Zero Defect and Risk Mitigation with
Advanced Analytics
Joy Gandhi, CQO
Anil Gandhi, Ph.D. President and Chief Data Scientist
Qualicent Analytics, Inc
Agenda
Qualicent Introduction
Relevant Trends in the Automotive Industry
Role of Data and Advanced Analytics
Technical Goal and the Analytics Process
Case studies
Advanced Analytics for Risk Mitigation in the APQP
Enhanced Manufacturing Anomaly Detection through
Analytics
Summary
Qualicent Introduction
Services
Advanced Analytics
Quality Engineering/Failure Diagnostics
Big Data/IoT Data Integration
Software
ZeroDefectMiner software for all industries
ZeroXMiner for healthcare and IoT
ABATE Risk Software-Service package
Electronics in Automotive
are pervasive
Source: McKinsey
http://www.mckinsey.com/client_service/semi
conductors/latest_thinking
A modern navigation and control panel in a high-end connected car.
These type of cars have as many as 50-75 ECUs making them truly distributed computers on
wheels. (T.Johnson et.al, Univ of Texas Dept of Computer Science and Engineering)
Recall Trend
Significant increase in the recalls
Source: 2015 Automotive Warranty and Recall Report, Stout, Risius and Ross
Note: SRR defines electricals as ignition module and switch, starter assembly, battery, instrument panel, various wiring
EWR Trend - Electricals
Significant increase in the EWRs with injury/death
from electricals
Source: 2015 Automotive Warranty and Recall Report, Stout, Risius and Ross
Confirmed by other Researchers:
T. Johnson, et.al. University of Texas Dept of Computer Science and Engineering) and University of Waterloo (Dept of ECE) Paper confirms the clear
rise in the electronic/electrical hazards and risk related notifications in motor vehicles in US, Canada and Europe.
The Solution is in your Data
Evolution of Advanced Analytics
Why is it difficult to achieve Zero Defects?
Business Problem = Reducing Risk
RISK COST
Field Failure Manufacture
Risk from bad design
Risk from manufacturing
Human Errors
Systemic this presentation
Reducing Risk from Manufacturing
RISK COST
Field Failure Manufacture
Prevent @Process @ Suppliers
Contain @Manufacturing
Marginalities = Units that pass SPC for each and all tests
but with all tests taken together the unit might be at
Predictors for large excursions / large effects not difficult to sourceBUT
Biggest field failure losses are from marginal effects and/or intermittent
deviations over extended periods
Marginal effects are difficult to detect with standard methods because of
high dimensionality, noise, small # of fails,
14
2s 2s
6s 2s
__
OUTLIER
6s 6s ?
Source: Mentor Graphics, 2012
o 1000s of components 10,000s of solder points, 100s part SKUs & suppliers
o Each parameter could be within tolerance but combination of parameters may be an outlier
o Lots of available multi-variate combinations which can make the unit an outlier
Need advanced methods to detect anomalous parts
Complex devices = Large number of influences / dimensions
Interactions
reflectivity
settling Small number of fails
time
thickness Impossible to model on physics
(too many interaction possibilities)
(skewed dataset)
resistance
capacitance
Tolerances based on individual parameters
.
2s 2s
6s 2s
__
6s 6s ? Ship marginal product
Yes!
