Sei sulla pagina 1di 36

4.

0 Modeling and Analysis of


Historical Batch Data

ProSensus Advanced Batch Analysis Course


(c) 2004-2012, ProSensus, Inc.

Purpose of Batch Data Analysis


• Industry periodically has problems with batch processes
– Periods of poor quality or poor operation
– Often at a loss for why these have occurred
– Fundamental understanding may not clearly explain these

• Industry has a lot of historical data on the batch process

• Multivariate analysis of these data can provide a lot of


information
– Where is the major variation coming from?
• Raw materials, initial conditions, transient batch operation?
– Which variables or combination of conditions are related to the
problem?
– Suggestions as to how to fix the problems.
(c) 2004-2012, ProSensus, Inc.

1
Multivariate Approaches to Modeling Batch Data

• PCA approaches all involve an unfolding of the 3-way array


of trajectory data (X) in some way.
– Several alternative ways of slicing and unfolding the 3-way
array
• (Nomikos & MacGregor laid out all these alternatives in early 90’s)
– Each way looks at a different source of variation
• But only 2 unfoldings were shown to consider reasonable sources of
variation
– Batch-wise Unfolding (BWU) – examines variation among batches
– Observation-wise Unfolding (OWU) – examines variation between
observations at different times
– So called 3-Way methods (PARAFAC and Tucker models) have
been suggested, but make assumptions that are not generally
realistic for batch data.
• Unfolded PCA/PLS approaches are also 3-way methods but with fewer
assumptions
(c) 2004-2012, ProSensus, Inc.

Observation-Wise Unfolding (OWU)


Variables/Tag
s

• Observation-Wise
Batch time
Unfolding (OWU) observations Batch
1
– method by which the data for
each batch is stacked on top
of one another
Batch
– Manner in which ProMV and 2

Simca_P read in batch data


Time, T

Variables, K

N
Batch
Batches Trajectory
Batch N
Data

(c) 2004-2012, ProSensus, Inc.

2
Batch-Wise Unfolding (BWU)
• Batch-Wise Unfolding (BWU)
– The method by which batch observations are extracted
horizontally in a time-wise fashion
– Each batch becomes a single row of data in the model

Time, T

Variables, K
N Batch
Batches Variable 1 Variable 2 Variable 3 Variable 4 Variable K
Trajectory
Data

Variables x Time, TxK


N
Batches Unfolded trajectory data

(c) 2004-2012, ProSensus, Inc.

Kernel algorithms for PCA & PLS with unfolded data

OWU – Long, thin matrices (K much smaller than N)

K K
XTX
K

(c) 2004-2012, ProSensus, Inc.

3
Kernel algorithms for PCA & PLS with unfolded data

BWU - short, fat matrices (N much smaller than K)

K M N
X Y XXT
N N
N
YYT
N

N
XXTYYT
N

(c) 2004-2012, ProSensus, Inc.

Adv: Kernel algorithms for PCA and PLS


For large asymmetrical matrices

• Fast kernel algorithms work with the smaller kernel matrices:


XTX (N >> K) OWU (PCA analysis)
XX T YY T T
XX YY T (N << K) BWU (PCA/PLS analysis)

• Compute eigenvalues and eigenvectors of these small kernel


matrices

References: - Lindgren, Geladi, Wold, “The kernel algorithm for PLS”, Journal of Chemometrics, 1993.
- Dayal, MacGregor, “Recursive exponentially weighted PLS and its applications to adaptive
control and prediction” , Journal of Chemometrics, 1997.

(c) 2004-2012, ProSensus, Inc.

4
Adv: Exponentially weighted recursive PLS
Based on the kernel algorithms
• New data often coming available periodically.
• Want to update the PLS model by simply correcting it for
new data and not rebuilding the entire model from scratch
each time (e.g. soft sensors).
• May also want to adapt model by exponentially discounting
past data.
• Can update the Kernels directly:
– Have kernels at time t-1 and new vectors of data become
available at t: xt, yt
– (XTX)t = λt (XTX)t-1 + xtTxt
– (XTY)t = λt (XTY)t-1 + xtTyt
• Update Kernels this way and then use fast Kernel algorithms
to update model
• λt is exponential forgetting factor (0 < λt ≤ 1.0)
(c) 2004-2012, ProSensus, Inc.

Approaches to Modeling & Analysis of Batch Data

• Will present 3 approaches in detail and will have workshops


on each
– Landmark approach to extracting trajectory information
• Simplest approach (should be considered first)
• Not always easy to apply
– Batch-Wise Unfolding (BWU) approach
– Observation-Wise Unfolding (OWU) approach

(c) 2004-2012, ProSensus, Inc.

