Special Topics in Statistics and

Quality Management
IE519 _Summer semester
2023_GJU
Ground Rules
• Learn! Apply!!
• Participate - ask questions.
• Take notes!
• Help each other.
• No PC multi-tasking - email, web surfing, etc. during class.
• Mobile phone on silent/vibrate mode.
• Be on time.
• What else?
Homework and Attendance
Homework Attendance:
• 100% completion required for • Please notify instructor prior to class if
credit. you cannot attend.
• Submit hard copy to instructor. • Unexcused absences are not acceptable.
• This icon will appear • Please sign-in at each class.
whenever a homework is • If you miss a class it is your responsibility
assigned: to catch-up by reading the notes, doing
the in-class exercises/examples, and
homework.
• Show exercises and examples to
instructor.
 Outline
• Module 1: Introduction
Assign HW#1

• Module 2: Process and PCS Health

Assign HW#2

• Module 3: Capability Assessment

• Module 4: Multiple Tools
Assign HW#3

• Module 5: RFC Health

• Module 6: Advanced Control Charts
• Module 7: Advanced Computing Lab
Assign HW#4
Module 1: Introduction
Module 1: Outline
Types of Control Charts
Types of control charts
Objectives
• At the end of this lesson you should be able to:
• Describe the key factors to consider for determining the type of control chart
to use.
• Map these factors to the different types of control charts.
• Identify these factors for parameters in your module.
• Describe the relation of Process Health and PCS Health assessment to the
control chart factors.
Importance of Control Chart Type
• The methods for assessing PCS Health depend primarily on the type
of control chart.
• The methods for assessing Process Health depend primarily on the
type of data.
• Because of this dependency:
• PCS courses are organized around the types of control charts.
• Then for each type of control chart, the methods for Process and PCS health
assessment are described.
Types of Control Charts: Factors
• Selecting the “best” control chart depends on careful consideration of many
factors.
• Primary Factors:
1. Data Type – Continuous, Binomial
2. Monitoring Level – Lot/Group Level, Unit Level
3. Summary Statistics – Process Mean, Process Variation
4. Detection Objective – Equipment/Entity, Differential

• Secondary Factors:
5. Sample Size – Constant or Varying
6. Matching – Matchable or Matchable with Delta
7. Entities – Single or Multiple.
8. Sidedness – One Sided or Two Sided limits
1. Data Type
• Continuous
• Values of continuous variables fall at any point along a continuum. Theoretically may take on
any value in a range.
• Continuous variables always have numeric values.
• Examples: Temperature, Viscosity, Thickness, Contact Resistance, Offset, Coplanarity

• Binomial
• Values of discrete variable only have two possible values.
• Pass/Fail, Good/Bad
• Examples: Yield, Loss Codes, Test Bins, Recovery, Retest
2. Monitoring Level
• Group Level
• Monitoring lot or group-level summary statistics.
• OOC decisions are not made until the entire sample is collected and available.

• Unit Level
• Monitoring individual “unit” (package, die, bond, etc.) level data points within
a product.
• OOC decisions are made after the data for each “unit” is collected.
• “Real Time” monitoring
3. Summary Statistics
• Process Mean
• Monitoring to detect changes in the location of the process mean.

• Process Variance
• Monitoring to detect changes in the process variation.

Note: Binomial data required different charts discussed late in the semester
4. Detection Objective
• Equipment/Entity/Tool Level
• Monitoring each link, tool, entity, or sub-entity independently to detect
instability.
• In contrast to the differential below.

• Differential
• Monitoring the relative performance of links, tools, entities, or sub-entities to
only detect differences in performance.
Entity/Tool Level Definition
• Need to carefully define the “entity/tool level” for all control charts.
• Most processes or links have multiple tools, entities, or sub-entities.
• For example:
Primary Types of Control Charts
• For each combination of these primary factors, there is a
recommended control chart.

• The methods for determining control limits and the assessment of

Process and PCS Health will be taught together for each type of
control chart.
Parameter Classification
• Depending on the parameter measured, the control chart may or
may not have specification limits.

