283153_en-US

liliconcom

Building Predictive Models for Spectral Data

 

How it's Done:

 

1. Examine data using Graph Builder to get idea of what different spectra look like

2. Use Multivariate analysis to examine all wave lengths and resulting Correlation Coefficients to confirm multicolinearity

3. Use Model-Driven Multivariate Control Charts to examine all wave lengths variables and drill into runs that are out of control
   
   1. ID how many PC you need to build model
4. Take spectra of the desired samples.  No need for output (Y) information at this point.
5. Identify most prominent dimensions in the spectra by Row with Functional Principal Components, f(PC’s), from Functional Data Explorer – not for use in the traditional sense
   1. Use Functional Principal Component Profiler to get an idea how your spectra are changing as variable of interest changes
   2. Save Functional Principal Component scores to data table to use in experimental design
6. Create an experimental design using the f(Principal Components) as factors (covariates)
7. Run the experiment – gather the output of interest
8. Model the results via PLS, and/or Generalized Regression (or other methods able to handle correlated factors)
9. Determine the overall optimum solution
10. Use this sustainable model to determine the “outcome” for all future sample*  the model will hold true for samples analyzed using the same calibrated instrument) 
11. 11. Use Score plot to examine Categorial Data

Summary:

• Compress the available information regarding spectral wavelengths or mass of different options via functional principal components
• Use covariate DOE to select the “corners of the box” for testing representative sample  spectra based on Design of Experiments
• Model the data via PLS – Generalized Regression is an excellent alternative option or you  may need to use more sophisticated techniques including Neural Nets
• Find the overall optimum solution
• As new samples are tested the spectral data can be input into the data table to determine level of active or desired component.
  • Highly efficient experimentation
  • Sustainable empirical model based on spectral data/wavelengths