5
The Problem

• Extract a few measures that characterize the important


differences among the batches
– Use these to summarize and compare batches
– Use these to monitor future batches
– Use these to predict the final product quality
– Use these to control the batches

• Can get these measures from:


1. Manual selection of key trajectory features (landmarks)
2. PCA or PLS latent variables

(c) 2004-2012, ProSensus, Inc.

A. Landmark Approach

• Illustrate with FMC batch reactor/drier data:


– Process variable trajectories for 10 variables (X)
– Initial chemistry analysis and charges (Z)
– Final quality data on 12 variables (Y)

(c) 2004-2012, ProSensus, Inc.

6
Extraction of Landmarks from batch trajectories
Extract landmark features from the trajectories that summarize the major sources of
variation from batch to batch

Temp 2

Level 1

Temp 1 Temp Slope

∆ Time 4

∆ Time 1 ∆ Time 2 ∆ Time 3

Note that timing differences are also important features.


(c) 2004-2012, ProSensus, Inc.

Batch trajectories

• For each batch


– Collect key trajectory features (landmarks) in X
matrix
– Collect initial conditions in Z matrix
– Collect product quality data in Y matrix

• Perform PCA or PLS in usual way

batches Z X Y

(c) 2004-2012, ProSensus, Inc.

7
Score Plot for FMC Landmarks

Separation of good and bad (red) batches in score space

(c) 2004-2012, ProSensus, Inc.

Loading Plot

(c) 2004-2012, ProSensus, Inc.

8
VIP Plot – ordered importance to model

(c) 2004-2012, ProSensus, Inc.

Workshop #1: Landmark Analysis of FMC data


1. Introduction to ProMV Software
a) Only the traditional non-batch features of the software
2. PLS on initial conditions (Z), extracted landmark features
(X), and final quality (Y)
a) PCA’s on Y, Z, and X only to see the multivariate separation of the batches
into good and bad based on the information contained in each matrix
b) PLS on Z, X, Y data
c) Examine R2 and Q2 summaries
d) Examine SPE plots for the overall data and for X and Z blocks separately
and T2 plot for over model
e) Examine the BIP for relative importance of each block in each dimension
f) Examine score and loading plots for each matrix
g) Examine VIP and coefficients plots for some of the better predicted y’s
h) plot observed vs predicted for y’s
i) Examine contribution plots for differences between batch clusters

(c) 2004-2012, ProSensus, Inc.

9
B. PCA or PLS to extract features of batch trajectories

Often the measured variable trajectories are continuously


varying
– Hard to select a small number of features to adequately
characterize them
– Eg. DuPont batch trajectories
for one batch (good landmark
features not obvious)

• Use PCA or PLS to extract a minimal set of features - latent


variables
(c) 2004-2012, ProSensus, Inc.

Batch-Wise Unfolding of the data array


– The batch trajectory observations are extracted horizontally in
a time-wise fashion
– Append both Quality (Y) and Initial condition data (Z)
– Each batch becomes a single row of data in the model
– PLS models built on these unfolded data matrices

Time, T

Variables, K
N Batch
Batches Variable 1 Variable 2 Variable 3 Variable 4 Variable K
Trajectory
Data

Variables x Time, TxK


N
Z Batches Unfolded trajectory data Y

(c) 2004-2012, ProSensus, Inc.

10
Mean Centering and Scaling

• Mean center by subtraction of means of each column


– For Z and Y this is the standard way of centering data
– For unfolded X matrix this removes the average trajectories
from each variable (a major source of non-linearity)
– Resulting PLS model relates the deviations about the means in
Y to the deviations about the means in Z and the deviations of
the trajectories (X) from the average trajectories.
• Scaling is usually to unit variance
– But alternative scaling can be used.

(c) 2004-2012, ProSensus, Inc.

Mean centering of the trajectories


It is the variation of different batches about the mean trajectories
that one is generally interested in.

(c) 2004-2012, ProSensus, Inc.

11
PLS for batch-wise unfolded data

• Models the time varying covariance structure among all the


process variables over the entire time history of the batch
– Different w weights on each variable for every time period
– Hence it provides an effective nonlinear model of the batch process
• Every batch summarized by a few LV scores (t1, t2, t3)
• Relates the IC’s (Z) and time varying trajectory information
(X) to the final product quality (Y)
(c) 2004-2012, ProSensus, Inc.

