function we need to know the co loss at the limit of functional specification.
TARGET
LS.L
U.S.L.
"' o red Valwe of the Cl\ara tereatc
TV Example
Consider a comparison between the quality of color television sets produced by two factories belonging to the same manufacturing company. One factory (A) is located in Japan, and the other factory (B) in America. Suppose the comparison was based on color concentration, which relates to the color balance of the television sets. Although both factories used the same design, the television sets produced in the American factory had lower quality, and consumers consequently preferred products made in Japan.
10/08/2007
10
TV Example (Cont.)
The figure given in the next slide shows the differences in quality characteristic (i.e., color concentration) distributions. The figure shows that the quality distribution of the Japanese-made television sets (shown by the solid curve) is approximately a normal distribution with a target value at the center; its standard deviation is about 1/6 of the tolerance, which in this case equals 10. In quality control, the index of tolerance divided by 6 standard deviations is called the process capability index, denoted by Cp. Cp=tolerance/(6*standard deviation)
11
10/08/2007
TV Example (Co
Tightened tolerance
Tolerance
Factory B
,I
'\
I I r
I I
l
I
m-a
I
m
'
m+
Distribution of color concentration
TV Example (Cont.)

The process capability of the Japanese-made television sets is 1. On the other hand, the quality distribution of the Americanmade television sets (shown in the figure by a dotted curve) has less out-of-specification products than the Japanese made products and is quite similar to the uniform distribution for those products that are within the tolerance limits. Since the standard deviation of the uniform distribution is given by 1/ 12 of the tolerance, the process capability index for these sets is given by Cp=tolerance/(6*(tolerance/ 12))=0.577
10/08/2007
13
QUALITY LOSS FUNCTION Quadratic quality loss function relates quality loss in dollars L(y) to the deviation away from a targeted value (m) of a measured response value (y) such that i.e. |y-m| L(y) = k(y-m)2
If m is achieved. . loss is zero
10/08/2007
14
Derivation of the Loss Function

Assume the loss due to a defective part (because of discarding, repairing, or downgrading) is A. then denote the loss function by L(y) and expand it in a Taylor series about the target value m: L(y)= L(m+y-m)
L (m) L (m) or L(y)=L(m)+ 1! (y-m)+ 2! (y-m)2 +
10/08/2007
15
Derivation of the Loss Function (Cont.)

Because L(y)=0 when y=m (by definition, quality loss is zero when y=m), and the minimum value of the function is attained at this point, its first derivative with respect to m is zero. The first two terms of the equation then, are equal to zero. When we neglect terms with powers higher than 2, the equation reduces to
L (m) L(y)= (y - m)2 2!

or L(y)= k (y - m)2 where k is a proportionality constant.
10/08/2007
16
Relationship between quality loss and deviation from the target value (m)
10/08/2007
17

When the deviation products functional characteristic is an amount o from the target value m, the loss equals Ao. Then, Ao=ko2 k=
A0 2 0
10/08/2007
18
- unifying concept of quality and cost - relates engineering and economic terms in one model - allows for easy cost optimization strategies k = quality loss coefficient m+o = functional limits beyond which 50% of system product needs customer maintenance i.e. average customer tolerance L(y) = Ao at y = mo Ao = cost to replace/repair product
10/08/2007
19
Example Assume that the cost of repairing a failed telev is $2 per unit. Compare the losses caused by deviations fro two television sets, one produced in Factory A and the othe B, as described earlier in Fig. 2-1. Recall that the tolerance m -atom + A, where !l = 5.
--
Solution. In order to calculate the losses caused by deviatio need to determine the constant k of Eq. 2.5. Since !l = 5 and result is obtained from Eq. 2.6:
k
==
(5)2 = 0.08
2.0
in Factory A is obtained by taking the expectation of L(y) in
kv 2 ($/unit)
where v 2 is the mean squared deviation from target m. For and

2
= 0.08 ( 10) =
$0.222/unit
The loss caused by deviation in the production of the tel B is also
but now v
= 10/ .Jf2 and

2
L = 0.08
10
= $0.667/unit
.Jf2
f/1/lltil
Quality comparison between two manufacturers o

Mean . squared deviation v2 100 36 100
12
' Mafiufacturer
F)ctory' A
,
. '
Target m
Expected l per unit, $0.222 $0.667
Factory B
CASE STUDY T1
A spring is used in the operation of a camera shutter. The manufacturing process suffers from a degree of variability, in terms of the spring constant (measured in oz/in), which significantly effects the accuracy of the shutter times. The functional limits for this spring constant are m0.3oz/in (m=0.5oz/in), and the average cost for repairing or replacing a camera with a defective spring is $20. What is the loss function? Hence, what is the loss associated with producing a spring of constant 0.25oz/in versus the loss associated with one at 0.435oz/in.
10/08/2007
23
Uses of the Loss Function

The loss function approach can be used in evaluating the effect of quality improvement. For example, assume that Factory A has improved the process so that a new standard deviation from target of 10/8 (the previous one is 10/6) is attained. What would be the losses caused by deviations from the target value? L=
2 k ( y m) 2 = k
L=0.08(10/8)2=$0.125 The loss per unit of production would decrease from $0.222 (current process) to $0.125, resulting in $0.097 savings per unit.
10/08/2007
24
Loss Function/Process Capability Index