Analytics Process Summary
Traditional: ANOVA, t-test
screen / coarse reduce
Machine learning model Composite distance
cluster analysis
1. Operating and exclusion
zones for design
2. Anomaly detection
visualization / client
Case Study 1
Who Automotive Semiconductors
KPI Field Failure
How Composite Distance
Result Detect field failures with high class
purity
VarX
VarY
VarZ
no yes
Outlier
Composite Distance
@6 @7
predicted predicted
pass fail pass fail
actual
actual
pass 6,974 15 6,989 pass 6,981 8 6,989
fail 0 2 2 fail 1 1 2
6,974 17 6,991 6,982 9 6,991
Yield Hit = 0.2%
Composite Distance
@6
pass fail
pass 6,974 15 6,989
fail 0 2 2
6,974 17 6,991
Yield Hit = 0.2%
Topmost parameter
Composite Distance
Incumbent Method Risk Assessment
? Project the number of units that will likely fail in the field in the next 10 years
Distance Method Risk Assessment
These units that will likely fail in the field in the next 10 years
Purity
Accuracy
Composite distance
Top Parameter
UCL = Median + x * robust sigma
Case Study 2
Who Electronic Manufacturing
KPI Field Failure
How Composite Distance
Result Detect almost all field failures with
high class purity
median + 6*robust s
Composite Distance
USL
Topmost parameter
Five out of seven field failures are detected by Composite Distanceat low cost
Composite Distance
predicted
pass fail
actual
pass 18,399 5 18,404
fail 2 5 7
18,401 10 18,411
predicted
Topmost parameter
pass fail
actual
pass 18,288 116 18,404
fail 3 4 7
18,291 120 18,411
Composite Distance offers significant improvement over single parameter controls
Pattern Discovery
Deductive 1. Make a hypothesis based on prior knowledge
Reasoning 2. Test the hypothesis
Traditional Statistics
DISCOVER PATTERNS IMPOSSIBLE TO HYPOTHESIZE
Machine Learning
Inductive
1. Discover patterns, discover hypothesis
Reasoning
2. Check if patterns have material meaning
When: Development, Pre-launch, Early production, HVM
Why: Process optimization, Exclusion Zones
How: Exact Models based on machine learning
Class: Supervised
Model Discovery
INPUTS Strategic OUTPUTS
y = f(x)
Field fail, yield, quality,
Thickness, resistance safety and
capacitance, time, effectiveness metrics,
x = x
Tactical
Anomaly Detection
When : HVM
Why: Containment, feedback to suppliers for prevention
How: Iterative Distance methods
Class: Unsupervised
Case Study 3
Who Large Semiconductor Company
KPI Yield
How Machine learning algorithms
Result Revenue increase by > $ MM/quarter
Rule Discovery
Variable M Variable Q Variable T
Variables M, Q and T individually have no influence on Metric of Interest (MOI)
Data is normalized, scaled and transformed
Rule Discovery / Machine learning
Variable M Variable Q Variable T
0
1
100 150 200 250 300 700 750 800 850 900 9950 10000 10050 10100
+ +
100 150 200 250 300 700 750 800 850 900 9950 10000 10050 10100
Variables M, Q and T interactively strongly influence the output
0.8 RESULT:
0.6 EXCLUSION ZONE
0.4 M < 191
0.2 Q < 812
0.0
T > 10,006
Yield = 0 Yield = 1
Case Study 4
Who PV Solar Company
KPI Cell Efficiency
How Machine learning algorithms
Result Prevent cell efficiency loss by 30%
Solar Panel Line Flow
Measurement at four sites all passing inspection but low cell efficiency
Algorithms discovered that its the ratio that matters
= PATTERN DISCOVERY
Parameters A, B, C, D fully in control and within normal distribution
I E A
J F B
K G C
L H D
Case Study 4
Before Date X
After Date X
A
C
Machine learning algorithms discover ratio of A/C as critical parameter (not predicted
by domain experts, but later successfully explained by experts)
Case Study 4: Solar
EXCLUSION ZONE:
Y - low process metric readings (< 24.5)
X -low in line measure(< 81)
Z (date) > something
Machine learning model predicts ~31% reduction in EFF in exclusion zone
Advanced Analytics in the Entire
Product Lifecycle
Proactive Analytics for Risk Mitigation at every APQP Phase
Solution: Analytics in the Product Lifecycle
Warranty and Advanced Product and
supplier data Analytics Process data
Pattern discovery Anomaly detection
Pre-prototype Phases
Product Process
Concept Verification
Design Design
