Sei sulla pagina 1di 30

Capability Analysis using JMP

CHONG SOOK FERN


14 AUG 2014
2014 Finisar Corporation, Confidential

Discussion Outline

2014 Finisar Corporation Confidential

SPC Goals
When special cause variation is present...
Unpredictable
???
Target

Target

Target

Target

Day 1

Day 2

Day 3

Tomorrow

After achieve a predictable process, compare the distribution with


specification limits & reduce the product variability

Target

Target

Target

Yesterday

Today

Tomorrow

Target
LSL

Target
USL

2014 Finisar Corporation Confidential

LSL

USL
3

General Flow for Cpk Analysis


Phase Exit Analysis

Yield Improvement Initiative

Test Time Reduction Initiative

Data Collection

Data Collection

Data Collection

Take cumm data that prior to any


rework
1 record per module
Include fail data (that require rework
in order to pass)

Take 1 data
1 record per module
Include fail data that is product
performance related
(just a guideline, focus should be
more on failure pareto analysis)

Take 1st pass data


(there should be minimal amount of fail
data as only stable process should be
considered for this option)
1 record per module
Minimum 100 individual data points
Exclude fail data that due to assignable
causes
(tester wrong reading, typo)

st

Distribution Study
Normality Assessment:
Assess & locate the best fitted distribution
Outlier:
Remove outlier that due to assignable causes
Typo error, tester wrong reading, defective component
that is batch related issue
Remain outlier that is product performance related
Defective / bad performance component that unable to
be resolved for time being, tuning related issue

Cpk Analysis
Calculate Cpk based on the best fitted distribution
Calculation may be done based on diff channel / temp depends on the objective of study
2014 Finisar Corporation Confidential

Some sharing from Phil Kiely on data collection for capability study:
Data could be collected using the following options:
1st pass data.this will have one record per module, but the data will include any issues
associated with tester capability, etc
Cum data, with rework..this will have one record per module, but the data is distorted as
modules can be reworked many times to make work (but its hard to call modules that require lots
of rework capable)
All data.failing modules may have many records, and so will distort this dataset it will be
excessively pessimistic as to the actual capability
Passing data only..this will have at most one record per module, but on the whole will likely
give an excessively optimistic view of the actual capability
Cum data, prior to any rework.this will have one record per module, and will include failing
data for any module that required rework to make it ultimately passthis is the preferred dataset,
and is aligned with the Cum yield, no rework requirement that is set for phase exits

Example data

The attached query is an example of how to get this data (example is for QSFP LR module)
The attached data is an example output from that query
Note: the drawback of the query as written is that if the module was reworked as part of 2 nd
ops (ie:prior to any testing), no data will be shown for that module. This should not be too
much of a problem, as it is expected those modules should perform the same in testing as
those which did not get reworked in 2 nd ops
Example query

2014 Finisar Corporation Confidential

1. Copy data from Excel &


paste into JMP data file

2. Click Analyze > Distribution

3. Put the study column into Y box


2014 Finisar Corporation Confidential

Click on the red button > continuous fit > Normal to


check whether distribution is normally distributed

Parameter for normal distribution


2014 Finisar Corporation Confidential

Perform Diagnostic Plot &


Goodness of Fit test to
assess the normality of
data

Parameter for normal distribution


2014 Finisar Corporation Confidential

95% confidence
interval

Normal:
P-value > 0.05

2014 Finisar Corporation Confidential

Alternatively, distribution study can be


performed by fitting all possible
distributions & compare all in 1 view

2014 Finisar Corporation Confidential

10

JMP will display the best fitted distribution


by default
In this example, the data is normally
distributed

2014 Finisar Corporation Confidential

11

2. Input the LSL & USL


* The default sigma used is
Long Term Sigma

3. Choose the best fitted


distribution

1. Click Capability Analysis


from the red button

2014 Finisar Corporation Confidential

12

2014 Finisar Corporation Confidential

13

Eg 1: Non-normal Data
The crossing is tuned just within spec if it fails on
the lower side, not tuned to center of spec
range

Cpk computed based on


normal distribution

Cpk computed based on


non-normal distribution

In this case, the normal fitted distribution


(right) can be seen not to be very
representative, the non-normal fit (left) is
much more accurate & gives better Ppk.

2014 Finisar Corporation Confidential

14

Eg 2: Non-normal Data with outliers


The ER is tuned based on other params
basically there are 3 distributions

Outlier:
> Upper Quartile + 1.5*IQR
< Lower Quartile 1.5*IQR
Cpk computed based on
normal distribution

Cpk computed based on


non-normal distribution

In this case, the normal fitted distribution (left) can be seen to not be
very
representative, the non-normal fit (right) is much more accurate.
2014 Finisar Corporation Confidential

15

Eg 3: Non-normal Data with outliers


The ER is tuned based on other params
basically there are 3 distributions
Outlier:
> Upper Quartile + 1.5*IQR
< Lower Quartile 1.5*IQR

Cpk computed based on


normal distribution

recommend either staying with normal statistics for calculation of Cpk,


and/or use a filter to remove outliers in JMP, use the robust mean and
standard deviation, see the next slide
2014 Finisar Corporation Confidential

Cpk computed based


on non-normal
distribution

16

Eg 3: Non-normal Data with outliers


This is the same dataset as slide 16
but we have now included robust values
for mean and standard deviation

Cpk computed based on


normal distribution
Regular mean & std dev

Remark (reason code) should be provided to audience if outliers are removed


Cpk should always be reported together with distribution plot for audience
better understanding
2014 Finisar Corporation Confidential

