ATT ar)
a
Sear ea eT a wes
DUE Ce a Cae Ly rent Ty DNL)A Convergent Approach for Problem Solving
Ha Dao, SSA & Company and Bill Maxon, Shai
Abstract
Ithas been argued thatthe best approach to
problem solving involves powerful statistics
Tequiring a good amount of deta, attr all,
‘day's problems are more complex and
require more sophisticated techniques to
‘elo thom. Mast problome salvars are
engineers and engineers work in the realm
of physics and geometry. The laws of physics
are not any different today than they were
‘years ago wnen te so cated simpa
probleins weve a1uund. Whit ake twday’s
tochnioal prablome eoom co tough ie the
inabilty ofthe engineer to think through
th science driving the probe, and
getting distracted with unnecessary
statistical analysis doesn’thelp them,
Instead of quessing at potential causes,
engineers should use a more efficient
stratogy and oliminate whats not causing
tie problem. This paper discusses a
convergent approach to problem solving
with 2 caso study ilustrating how using 2
proaressive search. auicklv and efficient.
eliminated potential causes fora lifted
wire bonds on circuit boards,
Baokground
If you speak to any executive about problem
solving, they wil tell ou they are looking for
5 things: Speed tothe rot cause, Efficiency
with resources and te, Focus to maximize
ity of techniques and results,
ce Work tu eisure Cunsisivity.
Howaver ifyou closely examine your problom
solving strategy. your teams might not he
meeting these § attributes. Speed — are the
teams making progress on a daly basis
solving even the mast dificult problems in
a few weeks if nota few days? ericlency
How many people ore on your teams? If
its more than 2 dedicated problem solvers,
itis too big and you are wasting resources.
Focus are the teams converging on the
root cause or diverging? Simplicity ~ Does
‘he progress ote team make sense’ Ist
easily followed and would the custome be
able to follow along? Standardized work
are al the teams using the same approach
in, LLC
Xx
and are they documenting their work ina
‘common tormat to aid communication and
‘w cuuivey, in # common language, progress
‘and roaulte?
Ytox
Tie ¥ lw X approach starts witha well-
dofined ¥, sometimes referred to a3 the
Green This metric is a measured
response that has engineering leverage
and reveals insight into the physics for the
failure. A good Y to X approach relies on
Ienttying contrast. nis contrast can be
explained by identifying the wause OX)
through a prograssive and convergent
strategy.
This approach eliminates the need to lst
potential X's through brainstorming and
engineering judgment inen tasting eacn to
see ifthey have an effect an the ¥, During
‘2 progressive search investigation, oach
stage is divided into natural buckets or
spits. By leveraging contrast with simple
efficient strategies, these buckets are
progresswvely eliminated and all further
‘wuikis focused on what remains. Now
this contrast may reside between parts,
between processes, or other areas that
may not be so apparent. The trekto
{00d problem solving is identifying the
Mot-so-obvious contrasts, especially
wire yous sick or we You las
no loveragoablo contrast wth parts or
processes. By leveraging contrast between
extremes, @ multitude of X's can be quickly
and eficiently eliminated leaving only few
suspects forthe root cause, The candidate
can then be tested with an
efficient confirmation test.
Why io 9 Vt0 X otratogy 20 efficient? It is
haraisa af the Parata principle, square
root of the sum of the squares, and talking to
the parts. Combining the three provide
‘8 powerful problem solving strategy.
There are not many root Causes
‘Many of focays problems appear to be
complex and overly complicated leading
to the conclusion that there mustbe many
causes or XS. This is simply not true.
