Professional Documents
Culture Documents
T. Gutowski
Ref:
1. Random Variables, Lipschutz
2. Hogg & Ledolter (Control Charts)
pp 50-56, 339-342 + p386 Table C.1 (also see Kalpakjian)
3. A Brief Intro to Designed Experiments, Taken from Quality
Engineering using Robust Design by Madhav S. Phadke,
1
Outline
I. Looking for problems
Types of Problems
Minimizing Random Effects
Minimizing Assignable Causes
2
Looking for Problems
3
Where you find the problem
matters
• Customer complaints
• Warranty claims
• Dealer returns Increasing
• In-house inspection Cost
• Assembly
• Observed at the process
4
Taguchi Quality Loss Function
Deviation, δ
Repair and upgrade of the Hubble Telescope
f ′′(0) 2
QL(δ ) = f (0) + f ′(0)δ + δ +L QL = k δ2
2!
Fix problems close to home! 5
Quality Systems
• Make problems obvious
– Poke yoke at the process level
– Clear flow paths and responsibility
– Andon board
– Simplify the system
• Stop operations to attend to quality problems
– Stop line
– Direct attention to problem
– Involve Team
• Reduce inventory
– Reduces rework, shorter trail to the problem
6
Types of Problems
7
Reducing Randomness
8
To do this…
• reduce variation
– How does variation propagate?
– Choose the right processes
– Control actions
• make product robust- insensitive to
variation
– Design for mfg
– Design for function
9
Random Variables
1. Expectation = mean = average
2. Variance = (std deviation)2
3. Properties of E() and Var()
4. Propagation of Errors
Ref. Lipshutz
10
Process Variation
Process measurement reveals a distribution in output
values.
35
30
25
20
15
10
0
1 2 3 4 5 6 7 8 9 10 11 12 13
Discrete probability
Continuous “Normal”
distribution based upon
distribution
measurements
µ = E(X) (1)
E( X) = ∑ xi p( xi ) (3)
i
12
Calculation of E(X) = µ and Var (X) = σ2
13
Properties of the Expectation
1. If Y = aX + b;
14
Properties of the Variance
1. For a and b constants,
(7)
15
Propagation of errors
• Dimensional Effects y
θ
x
– Abbe error: y ≈ θ x
– thermal expansion: δL = L α ∆T L
16
Propagation of errors
y = y + δy = (θ + δθ )( x + δx)
δy ≅ θ δx + x δθ
Var ( y ) = E[(δy ) 2 ]
2 2 2
δy ≅ (θ δx) + 2θ δx ⋅ x δθ + ( x δθ )
recall E ( x) = ∑ xi p( xi )
2 2
Var ( y ) ≅ θ Var ( x) + x Var (θ )
17
Propagation of errors
• this result is called “quadrature”,
in general, if y=θx, with θ, x independent
random variables with small variation, then
with Var (x) = σx2
2 2 2
σ y σθ σ x
= +
y θ x
18
Propagation of errors
• A more general result is for any
relationship like y=zαxβ, with z, x
independent random variables with small
variation, then
2 2 2
σ y 2σ z 2σ x
= α + β
y z x
19
example: errors due to thermal
expansion
• Say in a machine shop the conditions
fluctuate such that the means and
standard deviations (µ, σ) are for stock
size (10, 1), linear expansion coef.(1000,
1) and temperature (20, 3).Quadature tells
us that the variation in thermal expansion
of the work pieces is dominated by the
temperature and is approximately 0.39
times its mean value.
20
The out of specification parts are 2(0.5-φ(2σ))
= 2(0.5 -0.4772) = 0.0456 or 4.56%
Lower Upper
Target
Specification Specification
Limit Limit 21
22
Some propose a process capability index Cp that compares
the tolerance interval USL-LSL vs the process variation 6σ.
For the previous case this would be 4σ/6σ =0.67
USL − LSL
Cp =
6σ
Cp % out
⅔ 4.55
Lower Upper
Target
1 0.27 Specification
Limit
Specification
Limit
1⅓ .0063
Kotz & Johnson
23
In general the mean and the target do not have to
line up. In this case the Cp is misleading. A better
question is, how many parts are out of spec?
