Spectral data analysis approaches for improved provenance classification
In the last 10 years various chemometric methods have been developed and used for the analysis of spectra generated by Laser Induced Breakdown Spectroscopy (LIBS). One of the more successful and proven methods is Partial Least Squares Discriminant Analysis (PLS-DA). Recently PLS-DA was utilized for purposes of provenance of spent brass cartridges and achieved correct classification at around 93% with a false alarm rate of around 5%. The LIBS spectra from the cartridge samples are rich in emission lines from numerous mostly metallic elements comprising the brass and the cited results were based on the analysis of the full broadband high resolution spectra. It was observed that some of the lines were clearly saturated in all spectra, while others were sometimes saturated due to pulse-to-pulse variation. The pulse-to-pulse variation was also evident in the intensity variations of the spectra within cartridges and between cartridges. In order to improve on the accuracy of the classification we have developed some preprocessing strategies including the removal of spectral wavelength ranges susceptible to saturation and normalization techniques to diminish the effects of intensity variations in the spectra. The results indicate incremental improvements when applying additional preprocessing steps to the limit of 100% True Positives and 0% False Positives when utilizing selected wavelengths that are normalized and averaged.
Sorauf, Kellen J.; Bauer, Amy J.R.; Miziolek, Andrzej W.; and De Lucia, Frank C., "Spectral data analysis approaches for improved provenance classification" (2015). Regis University Faculty Publications. 577.