Let us see how the loss function is related to the process capability index Cp. For the current process and the improved process discussed above,
2 L1=k 1 2 L2=k 2
(loss with current process) (loss with improved process)
Divide the first equation by the second to obtain
L1 12 = 2 L2 2
But Then C p1 = tolerance and 6 1
Cp2 = tolerance 6 2
2 L1 C p2 = 2 L2 C P1
This equation implies that the losses caused by deviation are reciprocally proportional to the squares of the Cp indices.
10/08/2007
25
Economic Consequences of Tightening Tolerances as a Means to Improve Quality

As illustrated in the following example, the loss function approach can be used to determine the economic impact of tightening the tolerance to improve product quality. In order to reduce the difference in quality and process capability indices between television sets produced in Factories A and B, the management of Factory B tightened the tolerance from m5 to m5*(2/3). The cost of repairing an out-of-specification unit is still $2. What is the economic impact of tightening the tolerance?
10/08/2007
26
Economic Consequences of Tightening Tolerances as a Means to Improve Quality (Cont.)
With the original tolerance, the expected loss is 2 L=k =$0.667. The expected loss after tightening the tolerance is 2 L=k = 0.08[(2/3)*(10/ 12 )]2= $0.296/unit If improvement of the process was obtained by repairing the failed units (units outside the new tolerance m5*(2/3)) at a cost of $2 per unit, then the average cost of repair is as follows: Average cost of repair per unit = percent of production that needs repair to meet the tightened tolerance * repair cost per unit = 0.333*2 = $0.667
27
10/08/2007
Economic Consequences of Tightening Tolerances as a Means to Improve Quality (Cont.)

A summary of the results of this example is shown in the table given b be ello ow w.. I In nt th hiis sc ca as se e,, t tiig gh ht te en niin ng gt to olle er ra an nc ce e is an u un ne ec co on no om mic ca al a alt te er rn na at tiiv ve e because the expected total loss of tightening tolerance and repair (0.667+0.296=$0.963) is greater t th ha an n the expected loss using the original tolerance ($0.667).
10/08/2007
28
The Loss Function and Justification of Improvements
The loss function can also be used to justify improvements of the process, as illustrated in the following example. Assume that Factory A wishes to improve the quality of its television sets by reducing deviations from the target value so that the new standard deviation will be 10/8. This improvement can be technologically achieved at an additional cost of $0.05 per unit of production. Should the factory improve its process? (Assume that no inspection is performed.)
29
10/08/2007
The Loss Function and Justification of Improvements (Cont.)
Total loss per unit of the current process: 2 L= k = 0.08(10/6)2 = $0.222 Total loss per unit after improving the process: L=0.08(10/8)2 = $0.125 Additional cost of improvement = $0.05/unit Additional cost plus loss per unit = 0.05+0.125=$0.175 The net gain resulting from improvement in the process capability is 0.222-0.175= $0.047 per unit of production. If the production rate of this factory is 100,000 units per month, then the expected savings will be $4700 per month, or $56,400 annually.
10/08/2007
30
The Loss Function and Inspection

The loss function approach can be used effectively to determine whether 100-percent inspection can be justified or not. It should be noted that the objective of inspection is to screen or repair defective products that cannot meet the given specifications. Therefore, inspection cannot be used to improve the quality of items within the specifications. The improvement of the process can only be accomplished through improved manufacturing techniques or product design, not through screening or 100-percent inspection.
10/08/2007
31
The Loss Function and Inspection (Cont.)
Consider the case where the diameter of a stainlesssteel bar is m5m. The cost of repairing a defective bar is $6, and the cost of inspection is $0.03 per unit. Would a 100-percent inspection of items be justified? The estimated standard deviation of the process is 10/6. The expected loss without inspection is 2 L=k where k = A/2 = $6.00/52 = $0.24 therefore L = 0.24(10/6)2 = $0.667/unit
32
10/08/2007
Assuming that the characteristic of the product follows a normal distribution, the proportion of the products falling outside the specification m5 is 0.27 percent. The variance after screening defective products by using 100-percent inspection ( out ) is obtained using the procedure shown
2
below.
After the total inspection, the out-of-specification products are removed. The probability density function of those items that have passed the screening (acceptable items) is given by dividing the probability density function of the normal distribution by Q, the proportion of acceptable items.
33
10/08/2007

Let f(y) be the density function of the normal distribution, which is given by
10/08/2007
34
The Loss Function and Inspecti

Therefore, the variance of the passed items, v ut'is
,
IOJ2C
1
=
Jm+5 m-5
..,
1
.J27T
(v -
m)2e-<JJ2 ) (6/
Vout
0.9973
10
Using integration by parts, v
ut is
obtained as
1
"out =
(' 10
6.
x (0.986)-
'l
The expected total loss in the case of 100-percent inspectio L = inspection cost per item + (loss of a defective f in the inspection x fraction defective) + k v u
=.0.03
+ 6.00
0.0027
0.24 X (0.986) X
( 10
= $0.694/unit
Since the loss in the case of 100-percent inspection is higher than inspection is performed (0.667/unit), tOO-percent inspection is not
One might conclude that in this case 100-percent inspection is useless in improving quality, because the fraction defective is only 0.27 percent. It is different, of course, when the purpose of 100-percent inspection is to find serious defectives. In the case of a normal distribution with a standard deviation that is of the tolerance, the loss without inspection, L, is L = 0.24*(10/4)2 = $1.50/unit The proportion of the product falling outside the specification is 4.55 percent, and the variance of the outgoing items is (0.88)2 times that of the original value. The total loss in the case of 100-percent inspection when equals 10/4 becomes L = 0.03+6.00*0.0455+0.24*(0.88)2*(10/4)2 = $1.465/unit
10/08/2007
36
This result is an improvement of $0.035 per item. If there are 200,000 items produced each month, the amount of improvement is $7,000 each month. Assuming that the standard deviation is the tolerance and the production output still follows a normal distribution, the portion of the product falling outside the specification is 31.7 percent. Even if all the products are inspected and the defective ones screened out, the standard deviation of the outgoing quality is reduced to only 53.9 percent of the original value (=tolerance/2). Therefore, the loss caused by variation is L=0.24[0.539*(10/2)]2= $1.743
10/08/2007
37