PCA Analysis: DuPont Autoclave trajectory data (X)


10 trajectories for one batch
Time
One batch

Batches

Variables

(c) 2004-2012, ProSensus, Inc.

12
Industrial example: PCA of DuPont historical data

Batches from a nylon polymerization process


(c) 2004-2012, ProSensus, Inc.

Score plot reveals differences among batches

(c) 2004-2012, ProSensus, Inc.

13
Loading Plot - p1
-Time varying loadings for all variables (grouped by variable)
- Reveal the relative importance of variables and their negative or
positive impact as a function of time.

(c) 2004-2012, ProSensus, Inc.

SPE Plot

Why are batches 50-55 not showing up as outliers?

(c) 2004-2012, ProSensus, Inc.

14
Summary of analysis of nylon autoclave data

• 3 major groups of bad batches observed

• Contribution plots showed possible reasons for the different


groups of bad batches

• These were fixed through mechanical and operating policy


changes

• Improvements implemented world-wide

• New data then collected and on-line monitoring scheme has


been in place for many years

(c) 2004-2012, ProSensus, Inc.

Workshop #2a: PCA on DuPont Nylon Autoclave (X)

PCA on DuPont nylon data (X)


a) See how to use software to read in batch trajectory data
b) PCA on BWU trajectory data. Look at:
1. Plots of the trajectory data: single variable for several batches, and all
variables for one batch
2. Model Performance
3. Score plot to see clustering of batches
4. Loadings
5. Contribution plots for outlying cluster(s) – interpret difference among
clusters

(c) 2004-2012, ProSensus, Inc.

15
PLS Analysis of FMC Reactor/drier data

Multi-block Structure

Z1 Some Landmark Features

Z2 Initial Conditions
– Initial chemical analysis
– Weight of wet cake

X Process Trajectories
– 10 trajectories
– Batch duration

Y Final Quality
– 11 Quality variables from QC lab
– Solvent concentration
(c) 2004-2012, ProSensus, Inc.

Recommended steps to follow in general

1. Plot trajectory data


2. PCA Analysis of final properties (Y) and classification of y’s
from a multivariate perspective
3. PCA of raw material and initial condition data (Z)
4. Batch trajectory synchronization (if batches are all of
different durations) Will treat this later. For now we will
consider the data with aligned trajectories
5. Model building
6. Interrogate the model via score, loading and contribution
plots to understand process variability

(c) 2004-2012, ProSensus, Inc.

16
Alignment of batch trajectories

• Batches can often be of different durations


Examples:
– Length of certain phases determined by an operator or by
attainment of a specified value of a measured variable
– Cooling capacity differs between summer and winter – Batches
with an exothermic reaction can be run faster in winter

• If batches have very different durations (overall or in certain


stages), then need to align the trajectories prior to doing
multi-way batch analysis.

• Knowing the time differences among batches for each stage


is also important information to retain.

(c) 2004-2012, ProSensus, Inc.

Approaches to alignment
1. Linear interpolation of the time scale.
• Choose an average final time – resample the trajectories at a fixed
number of intervals for all batches
• Retain the varying batch duration information by including it in Z
• Often poor since most batches proceed faster or slower only in
certain phases.
• Could also fix time of each phase and resample to give fixed
number of points.
2. Use of an indicator/maturity variable (Nomikos & MacGregor)
• Indicator variable – e.g. conversion – resample at equal conversion
intervals
• Indicator variable must be monotonic with time in any phase (can
use different indicator variables in each phase)
• Other examples – cumulative amount of a key reactant fed to
batch, or cumulative heat release
• Compute a new time usage trajectory to retain the time variations
3. Dynamic time warping (Kassidas, MacGregor, Taylor)
• Speech recognition literature
(c) 2004-2012, ProSensus, Inc.

17
Trajectory alignment

(c) 2004-2012, ProSensus, Inc.

Trajectory Variability over all Batches


Collector Level Diff. Press Press Power Speed

Jtemp SP Jtemp PV Dtemp SP Dtemp PV


Torque

Garcia, Kourti, MacGregor (2003) Ind. Eng. Chem. Res. 42, 3592-3601.
(c) 2004-2012, ProSensus, Inc.

18
Trajectory Alignment
• Indicator Variable Collector Level
Approach:
– Re-sample the batch
samples against a new
variable different from time
(e.g. conversion, reactant Dryer Temperature
feed). Monotonically
increasing/decreasing

• In this case there is no


single variable with such Agitator Speed
characteristics

37
(c) 2004-2012, ProSensus, Inc.