13.6
+2.1
0.135% +0.135
15.835
+0.135
15.97
Lower Upper
Target
Specification Specification
Limit Limit 24
In this case an alternative process capability can be
used called the Cpk
min(USL − µ , µ − LSL)
C pk =
3σ
25
Comparison
Case 1 (µ on target) Case 2 (µ drift)
Cp = 4σ/6σ = 2/3 Cp = 4σ/6σ = 2/3
Cpk = Cpk =
Min(2σ/3σ,2σ/3σ)=2/3 Min(1σ/3σ,3σ/3σ)=1/3
26
Why the two different
distributions at Sony?
27
“Tolerance Stack up”, really about variance,
X1 Xn
recall that
E(X1 + … + Xn) = E(X1) +…+ E(Xn)
but how about
Var(X1 +…+ Xn) = ?
28
If X1 and X2 are random variables and not necessarily
independent, then
this can be written using the standard deviation “σ”, and the
correlation “ρ” as
2 2 2
σL = σ1 +σ2 + 2σ1σ 2 ρ
(9)
where L = X1 + X2
29
If X1 and X2 are correlated (ρ = 1), then
(14)
σ L2 2
= σ1
2
+σ2
2
+ 2σ1σ 2 = (σ 1 + σ 2 )
for X1 = X2 = X0
2 2
σ = 4σ
L 0
(15)
2 2 2
for N σL =N σ0 (16)
or (17)
σ L = Nσ 0
30
Now, if X1 and X2 are uncorrelated (ρ = 0) we get the result as in
eq’n (7) or,
2 2 2
σL = σ1 +σ2 (10)
N
and for N
σ =∑ σ
2
L i
2 (11)
i =1
2 2
If X1=X =Xo σL = Nσ 0 (12)
σ L = Nσ 0
Or (13)
31
“Tolerance Stack-
Stack-up”
As the number of variables grow so does the
variation in the system;
correlated
σ L = Nσ 0
uncorrelated
σ L = Nσ 0
32
Control Actions
1. Measure parameter
33
How to control variation
∂Y ∂Y
∆Y = ∆α + ∆u
∂α ∂u
sensitivity
output variation control action
disturbance
34
Real Time Control
∂Y ∂Y
∆Y = ∆α + ∆u
∂α ∂u
sensitivity
output variation control action
disturbance
35
Statistical Process Control
∂Y ∂Y
∆Y = ∆α + ∆u
∂α ∂u
sensitivity
output variation control action
disturbance
36
Designed Experiments
∂Y ∂Y
∆Y = ∆α + ∆u
∂α ∂u
sensitivity
output variation control action
disturbance
37
Statistical Process Control and
Process Control Charts
Ref. “Engineering Statistics”
Robert V. Hogg, and Johannes Ledolter
1987 MacMillian Publishing
38
Measurement Process
Centerline
Lower Control
Limit
Sampling period
39
Statistical Control Methods
Strategy:
1. Determine Centerline, UCL, and LCL
(from past data sampling when process is
under control)
2. Monitor stability of process
3. Data outside of UCL/LCL indicates mean
shift
4. Investigate and eliminate causes of shift 40
Statistical Control Methods
Factors that determine the appropriate
sampling frequency:
– Stability of process
– Potential loss
41
“x-bar charts”
Mean and Standard Deviation
• Mean
n
xi
x =∑
i =1
n
1 k
x = ∑ xj
k j =1
UCL, LCL = x ± A2 R
Where, n = sample size
k = number of samples
A2 = constant from Table C.1
42
“R-charts” Range = high - low
• Standard Deviation
R = max( x1 ,K, xn ) − min( x1 , K, xn )
1 k
R = ∑ Rj
k j =1
LCL = D3 R
UCL = D4 R
Where, n = sample size
k = number of samples
D3 , D4 = constants from Table C.1
43
44
p-chart
Proportion of Defects (binary)
d1 + d 2 + L + d k p(1 − p)
p= UCL, LCL = p ± 3
nk n
Where, d = number of defective samples
n = sample size
k = number of samples
45
Note: LCL is neglected because it is not possible for negative defects
c-chart
Count of Defects (in each unit)
1 k
c = ∑ ci UCL, LCL = c ± 3 c
k i =1
Where, c = number of defects per sample
k = number of samples
Note: LCL is neglected because it is not possible for a 46
negative number of defects
Robustness Vs Sensitivity
• Mean shift by 1σ requires 44 data points to
observe
• Alternatives
– ±2σ; two consecutive observations
– ±1σ; three consecutive observations
– other charts (e.g. “Cusum”)
47
Source Identification; Ishikawa
Cause and Effect Diagram
Man Machine
Effect
Material Method
48
SPC - Statistical Process Control in Injection
Molding and Extrusion By Chris Rauwendaal
49
Designed Experiments
50
Designed Experiments
∂Y ∂Y
∆Y = ∆α + ∆u
∂α ∂u
sensitivity
output variation control action
disturbance
output, Y
input, α 51
CVD Process Example
4 factor, 3 level experiment
• Temperature
“Factors”
• Pressure
Effect on Defects
• Settling time
• Cleaning Method
52
Designed Experiments
Temp “T” (3 settings = levels)
Pressure “P” (3 settings)
Time “t” (3 values)
Cleaning Methods “K” (3 types)
How Many Experiments?
