Representative subset selection and standardization techniques.A comparative study using NIR and a simulated fermentative process UV data

GABRIEL G. SIANO; HÉCTOR C. GOICOECHEA

CHEMOMETRICS AND INTELLIGENT LABORATORY SYSTEMS

Elsevier

Año: 2007 vol. 88 p. 204 - 212

0169-7439

Several algorithms related to the calibration transfer process were tested with the objective of comparing them in terms of its performance.Firstly, three algorithms for representative subset selection (Kennard–Stone, Leverage and OptiSim) were used and compared to select appropriatesamples to make the calibration transfer. Secondly, three algorithms of spectra standardization were tested using different values for inputparameters. The three methods tested were direct standardization (DS), piecewise direct standardization (PDS) and wavelet hybrid directstandardization (WHDS), the latter was implemented testing a large number of available filters.Two data sets were used: a) NIR data acquired from different NIR instruments, previously employed in other publications and available fromthe PLS Toolbox, and b) UV data obtained in our laboratory from samples of penicillin dissolved in a culture medium representative of anindustrial process of penicillin production. The spectra were recorded at different rates and from samples undergone to drastic modification on thepH, originating diverse data sets.Results show that OptiSim can be recommended on Kennard–Stone and Leverage for representative subset selection for data with similarstructure that the analyzed herein. On the other hand, the performance of WHDS was remarkably better than the other tested methods when NIRdata were analyzed. Finally, a strategy combining PDS optimization with WHDS results a good choice for more complex data. In addition, whenthe data become more complex, using wider windows improve the PDS performance.