Complex problems become understandable
with imple yet very effective convergent
strategies. Itisnt necessary to control or
even idenily many Xo solve difficult
probloms; ity yoars of practice! experience
‘and Pareto tellus thisns of Dollars)
Net Worth (Bi
Pareto 1s me BU/ZU rule, When states for any
ivan effect uly 2096 ofthe X's have ony
noticoable contribution. Vilfrado Parsto, mado
this discovary after studying the distribution
cof wealth of nations under different paltcal
systems and time periods. He concluded
there is always a non-random cistribution of
‘wealth, Not only is your piggy bank affected
but your problem solving strategy s also
affected. For problem solvers, it separates
‘he trivial many from the vital fev
AY = f(AX)
{tho problom ie described as a change in
Y.thera ig ana, mayha two, X's that account
for most ofthis change. Tis isthe X (often
referred to asthe Red X®) you want to
discover first. The real question comes down
tothe techniques. Are your tools designed tv
flush out these daminant X's or are they going
toflugh out any X that has some statistical
significance or correlation tothe ¥? If vou
find yourself disappointed after implementing
corrective actions or contrling the “root
‘cause then your tools are not designed to
find the true root cause or the dominant X
Serre TT
‘Which x has more power
“Th square 1uut of re sum uf Ure squares is
8 statiticl law. It howe us the non-linear
relationship between Xin Y terms. The
power ofthe dominant Xis exponentially,
‘more powerful than any other X!
‘When two independent distributions are
ccumbined, their standard deviations do not
add, but combino ac equaroe. The standard
deviation af tha roeulting distribution wil he
the square root ofthe sum ofthe squares
of the standard deviations of the source
distributions,
{he dominant Xis causing 5 units of
variation in an output and another Xie
contributing unite th eelativn affact is
not 6:2 but 25:4. The total amount of
variations not (5+2) but about 84
{the square root of 25+4),
\ (52+22) = 5.38
What happens when you attempt to control
variation by addressing e root couse other
Uva te domino X® In the prior exemple
if you romove ono unit of variation from tho
smaller cause, the result will be @ reduction
of total variation from 54to about 5.1
V (52412) = 5.09
Alternatively had you removed 1 uni uf
‘variation from the dominant X and lft the
\woaker ¥ lon, th total variation would
have fallen to about 45 units
\ (424-22) = 4.47
Both the dominant ané smaller XS are a true
cause of variation, but the dominant X has
‘much more power over the effect. Itmay have
been statistically significant bu practcally
‘speaking itis worthless end therefore adding
cost to control with litle to na improvement in
vyour metric.A progressive search
The fastest route to finding the dominant X is
through a progressive seerch. This requires,
leveraging contrast between extremes of
the ¥ distribution. By leveraging the extremes,
weak X's are avoided leaving only the
dominant X as the only possibilty of causing
such a large variation nthe ¥. this means
very few parts are needed tu suv problems,
and the otatiatea uoed to inveatigate the
fitferanes are simple and straight forward
letin the prablem solver focus on strateay
and tool selection not statistical analysis
Once the dominant X candidate is identified,
itis now ready forthe tra
XtoY
Confirmation is the only time for an X to Y
approach. Wit a well-structured progressive
scorch, many suspects will be eliminated
leaving only a few and often one candidate
This means you can run a more efficient
confirmation test. Only after effective clue
‘generation (screening experiments and DOES
4are not eficient clue generation tecnniques)
isXtoY appropriate, Atthis point you should
have an understanding of tha physics sllowing
You to make a decision on what to control
oreliminate. The confirmation test now
becomes a way to demonstrate that you truly
understand the problem and can predict the
response with the knowm true root cause.
Conclusior
For any outout, there are a number of input
factors X that affect variation. Given enough
resources and time, you could discover al of
these XS. As a practical matter, however, there
isno need todo this. Pareto tlis us there is,
anly one dominant and tha square root of
the sums of squares says thatthe power of
this Xis far greater than many ofthe other X's
combined. Therefore, « progressive search
is the quickest and most efficient way to find
the dominant X. This is a combination af good
strategy, a balance between engineering
and statistics, and evidence based problem
solving
Case Study Ton Yoor Old Problem
Focus
Electronics manutacture was losing
credibility with ts customer due to several
fallad attompts to colve a prablom invohing
lifted wies hands a eteuit haard eosting
them $2 milion per year. Initially the customer
‘wanted to find the cause for excessive tin
contamination on the gold pads, however,
tin didnot explain al of the contrasts and
the project was redefined as finding the
ro2t cause for lifted wiro bond.