• Parameters are classified as:

• Key Process Parameter
• Control Process Parameter
• Data Pass Thru
Parameter Classification: Definition

• Key Process Parameter (KPP):

• Direct relation to one or more output parameters (i.e., yield, quality, reliability, or performance)
• Has spec limits
• Online system must have control limits and rules
• Control Process Parameter (CPP):
• Indirect relation to one or more output parameters (i.e., yield, quality, reliability, or performance)
• Directly affects other process parameters
• May act in tandem with other process parameters to affect output.
• Does not have spec limits
• Online system must have control limits and rules
• Data Pass-Thru
• A parameter for which data is collected, but there are no online controls
• Does not have spec limits
• Online system does not have control limits or rules
5. Sample Size
• Constant
• The number of data points sampled within a lot is the same for all lots and
entities.

• Varying
• The within lot sample size varies for each lot and/or entity.
6. Matching
• Matchable
• Entities produce “identical” output.

• Matchable with Delta

• Entities do not produce “identical” output. Delta is used to compensate for
equipment-to-equipment differences.
7. Equipment/Entities
• Single Entity
• Monitoring only apply to a single entity.
• Early in development, only a single entity may exist.

• Multiple Entities
• Monitoring multiple entities.
• Typically, after development, all processes have multiple entities.
• The BKM requirement is CE! control limits across a fleet of tools or entities, or sub-
entities within a tool.

• Note: In this course, the single entity/tool case is covered primarily for illustration and the focus is on setting control limits
across multiple entities.
8. Sidedness
• Two Sided Control Limits
• For process means, if both LCL and UCL limits exist, use two-sided control
limits.
• Examples: Placement offset, thickness, weight

• One-Sided Control Limits

• If only LCL or UCL, use one-sided control limits.
• Examples: Pin radial offset, LGA coplanarity, High-speed shear, Voids
• Sigma, range, max, and min are typically one-sided.
Summary: Control Chart Types
• Determine the type of control chart to implement by considering the
4 primary factors:
• Data Type, Monitoring Level, Summary Statistics & Detection Objective

• Adjust the PCS and Process Health assessment methods by

considering the 4 secondary factors:
• Sample Size, Matching, Entities & Sidedness
Process & PCS Health
Module 2: Outline
• 1: Setting Control Limits for a Single Entity
• 2: Stability Assessment
• 3: Control Limit Validity
• Appendix: Imputed Values
Statistics Review
Statistics Review
Statistics Review
Statistics Review
Statistics Review
Statistics Review
Statistics Review
Statistics Review
Setting control limits for a single
entity
Objectives
• At the end of this lesson you should be able to:
• Justify the use of 3 sigma limits to determine if a process is in or out of
control.
• Understand the differences between individual, 𝑋ത and s control charts.
• Select the correct method for calculating continuous control limits for a single
entity.
• Understand the purpose of outlier screening and median moving range
computation.
• Define the minimum data required for setting control limits.
Types of Control Charts
• There are different types of control charts for continuous data depending on the
sampling plan.

Control Chart Application

• Sample size n = 1.
Individual Chart
• Each raw data point is plotted on the control chart.

• Sample size n > 1.

ഥ Chart
𝑿 • Group means (𝑋ത chart) and within group standard deviations
s Chart (s chart) are plotted.

• Both charts are used together.

Control Charts
• A control chart is a trend chart with the following elements:
• Center Line (CL)
• Control Limits
• Lower Control Limit (LCL)
• Upper Control Limit (UCL)

How do you set the centerline and control limits?

Center Lines
• All control charts must have center lines.
• The center line is either:
• Process target
• Process mean
• The choice of center line depends on:
• Chart sidedness
• One sided or two sided control/spec limits
• Is the target achievable?
Center Line Examples
• For each parameter, decide whether to set the centerline at the target
or mean.
Target
Parameter Limits Ideal Target Center Line
Achievable?
Placement
two sided 0 yes target
Offset

Flux Weight two sided 20 yes target

one sided
Voids 0 no mean
upper
High Speed one sided
infinite no mean
Shear lower
Coplanarity two sided unknown no mean
Control Limits
• Statistical control limits are set at 3 standard deviations from the
centerline.
𝐿𝐶𝐿 = 𝐶𝐿 − 3𝜎

𝑈𝐶𝐿 = 𝐶𝐿 + 3𝜎

• Limits can be either one sided or two sided.

