Statistical Process Control Statistical Process Control

D.A. MAISANO

Statistical Process Control D.A. Maisano

Quality Engineering
SUBJECTS

Quality in design Statistical Process Control Performance Indicators Domenico MAISANO Ph.D. - Franceschini. Advanced Quality Function Deployment, St. Lucie Press/ CRC Press LLC.

TEACHER

REFERENCE BOOKS

- Montgomery. Introduction to Statistical Quality Control. 5th Ed., John Wiley & Sons. - Franceschini, Galetto, Maisano, Management by Measurement, Springer.

TIMETABLE

Monday Thursday

10.00-13.00 14.30-17.30

(room 11) (room 9I)

2

Statistical Process Control D.A. Maisano

QUALITY

Quality of a product or service refers to the degree to which the product or service is able to satisfy (stated or implied) needs. In other words, Quality is the degree to which a product/service is conform to its requirements.

Statistical Process Control D.A. Maisano

Every product posses a number of characteristics that are critical to quality (for the user/consumer): quality characteristics. e.g. - length of mechanical components - duty of batteries - thickness of the coat of paint - amount of material in a tube of toothpaste Most organizations find it difficult (and expensive) to provide the customer with products that have quality characteristics, which are always identical from unit to unit.

Statistical Process Control D.A. Maisano

VARIABILITY
A major reason for this is VARIABILITY: two products cannot be ever identical (e.g. the diameter of a screw). if the variation is large, the customer may perceive the unit to be undesirable and unacceptable. beyond this, if the variation is large, these units cannot be interchangeable (e.g. problems in the assembly process). Most common sources of variability: - differences in materials; - differences in the performance of the manufacturing equipment; - differences in the way operators perform their tasks.

Statistical Process Control D.A. Maisano

Example: the diameter (D) of a workpiece manufactured hole cannot be identical in all the products

D [mm] D [m] 30,70 30,60 30,50 30,40 30,30 30,20 30,10 30,00 1 2 3 4 5 6 7 8 9 measure
6

Statistical Process Control D.A. Maisano

SPECIFICATIONS Quality characteristics are evaluated relative to specifications: - a value of a measurement that corresponds to the desired value is called the NOMINAL (or target) VALUE; - these values are usually bounded by a range of values that we believe will be sufficiently close to the target so as to not impact the function or performance of the product if the quality characteristic is in that range.
USL (Upper Specification Limit): the largest allowable value; LSL (Lower Specification Limit): the smallest allowable value.

Statistical Process Control D.A. Maisano

LSL

NV

USL

X [mm]

Statistical Process Control D.A. Maisano

NATURAL VARIABILITY: process tendency towards producing (in normal conditions) products with quality characteristics different from target values. it is an internal characteristic of the process
S

LSL

USL

different natural ranges of variability

Statistical Process Control D.A. Maisano

STATISTICAL PROCESS CONTROL methods make it possible to control quality characteristics during production (on-line), in order to maintain the process under-control and to detect and correct possible abnormalities.

Statistical Process Control D.A. Maisano

10

SPECIFICATION AND NATURAL TOLERANCE LIMITS Natural tolerance (NT) range is a measure of the natural variability of the process. The process variability is usually measured by the standard deviation (). is an index of the natural dispersion of the process. Specification range (S) are determined externally (usually set by product designers).

S
NV

S
nonconforming products (scraps)
NV

NT
if NT is included within S limits, there is (almost) no scrap
Statistical Process Control D.A. Maisano

NT
11

SPECIFICATION AND NATURAL TOLERANCE LIMITS For every product quality characteristic (e.g. geometrical dimensions) we define the specification limits (USL, LSL).