The wire bonder places a ball bond
on the ASIC, then puts a wedge or
stiteh bond on the bond pad
Approach,
The destructive event med it easy Ww Wack
the location of filad wire bonds, however
the project quickly moved to measuring peak
force in grams using a wire pull tester.
Converge
‘The team began the investigation by analyzing
the field returns. A concentration diagram
auickly revealed a nonrandom pattern
between the nine bond sites. The twa lower
bond sites were the ony sites to exhibit ited
wire bonds. A muit-vari revealed a similar
pattern with current production units
Those two nonrandom patterns mado tho
critical split hetween strength and eneray
‘The problem for this particular case is due
to poor strangth ofthe bond, not excessive
nergy in the application, whether it be
énvironmental, mechanical, or electrical
Forensic analysis of tild returns show
ifted wire bonds at sites 1 and 2
‘The next split wes betwen (Wu lle energy
from the wire bondor, contamination on the
hand pad, or inefficient weld ofthe gold wire
tothe gold pad. Furthor forensic analysis of
the field returns provided clues to the answer.
sites, and twas also found at the non-lfted
bond sites. This meant tin was either part of
‘an interaction or nota cause for week bands.
‘SEM analysis showed streaking which is
indicative of an inefficient bond. The transfer
of energy trom the wire bonder to the weld
was beiny compromised, The gold ped was
deflecting under the pressure of the wire
bonder
“in contamination was found atthe raiea pond
Test
Now that we understood the physics ofthe
failure, a model was created and vorifed that
excessive deflection was occurring at sites,
1 and 2. A Spike B vs. C Test confirmed with
95% confidence that thera was an interaction
between tin contamination and pad deflectionUnderstanding the Physics
For ten years, the customer was convinced
‘that in contamination was the sule reason
{or lied wiro bonds. Thoy falled to recognize,
thattin did not oxplain the eantraethetwasn
thar contrasts This ultimately led to the
discovery of an interaction. By revealing
‘and understanding that lifted wire bonds
required high tin and high deflection this
ave tie custunter a much eesier solution
toimplomont and there wae no eoct
ascnciatad withthe corrective nation
Apply
AA simple program change to the wire bonder
‘moved the bong site on stes 1 and 20
Fucativt ar ne ped Usa uid wut delet
Leverage
There were mutiple lines manufacturing this
product, Alllines were updated as well as the
PEMEA and control plan. Actual savings were
‘8 ion annually
Untrasoniowi
ceran
ess waves and streaks from
¢ capillary throughout wedge face.
High bow Derlection inthe Z Axis,
Pull
Strength
(grams)
Minimum Specification
Zero
% Tin
Low High
(< 8%) (> 8%)Contact:
Questions and comments can be sent to:
Ha Dao at: hdao@SSAandoo.com
Bill Maxson at: wrmaxson@shainin.com
Defini
ns, Acronyits, Abbreviations:
80/20 Principle:
An inbuilt imbalance between causes and results,
inputs and vutputs, and effurl and reward. *
BOB:
Best of the Best and Worst of the Worst are tho
‘extreme eamplos that can bo roadily takon from a
distribution of tho Groon Y®.!
Green Y°:
Is a response that has engineering leverage and
reveals insight into the physics for the failure.”
Red X®:
{s the product feature or process setting that drives
‘the Green Y® from BOB to WOW. !
wow:
‘Worst of the Worst and Best ot the Best are the
extreme samples that can be readily taken from
a distribution of the Graen Y®.!
References:
1. Shainin LLC, Problem Solving for Engineering,
vi.1g 2006, Page: 1-2, 1-5, A-32, A-34
2, Richard Koch (1998) The 80/20 Principle
the Secret lo Success By Achieving More
With Less (Currency and Doubleday),
pages: 4, 264
3, ttpy//www.shainin.com
Required
Before
Ee ae
DP DM
——
Stronath of | Ze
Bond ea
Sl
pee \eSion| | Inefficient
2)
Enetgy (C1)
ariatior sal
Dynamic (Bond
Pad is Flat But
Has Bounce)
Bond (B,)
Static (Bon:
Sao
r
Secure
Lis Warped, ete
Red X
OA
The Fall Edition of
Automotive Excellence
Coming Soon!