Maximum likelihood fusion for detection of lED precursors using laser-induced breakdown spectroscopy

"Improvised Explosive Devices (IEDs) pose an increasing threat to the safety of soldiers and civilians in embattled areas. The use of compounds such as RDX complicates the detection of IEDs, as these compounds are often hard to identify in the presence of contaminants such as alcohol, oils and dust. Spectroscopic techniques such as Laser-Induced Breakdown Spectroscopy (LIBS) and Raman have shown promise for the detection of explosive compounds; however, their accuracy is limited, as noise can cause variations in the peak strengths and locations that they utilize for classification.The research presented in this thesis applies signal processing to the LIBS spectra of a sample to detect the presence of IEDs in trace quantities, on the order of micrograms, from a distance of up to 20 m. We use independent component analysis to determine the locations of elemental peaks, and partial least squares -- discriminant analysis to identify the peaks providing discriminatory information about the presence of explosives. Our algorithm captures variations in the peak energies, and not peak strengths, in a region, rather than at specific locations, by fitting Lorentzian or Gaussian curves about the location of the peaks. The peak energies are then normalized using peaks not affected by sample type, and are used for classification, often with consideration of the energies of multiple peaks. Multi-spot fusion is performed in the area of interest, based on the maximum likelihood of finding a sample under given test conditions, to further increase the probability of detection.The method effectively detected the presence of explosives at a very low false alarm rate for the training data, and at a higher false alarm rate for test data. The reasons for this increase in false alarm rates, as well as possible solutions, are discussed in this thesis. Overall, the discriminating methods presented here perform better than most existing methods, especially in the case of real-time test data"--Abstract, page iii.