introduction to instrument standardization and calibration transfer · calibration models for...

Eigenvector Research, Inc.

Introduction to InstrumentStandardization andCalibration Transfer

Barry M. Wise

Eigenvector Research, Inc.830 Wapato Lake Road

Manson, WA 98831

[email protected]

©Copyright 1996-2000Eigenvector Research, Inc.No part of this material may be photocopied or reproduced in any form without prior written consent from Eigenvector Research, Inc.


Motivation

Calibration models for quantitation or classification often take advantage of relatively small changes in spectraInstrument to instrument differences can be substantial, i.e. samples look differentInstruments may drift over time Renders models invalidInconvenient to recalibrate instruments or may want to utilize a historical database


Two Main Approaches

Find a transformation that maps the response of the field instrument onto the standard instrument

Direct and piece-wise direct standardization

Neural network and other variants

Process the data from both instruments in a way that makes the differences disappear

baselining and derivatizing

multiplicative scatter correction, FIR filtering

orthogonal signal correction

prediction augmented classical least squares

generalized least squares

explicit deresolution


Piece-wise DirectStandardization (PDS)

Develop models which use windows on field instrument to predict single channels on standard

800 900 1000 1100 1200 1300 1400 1500 16000

0.5

1

1.5

2

Wavelength

Abs

orba

nce

800 900 1000 1100 1200 1300 1400 1500 16000

0.5

1

1.5

2

Wavelength

Abs

orba

nce


Develop Transfer Matrix Fb

00F =b

Difference between instrumentsmodelled as:

S1 = S2Fb + 1bsT


Data Arrangement for PDS

Standard

Field

Window width


Data Arrangement for DoubleWindow PDS

Standard

Field

Window width

Window 1 = 5 Window 2 = 3

Second window can be spectrum full width =single model PDS


Direct Standardization

Similar to PDS except Fb matrix is full:

Fb = S2+S1

Many more parameters in DS compared to PDS


Variations on PDS

Single model PDSwiden second window in DWPDS until it is the width of the

entire spectrum

model is the same for each channel in master instrument

transfer function not a function of wavelength

Single model PDS with indexinclude the channel number as the parameter in the model

use non-linear model such as ANN

transfer function is a function of wavelength


Orthogonal SignalCorrection

OSC attempts to remove extraneous variation unrelated to the property of interest from the predictor variablesPrincipal components are calculated for the predictor variables then orthogonalized against the variable(s) to be predictedWeighting vectors are determined with PLS which reproduce the orthogonal directions on new dataTo use in standardization, apply to data measured on both instruments


Example From NIR, PseudoGasoline Mixtures

800 900 1000 1100 1200 1300 1400 1500 1600-0.5

0

0.5

1

1.5

2NIR Spectra

Wavelength (nm)

Abs

orba

nce

Instrument number 1 spectras shown in red

Instrument number 2 spectras shown in blue


Difference BetweenInstruments

800 900 1000 1100 1200 1300 1400 1500 1600-0.2

-0.15

-0.1

-0.05

0

0.05Difference between NIR Spectra from Instruments 1 and 2

Wavelength (nm)

Abs

orba

nce

Diff

eren

ce


Instrument 1 Calibration

0 5 10 15 20 25 30 35 40 45 500

5

10

15

20

25

30

35

40

45

50Actual versus Fit Concentrations Based on Instrument 1

Actual Analyte Concentration

Ana

lyte

Con

cent

ratio

n F

it to

Mod

el Each analyte shown as different color


After Standardization

800 900 1000 1100 1200 1300 1400 1500 1600-0.2

-0.15

-0.1

-0.05

0


Wavelength (nm)

Abs

orba

nce

Diff

eren

ce

Difference before correction shown in red

Difference after direct correction shown in green

Difference after piecewise correction shown in blue

Difference after OSC shown in magenta


Instrument 1 Calibration onUnstandardized Instrument 2

0 5 10 15 20 25 30 35 40 45 50-10

0

10

20

30

40

50Actual Concentrations vs. Predictions Based on Instrument 2


Pre

dict

ed A

naly

te C

once

ntra

tion

Each analyte shown as different color


Instrument 1 Calibration onStandardized Instrument 2

0 5 10 15 20 25 30 35 40 45 50-10

0

10

20

30

40

50Actual Concentrations vs. Predictions Based on Standardized Instrument 2


Pre

dict

ed A

naly

te C

once

ntra

tion

Each analyte shown as different coloro Piecewise direct standardized samples

* Direct standardized samples+ OSC standardized samples

Instrument 1 Fit Error 0.33 Unstandardized 7.49 Piecewise Standardized 0.70 DWPDS Standardized 0.77 Direct Standardized 2.50 OSC Standardized 1.45