Not only is this worse than the loss of $0.667 for =10/6 and no inspection, but it is also worse than the loss of $1.50 for =10/4 and no inspection, with 4.55 percent defective products. Thus, the solution to the quality problem is, in this case, through improvement of the process and not through 100-percent inspection.
10/08/2007
38
Tlhe Loss
Case
Mean
m
Function and
=
Standard de-viation 10 2 10 2
10
Inspecti
=
0 .24v dollars
Distribution of quality characteristic and its loss

Specification: m:!: 5}U1t, loss by unit defective (A) $6, loss function L
Variance Screening no
v ,
Fraction defective discovered 0.00
(dollars) 6.000
(ot
(1:f
eo 2
0.00
2 2 X ()
3 4 5 6 7
4
10
no yes no yes no no no no
1.500 1.162* 0.667 0.648* 0.375 0.094 2.000 2.167 1.667 1.594 1.500 6.000
m m m m m m
m- 2.5
4 6
10
0.880 6
4.55* 0.00
2
6 10 8
10
0.986
8
9 10
11
16
(lOt 16
2.52 2.52 2. 52 2.5
2 2
(r
X ()
0.27* 0.00 0.00
-10 .m
10
( ;.)
+ ( 160 r + +
( 10' 2
0.00
0.00
m - 2.5
m -2.5
12 13 14
6 10 12 10 16 0 0
no no no no
12)
0.00 0.00 0.00 0.00
Cf
m - 2.5 m - 5.0
5.0
t Based on Taguchi (1981) .
* Nonnal distribution
** Unifonn distribution
(no mark) Applicable to any type of distribution

The table shows a summary of the expected losses caused by variation for different probability distributions. These expected losses do not include the cost of inspection or loss caused by defective products found by inspection. Cases 1 through 6 demonstrate how screening reduces total losses for the given parameters. A detailed analysis of Case 2 follows, in order to illustrate how the results of the table are obtained.
10/08/2007
40
Since L=k
10/08/2007
41
10/08/2007
42
Determinations of Tolerances
Loss function can also be used to determine tolerances of the quality characteristics. The determination of tolerance is illustrated in the following example.
Example: Consider the production of high-voltage transformers. During the life of this kind of transformer, output voltage might change because of the deterioration of transistors in the power circuit. Assume that a transformer is not suitable for its intended function when its output voltage exceeds the tolerance limits of 11525V. Exceeding the limits results in a loss (denoted by A) of $300. Before shipping to a customer, the manufacturer can adjust the voltage in the plant by changing a resistor at a cost of $1. What should the manufacturers specifications be?
Solution: The loss caused by product variation from the target value, L(y), is L(y)=k(y-m)2 where m is the target value (115V in this case) and k is the proportionality constant.
10/08/2007
43
k= The loss function
A 300 = = 0.48 2 2 0 (25)
L( y ) = 0.48( y 115) 2
It is assumed that the allowable varying range of the output voltage for the customer is 11525V. The allowable varying range in the plant will be different, because it is easy to adjust the voltage to the target value by changing a resistor in the circuit. The loss or cost of adjust to the manufacturer is $1. Substitution of this value in Equation above yields
1.0 = 0.48( y 115) 2 y = 115 1 / 0.48 115 1.4V

As shown above:
300 = k ( y cus. m) 2 for y cus. m 0 300=A0 = k ( 0 ) 2 1 = k ( y manu . m)2 for ymanu . m 1=A = k ()2
A 0 A0
44
A0 A k = 2 = 2 = 0
10/08/2007
and the manufacturers tolerance is
= 0 /
A0 0 (functional orcustomer limit) = A (safety factor)
=manufacturers tolerance limit A=manufacturers loss when the product does not conform to the specification limits A0=loss to the customer caused by the failure of the product
10/08/2007
45
\ \ \
Customer's allowable range (Functioning range)
300
Loss to customer
200
100
Optimaltolerance
90
m-
115
m+
m-
m
Transformer output voltage
FIGURE 2-7 Tolerance determination for Example 2.12.
Main function of quality loss function: = define manufacturing tolerances or more generally define system variability
QUALITY LOSS FUNCTION 4 types

In the following illustrates the evaluation of the quality level of products by using the loss function approach for four types of tolerances. The three types are listed below:
1 2 3 4
10/08/2007
target-is-best smaller-is-best larger-is-best asymmetric target-is-best

47
e.g. previous example!
Target-is-best
-quality characteristics is usually a nominal output, for example --most parts in mechanical fittings have nominal dimensions --Ratios of chemicals or mixtures are nominally the best type. --Thickness should be uniform in deposition /growth /plating /etching..
Average quality loss

This type of tolerance is required for many products, parts, elements, and components when a nominal size (or characteristic) is preferred.
calculated in terms of MSD (mean square deviation) for n observations:

