26
Bradley Novic, Ph.D. PhaseTwo Analytics, LLC Data … Knowledge … Intelligence [email protected]

Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

  • Upload
    others

  • View
    2

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

Bradley Novic, Ph.D. PhaseTwo Analytics, LLC Data … Knowledge … Intelligence

[email protected]

Page 2: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

Reluctant to analyze manufacturing data ◦ Requests to analyze plant data (circa 2000)

◦ Influence of J. S. Hunter … PARC Analysis … skepticism

Why am I seeing useful relationships? ◦ The bad news & the good news

◦ Bad news: Processes are “out of control”

◦ Good news: Lots of strong signals in the data

◦ Bad news: Data assembly is painful … so no one is looking at it

◦ Good news: LOTS of low hanging fruit!!! Big opportunity!!!

Probability of successful discovery higher than I anticipated

High return on investment -> increased demand

Realization: Mining mfg data has perils … mitigated by JMP®

Case studies will highlight perils related to: ◦ Data: Time scale issues … manufacturing specific

Data integration with non-matching timestamps

Integration of static & time-based data

◦ Known relationships UNKNOWN TO YOU! … value of visualization!

◦ Prediction vs Control … the need for guidance in tree building!

PhaseTwo Analytics

Page 3: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

Data Mining Would be Great … if it weren’t for the Data

70-80% of time is spent wrestling with data issues ◦ Data Access … multiple, disparate databases

◦ Completeness … having all the right data … observability!

◦ Data cleaning

◦ Missing data

◦ Data integration & alignment … joining, concatenating, stacking

◦ Time scale differences

Aligning datasets with non-matching time stamps

Integrating Static data + time-based data (batch processes)

PhaseTwo Analytics

Page 4: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

A rolling process produces coils “about” 1 every 20 minutes

The rolling process timestamp reflects the start of coil processing

The rolling process dataset contains process variables, process performance metrics & product quality

Lubricant, sprayed on rolls and the coil, is influential … but quantification of effects is needed

Problem: Lubricant chemistry data are measured less frequently

Lube timestamp will not match the coil process timestamp but it varies smoothly

Alignment is required to model results as a function of both process and chemistry

Here’s a way of aligning data with non-matching timestamps in JMP® using spline predictions

PhaseTwo Analytics

Page 5: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

Objective: Align variable LubeChem1 ( measured 1/day ) with Rolling Process data measured 3/hr

LubeChem1 raw data flexible spline fit

Next: Save the prediction formula

Note: use the same name for the time stamp as in the rolling data set

Create a new variable in the rolling data set ( LubeChem1_Pred )

Use the spline prediction formula as the formula for LubeChem1_Pred

Here are the aligned (predicted) LubeChem1 data:

PhaseTwo Analytics

Page 6: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

An Ingot Casting Example ◦ Static Data (1/cast) … static conditions … single point meas

Holding furnace temp

Coolant Additive %

Grain refiner set point

Drop rate @ cast start

Steady state drop rate

Ingot head ml set point

Metal treatment set point

Ambient dew point,

% humidity

Temperature

◦ Time-based Data … per cast trajectories … about 1,000 data points per cast Filter Temperature

Cast trough temp

Coolant temperature

Coolant flow

Dist Trough Metal level

Mold level (5)

Mold Controller(5)

Casting rate

How do we combine the data to model the impact of all of these data on some response???

PhaseTwo Analytics

Tim

e-based v

ari

able

s s

uch a

s:

Sta

tic v

ari

able

s s

uch a

s:

Page 7: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

An Ingot Casting Example

◦ Static Data (1/cast) Holding furnace temp

Coolant Additive %

Grain refiner set point

Start drop rate

Steady state drop rate

Ingot head ml set point

Metal treatment set point

Ambient dewpoint,

% humidity

Temperature

◦ Time-based Data (cast trajectories)

FilterTemperature

Cast temp

Coolant temperature

Coolant flow

Dist Trough Metal level

Mold level (5)

Mold Controller(5)

Integrating Static Data (1/batch) with Time-based batch trajectories

PhaseTwo Analytics

Filte

r te

mp

C

ast

Tem

p

Co

ola

nt te

mp

M

old

Level

1000 data points but not 1000 pieces of information!!!