Define requirements Design product, process Test device function
Select materials, FMEA and DFX Predictive modeling
suppliers Supplier qual data functional and
Identify similar products application data;
Anomaly detection on anomaly detection
Get relevant historical supplier material and
data process data Optimize product and
process design based
Model Discovery CA on material and
process Corrective action on
Rules discovery anomalies
Adjust process or design Optimize product and
to rules for zero defect process design
Data Sources and Outcomes
Special/Critical
Materials
Special/Critical Supplier Data DFMEA special
Parts/Process function/dimensi
Data ons
Relevant
Historical Process FMEA
Warranty/Field special process
Failure Data characteristics
Datasheet
DFM, Process and
Design CA/CI Analytics Specification
Outcome Control
Outcome
Techniques Machine Learning Composite Distance
Validation, Safe Launch and HVM
Safe High Volume
Validation
Launch/SOP Production
Customer Qualification Optimize Yields Ongoing Production Test and
Validation/Application Anomaly detection on pilot Inspection
Testing Predictive Model Anomaly detection on test
Predictive modeling Refinement data
Rules discovery Corrective Action on Anomaly Detection on
material and process supplier data
Anomaly detection
Optimize product and Corrective action on
Optimize Design anomalies
process design based on
Manufacturing process predictive model Corrective Action on
corrective actions maverick/high DPPM lots
Worldwide Studies
The Composite Distance technique has been proven to accurately detect field
failures from manufacturing data.
COMPOSITE DISTANCE CHART
Data involved key OEMs, Tier 1, Tier 2 and Tier 3 suppliers
Composite Distance: Cost Impact
7 out of 10 field failures have been detected
Cost Analysis per Part in a Tier 2 supplier
Typical electronic board
1 year period, 91 failures at Tier 1 and OEM
Estimated total cost of failure handling =$1.7M
Cost savings from detection of 70% of
failures~$1M
Impact to reputation and loss of business are
not included
Composite Distance Use Cases for SQM
Supplier 1 Supplier 2
Data Mirror Data Mirror
IQC In-process Test
Engineering Mfg
Server
Early Warning Process for Containment
Sample OOC Action Plan
OOC
detected
Put product
on hold
Onsite Eng Medium
Risk? Stress to fail
Dispositions
Low
High
Variables of Perform FA
Importance
Eliminate Root Process is crucial for full
Cause Issue Resolution
Purge or Recall
Design Verification Validation HV Production
Automotive
Sample Pre-proto-type A, B samples C, D Samples Production
Phase
Model Historical Data Model with A,B data Model with C, D data
Extract operating and Extract operating and Ongoing Outlier
Extract operating and
exclusion zones exclusion zones Detection
Advanced exclusion zones
Improve product and Outlier Detection for Safe Continuous
Analytics Calculate DPPM
process design Improve product and Launch improvement of
Anomaly detection on process design Improve process for Safe Process/product
supplier data Launch
Goal Prevent Prevent Prevent
Contain Contain
Resolve Resolve
Predictive Models Predictive Models Predictive Models Explanatory Models
Anomaly Detection (Supplier Anomaly Detection Anomaly Detection
Rule Discovery
Data)
Summary
Zero defect can be achieved using Advanced Analytics
Anomaly Detection unsupervised learning
Machine Learning supervised learning
Contain high probability field failures using composite
distance analysis
Defect reduction and yield improvement can be
achieved with predictive models
Root cause identification with explanatory models
Advanced Analytics can be employed in the entire product life-cycle.
THANK YOU!
BACK-UP
Sample Data Stack for Analytics
Process Device Defect Inspection Final Test Field
Unit # Solder Reflow Gas R12 C48 Shorts Bridging Idd Leakage Func Field
Volume Temp Flow
1
0
1
Model/Pattern
Discovery
What are the predictors of DPPM?
Supervised Learning
What are the best operating or process Supervised Learning
Rules Discovery
conditions to achieve low field DPPMs
Anomaly Which parts are highly likely to fail in the
Unsupervised Learning
Detection field?