10/08/2007
1 n MSD = ( yn m) 2 n i =1
48
From this the average loss function for multiple lots:

1 n 2 y y ( ) S2 = variance = n 1 i =1 i
L ( y ) = k S 2 + ( y m) 2
CASE STUDY T2
Continuing from above, it was thought that if a new machine was purchased that the losses would reduce. To test this, 8 springs were tested from each machine, as detailed below. Which machine is best, and why? New machine: Old machine:
Data table Winder New Old
10/08/2007
0.37, 0.41, 0.37, 0.43, 0.39, 0.35, 0.40, 0.36 0.55, 0.67, 0.70, 0.54, 0.41, 0.32, 0.46, 0.66
S2
(-m)2
L(y)
0.0007 0.385 0.0184 0.539
0.0132 3.08 0.0015 4.41

49
New machine: Old machine:
lower variance but off-target higher variance but on-target
Loss incurred influenced by variability more than target value To reduce loss further (new machine) use an adjustment parameter i.e. reduce variability then adjust average response A 2-stage optimization!
10/08/2007
50
Target-is-best
Example: A manufacturer of ball bearings used in gas turbines requires that tolerances of the diameter and hardness of each ball be as follows:
Tolerance of diameter Tolerance of hardness m1 0.6 m m2 2.0 (Brinell hardness)
where m1 and m2 are the target values of the diameter and the hardness, respectively. The production rate is 80,000 balls per day at a cost of 30 per ball. Defective balls cannot be reworked and are scrapped. The following deviations from the diameter and the hardness target values were recorded.
10/08/2007
51
Target-is-best
Deviations from the target diameter:

0.3 0.0 -0.1 0.0 0.3 0.2 0.1 -0.2 0.6 0.4 -0.2 0.1 0.0 -0.4 0.5 0.4 -0.2 0.0 0.0 0.2
Deviations from the target hardness:

-1.0 -1.6 -0.4 -1.0 0.6 0.4 -1.2 -1.3 -0.2 -0.4 -0.4 0.5 -0.3 0.6 0.6 -0.9 -0.7 -0.9 -0.7 -1.3 Based on the diameter and hardness measurements recorded above, determine the quality levels of the production process for the diameter and hardness attributes of the balls.
10/08/2007
52
Target-is-best
10/08/2007
53
Target-is-best
The quantity vl. is usually estimated based. on available 2 estimate of v 2 has been referred to as v and is obtained as follo
where n is the number of measurements available, and y; is the v ment i. Equations 2.18 and 2.19 are used to determine the quality lev of the balls. Actually, this quality level (expressed in dollars) is the to deviations of the particular characteristic involved (in this c diameter) from its specified target value.
A == 30
A= 0.6
2[(0.3)
+ (0.0)2 + ( -0.1)2 + ... + (0.2)2]
30 (0. ) 2 X 0.075
6.25
Similarly, the quality level of the hardness of the balls, as m function, cn be obtained as follows:
A= 30
2.0
i/
= 2[( -1.0)
30 .
+ ( -1.6) 2 + ( -0.4)2 + ... + (-1.
L = ( .Q):! X 0.704 2
= 5.28
The difference between the quality levels relating to the re diameter of the balls becomes evident when the yearly production assuming 250 working days per year. Difference in quality levels
(0.0625- 0.0528) x 2
= $194,000 per year
As shown above, the loss function approach can be u quality levels of the various processes underlying the differe It should be noted that the mean squared deviation of y f 1n, as given by Eq. 2.20 below, is decreased by reducing ei y or the term [E ( Y) - 1n ]2 The reduction of variance can b both off-line and on-line quality control approaches. Mean squared deviation = E ( Y - 111) 2
= var ( Y ) +. [E ( Y )
--,.
where E (Y) is the expectation of Y , and var(Y) is the varia
.. Let y be he estimated expectation of Y . No adjustmen needed hen predicted y = m; however, if predicted y : 1n adjustment equal to y - m can be made on y. If the amou denoted as e, then
=-
I<Y;m) 11
i= 1
= predicted deviation from target
In other words, if predicted y does not equal m, as shown process should be adjusted (if possible) so that predicted y c obtain a significant decrease in the mean squared deviation.
Tolerance
f(y)
m-
....
m
......
y .,
Adjustment = y-m
A process without adjustment
Example: By examining the diameter data we find that more positive deviations than negative ones, whereas the hardness data show more negative deviations than positive ones. Assume that the manufacturer can shift the means of the data to the target values. What are the quality levels of the diameter and the hardness after the adjustments? Solution: The process should first be adjusted so that the value of every diameter is adjusted by an amount e*= ( y m) , the predicted deviation of the diameter from the target value. The new deviation (after adjustment) from the target value (m+ e*) is
Deviation after adjustment = y m ( y m) = y y
10/08/2007
59
Let the sum of squared deviation after adjustments of the diam A2 denoted as v d. Then
2
n "2( f d" d a tr a u stment) v
1 L y;2 - -n- 1 ;=I n
1[
19
o.3 + o.o + ..+ o.