• Two sided limits are typically symmetric about the centerline.
Why 3 Sigma Limits?
• Historically, control limits are placed at three standard deviations from the mean, m ± 3s, for the
following reasons:
• Represents the typical range of process performance.
• Effectively signals ‘rare’ events.
• Minimizes time spent unnecessarily looking for assignable causes.
2 Sigma Limits 3 Sigma Limits 4 Sigma Limits

Too Tight: Response would result Just Right! Minimize Tampering Too Loose: No Response would
in Tampering and Ignoring result in Ignoring
False Alarm Rates
The normal distribution can be used to estimate false alarm rates under
stable “common cause” conditions.

• 2 sigma limits will falsely trigger 5% (50 out of 1000) of the time.
• 3 sigma limits will falsely trigger 0.3% (3 out of 1000) of the time.

Can you live with a 0.3% false alarm rate?

Exceptions
• Do all control charts need to have statistical limits?
• Statistical limits are always the default assumption for any chart.
• Control limits should identify when to respond to the process.
• Potential exceptions:
• Tool to tool differences: Inherent variation between tools that cannot be eliminated. (Module 6)
• Degradation: Some parameters are expected to degrade over time.
• Examples: bath concentration, laser power
• Uncertain response: Some parameters do not easily allow for small adjustments to the process.
• Examples: Minimum adjustment > statistical limit.
• Data pass-through: Information charts do not require limits.
• All exceptions must be justified.
Exceptions: Examples
Case Example Control Limit Type
Minimum adjustment > Potentiometer can make adjustments to 0.5 Engineering
Statistical limit and my statistical limits are +/-0.2

Wafer scratch on 5 die in specific location then

scrap the wafer.

Parameter degrades over Tool wear causes the performance to degrade. Engineering
time Parts are then replenished.

Defects increase to a point and then a PM is

performed or RFC executed.
Legacy product Builds for product occur for 6 weeks then not Engineering
manufactured in batches again for another year.

Tool to Tool differences that Metrology tools not able to be matched TTD (Tool to Tool
cannot be eliminated across sites due to incoming facilities Delta)
differences.
Setting Control Limits - How
• Setting control limits for a new parameter:
• Set on first tool Initial Limits:
• Start with legacy limits if applicable.
• Otherwise, leave empty.
• Collect data, evaluate excursions, and define responses.
• Refined limits:
• Minimum of 16 runs.
• 2-3 weeks to capture noise (recommended).
• Use methods in the remainder of this lesson.
Computing Control Limits
• Three methods are available in JMP for control limits:
• Moving Range (MR) Method: “Local” Sigma Method
• Levey Jennings (LJ) Method: “Total” Sigma Method
• When evaluating different methods, ask:
• “How well do the limits separate common cause from special cause variation?”
• The most appropriate limits are the ones that best contain common cause variation and best
identify special cause variation.

All methods use “3s Limits”. The difference is how s is estimated.

Individual Chart Limits
• Example:
• Thickness measurements made to monitor a plating
process.
• Single measurement per lot.
• n=1
• Set limits comparing 2 methods:
• Levey Jennings (LJ)
• Moving Range (MR)

• Set centerline = mean for illustration.

Individual Chart: Levey Jennings
• Levey Jennings method uses the sample standard deviation
to estimate s..
ത 2
σ𝑛𝑖=1(𝑋𝑖 − 𝑋)
𝑠=
𝑛−1

• Reported in distribution platform in.

• Levey Jennings control limits:

𝐿𝐶𝐿 = 𝑋ത − 3𝑠 = 99.68 − 3 ∗ 5.7 = 82.59

𝑈𝐶𝐿 = 𝑋ത + 3𝑠 = 99.68 + 3 ∗ 5.7 = 116.77
Individual Chart: Moving Range
• The moving range is calculated as the absolute
value of the difference between successive
observations.
MR1 = undefined (missing)
MR2 = |106.93 – 101.3| = 5.63
MR3 = |92.86 – 106.93| = 14.07

• Moving Range method of estimating sigma:

𝑀𝑅 6.66
𝜎= = = 5.9
1.128 1.128
Note: 1.128 is a constant.