Cpk computed based


on normal distribution
Robust mean & std dev

17

Process Capability Cpk vs Process Performance Ppk

2014 Finisar Corporation Confidential

18

Process Capability Cpk vs Process Performance Ppk

2014 Finisar Corporation Confidential

19

Theoretical Relationship Between Cpk, Process Yield & PPM


PPM

Sigma Level

Yield

X-Bar 0.5
X-Bar 1
X-Bar 1.5
X-Bar 2
X-Bar 2.5
X-Bar 3
X-Bar 3.5
X-Bar 4
X-Bar 4.5
X-Bar 5
X-Bar 5.5
X-Bar 6

38.30%
68.27%
86.64%
95.45%
98.76%
99.73%
99.95%
99.9937%
99.99966%
99.999943%
99.999996%
99.9999998%

DPMO

(Part per million)

(defect per million


opportunity)

617,000
317,310
133,612
45,500
12,419
2,700
465
63
3.4
0.57
0.038
0.002

841,345
691,462
500,000
308,538
158,655
66,807
22,750
6,210
1,350
233
32
3.4

Cpk
0.27
0.33
0.50
0.67
0.83
1.00
1.17
1.33
1.50
1.67
1.83
2.00

** All values above are estimation based on standard normal distribution


** For bilateral tolerances, ppm computed based on Cpk / Ppk tends to overestimate defectives
** PPM may or may not = DPMO. If 1 defective part have > 1 defect opportunity, DPMO >>> PPM
Question to ponder: Is the above chart applicable for Ppk??
2014 Finisar Corporation Confidential

20

Standard Normal Distribution

68.26%
95.46%
99.73%
99.993%
99.999943%
99.9999998%
2014 Finisar Corporation Confidential

21

Theoretical Relationship Between Cpk, Process Yield & PPM

2014 Finisar Corporation Confidential

22

Case Study 1: Mean shift > Cpk improved > Yield improved
Eg1: bad sensitivity on 12RZOL product code
Cpk = 0.081

In WW2014
the 1st pass yield in Final is 29.17%
Major contributor is bad sensitivity failure
Failure rate is 9/25 = 36%
The average sensitivity = -25.8 dBm
Upon analysis, the best fitted distribution is
Johnson SI
Cpk estimated is 0.081

2014 Finisar Corporation Confidential

23

Case Study 1: Mean shift > Cpk improved > Yield improved
Upon evaluation, observed apd was over
biased. Action taken to optimize apd manually.

Eg1: bad sensitivity on 12RZOL product code


Cpk = 0.081

In WW4314
the 1st pass yield in Final improved to 52.38%
Failure rate drop from 36% to 7/42 = 16.67%
The average sensitivity = -26.3 dBm

Cpk = 0.252

Upon analysis, the best fitted distribution is


Johnson SI
Cpk improved from 0.081 to 0.252

eg1

Mean shifted
2014 Finisar Corporation Confidential

24

Case Study 2: Spec Relaxation > Cpk improved > Yield improved
Eg 2: OSNR failure on 12RZOL product code

OSNR condition loosen from 13dB to 14dB


Original spec = 2e-7
Cpk = -0.003

eg2

2014 Finisar Corporation Confidential

Revised spec = 3e-5


(as per OCA)
Cpk = 1.101

25

Case Study 3: Distribution change > Cpk improved > Yield improved
Eg3: Rxpowalrmthresh failure on 12RZZT product code

In WW814
the 1st pass yield in corner is 21.8%
Rxpowalrmthresh failure contribute to 10.3% of
total test failure
The big drift between diff temperatures contribute
to big delta that resulted in poor Cpk
Upon analysis, the best fitted distribution is
Johnson SU
Cpk estimated is 0.236

2014 Finisar Corporation Confidential

26

Case Study 3: Distribution change > Cpk improved > Yield improved
Eg3: Rxpowalrmthresh failure on 12RZZT product code
Cpk = 0.236

In WW4714
After cut in improvement action on tuning optimization,
the 1st pass yield in corner improved to 70.3%
Rxpowalrmthresh failure contribute to 4.5% (improved
from previous 10.3%) of total test failure
The big drift between diff temperatures reduced

Cpk = 0.92

Upon analysis, the best fitted distribution is Johnson SI


Note that the mean value before & after improvement
action remained the same as -28dBm
Cpk estimated improved from 0.236 to 0.92

eg3

2014 Finisar Corporation Confidential

27

Appendix A: Elements of Control Chart


Upper Control Limit (XUCL)

Lower Control Limit (XLCL)

Center Line (X-Bar-Bar)

9th Subgroup
X-Bar (9) = 598

Subgroup

Center Line (R-Bar)


Upper Control Limit (RUCL)

1st Subgroup
R (1) = 5

Lower Control Limit (RLCL)


Sample size = 3

Subgro
up
X1
X2
X3

# of subgroup = 20

10

11

12

13

14

15

16

17

18

19

20

XXXXXXXXXXXMean X-Bar1 X-Bar2 X-Bar3 X-Bar4 X-Bar5 X-Bar6 X-Bar7 X-Bar8 598
Bar10 Bar11 Bar12 Bar13 Bar14 Bar15 Bar16 Bar17 Bar18 Bar19 Bar20
Range
5
R2
R3
R4
R5
R6
R7 2014
R8Finisar
R9Corporation
R10 R11
R12 R13 R14 R15 R16 R17 R18 R19
28 R20
Confidential

Appendix B: Example of Different Distributions

2014 Finisar Corporation Confidential

29

Q&A Session

2014 Finisar Corporation Confidential

30

Potrebbero piacerti anche