Choice of Indicator Variables

Peak Temp Different Levels


100% per Batch
0%

Final Level
100%
Collector Tank Level
Different Peak
Temp per Batch
0% 100%

Initial Temp

0%
STAGE 1 STAGE 2 STAGE 3 Dryer Temperature
Different Initial
Temp per Batch

Possible Indicator Variables: Stage 1: Collector tank level


Stage 2: Drier Temp
Stage 3: time
(c) 2004-2012, ProSensus, Inc.

19
Trajectory Alignment
Collector Level Collector Level

Dryer Temperature Before Alignment


Dryer Temperature

After Alignment
Agitator Speed Agitator Speed

(c) 2004-2012, ProSensus, Inc.

Preservation of timing information – Calculate a time


usage variable & use as new trajectory variable

Before Resample time in the same


way as the other
variables

After

(c) 2004-2012, ProSensus, Inc.

20
Alternative preservation of timing information
– Record phase times
Put this timing variable information in the Z matrix
- Used in the Landmark approach

Temp 2

Level 1

Temp 1 Temp Slope

∆ Time 4

∆ Time 1 ∆ Time 2 ∆ Time 3

(c) 2004-2012, ProSensus, Inc.

PLS Modeling and Analysis of FMC data via BWU

• Consider PLS modeling of the FMC historical data


– Z1, Z2 data
– aligned trajectory data (X)
– Final quality data (Y)
• Analysis of the results
• Improvement of the process

(c) 2004-2012, ProSensus, Inc.

21
The full dataset

• Initial conditions (Z2)


– Initial chemical analysis
– Weight of wet cake
• Warping information and some landmarks (Z1)
– Time1, Time2, Time3, Time 4, Temp1, Temp2, TempSlope,
Level1
• Process (X)
– 10 Trajectories (9 Process + Time Usage)
• Final (Y)
– Chemical Analysis
– Solvent Concentration
– On-Spec/Off-Spec/Hi-solvent classification

• Approximately 250,000 data points


(c) 2004-2012, ProSensus, Inc.

Analysis step
Trouble-shooting the industrial batch reactor/dryer process

71 batches
(c) 2004-2012, ProSensus, Inc.

22
Batch-Wise Unfolding of the data arrays
– The batch trajectory observations are extracted horizontally in a time-
wise fashion
– Append both Quality (Y) and Initial condition data (Z)
– Each batch becomes a single row of data in the model
– PLS models built on these unfolded data matrices
– Scores for each batch – score plot shows differences among batches

Time, T

Variables, K
N Batch
Batches Variable 1 Variable 2 Variable 3 Variable 4 Variable K
Trajectory
Data

Variables x Time, TxK


N
Z Batches Unfolded trajectory data Y
X

(c) 2004-2012, ProSensus, Inc.

Block Importance to the projection - BIP

BIP is same as VIP except it treats each block as a unit

(c) 2004-2012, ProSensus, Inc.

23
Analysis of combined Z1, Z2 blocks
Score plot Analysis reveals obvious clustering
among on- and off-spec batches.

The model also allows an understanding


t2 of why off-spec batches occur.

The critical to quality variables were


process related, not chemistry related.

t1
VIP’s for Z model

(c) 2004-2012, ProSensus, Inc.

Analysis of X trajectories
• Clustering of good/bad batches
•Good batches have high t1

•Loadings w1* show that for high t1:


a. Low tank level and used time for
entire batch
b. High Press, Jtemp, Dtemp in first
half only

Score plot for X

Loading vector w*1


for X model

(c) 2004-2012, ProSensus, Inc.

24
Score plot for Z1,Z2 Score plot for X

VIP’s for Z model Loading vector w*1 for X model

(c) 2004-2012, ProSensus, Inc.

Good Batches

• High quality product is associated with


– A high temperature slope in stage 2 and low values of the
length of stage 2 (Time 2) and low high-speed agitation time
(Time 4).
Fast evaporation stage 2 is desirable.

– High quality product is also associated with low solvent level in


the collector tank at the end of stage 1 and a low initial charge
of wet cake.
Smaller charge of wet cake is desirable.

50
(c) 2004-2012, ProSensus, Inc.