Levey Jennings estimate ≠ Moving Range estimate of sigma

Individual Chart: Moving Range
• Control limits based on Moving Range:

𝑀𝑅
𝐿𝐶𝐿 = 𝑋ത − 3 = 99.68 − 3 ∗ 5.9 = 81.97
1.128
𝑀𝑅
𝑈𝐶𝐿 = 𝑋ത + 3 = 99.68 + 3 ∗ 5.9 = 117.39
1.128

• Note: The computation of MR requires that the data is in process time order.
Example: Setting Limits using Minitab
Thickness measurements made following a plating step.
• Response: Thickness

• Sampling Plan:
– Single measurement per lot (n=1)
– 30 lots

Task: Create an Individual Control Chart.

Compare moving range limits to Levey Jennings limits.
Example: MR Method
 Suggested Limits use the Moving Range Method:
𝑀𝑅
𝐿𝐶𝐿 = 𝑋ത − 3 = 99.68 − 3 ∗ 5.9 = 81.97
1.128
𝑀𝑅
𝑈𝐶𝐿 = 𝑋ത + 3 = 99.68 + 3 ∗ 5.9 = 117.39
1.128

Do the control limits appear to distinguish between

common and special cause?
Example: LJ Method
 Compare the moving range method with Levey Jennings:
– Change the “Control Limit Width” by typing in the Levey Jennings estimate of 3 sigma.
(17.09 = 3*5.7)

– Now the “User Defined” limits reflect the LJ method:

𝐿𝐶𝐿 = 𝑋ത − 3𝑠 = 99.68 − 3 ∗ 5.7 = 82.59

𝑈𝐶𝐿 = 𝑋ത + 3𝑠 = 99.68 + 3 ∗ 5.7 = 116.77

How do the MR and LJ methods compare?

MR vs. LJ Methods
• For a stable process, the MR method and the LJ method will produce
nearly identical estimates of s and corresponding control limits.

• But for an unstable process, the control limits can be very different.
MR vs. LJ Methods
Moving Range Method Levey Jennings Method

• Local estimate of s. • Total estimate of s.

• Limits are narrower than Levey • Limits accommodate all variation in
Jennings method in the presence the timeframe, including shifts and
of trends and shifts. trends.
• Better able to detect trends • Highly dependent on
and shifts. timeframe selected.
Outliers
• The MR method is generally “robust” to process instabilities such as
trends and shifts, but not in the presence of large outliers.
• Estimate of s will be inflated.
• LJ method has the same problem.
• To minimize the affect of outliers, two additional features are added
to the BKM:
• Outlier screening
• Median moving range
Outlier Screening
• Outlier screening detects outliers that are beyond the “mean +/- 4
sigma”.
• Use the Sigma Method to identify these outliers.

• Since outliers will inflate estimates of the mean and sigma, robust
estimates are used:
• Replace mean with median.
• Replace sigma with pseudo sigma based on percentile and median.
Outlier Screening (cont)
• Calculate Upper Screen and Lower Screen values:

75𝑡ℎ 𝑃𝑒𝑟𝑐𝑒𝑛𝑡𝑖𝑙𝑒 − 𝑀𝑒𝑑𝑖𝑎𝑛

𝑈𝑝𝑝𝑒𝑟 𝑆𝑐𝑟𝑒𝑒𝑛 = 𝑚𝑒𝑑𝑖𝑎𝑛 + 𝑘 ∗
0.6745
𝑀𝑒𝑑𝑖𝑎𝑛 − 25𝑡ℎ 𝑃𝑒𝑟𝑐𝑒𝑛𝑡𝑖𝑙𝑒
𝐿𝑜𝑤𝑒𝑟 𝑆𝑐𝑟𝑒𝑒𝑛 = 𝑚𝑒𝑑𝑖𝑎𝑛 − 𝑘 ∗
0.6745

• Default value for k is 4.

• Exclude runs from control limit computations if above Upper Screen or below
Lower Screen.
• Excluded runs are still included in computing OOC rates.
Example: Effect of Outliers
 Thickness data with one outlier.
 MR limits using all data:

How much influence does the outlier have on control limits?

