MCCTTool

Introduction

No two instruments are identical. Data generated from different instruments can rarely be used with the same model. It may be prohibitively expensive or otherwise impossible to run an entire complement of calibration samples on new instrument. A calibration transfer step can be developed to transform "slave" instrument data to look like it came from a "master" instrument.

Model-centric Calibration Transfer (MCCT) refers to the situation where a model has been developed on a "master" instrument and is the basis of creating a "slave" model that includes a calibration transfer step in the preprocessing structure. It is assumed there data sets used for calibration transfer for each instruments are matched samples.

How data is preprocessed can affect the standardization. The location of calibration transfer step within the preprocessing can have a large impact on how well the slave model will perform. The MCCT Tool allows the testing of different combinations of calibrations transfer methods and parameters, and where the calibration transfer step is inserted into the preprocessing.

MCCT Tool Interface

Getting Started

The goal of the tool is to create and test multiple calibration transfer scenarios for a given "master" model. Each calibration transfer scenario consists of:

• A calibration transfer model (see caltransfer) as a preprocessing step.
• Insert location in the preprocessing steps of the master model.

A given calibration method (DS, PDS, PWDS, SST) may have one or more parameters to "tune" for a given instrument. For example, window size in the case of PDS. The ultimate goal is to produce a new "slave" model based on the "master" model with a calibration transfer preprocessing step inserted to account for instrument variation.

Adding Data or Model

To add data or a model to the MCCT Tool, drag and drop it from the Workspace Browser, Matlab Workspace, or from a file. The data will automatically be loaded into the tool. The File menu can also be used to load items.

Calibration Transfer Methods

Select the methods to try and the combinations of parameters to use. It may be useful to choose large "step" sizes when surveying over a large range of values. PDS and DWPDS step sizes are odd.

Preprocessing Insert

Choose what positions the calibrations transfer step will be inserted at. More than one value can be used. A value of 0 indicates the transfer step will occur before all preprocessing. For example, if a model has 3 preprocessing steps ('Derivative','GLS Weighting','Mean Center') then:

• 0 - Before all preprocessing.
• 1 - After derivative.
• 2 - After GLSW.

Note: By default, the calibration transfer models automatically created will include the preprocessing steps prior to the insert point. With the example above and the insert location at 1, the slave model will have the following preprocessing steps:

• Calibration Transfer model (derivative preprocessing embedded)
• GLSW
• Mean Center

At the insert point, slave data is processed to "look like" it was measured on the master instrument. Preprocessing steps in the master model prior to the insert point are copied by default into the calibration transfer model object. Customization is possible by editing in the calibration transfer GUI (see caltransfergui). Preprocessing steps in the slave model after the calibration transfer model are copied and not changeable.

Interpreting Results

Use the calibration and validation difference ratios for a general guide to performance on the x-block. Smaller ratios indicate that the slave data - post application of the calibration transfer model and all preprocessing - is similar to the master data. <need equation here?>

For regression models (that use a y-block) various RMSE values are calculated. Please note that at this time only regression models with univariate Y are supported. The abbreviations for RMSE comparisons are:

• CalM - Master Model Prediction of Master Calibration Data
• CalS - Slave Model Prediction of Slave Calibration Data
• CalY - Calibration Y Data

• ValM - Master Model Prediction of Master Validation Data
• ValS - Slave Model Prediction of Slave Validation Data
• ValY - Validation Y Data

For single block methods (PCA) the RMSE(CalM,CalS) and RMSE(ValM,CalS) column become a difference summary based on the following calculations:

T_m = scores from master T_s = scores from slave

m = number of samples

k = number of principal components

RMSE = ${\sqrt {\frac {\sum \sum ((T_{m}-T_{s}).^{\wedge }2)}{m\,*\,k}}}$

NOTE: Clicking a column then right-clicking the header will display a menu where sorting is available.