Page 8: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

An Ingot Casting Example

◦ Static Data (1/drop)

Holding furnace temp X1

Coolant Additive % X2

Grain refiner set point X3

Start drop rate X4

Steady state drop rate X5

Ingot head ml set point X6

Metal treatment set point X7

Ambient dewpoint, X8

% humidity X9

Temperature X10

◦ Time-based trajectory features

Filter Temperature F1, F2, F3, F4

Cast temp F5, F6, F7

Coolant temperature F8, F9

Mold level F10, F11, F12, F13

Coolant flow etc

Dist Trough Metal level etc

Mold Controller(5) etc

Integrating Static Data (1/batch) with Time-based batch trajectories through feature extraction

MaxCASTTemp( L>4)

F6

MinCASTTemp( L>4)

F7

MaxCASTTemp( L<4)

F5

MaxFiltTemp( L<4)

F1

MaxFiltTemp( L>4

F2

Init_FiltTemp( L<4)

F3

MinFiltTemp( L>4)

F4

PhaseTwo Analytics

MaxCoolTemp

F9 MinCoolTemp

F8

Lvl Slope1

F12

Lvl Slope2

F11

Lvl Max

F10

Lvl Final

F13

Filte

r te

mp

C

ast

Tem

p

Co

ola

nt te

mp

M

old

Level

Model Y = F(X1, X2, …X10, F1, F2,…,Fi)

Page 9: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

“You Can Observe a Lot Just By Watching ” … Yogi Berra

Clean, merge, align, characterize your data … THEN EXPLORE IT!!!

Fit Y by X, time plots, Scatterplots, Graph builder, PCA

Case Study 3: Hidden Effects of Lubrication on Extrusion Process Friction

◦ Parts are extruded through a die and lubricant is sprayed on the part as it enters the die. Friction variation can occur and cause part surface damage & scrap. Lube components are measured but are difficult to control tightly. Lube effects are suspected qualitatively but never quantified

◦ A plant study was conducted. Key lube components were set at specific levels for periods of time with changes made from time period to time period.

◦ Data were “thrown over the fence” for analysis … ~9000 obs., ~30 predictors

◦ Research question #1: Do we see significant changes in mean measured die friction from time period to time period?

◦ Naïve answer (based on a Fit Y by X of the response versus test periods: Not much!

◦ Unusually high level of variability

But wait! There’s More! PhaseTwo Analytics

Tool Fri

cti

on

Page 10: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

Closer Inspection/Visualization of The Data Reveals More

Plot of the response vs Time

Magnification reveals a repeating systematic downward trend

Further visual exploration reveals the culprit is Die Life.

• Die Life is the primary driver of the response • The effect of Die Life was known(qualitatively) but not communicated to

me • It’s effect obscures the effects of studied variables • If partitioning were used to model Die Friction, the tree would primarily

split on die life and miss other important variables • Solution??? “Regress out” the known die life effect

• A smoothing spline was used to regress out the effect • CompareTest Periods and the effects of other studied variables using Residuals

Comparisons are now clearer Facilitated follow-on modeling PhaseTwo

Analytics

Die

Fri

cti

on

Die

Fri

cti

on

Die

Fri

cti

on

Die

Fri

cti

on

Die

Fri

cti

on R

esid

s

Page 11: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

Peril #3: Decision Trees Selecting Candidates that are Predictive But not Useful or Provide Incomplete Control Guidance

In manufacturing prediction is not good enough!

The end game in manufacturing is control

For control purposes, a predictor must be causative

Decision trees can select variables that are predictive but make no sense or are not useful to a process engineer from a control standpoint

Tree growth needs to be guided by process knowledge

JMP® permits guiding the partitioning process … enabling reality checks!!!