Example: Outlier Screening
 Calculate Upper Screen and Lower Screen values:
105.6 − 99.4
𝑈𝑝𝑝𝑒𝑟 𝑆𝑐𝑟𝑒𝑒𝑛 = 99.4 + 4 ∗ = 135.7
0.6745
99.4 − 94.7
𝐿𝑜𝑤𝑒𝑟 𝑆𝑐𝑟𝑒𝑒𝑛 = 99.4 − 4 ∗ = 71.7
0.6745

 Identify points beyond screening limits.

Example: Outlier Screening
 Recalculate control limits with outlier excluded.

Limits without outlier better represent common cause variation.

Calculating Limits with Excluded
• Consider the outlier for lot 18.
Data
 2 methods for calculating MR by excluding
lot 18.
 Method 1
– Set both affected MR’s to missing (lots 18 and 19)
 Method 2
– Set MR for lot 18 to missing.
– Calculate MR for lot 19 by skipping over lot 18.
– MR19 = 97.42 – 93.19 = 4.23
Median Moving Range Method
• Note that in the previous example, if the outlier value had been 135
instead of 140:
• Then it would not have been removed by outlier screening.
• Resulting control limits would be too wide.

Any other methods robust to outliers?

Median Moving Range Method
• Recall estimate of sigma using Mean MR.

𝑀𝑅
𝜎=
1.128

• Estimate of sigma using Median MR.

𝑀𝑅50
𝜎=
0.954
Note: 0.954 is a constant.
Median is less influenced by outliers
Example: Mean vs Median MR
𝑀𝑅 8.94
𝜎𝑀𝑅 = = =7.92
1.128 1.128

𝑀𝑅50 6.68
𝜎𝑀𝑅50 = = =7.00
0.954 0.954
Exercise : Dispense Weight
• The dataset DispenseWeight contains dispense weight measurements from 45
lots.
• The process target is 15 mg.
• Note: Be sure to specify the target in SOS.
• Note: Use the default of k=4 for outlier screening.

• Set control limits for this process:

• How many lots were excluded as outliers?
• What are the suggested control limits?
• How many lots are OOC?
• What is the predicted %OOC for the new control limits?
• Do these control limits seem reasonable?
Exercise (cont)
• Run Minitab again without outlier screening.

• How do these control limits compare to the previous?

• Which control limits would you implement?
𝑋ത and s Chart (n >1) Methods
• Control limits for an 𝑋ത and s chart can be calculated
using the following methods:
• Moving Range
• Levey Jennings

• Example:
• Thickness measurements made to monitor a plating
process.
• Five thickness measurements per lot.
• n =5
𝑋ത Chart: Moving Range
• The same moving range formulas for an
individual chart are applied to the column of Lot
Means:

𝐶𝐿 = 𝑋ത = 99.07

𝑀𝑅𝐿𝑜𝑡 𝑀𝑒𝑎𝑛
𝐿𝐶𝐿 = 𝑋ത − 3 = 99.07 − 3 ∗ 6.78 = 78.73
1.128
𝑀𝑅𝐿𝑜𝑡 𝑀𝑒𝑎𝑛
ത
𝑈𝐶𝐿 = 𝑋 + 3 = 99.07 + 3 ∗ 6.78 = 119.41
1.128

𝑀𝑅𝐿𝑜𝑡 𝑀𝑒𝑎𝑛𝑠
• Where = 7.65/1.128 = 6.78 an estimate of
1.128
s.
s Chart: Moving Range Lot Std Dev(Thickness) Moving Range

• The same moving range formulas for an

1 1.03
2 2.17 1.14
3 1.34 0.84
individual chart are applied to the column of Lot 4
5
1.63
3.39
0.3
1.76
Standard Deviations: 6
7
1.57
1.38
1.82
0.19
8 1.02 0.36
9 1.95 0.93