Si
LSLi NVi USLi

A LSLA, NVA, USLA B LSLB, NVB, USLB C LSLC, NVC, USLC D LSLD, NVD, USLD E LSLE, NVE, USLE
12

Statistical Process Control D.A. Maisano

A process operating with only chance causes of variation (not other assignable causes) generally shows a random pattern (also defined as white noise). typically it follows a NORMAL DISTRIBUTION

x t

Statistical Process Control D.A. Maisano

13

Example: non random pattern (RUN), due to the presence of assignable causes (e.g. thermal expansion, tool wear)

Statistical Process Control D.A. Maisano

14

Example: non random pattern, due to the presence of two points related to assignable causes (e.g. failures in the process)
5062,9 5062,7 5062,5 5062,3 5062,1 5061,9 5061,7
Statistical Process Control D.A. Maisano

shaft diameter [ m]

FAILURE of the TURNING LATHE!!

15

It is customary to define the upper and lower natural tolerance limits, say UNTL and LNTL, as 3 above and below the process mean. NT 6

UNTL = + 3 3 LNTL = 3

This is due to the fact that for a normal distribution these limits include 99.73% of the process output.

x
LNTL UNTL

Statistical Process Control D.A. Maisano

16

For other distributions (for example the exponential), the 6 interval does not necessarily include 99.73% of the process output (population). f(x)

f(x) = e - x
=1/

x exponential distribution (f.e. time to failure of a component)

Statistical Process Control D.A. Maisano

17

To calculate the natural tolerance NT 6, we should know the standard deviation () of the population. can be estimated by using the sample standard deviations (s) or the sample ranges (R), related to several samples extracted from the population.

Statistical Process Control D.A. Maisano

18

Example: m samples are extracted from the population; each sample is made of n observations. 1) x11, x12, , x1n 2) x21, x22, , x2n . . . m) xm1, xm2, , xmn

x1
. . .

s1 . . .

R1 . . .

x2 s2 R2

xm sm Rm

sample standard deviation s j = sample range

i =1

( x ji x j ) 2

n 1 R j = max( x ji ) min( x ji )

j i

generic sample number generic observation number

Standard deviation of the population () can be estimated using sj or Rj. In the following we will see how.
Statistical Process Control D.A. Maisano

19

TOLERANCES OF ASSEMBLED COMPONENTS What is the NT of a product made up of different parts, each of those having its own NT? Example: I = A + B + C
(dimensional quality characteristics of 3 components assembled together) A) 1.000 0.001 cm

B) 0.500 0.0005 cm C) 2.000 0.002 cm

NT A B I C

The average value of product I is given by the sum of A, B and C average values I = 1.000+0.500+2.000 = 3.500 cm But how to calculate the NT of product I?
Statistical Process Control D.A. Maisano

20

A possible way is to use an Additive method: NTI = NTA+NTB+NTC = 0.001 0.0005 0.002 = 0.0035 cm Is this correct????

Statistical Process Control D.A. Maisano

21

NO! NTI = NTA+NTB+NTC = 0.001 0.0005 0.002 = 0.0035 cm Probabilistic method: If the different quality characteristics (A, B, C) are statistically independent (the occurrence of one event occurs does not affect the outcome of the occurrence of the other event) and normally distributed, then we can use the following formula: I2 = A2 + B2 + C2
+2cov(AB) +2cov(AC) +2cov(BC)

As a consequence, considering that NTi = 3i : NTI2 = NTA2 + NTB2 + NTC2

NTI = NTA 2 + NTB2 + NTC2 = 0.0023 cm
this value is lower than the one calculated adding the single Natural Tolerances (NTi) up
Statistical Process Control D.A. Maisano

22

EXAMPLE

x (, )

being: = 10 mm = 0.1 mm

Assuming that a chain is made of 100 chainrings, the chain average 100 length is given by:
chain =

= 100 10 = 1000 mm
i i=1

Additive method: If we calculate chain adding up the single st. deviations, we obtain: 100
chain =

= 100 0.1 = 10 mm
i i=1

NO!
EXCESSIVE PLAY!!
23

consequently NTchain = 3chain = 30 mm

Statistical Process Control D.A. Maisano

Probabilistic method: chain can be correctly calculated by applying the following formula (probabilistic method) 2chain = 2 chain = 100 2 = 10 = 1 mm
The lengths of the parts can be assumed independent. When the process is operating in regular conditions, the length of one part is not influenced by the length of the previous

consequently

NTchain = 3chain = 3 mm

Let notice that in this case the global variability (chain) is much lower than in the previous one. This is due to a sort of compensation among the variations in the parts assembled together.
Statistical Process Control D.A. Maisano

24

In a more general case: Z = nX + mY (X and Y statistically independent) Z2 = n2 X2 + m2 Y2

If X and Y are not statistically independent, all the previous equations are not valid. Additional terms should be introduced (COVARIANCES).