Decision trees may also miss important variables. They may be correlated with the top choice in a split, can share prediction/control capabilities but partitioning may ignore them

In non-manufacturing classification tree applications we’re not looking to control but to classify … even if the predictor isn’t “root cause” it may be good enough!

Common Issues: ◦ Nonsense correlations

◦ Non-actionable predictors

◦ Predictors that are actually responses … “After further review…”

◦ Important control variables are missed because of correlation between key predictors

PhaseTwo Analytics

Page 12: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

Storks vs population in Oldenburg, Germany 1930-36

So, would getting rid of storks be an effective mechanism for population control ???

The reality is that increased population causes the increase in the number of Storks

Does watching TV increase life expectancy? "Televisions, Physicians, and Life Expectancy" in the _Journal of Statistics Education_ (Rossman 1994)

Data from a number of countries were gathered from The World Almanac and Book of Facts 1993

So, to improve life expectancy do we ship boat loads of TVs to those countries with low life expectancy ???

Nonsense Correlations

PhaseTwo Analytics

Page 13: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

Sheet scrap was a >$1MM/yr problem

The problem was seasonal

Engineers weren’t sure if the cause was related to ingot casting, heat treatment or rolling

A swat team of metallurgists, process engineers and a data scientist was formed

Data were being recorded throughout the manufacturing flow path

Data collection, assembly, characterization/feature extraction was a HUGE task

Shortly after assembling & cleaning the data a quick partitioning exploration yielded the following:

Case Study 4: Sheet Scrap Integrating Process Knowledge With Partitioning

PhaseTwo Analytics

Page 14: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

Occurrence rate ~10%

Partitioning Step 0

Partitioning Step 1: Problem Solution Emerges!

Although the root cause emerged quickly and required verification some additional exploration of the data took place that provided further discovery insights

PhaseTwo Analytics

Page 15: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

Sheet Scrap (cont.)

Step 3 Over-ruled by the process engineers and MC1 is locked out Interesting … but can’t be implemented

Step 3.1 Tells us that the remainder of the problem would go away if dew point<63.3 BUT we can’t control dew point in a plant

Step 4 Also over-ruled by the process engineers and MC2 is locked out Step 4.1 Produced an interesting countermeasure

Step 5 Was a “what-if” scenario where we locked out the root cause to explore additional process countermeasures

Step 6 identified a temperature countermeasure that made total sense to our metallurgists!!!

PhaseTwo Analytics

Page 16: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

Sheet Scrap Case Study Summary Points

1. Data assembly was a PAINFUL task!!!

2. Once data assembly was completed … discovery was FAST !!!! • This problem was ongoing for more than 2 years!!!

• Once the data were assembled discovery occurred within minutes!!!

3. Discovery was validated through a designed experiment

4. JMP® permitted control of the partitioning steps with the ability to override split selection based on engineering judgment and 1st principles • Nonsense correlations can be easily dismissed using JMP®

• Splits that are not practically useful (but perhaps interesting nonetheless) can be noted and then locked out

5. Interaction between the Data Scientist and Engineer during the partitioning process is critical to getting to root cause understanding

PhaseTwo Analytics

Page 17: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

Case Study 5: Root Cause Discovery In Part Fabrication with Correlated Predictors

**Guiding The Discovery Process, Part II**

Key Learnings are:

◦ The Value of JMP® Visualization

◦ Must be attentive to correlated Predictors when partitioning

May be identified in “Candidate” list ….. pay attention!

Important (root cause) predictors can be missed

Partitioning can obscure their contribution

Only one may be selected as Predictive BUT More than One may actually be Needed for Control

◦ Engineering judgment/1st Principles Should guide Partitioning

◦ Complement Partitioning with Multivariate Methods

◦ Once the data are assembled, Discovery is FAST!!!