2 2
..
= 0.0684
The denominator n - 1 is used instead of n in the abov
one degree of freedom is used to estimate the mean. The quality after the adjustment is
L
30
(0. ) 2 6
X
0.0684 = 5.70
Thus, the annual improvement is

(0.0625 - 0.0570)
X
80,000 X 250 = $110
Similarly, by using Eq. 2.23 and denoting the squared ing the hardness of the balls by vj;: .
.,
V;;(after adjustment) =
r[
19
( -1.0)
+ ( -1.6)
+ + (
,..,
vj; = 0.499
The quality level of the hardness after the adjustment is 30 L = 2 0 ., ( . )X
0.499 = 3.74
The annual improvement resulting from this adjustment is (0.0528- 0.0374)

X
80,000
250 = $30
Also, a comparison of quality levels between the diam after the adjustment indicates that the quality level of the har 1.5 times better than that of the diameter.
The above procedure can also be used to evaluate the same product when provided by different suppliers.
Example An automobile manufacturer requires that the s disk brakes be made of two separate parts, the knuckle and the be assembled by shrink-fitting the spindle into the knuckle, t more desirable stress distribution. The specification of the diam is 1n + 20JLm. The loss caused by a defective spindle is $24. observed the deviations shown in Table 2.4 from three differe are the quality levels of their spindles? If adjustments could be the quality levels be after adjustments?
TABLE 2.4
Suppliers' observed data

Supplier
1
Deviation from target
-5 -2 -6 -5
8 0
-4
8
.3 -2 -4 -6 -7 -5 -3 -8
4 5
-3 -5 -4 -8
-6 -6
-3 -7 -6 -5
-7 18 0 15 -16 -7 -10 -9 -1 -8 -10 12 -10 -6 -9 4 10 1
Solution. By using Eq. 2.18 we obtain the suppliers' losses befo

made as follows:
A
k =
24 2 = (20)2 = 0.06
and
L = 0.06v
where
V =
[(YI- nt)
(Y2- nt)
+ +
(y,r-
A summary of the suppliers' losses before adjustments is given i
TABLE 2.5
Suppliers' losses before adjustnten

Supplier
1 2 3
"'2 v
L =
o.o
17.60 34.75 122.40
$1.06 $2.08 $7.34
Adjustments could be made so that y 1 = y 2 = y3 = m deviations after adjustment for Suppliers 1, 2, and 3 are obta as follows:
"2
V
]
'iy;)' (
I=
n- 1
"""" y- -; =1
'
1'
Thus,
vi = ._!_ [< -5) 19

v 2 = 2.98
v3
...... 2 ...... 2
+ (8) 2 + . + co) 2 -
08 2 ) ] 20
122.26
The suppliers' losses, if adjustments - could be made, are The above data suggest that the manufacturer should choose Sup for the spindles if Supplier 2 is able to keep the distribution centered at the target value m .
TABLE 2.6
Suppliers' losses after adjus

Supplier
1
2 3
"2 v
17.67
2.98
122.26
$1. $0. $7.
Smaller-is best
here ideal response = zero L(y) = k(y)2 examples? Background density on a text image Radiation leakage Corrosion of metals Signal to noise ratios! Defective components
-quality characteristics is usually an undesired output, for example
--Defects like pin holes, particulates in deposition processes --Unwanted by-product or side effect
10/08/2007
68
Features: The-Smaller-The-Better type tolerance involves a nonnegative characteristic, whose ideal value is zero. A typical example of such a characteristic is impurity. Wear, shrinkage, deterioration, and noise level are also examples of this type. Under The-Smaller-The-Better (S-type) tolerance, the characteristic value is y0, the target value is m=0, and the upper tolerance limit is .
10/08/2007
69
CASE STUDY T3
In the copier industry, one measure of the acceptability of a copy is the amount of background toner particles that adhere to the portion of the copy that is intended to be white. Minimizing the residual toner in white areas is a smaller-is-best objective. It has been determined that approximately half of the customers will not tolerate a background level beyond 1.2 background units. Beyond that, a service call is placed at a cost of $200 plus the cost of the down time of the copier - $150 per hour. If the average copier down time is 2.5 hours, what is the associated loss function? As time progresses, it is apparent that one copier is not enough, so a second one is introduced. 8 sample background measurements are taken from each machine to compare and contrast their performances. From the data below what conclusions can you draw about the cost of this facility? Machine 1: Machine 2: 0.64, 0.56, 0.71, 0.55, 0.59, 0.75, 0.64, 0.76 0.55, 0.67, 0.70, 0.94, 0.71, 0.82, 0.86, 0.96
10/08/2007
70
Here: o=2 = 1.2, Ao = 200 + 150(2.5) = $575 So k = Ao/(o)2 = 575/1.22 = $399.30 L(y) = 399.30(y2) Average quality loss Data table = Machine 1 2 S2 0.0068 0.0203 0.65 0.77 S2+2 0.4293 0.6229 L(y) 171.41 248.70
Note: S2+2 not (-m)2 Clearly machine 1 is best but is probably sub-optimal?
10/08/2007
71
The-larger-The-Better (L T
There are cases where The-Larger-The-Better is applicable to ch as the strength of materials and fuel efficiency. In these ca predetermined target values, and the larger the value of the c better it is. Under this type of tolerance, the characteristic value is tolerance limit is A, and the target (or ideal) value is m = + caused by falling below the lower tolerance limit (i.e., y < A). of this type of characteristic is obtained by transforming the L an S-type tolerance as follows: Let
.., - -
The characteristic z > 0 has an S-type tolerance with the target the upper specification limit 1/ A. The loss function of the char
L( ) -
z -
(1/A)2
z2
Larger-is-better
here ideal response = max L(y) = k(1/y2) and k = Aoo2 examples? . . etc!
--quality characteristics is usually a desired output, for example --Bond strength --Critical Current
10/08/2007
73
CASE STUDY T4
The seal strength of a vacuum blower housing in an office copier is an example of a larger-the-better case. The better it can run under widely varying use environments, the better it is for minimizing loss. When the blower seal fails to operate, it costs $40 to replace: $20 in part costs and another $20 in installation labour costs. While the device that uses the vacuum blower sits idle, the cost to the customer is $340/hour. On average it takes 30 minutes to replace the blower. Seal integrity is measured by testing the seal adhesion strength. The seal level at which the vacuum loss becomes objectionable is 20 psi. What is the loss function and what is the loss incurred for a machine whose seal strength is measured as 13 psi?
10/08/2007
74
Ao = 340/2 + 40 = $210 o = 20 psi so k = 210 x 202 giving When y = 13 psi: L(y) = 84000/y2
L(y) = 84000/132 = $497
BUT/ the actual cost of repair is only $210! This illustrates that costs continue to increase beyond the acceptable value due to other consequences on the system Average quality loss
1 n 1 2 L( y ) = k ( MSD) = k n y i =1 i
75
10/08/2007
Example: Consider two types of cables, T1 and T2. The price and strength for either type are proportional to the cable s cross sectional area. The prices are P1=$1750/mm2 and P2=$2250/mm2, and the strengths are S1=220kgf/mm2 and S2=265kgf/mm2 for types T1 and T2, respectively. The lower tolerance limit of the cables breaking strength is 20000kgf, and the loss caused by falling below the lower tolerance limit is $58 million. Perform tolerance design and determine the tolerance limits for the better cable. Solution: We first calculate the total cost for each cable (price +quality loss). Let x be the cross-sectional area of the cable, which is the parameter being sought. Cable T1. The total cost C is obtained as the sum of the price and the quality loss.
2 A0 0 C = P1 x + ( S1 x) 2
= 1750 x +
10/08/2007
58, 000, 000 (20, 000) (220 x) 2
(3.13)
76
The total cost is minimized by taking the derivative of Eq (1) with respect to x and equating it to zero:
2 2 A0 dC 0 =P =0 1 2 3 dx S1 x
or
2 2 A 0 0 x= 2 S1 P1 1/ 3
= 818mm 2
The price of this cable is 1750818=$1.43 million
Cable type T2 : The corss-sectional area is