Prediction AugmentedClassical Least Squares

If CLS is used for predictive model, new spectra can be added to prediction step to account for differences between instrumentAugmented spectra can include known new components or estimates of changes such as a baseline offset or mean differenceEigenvectors of difference matrices can also be included


CLS: Predictions on Instrument 2with Instrument 1 Spectra

0 10 20 30 40 50 600

10

20

30

40

50

60Calibration Using Instrument 1 Spectra

-20 -10 0 10 20 30 40 50 60 70 80-20

0

20

40

60

80Predictions Using Instrument 2 Spectra


Estimated Pure ComponentSpectra and Additional Factors

800 900 1000 1100 1200 1300 1400 1500 1600-0.005

0

0.005

0.01

0.015

0.02

0.025Estimated Pure Component Spectra

800 900 1000 1100 1200 1300 1400 1500 1600-0.2

-0.1

0

0.1

0.2

0.3

Mean Difference and First 3 Eigenvectors


PA-CLS Predictions

-10 0 10 20 30 40 50-10

0

10

20

30

40

50Pure Spectra Augmented with Mean Difference

-10 0 10 20 30 40 50-10

0

10

20

30

40

50Pure Spectra Augmented with Mean Difference and First Eigenvector

Fit Error 0.48 Uncorrected 18.05 PA-CLS w/mean 2.46 PA-CLS 1 EV 1.78 PA-CLS 3 EVs 1.15


NIR of Corn Samples

1000 1500 2000 2500-0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9NIR Spectra

Wavelength (nm)

Abs

orba

nce


Calibration

2 3 4 5 6 7 8 9 10 11 122

3

4

5

6

7

8

9

10

11

12Actual versus Fit Concentrations Based on Instrument 1


Ana

lyte

Con

cent

ratio

n F

it to

Mod

el Each analyte shown as different color

Analytes are oil, moisture, starch and protein


Difference Before and AfterStandardization

1000 1500 2000 2500-0.06

-0.04

-0.02

0

0.02

0.04

0.06

0.08


Wavelength (nm)

OSC

Direct

Before

PDS


Effect of OSC on Spectra

1000 1500 2000 25000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Wavelength (nm)

Abs

orba

nce

Spectra Before and After Orthogonal Signal Correction

Original Spectra in BlueOSC Transformed Spectra in Red


Results of CornStandardization

2 3 4 5 6 7 8 9 10 11 122

3

4

5

6

7

8

9

10

11

12Actual Concentrations vs. Predictions Based on Standardized Instrument 2


Each analyte shown as different color

o Piecewise direct standardized samples

* Direct standardized samples

+ OSC standardized samples

Fit Error = 0.10 Unstandardized = 1.32 PDS = 0.33 DWPDS = 0.33 Direct = 0.45 OSC = 0.20


SummaryMethod Transforms? Standards Parameters Uses Y CommentsDirect Yes Real Lots No Many samplesPDS Yes Anything Few No Few samplesNN-PDS Yes Anything Moderate No Non-linearDerivative No None None No EasyMSC Yes Real, Few Few No EasyOSC No Real Few Yes Requires YPA-CLS No Anything? Few No InterpretableGLS No Real Moderate No NewDeresolution No None Few No FTIR


Conclusions

PDS still the method to “shoot for”DS more sensitive to number of transfer samplesOSC produces especially good results in some data, also useful as a preprocessing techniqueFIR not adequate in situations we’ve seen

PLS_Toolbox 2.0for use with MATLAB

PLS_Toolbox 2.0for use with MATLAB™

Barry M. WiseNeal B. Gallagher

• Version 2.0 for MATLAB 5 now available

• Wide selection of multivariate analysis tools

• Used in our Chemometrics Short Courses

Continuum ParameterNumber of LVs

PR

ESS

introduction to instrument standardization and calibration transfer · calibration models for...

Documents