Statistical Process Control D.A. Maisano

25

EXAMPLE
A shaft is to be assembled into a bearing. The internal diameter of the bearing is a normal random variable say x1 with mean 1=1.500 inches and standard deviation 1=0.002 inches. The external diameter of the shaft say x2 is normally distributed with mean 2=1.480 inches and standard deviation 2=0.004 inches. When the two parts are assembled, interference will occur if the shaft diameter is larger than the bearing diameter that is, if: y = x2 x1 > 0
x1 x2

The distribution of y is normal, with mean and variance

y = 2 1 = 1.480 1.500 = -0.020 inches

2y = (12)22 + (-12)21 = (0.004)2 + (0.002)2 = 0.00002

Therefore, the probability of interference is:

0 - y 0 + 0.020 P {interference} = P {y > 0} = P z > = 1- ( 4.47 ) = 0.000004 = 1- 2 0.00002 y
Statistical Process Control D.A. Maisano

26

NT of the assembled component NT of the parts Often, the overall NT of a component is assigned and it is necessary to define the NT of the single parts, in order to satisfy it. Example: I = A + B + C (dimensional quality characteristics of 3 parts assembled together) A) 1.000 ??? cm B) 0.500 ??? cm C) 2.000 ??? cm
A B I C

Statistical Process Control D.A. Maisano

27

Assuming NTI = 0.0035 cm, how to split it among NTA, NTB, NTC? 1) Depending on the production constraints: for some processes it is more difficult (or expensive) to achieve narrow tolerances than for others. 2) In the absence of particular reasons, NTI can be uniformly split up among NTA, NTB, NTC.

Statistical Process Control D.A. Maisano

28

Using the previous relation (NTI2 = NTA2 + NTB2 + NTC2 ), NTI can be uniformly split up among A, B and C (NTA = NTB = NTC). Consequently, in this case (probabilistic method):
1) NTA = NTB = NTC = NTI 3 0.002 cm

If we calculate the NT of the parts, by considering the sum of their values (additive method):
2) NTA = NTB = NTC = NTI 3 0.0011 cm

Statistical Process Control D.A. Maisano

29

However, calculating the NT of the parts using the first (probabilistic) method, entails the risk of producing (a small portion of) assembled components out of natural tolerance limits.

Probabilistic method is based on an implicit hypothesis of compensation Example: if for every part we consider its lower limit value: A = 10.002 cm B = 0.50.002 cm C = 20.002 cm I = 3.5 0.006 < 3.5 0.0035/2 cm
NT/2

LNTLI

in this case the assembled component length would be BELOW its LNTL.

Statistical Process Control D.A. Maisano

30

 let remember that if the distributions of independent variables (A, B and C) are normal, also the dependent variable (I) is normal. In this case we implicitly assumed A, B and C to be normal. 3 interval includes 99,73% of normal distributed data In other terms, since NTI = 3I, the probability of assembling components outside the natural tolerance limits is almost 0.3%.

I
0.15% LNTLI 0.15%

I 3 3

UNTLI

Statistical Process Control D.A. Maisano

31

Assuming NTI = 3 = 0.0035 cm, we have I = 0.00116 cm Hypothesis: considering the distribution of I as normal, the probability of assembling components out of NT limits, below the LNTLI, is given by:
LNTLI I 3.4965 3.500 = P z = P {z 3} 0.15% 0.00116 I

The same goes for the probability of assembling components above the UNTLI.
UNTLI I 3.5035 3.500 = P z = P {z +3} 0.15% 0.00116 I

Statistical Process Control D.A. Maisano

32

One more EXAMPLE Y = X1 + X2 + X3 + X4 The purpose is to obtain: with the following values: NT Y = 2.510 0.008 cm X1 = 0.370 X2 = 0.625 X3 = 0.750 X4 = 0.765 In the absence of particular reasons, NTY can be uniformly split up among NTX1, NTX2, NTX3, NTX4.