x= 2 58, 000, 000 (20, 000)
2 1/ 3
= 665mm 2
2250 (265) 2
10/08/2007
77
The price of cable T2 is 2250665=$1.50 million
Cable T1 is selected, since the price of T1 is less than T2. The tolerance of this cable is obtained using the formulation presented before:
k= A0 A = (1 / )2 (1/ 0 ) 2
It turns out
= A0 0 = A 58, 000, 000 20 metric tons force 1, 430, 000
=127.4 metric tons force
10/08/2007
78
Asymmetric target-is-best
i.e. when it is more harmful for the variable to be off-target in one direction over the other here 2 loss functions required: i.e. L+(y) = k+(y-m)2, y>m L-(y) = k-(y-m)2, ym
10/08/2007
79
as m where a! and a2 are the lower and the upper limits of respectively. The loss caused by deviation of a data point y from t is shown in Fig. 2-6 and is expressed as
fj.2
l
t! '
(y-m) .
A t
if "v
<
L(y) =
A2
-(y- m)
2
if v > m
r
where A 1 is the loss caused by y being below the lower limit of A 2 is the loss caused by y being above the upper limit of toleran When n observations are taken, the expected loss L is obtai 1 Ar ' 2 L = - 2L (y - m)
11
dl
A2'\'" 2 + 2L (y - nz)
2 .
Loss
m-tt. 1
FIGURE 2-6
Loss due to deviations from m.
- m)2 is the sum of the squared deviation for y I''(y - m)2 is the sum of the squared deviation for y; lar
where
I'(y
cylinder aiidthe piston of a six-cylinder engine be 3 +JLm. cylinder and piston assembly is $200, and the monthly produ Data showing deviation from the target value for the first two are shown below. What are the quality levels during these t the improvement, if any, of the quality level?
Month
1
Deviations
--2
0
3 4
5 -2
3 -2
0 4
3 -2 0 5
0 -2 0 -1
-1 -1
2 -1
0 4
3 2 6 -2
1 -1 -1 3 0 -2
3 2
Solution. The quality level of the production during Month using Eq. 2.25:
At = A2 = $200
A1
= -2
Az
=7
Lt =1 { 200 2 [(-2)2 ' 23 (-2)
(-2)2
+ (-2)2 + (-1)2 + (-1)

32
+( - 1)2] + 200 [32 + 42 + 52 +

72
+ 42 +
32 + 12 +
= 23{ 1400 + 612.24} = $87.50
The loss during Month 2, L 2 , is

L2
2{ (
2[(-1)2
+ (-1)2 + (-2)2 +
(-2)2
200 [32
72
+ 22 + 12 + 32 + 62 + 42 + 32 + 22
. $58.70/untt
1 {750 20
+ 424.49} =
The losses resulting from variation on the lower limit of the upper limit in each month. The improvement of qualit Month 2 is
87.5- 58.7
$28.80
which results in an improvement of