Full Factorial gives: 3x3x3x3= 34 = 81
53
But what if we varied
more than one factor at a time?
Our strategy would be (T2:P,t,K)
equivalent to ( 2:1,2,3)
measuring one of the ( 2:2,3,1)
factors, say
temperature, while ( 2:3,1,2).
“randomizing” the An orthogonal array
other factors. For satisfies this
example measure T2 requirement for all
with all combinations factors at all levels.
of the other factors
e.g.
54
“Orthogonal Array” for 4 factors at 3 levels.
Only 9 experiments are needed
temp pressure time clean
Exp
1 1 1 1 1
2 1 2 2 2
3 1 3 3 3
4 2 1 2 3
5 2 2 3 1
6 2 3 1 2
7 3 1 3 2
8 3 2 1 3
9 3 3 2 1 55
Matrix Experiments
1. Use orthogonal matrix to plan
experiments
2. Additive Model Approximation
3. Balancing property (orthogonality) Any two
columns contains all possible pairs e.g (A1, B2) etc
Levels*
Factors
1 2 3
A. Temperature [°C] To - 25 To To + 25
57
The Experiment Matrix
Factors
Observation
Exp. # Pressure
Temperature Settling Time Cleaning η* [dB]
(A) (B) (C) Method (D)
1 1 1 1 1 η1 = -20
2 1 2 2 2 η2 = -10
3 1 3 3 3 η3 = -30
4 2 1 2 3 η4 = -25
5 2 2 3 1 η5 = -45
6 2 3 1 2 η6 = -65
7 3 1 3 2 η7 = -45
8 3 2 1 3 η8 = -65
9 3 3 2 1 η9 = -70
m = 19 ∑ηi
i =9
Effect = mA1 − m
Note: The averaging for each level is done in a balanced manner
over all levels of the other three factors 59
Result Analysis
m = -41.67 dB
A1B1C2D2 or A1B1C2D3
61
Additive Model for Factor Effects
η (Ai , B j , Ck , Dl ) = µ + ai + b j + ck + d l + e
where: µ = over all mean of η for the experimental region
ai = deviation from µ by setting factor A at level Ai
e = error of additive approximation + experimental error
63
Example:
truck front
suspension
assembly
Problem:
warranty
rates
excessive
64
4 plants, 2 frame mfgs
POINTIAC N - - + DANA
OSHAWA N - N N DANA
Problem identified by
John Dagg as the
automatic torque sequence
on the upper control arm
67
Quality Homework
1. Can you come up with examples of products that have
requirements that exceed their Process Capabilities? Explain?
2. Can you design an orthogonal array for 3 factors at 2 levels?
3. Experience shows that when composites are cured by
autoclave processing on one sided tools the variation in
thickness is about 7%. After careful measurements
of the prepreg thickness, it is determined that their variation
is about 7%. What can you tell about the source of variation?
68
Homework; Consider the final dimension and
variation of a stack of n correlated blocks.
1, 2 ……
If USL – LSL = ∆, σ = σ’, and Cp = 1 for each block,
a) How many stacks are out of compliance? What is the
Cp for the stack? Assume that USL-LSL = n∆ for the
stack.
b) Now USL-LSL=∆, σ=10σ’, what is Cp for the stack?
How many stacks are out of spec?
c) Repeat a) with σ=100σ’
Assume that µ = target value, i.e. the variation is centered
on the target.
69
Addendum: Taguchi
Quality loss calculation
By Taguchi in
Introduction to Quality
Engineering
1986