25
Summary of trouble-shooting results

• Chemical composition variations of feed NOT a problem


• Operation strategies used ARE the problem
• High quality associated with:
1. Smaller initial wet cake charge (low CTLevel through batch)
• Shows up in both Z and X block analysis
2. Stage one and two: short duration and a high level of heating
 High trajectories for Press, Drier Temp, Jacket Temp during these stages
3. Short time duration for batch

(c) 2004-2012, ProSensus, Inc.

Investigate Batch Differences via Contribution Plots

• Next slides just illustrate the interrogation of differences


among individual batches or batch clusters using
contribution plots
• You will do these in the workshop

(c) 2004-2012, ProSensus, Inc.

26
Inconsistent Batches

Cluster of 4 good batches falling outside region of other good batches.


Why? Or are they really poor batches that were just misclassified by univariate
analysis of y’s?

53
(c) 2004-2012, ProSensus, Inc.

Contribution Plots

54
(c) 2004-2012, ProSensus, Inc.

27
Contribution Plots

• Cooling Time, and high speed heating in stage 2 appear “strong” enough
to compensate for other adverse processing conditions?

55
(c) 2004-2012, ProSensus, Inc.

Workshop #3: PLS Analysis of BWU - FMC data

1. Alignment of FMC trajectories


2. PLS Analysis of aligned FMC data
PLS on Z1, Z2, aligned X and Y
i. examine separation of good and bad batches in Z and X score spaces
ii. Examine BIP plots to see which blocks contribute to each dimension
iii. Examine loading plots, contribution plots in both Z and X score spaces
iv. Examine VIP and coefficient plots relating y’s to Z variables and to X
trajectories
v. Interpret reasons for bad batches

(c) 2004-2012, ProSensus, Inc.

28
Observation-Wise Unfolding (OWU) Analysis

Unfold so that each row contains one • Mean center and scale columns
observation at each time for one batch to unit variance (e.g. Umetrics)
• Each latent variable t = Xw
provides a linear combination of
the original variables over all
batch times and all batches that
explain the greatest variance of
the raw time histories.
• Scores (t) are local – value at
every observation during the
batch. One score value per
time point instead of one
score value per batch.

(c) 2004-2012, ProSensus, Inc.

Problems with OWU for batch data


• Models the variation about the grand mean of each variable
– not around the average trajectory (ie. It models the within-batch
variation.
– Model explains mainly the average trajectories (not difference among
batches)
– Better to remove the average trajectories first if use this approach

• Ignores the time varying behavior of batches


– Can overcome this somewhat by breaking the batches up into many
phases and then modeling each phase by steady state models

• Assumes the covariance structure among x’s is constant over


the duration of each phase.
– i.e. gives a steady state model that fits all observations throughout
the duration of any phase. Again can overcome by building OWU
models for many phases.
(c) 2004-2012, ProSensus, Inc.

29
Comparison of variance explained BWU vs OWU

After 8 components with


OWU still haven’t reached
the variance explained by
BWU with 0 components

(c) 2004-2012, ProSensus, Inc.

Score plots: BWU vs. OWU (Nylon data)


BWU- Only one set of scores per OWU- New scores at every time point.
batch, that summarize differences Can’t see clear differences between
among batches. Can see clearly that batches or any outliers because OWU
batches 50-55 are different from the models variation within batches.
rest (Observations are colored by batch
number.)

(c) 2004-2012, ProSensus, Inc.

30
BWU vs OWU Loading Plots (Nylon data)
BWU- shows the model loadings OWU- the model is built only on
for each time and each variable the variables. One loading per
in the batch. This plot reveals the variable. Does not provide any
major difference among the information on behaviour during
batches due to their variable- a batch.
time histories.

(c) 2004-2012, ProSensus, Inc.

Recommendations
Batch-wise unfolding (BWU) recommended
– Necessary for analysis of differences among batches
• BWU looks at between batch variance, OWU looks at within batch variance (example – score
plots on FMC batches)
– Necessary to incorporate batches into PLS models with initial Z and
final Y values
– Necessary for control and prediction of product quality
– Preferable for on-line monitoring
• Use same model as for analysis, prediction and control
• Don’t need multiple models for phases
• Focuses on differences in performance of current batch from common cause variation (training
batches) which is what we want
• But OWU will still detect many gross deviations
Observation-wise (OWU) unfolding
– Mainly useful as a preprocessing step to reduce #of variables when
have multivariate sensors (Eg. NIR), etc.
• Then use scores from OWU on sensor data as variables in BWU
– More on this after the lab
(c) 2004-2012, ProSensus, Inc.