28.8 x 50,000 = $1,440,000 per mon
CASE STUDY T5
Consider the temperature drift in a refrigerator. The standard target for most refrigerators is 40F. Consider the consequences of being above and below this targeted temperature. When the temperature gets above 50F, several things can happen that annoy the consumer. These include tepid food and drink that is not pleasant to the taste, and spoilage due to accelerated bacterial growth. Each of these can cause economic loss, losses due to discarding and replacement of food, and losses due to illness from ingestion of tainted food. When the temperature gets below 30F, there may be some damage due to ice crystals, but there should be little food lost. When too hot the losses incurred include $50 for lost food replacement and $100 for a service call. When too low, the losses incurred include $10 for lost food replacement and $100 for a service call. What are the loss functions in this case?
10/08/2007
85
here, k- < k+ k+ = Ao/O2 = $150/(10F)2 = $1.50/F2 k- = Ao/O2 = $110/(-10F)2 = $1.10/F2 L+(y) = 1.5(y-40)2 L-(y) = 1.1(y-40)2
10/08/2007
86
0 L (y) (c) Larger-the-better
'
.
e. \ _.
A b'!;. j L.
\.\\:-+cO
z_
. L (') -=0
= At)
\
60
L (y)
y .ft.e. .e.,...
. \.j"1
;t.,.
\ -t t- -\-
\osS' . ,.
k ( 1 I 91. J C. r -t- ) v
"z./r
y }
(d) Asymmetric
m-6 0 '
TAGUCHI S/N RATIOS Taguchi idea use signal/noise ratios as performance measures signal = target value noise = scatter around target value Performance measure when maximized variation is minimized loss is minimized
10/08/2007
88
Advantages of S/N Method
10/08/2007
89
When to Use the S/N Ratio for Analysis

Whenever an experiment involves repeated observations at each of the trial conditions, the S/N ratio has been found to provide a practical way to measure and control the combined influence of deviation of the population mean from the target and the variation around the mean. In standard ANOVA they are treated separately.
10/08/2007
90
When to Use the S/N Ratio for Analysis (Cont.)

S/N offers the following two main advantages: 1. It provides a guidance to a selection of the optimum level based on least variation around the target and also the average value closest to the target. 2. It offers objective comparison of two set of experimental data with respect to variation around the target and the deviation of the average from the target value.
10/08/2007
91
Signal to Noise Ratio

The relevance of the S/N ratio equation is tied to interpreting the signal or numerator of the ratio as the ability of the process to build good product, or of the product to perform correctly. By including the impact of the noise factors on the process or product as the denominator, we can then adapt the S/N ratio as the barometer of the ability of the system (product or process) to perform well in relation to the effect of noise.
10/08/2007
92
Signal to Noise Ratio (Cont.)

By successfully applying this concept to experimentation, we can determinethe control factor settings that can produce the best performance (high signal) in a process or product while minimizing the effect of those influences we can not control (low noise).
10/08/2007
93

To obtain a better understanding of how this approach works and what it means, lets discuss a practical example (car radio) illustrated below:
10/08/2007
94
Signal to Noise Ratio (

_ :(\l&t
d.t?J .
. '
....
'
. ' 1-... .. .: :"' \ _r ('),
:... f.:
'
'
({AtA;e .
\_9V\Q.\!
5 .,-oiuJty _weo
Jere- .
---
Leo .rt dm rQ
S\q tACL.Q_
>h:P"'i lVLit-:rJrrc

For improved additivity of the control factor effects, it is common practice to take log transformation of 2/2 express the S/N ratio in decibels.
2 Z = 10 log10 2
The range of values of 2/2 is (0,), while the range of values of Z is (-, ). Thus, in the log domain, we have better additivity of the effects of two or more control factors. Since log is a monotone function maximizing 2/2 is equivalent to maximizing Z.
10/08/2007
96
very important reason
Taking logarithm improves additivity. Why?

On taking logarithm these become additive For example Log10 ( B Xa . exp(-Ea/kT) ) = Log10(B) + a. Log10 (X) + (-Ea/kT) . Log10 (e) These are additive for variables (Log10 X) and (1/T)
97
Most functions in nature follow the power-laws Y = A exp (-Ea/kT) (EXPONENTIAL) Y = B Ax . . . . ( a RAISED POWER X) Y = B Xa . . . . ( X RAISED POWER a)
10/08/2007

Consider the following two sets of observations around the target and the deviation of the average from the target value. Let m=75 Observation A: 55 58 60 63 65 y = 60.2 Dev. Of mean from target = 75 - 60.2 = 14.8
y = 75 Observation B: 50 60 76 90 100 Dev. Of mean from target = 75 - 75 = 0
10/08/2007
98
Signal to Noise Ratio (

. . M::lf
.
. I:'
b 't .
((.
t_ '
. l
_.:.__"1_---------- -- -- .:..-t--..---- --:- - -- -t---
J--r -- --t-\. -- "
-- 1 t) .
lt t) '
:
{.c"
cr
c...--::-
f -)
ENGN8101 ModeJUng and Optimization
Conversion of Results into S/N Ratios (Cont.)
These two sets of observations may have come from the two distributions shown in the figure above. Observe that the set B has an average value which equals to target value, but has a wide spread around it. For the set A, the spread is smaller, but the average itself is quite far from the target. Which of the two is better?
10/08/2007
1 00
Conversion of Results into S/N Ratios (Cont.)