PhaseTwo Analytics

Page 18: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

Problem: Dimensional Tolerance In A Part Fabrication Process

Process entails production of a highly engineered part

Processing irregularities produced thick and off tolerance parts (6% scrap and 50% rework)

Process data were in multiple, disparate databases that required extraction, alignment, characterization and cleaning

Attempts over many months by engineering to identify root cause using standard “one variable at a time” statistical tools were unsuccessful and time consuming

68 different process parameters studied against the product data

PhaseTwo Analytics

Page 19: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

Initial Visual Exploration Critical part dimensions both with USL=3.0

PROBLEM LOCATION 6% Scrap 50% Rework

Different Location OK!

Problem improved over time but they didn’t understand why and were still incapable Furnace differences???

PhaseTwo Analytics

Page 20: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

Guided Partitioning

The 1st Splits Appear to be Promising BUT Engineering Questioned The Validity and Usefulness of Some Predictors selected

This was a correlated response … NOT a predictor

This was interesting but is not actionable

V_WDY and S_KW_P were locked for the rest of the analysis

PhaseTwo Analytics

Page 21: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

• Guided partitioning yielded actionable rules • The rules were easily implemented and the predicted results

were validated in production!!! • Partitioning required engineering guidance • Guidance provided a reality check on root cause • Yearly savings ~$300,000/yr. • Time to discovery … SEVERAL HOURS!!! • Return on investment VERY HIGH • Engineering spent months looking at data for a solution

Continue Partitioning With Guidance

PhaseTwo Analytics

Page 22: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

Correlated Predictor Considerations

Candidates That Are Close Choices Require Scrutiny !!! 1. Initial group of “Candidates” were very close choices based on Logworth/SS 2. Be aware that minor error (e.g. measurement) would affect their ranking 3. The group is closely ranked because they’re strongly correlated (in this case) 4. Partitioning only selects one of them although all are important !!! 5. Because of this, careful scrutiny is advised and candidate selection based on engineering judgment 6. Diagnostics from a PLS suggests they may be influential IN TANDEM 7. This should factor into partitioning rule implementation with engineering input

Are SW_8 and W_7 worthy of closer consideration? PhaseTwo Analytics

Page 23: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

A Case For Considering a Candidate Not Ranked 1st • W_8 and SW_8 were very close choices based on both Logworth and SS

• Can we be sure W_8 is more significant than SW_8? • Let’s examine what would have happened if W_8 were measured with a small amount of additional error

• Replace “W_8” with “W_8 err” where “W_8 err” is W_8 with .5% of error added • The Candidate list now reveals that SW_8 has moved into the 1 st ranking and W_8 is third

• Perhaps a case exists to give SW_8 consideration based on the sensitivity of the ranking to errors

PhaseTwo Analytics

Page 24: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

PLS Useful In Vetting Influence of Predictors

• Augment a partitioning study with PLS modeling to guard against missing important influences • For the Part Fab data set a Correlation Loading Plot can be used to identify influential predictors • Correlated predictors will tend to cluster • The distance from the center to a predictors normal projection onto the diagonal line measures its influence

• PLS Validated Partitioning Contributors But Augmented The List & Suggests the Combined Influence of W & SW variables

• Engineering agreed that the correlation between the Ws and SWs needed to be maintained in implementation

PhaseTwo Analytics

Page 25: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

There are Perils When Mining Manufacturing Data ◦ Data assembly

◦ Data with differing time scales… alignment & feature extraction

◦ Known effects, unknown to you & the value of visualization

◦ Prediction vs Control … getting to root cause & the need for engineering guidance (permitted by partitioning platform)

◦ Correlated predictor issues … important predictors can be missed

JMP® Tools & Platform Features help Mitigate the Perils

After data assembly discovery is fast It’s worth it!

PhaseTwo Analytics

Page 26: Bradley Novic, Ph.D. PhaseTwo Analytics, LLC

QUESTIONS

PhaseTwo Analytics