𝐶𝐿 = 𝑋ത𝐿𝑜𝑡𝑆𝑡𝑑𝐷𝑒𝑣 = 1.89 10
11
2.75
1.21
0.8
1.55
12 1.67 0.47
13 1.64 0.04
14 1.94 0.3
𝑀𝑅𝐿𝑜𝑡𝑆𝑡𝑑𝐷𝑒𝑣 15 1.46 0.48

𝐿𝐶𝐿 = 𝑋ത𝐿𝑜𝑡𝑆𝑡𝑑𝐷𝑒𝑣 − 3 = 1.89 − 3 ∗ 0.9 = −0.81 16 1.7 0.24

1.128 17
18
2.4
0.58
0.7
1.82
𝑀𝑅𝐿𝑜𝑡𝑆𝑡𝑑𝐷𝑒𝑣 19 2.93 2.35

𝑈𝐶𝐿 = 𝑋ത𝐿𝑜𝑡𝑆𝑡𝑑𝐷𝑒𝑣 + 3 = 1.89 + 3 ∗ 0.9 = 4.59 20 1.16 1.77

1.128 21
22
3.37
0.85
2.21
2.52
23 1.81 0.96
24 2.49 0.67
25 2 0.49
𝑀𝑅
• Where 𝐿𝑜𝑡𝑆𝑡𝑑𝐷𝑒𝑣 = 1.02/1.128 = 0.9 an estimate of s
26 3.76 1.76
27 2.19 1.56
1.128 28 1.65 0.54
for the control chart. 29 1.44 0.21
30 2.11 0.67
Mean 1.89 1.02
Std Dev 0.77
s Chart Lower Control Limits
• In practice, s charts typically use only an UCL since the objective is to
detect an increase in variation.
• If the computed LCL is < 0, then clearly it is not logical to implement a
LCL.
• In some cases, if the computed LCL > 0, LCL may be implemented to
detect:
• Metrology or data integrity problems, or
• Actual process improvement.
Control Limit BKM Summary
• defaults to start.
• ±3 sigma using mean moving range with ±4 sigma outlier screening.
• Evaluate the separation of special and common cause variation.
• Fine tune accordingly.
• Rounding for convenience.
• Consider other sigma estimates.
• Levey Jennings, Median moving range, different k multiplier for outlier screening
• Manual adjustment.
Example: 𝑋ത and s chart
Thickness measurements made following a plating step.
• Sampling Plan:
• Response: Thickness – 5 measurements per lot (n=5)
– 30 lots

In JMP:
• Open the data set: Thickness003
• Use Minitab- Stat- Control Chart
• Fill in the dialog box.
• Click OK

Task: Create an 𝑋ത Chart using the MR method.

Example: 𝑋ത and s Chart
• Variability chart of raw data
Example: 𝑋ത Chart MR Method

• Do the control limits appear to distinguish between

common and special cause?
Example: s Chart MR Method

• Does the UCL appear to distinguish between common and special cause?
• Note: Computed LCL < 0 is displayed but not used.
 ഥ and s Charts
Summary: 𝑿
Three graphs are required for setting statistical control limits for lot level data:
• Variability chart of raw data
• 𝑋ത Chart
• s Chart
Exercise : Trace Width
• The dataset TraceWidth. contains trace width data from 36 lots.
• The sample size is 4 units per lot.
• The target for the process is 100 um.

• Set control limits for this process.

Exercise (cont)
• How many lots were excluded as outliers?
• For the 𝑋ത chart:
• What are the suggested control limits?
• How many lots are OOC?
• What is the predicted %OOC?
• Do these control charts and limits seem reasonable?
• What are the control limits based on the median MR?
• How would you decide which control limits to implement?
Exercise (cont)
• For the s chart:
• What are the suggested control limits?
• How many lots are OOC?
• What is the predicted %OOC?
• Do these control charts and limits seem reasonable?
Predicted %OOC
• %OOC rates from suggested control limits typically overestimate the
actual %OOC rate that would occur if they were implemented.
Exercise : Thickness
• The dataset Thickness005. contains plating thickness data from 40
lots.
• Entity: Bath 005
• Sample size is 1 unit per lot.
• Target for the process is 100 um.
• Set control limits for this process.
• What is the %OOC rate based on these control limits?
• What %OOC rate would be predicted if these limits were
implemented?
Stability assessment
Objectives
• Describe the difference between online and offline control limits.
• Describe the BKM for assessing process stability using both online and offline
control limits.
• Understand the usage of trend rules.
• perform stability analysis.
Stability Assessment: %OOC
• Stability is assessed using %OOC.
OOC = Out of Control
• A very simple metric.
# 𝑂𝑂𝐶 𝑝𝑜𝑖𝑛𝑡𝑠 𝑜𝑛 𝑐𝑜𝑛𝑡𝑟𝑜𝑙 𝑐ℎ𝑎𝑟𝑡
%𝑂𝑂𝐶 = 100 ∗
# 𝑝𝑜𝑖𝑛𝑡𝑠 𝑜𝑛 𝑐𝑜𝑛𝑡𝑟𝑜𝑙 𝑐ℎ𝑎𝑟𝑡