Based on the average value the product shown by obs. B appears to be better. Based on consistency, product A is better. How an one credit A for less variation? How does one compare the distances of the averages from the target? Surely comparing the averages is one method. Use of S/N ratio offers an objective way to look at the two characteristics together.
10/08/2007
1 01
Computation of SJ
Ca5e.. ;. N
IAAS"!)
{(SS-1sJL-t
(I '
(&
-= . - z 3,
. . . -::- LJ
. . . 5/t\.1
. 5f/N.
o te
- '
o < rc/ tu
nJ
-- - '3 .\ / . ? .
.(1.1
L.
S(N =. 4-2
. 3brbc;-
Computation of S/N Ratio (Cont.)
10/08/2007
1 03
Most Common S/Ns for the Static Case

Nominal-is-Best (N.B.)
10/08/2007
1 04
Most Common S/Ns for the Static Case (Cont.)
y Z = 10 log 2 s
2
10/08/2007
1 05

Smaller-is-Better (S.B.)
10/08/2007
1 06
1 n 2 Z = 10 log yi n i =1
10/08/2007
1 07

Larger-is-Better (L.B.)
10/08/2007
1 08
n 1 2 i =1 y Z = 10 log n
10/08/2007
1 09
Target-the best
aim
= sample mean
y2 Z =10 log = 2 s
s = sample standard deviation
to maximize Z through parameter design
Smaller-is-best
here,
n y2 i Z = 10 log i =1 n
needs to be maximized
Larger-is-best
here,
n 1 2 i =1 y Z = 10 log n
needs to be maximized
10/08/2007
1 10
But what about quality characteristics that approach an ideal value?
Examples are
Efficiency : all efficiencies approach the ideal value of 100% Weld strength : approaches the ideal strength of the material Critical temperature or Critical current density for High Temperature superconductors (YBCO) : These approach ideal values, say 92K and 108 A/cm2
Which SN-Ratio is most suitable among the following ?
smaller-the-better
LARGER-THE-BETTER NOMINAL-the-BEST
1 11
10/08/2007
Taguchi Method is most effective when there is at least one quality characteristic that is sensitive to variations or NoIsE
Desirable Qualities e.g. Nominal-the-Best type are sensitive to NoIsE

S/N Ratio
Z=
10 Log10 ( mean2 / Variance )

2
Question: the large of mean, the better the quality is? Undesirable properties e.g. Smaller-the-better type are also sensitive to NoIsE
S/N Ratio Z = 10 Log10 ( 1/n = zero
10/08/2007
? 10 Log10 (Variance) . . . if ideal value
Yi2 )
1 12
Best approach for experimental design Maximize these ratios to minimize variability to maximize system robustness then adjust target value to desired value 2-stage optimization! Some drawbacks, however, to Taguchi approach e.g. assumes and s are of equal importance/influence + noticeable confounding issues
10/08/2007
1 13

Loss Functions-Lecture 5

Diunggah oleh

Informasi Dokumen

Judul Asli

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

Loss Functions-Lecture 5

Diunggah oleh

Hak Cipta:

Format Tersedia

THE QUALITY LOSS FUNCTION

ENGN8101 Modelling and Optimization

Loss Function and Quality Level

ENGN8101 Modelling and Optimization

Quality Loss Concept

Deviation from target results in loss.

Lower than target Greater than target Both lose

ENGN8101 Modelling and Optimization

How can we use Quality Loss Concept in real life?

ENGN8101 Modelling and Optimization

How can we use Quali concept in rea/life

Traditional Quality Metric

All products equally good

Loss Function (Cont.)

ENGN8101 Modelling and Optimization

loss Function (Con

"' o red Valwe of the Cl\ara tereatc

ENGN8101 Modelling and Optimization

Distribution of color concentration

ENGN8101 Modelling and Optimization

If m is achieved. . loss is zero

ENGN8101 Modelling and Optimization

Derivation of the Loss Function

ENGN8101 Modelling and Optimization

Derivation of the Loss Function (Cont.)

L (m) L(y)= (y - m)2 2!

ENGN8101 Modelling and Optimization

Derivation of the Loss Function (Cont.)

ENGN8101 Modelling and Optimization

Derivation of the Loss Function (Cont.)

ENGN8101 Modelling and Optimization

ENGN8101 Modelling and Optimization

in Factory A is obtained by taking the expectation of L(y) in

where v 2 is the mean squared deviation from target m. For and

The loss caused by deviation in the production of the tel B is also

= 10/ .Jf2 and

Quality comparison between two manufacturers o

Expected l per unit, $0.222 $0.667

ENGN8101 Modelling and Optimization

Uses of the Loss Function

ENGN8101 Modelling and Optimization

Loss Function/Process Capability Index

(loss with current process) (loss with improved process)

Divide the first equation by the second to obtain

ENGN8101 Modelling and Optimization

Economic Consequences of Tightening Tolerances as a Means to Improve Quality

ENGN8101 Modelling and Optimization

Economic Consequences of Tightening Tolerances as a Means to Improve Quality (Cont.)

Economic Consequences of Tightening Tolerances as a Means to Improve Quality (Cont.)

ENGN8101 Modelling and Optimization

The Loss Function and Justification of Improvements

The Loss Function and Justification of Improvements (Cont.)

ENGN8101 Modelling and Optimization

The Loss Function and Inspection

ENGN8101 Modelling and Optimization

The Loss Function and Inspection (Cont.)

The Loss Function and Inspection (Cont.)

The Loss Function and Inspection (Cont.)

ENGN8101 Modelling and Optimization

The Loss Function and Inspecti

Using integration by parts, v

The Loss Function and Inspection (Cont.)

ENGN8101 Modelling and Optimization