Statistical Process Control D.A. Maisano

33

If we use the additive method, we have: NTX1 = NTX2 = NTX3 = NTX4 = NTY / 4 = 0.008 / 4 = 0.002 cm If we use a probabilistic method, we have:
Y = X 2 + X 2 + X 2 + X 2 1 2 3 4

Statistical Process Control D.A. Maisano

34

To simplify, we assume: X1 = X2 = X3 = X4 ( = X ) consequently: Y = 4 X2 = 2 X since NTY = 3 Y and

NTX = 3 X

then NTX = NTY / 2 = 0.004 it is double the value calculated using the additive method!!

Statistical Process Control D.A. Maisano

35

Anyway, let notice that the probabilistic method - even under conditions of normal variability - entails a 0.3% risk of producing assembled components out of NT limits. assumptions: - normal distribution of the quality characteristics, - statistical independence of the quality characteristics (we
intuitively mean that knowing something about the value of one of them does not yield any information about the value of the other)

Statistical Process Control D.A. Maisano

36

ASSEMBLED COMPONENTS: NON LINEAR FUNCTIONS So far, we have been considering linear functions only (among the assembled component and the parts). In some problems, the dimension of interest may be a nonlinear function of the part dimensions (x1, x2, , xn)

Z = nX + mY

Z2 = n2 X2 + m2 Y2
variance

Statistical Process Control D.A. Maisano

37

For non linear functions, what is the relationship among variances? y = g (x1, x2, , xn)
non linear function

Considering a first order Taylor series (truncated) development of the previous function, in the neighbourhood of the mean values of the parts:
y = g(X , X , ... , X ) + 1 2 n

i=1

(xi -X )
i

g xi

+
X1,X2 ,...,Xn

...

partial derivative

second and higher order terms are neglected

Statistical Process Control D.A. Maisano

38

In order to develop the function, we implicitly assumed that: - it should be derivable - it should be continuos in the neighborhood of the point
(1, 2, ... , n )

Statistical Process Control D.A. Maisano

39

Considering the function

y g(X1 , X2 , ... , Xn ) +

i=1

(xi -Xi )

g xi , ,..., X1 X2 Xn

From statistics, we can apply the following (approximate) formulas

Y g(X ,X , ... , X ) 1 2 n
number of independent variables

g 2Y 2X x i i , , ... , i=1 X1 X2 Xn

TRANSMISSION ERROR FORMULA

Statistical Process Control D.A. Maisano

40

Let notice that the previous formulas can be applied assuming: - normal distribution of the quality characteristics, - statistical independence of the quality characteristics (we
intuitively mean that knowing something about the value of one of them does not yield any information about the value of the others)

Statistical Process Control D.A. Maisano

41

EXAMPLE
I V=RI R being: I = 25 1 A R = 4 0.06

(Ohm law)

Purpose: evaluate V and NTV of V Supposing the distributions of R and I to be normal, we can calculate:

V = g(R ,I ) R I = 100 V
To calculate NTV, we need: I = 1 A / 6 = 0.33 A R = 0.06 / 6 = 0.02 Consequently:
2 2

V V2 I , I R

V 2 + I R , I R

2 R
42

Statistical Process Control D.A. Maisano

 V 2 R , 2I + I , 2R = R2 2I +I2 2R I R I R
= 4 ( 0.33 ) + 25 ( 0.02 ) = 2.0277 V 2
2 2 2 2

therefore V = 1.41V and NTV = 3 1.41 = 4.23 V

Statistical Process Control D.A. Maisano

43