• Stability can be assessed using online or offline control limits.

• A process is considered stable if %OOC = 0%.
Online vs. Offline Control Limits
• Online Control Limits
• Actual control limits on the control chart in production on the factory floor to make real-time decisions:
• Equipment/Tool down
• Product/Lot on hold
• Response, Execute RFC
• Limits computed from historical datasets or legacy products.
• CE! across the VF
• Under change control SPC++ Control Chart
• Do not dynamically change as
data are collected.
• Not changed frequently.
• Require an update to the automation model.
• Use %OOC to assess stability
• Typically, no trend rules.
Online vs. Offline Control Limits
• Offline Control Limits
• Calculated on a dataset as an offline engineering activity.
• Use outlier screening.
• Depends on the dataset.
• Different data + Different time frame = Different limits.
• Use %OOC to assess stability; may also use trend rules.
JMP Control Chart
120
UC L=117.39
115
110

Thickness
105
100 Avg=99.68
95
90
85
80 LCL=81.98

12

15

18

21

24

27

30
3

9
Lot
Online and Offline Stability
• Let’s review a simple case to illustrate the potential difference in the
online and offline assessment of stability.
Is This Process Stable? (Online)
• How many points are out of
control relative to the online
limits?
• 0
• What is the %OOC rate?
• 0%

Online Control Limits

Conclusion: There were no out of control points signaled on the factory

floor. What does the offline stability assessment indicate?
Is This Process Stable? (Offline)
• What is the %OOC rate?
• 2/52 – 3.8%
• Describe the instability.
• Downward drift over time.

JMP Calculated “Offline”

Control Limits

Conclusion: The process is drifting downward over time.

Stability Assessment BKM
• In daily reviews and data reviews, the stability assessment is based on
online control limits.
• The typical requirement is %OOC ≤ 5%.

Flag: %OOC > 5%

Warning: 2% < %OOC <= 5%

• Note: Stability is still defined as %OOC = 0%.

Stability Analysis BKM
• Stability Assessment:
• Data review and cert requirements are assessed by %OOC based on online
control limits.
• The offline control limits are used to:
• Assess the validity of the online control limits
• Highlight any instabilities that were not detected by the online control charts

Recall:
• Online Control Limits: Used on the factory floor.
• Offline Control Limits: Calculated by JMP or Catalyst.

Use both online and offline control limits together!

Stability
• A stable process is consistent and predictable over time.

Same distribution over time

– Same mean (location)
– Same standard deviation (spread)

• Evaluated with a control chart.

Example: Stability
• A stable process:

• Appears random.
• Results from common cause variation.
• Centered around the center line.
• Variation is contained within control limits (0% OOC).
Trend Rules
• The different types of instabilities:
• Outliers
• Drifts
• Shifts
• Chaos (Any combination of the above)

Trend rules are used in offline analysis to help identify these patterns of
instability.
Types of Instability: Terminology
• There are 4 typical terms used to describe patterns in an unstable process.
1. Outliers

2. Shift (upward)
Types of Instability: Terminology
3. Drift (upward)

4. Chaos
WECO Trend Rules: Zones
WECO Trend Rules
• The following trend rules are used for offline analysis.
Rule Uses Example
#1. A single point is Detects very large, UCL
outside either sudden shifts, excursions,
control limit. or outliers. CL

LCL

#2. 9 or more consecutive Detects small shifts or UCL

points on the same trends.
side of the CL

centerline.
LCL
WECO Trend Rules
Rule Uses Example
#5. 2 out of 3 Detects large shifts.
UCL
consecutive points
+2s
are more than 2
CL
sigma from the CL,
-2s
on the same side. LCL

#6. 4 out of 5 Detects moderate-sized

consecutive points shifts.
on the chart are UCL

more than 1 sigma +1s

away from the CL. CL
-1s

LCL
WECO Trend Rules (Tests)
Test 1 One point beyond Zone A.