31
Umetrics – Simca_P software
• Uses both OWU and BWU
• Always leads user into OWU first
– As a result most users try to analyze differences among historical
batches at this level
– That is not good - OWU does not look at between batch difference
• In accompanying Comprehensive Chemometrics chapter on batches
Svante Wold also states that this is not good.
– On-line monitoring done on these scores (cf. Hahn’s approach)
• Then takes LV’s from OWU and refolds them batch-wise.
– First aligns them by linear time warping
– Then have aligned BWU scores from first OWU level as new variables
– Allows for PLS but variable loadings, contributions, etc are in terms of
these OWU level scores and not original variables – interpretation?
– Umetrics has introduced a way to put original variables into this BWU
matrix – We recommend that one should do this as first step in any
batch data analysis (ie. bypass OWU level).
(c) 2004-2012, ProSensus, Inc.

Observation-Wise Unfolding (OWU) in ProMV


Variables/Tags
• Data are read into ProMV (and Simca_P) with
batches stacked on top of one another
Batch time
• For OWU analysis in ProMV just treat data as observations Batch
normal (non-batch data) 1

• PCA scores (T) on this matrix summarize SS


information at each observation
Batch
2 T


Time, T

Variables, K

N
Batch
Batches Trajectory
Batch N
Data

(c) 2004-2012, ProSensus, Inc.

32
Workshop #2b: Comparison of PCA on BWU and
OWU Nylon data
• The purpose of this exercise is to look at the two methods for unfolding
batch data to determine how data can be interpreted using each
method:
• BWU – Batch wise unfolding
• OWU –Observation wise unfolding

• Open two instances of ProSensus Multivariate and import Nylon


autoclave data from Nylon_Unfolding.xls. In one import the BWU
sheet, in the other the OWU sheet.
A) What do the score plots of both unfolding methods show?
B) What do the loading plots of each unfolding methods show?
C) Which batches are outliers? Why is batch 49 different from the rest?

(c) 2004-2012, ProSensus, Inc.

Extracting information from on-line NIR signatures

There are approximately 31,000 data points per batch run in this
example.
(c) 2004-2012, ProSensus, Inc.

33
OWU makes sense to summarize Spectral variations
during batch
• PCA on OWU spectral data gives principal components that
apply at every instance of time over all batches
– Instantaneous (steady state model)
– This is what one wants for spectral data – no matter how the process
has reached this point, the spectra reflect the current composition of
the contents of the batch at that instant
– The new latent variables (t1, t2, …) reflect the number of detectable
constituents and their current concentrations
– These latent variables (t1, t2, …) just provide a small number of new
variables (eg. 3 or 4) that summarize the variations of the 1200 or so
spectral absorbances of the raw spectra.
– These principal components can then be used as other process
variables in a BWU model that captures the compositional history of
the batch over the entire batch duration and the effect on final Y

(c) 2004-2012, ProSensus, Inc.

Combining NIR with other process data


3 LV’s from PCA

Product temp
Pressure

290 columns: 4752cm-1 to 9211cm-1


etc

PCA

Summarizes ~99% of
Each row is a spectrum, taken the original spectra with
3 scores.
during the batch process (OWU) Each run reduces
down to 318 data points
(c) 2004-2012, ProSensus, Inc. (~100-fold reduction).

34
Time behavior of scores summarize spectral
information during batch
Progress of Spectral scores for a good batch

(c) 2004-2012, ProSensus, Inc.

Time behavior of scores summarize spectral


information during batch
Progress of Spectral scores for a bad batch

(c) 2004-2012, ProSensus, Inc.

35
Interpretation of latent variables from OWU

• The LV’s (t1, t2, … tA) are simply A new variables that explain the most
variance in the original J raw trajectory variables.

• Since most batches vary widely over their duration, one often needs
almost as many new LV’s as one has original measured variables.
– If use OWU it would be better to subtract off the average trajectories first –
will get many fewer LV’s since no need to model the average trajectories

• If want to use the OWU approach for monitoring, it is simpler and often
just as good to use monitoring charts on a few of the key process
variables (i.e. univariate approach)
– (Gerald Hahn, Principal GE statistician in late 80’s presented this approach)
– OWU approach essentially is Hahn’s approach, but it uses PCA to select some
combinations of the variables instead of key original process variables

(c) 2004-2012, ProSensus, Inc.

Contact details

Phone: (905) 304-9433

Email: info@prosensus.ca

Web site: http://www.prosensus.ca

(c) 2004-2012, ProSensus, Inc.

36

Potrebbero piacerti anche