Test 2 Nine points in a row in a single (upper or lower) side of Zone C or beyond.

Test 3 Six points in a row steadily increasing or decreasing.

Test 4 Fourteen points in a row alternating up and down.

Two out of three points in a row in Zone A or beyond and the point itself is in
Test 5
Zone A or beyond.
Four out of five points in a row in Zone B or beyond and the point itself is in
Test 6
Zone B or beyond.

Test 7 Fifteen points in a row in Zone C, above and below the center line.

Test 8 Eight points in a row on both sides of the center line with none in Zones C.
Trend Rule BKM
• All online control charts use rule #1.
• Offline control charts use rule #1 and WECO (Western Electric) rules #2, 5 and 6
to highlight instability.
• Offline %OOC computation is based only on rule #1.
• Note: WECO rules aren’t applied when engineering limits are used.
• On an exception basis, online charts may implement selected WECO rules to
more quickly detect shifts and drifts.
• Experience has found that WECO rules are rarely required online if the control limits are statistically
valid.
• EWMA charts (discussed in another lesson) are often a better approach for quickly detecting shifts and
drifts.
Checkpoint : Trends
• Assess the stability in each offline control chart. Trend rules #1, 2, 5, and 6 have been
applied. Describe the nature of the instability where applicable.

Control Chart Interpretation •U

A.

B.
Checkpoint 2.1: Trends (cont)
• Assess the stability in each offline control chart. Trend rules #1, 2, 5, and 6 have been
applied. Describe the nature of the instability where applicable.
Control Chart Interpretation
• Unstable.
C.
– Single o

• Unstable.
D.
– Chaos.
Example: Laser Scribe Placement
• Laser Scribe Process
• A laser is used to blast away (ablate) material in the streets of a wafer prior to saw.
• Key Parameter: Laser scribe placement.
• Measured in the x direction and y direction (horizontal and vertical streets); on both sides of
the scribe (x1 and x2).

Laser Scribe Trench Saw Cut

Wafer Street Active Die

Example: Stability Analysis
• Data Set: LaserCutPlacement
• Response: Laser Cut X1 Placement
• Tool: V WXY 003
• Sampling Plan: 52 lots with 5 data points per lot
• Target = 13 um Chart Type LCL CL UCL
• Online Control Limits (microns): Means 7 13 19
Std Dev 2 5

Task: Perform a stability analysis.

Example: Stability Analysis
• Select Minitab>> >> Stability Analysis.
• Select Laser Cut X1 Placement as the Response Variable.
• Select TXNDATE as the Order By variable.
• The LOT, FACILITY and ENTITY columns should already be pre-selected.
• Note: SOS recognizes some common column names and automatically pre-selects them.
Example: Stability Analysis
Your dialog box should look like this:

Enter the online control limits as follows:

Example: Stability Analysis
• creates 4 sections of output:
• Stability Analysis Indicator Summary
• Control Charts by Entity
• Trend Charts by Facility
• Control Charts (Combined Entities)

• Select the Control Charts by Entity tab.

• Note: Leave the SOS window open for assessing control limit validity in Lesson 3.
Example: Analysis
What is the online %OOC rate?
Conclusions
• 0% for both charts.

Is the process stable based on

the offline control limits?
• No, means chart is OOC.
• 2 out of 52 (3.8%) points
OOC relative to offline
limits.
• Downward shift over
time identified by trend
rule violations.
Control limit validity
Lesson 3: Objectives
• At the end of this lesson you should be able to:
• Describe how to assess control limit validity using CLSR and OCI indicators.
• Use JMP SOS to assess control limit validity.