Elaboração de protocolo para identificação de resíduos de disparo por arma de fogo utilizando espectroscopia LIBS e aprendizado de máquina

Abstract

Gun violence causes thousands of victims annually in Brazil. This occurs, among other factors, due to the low homicide resolution rate in several regions of the country. Efficiency is essential for investigation, as almost all crimes that are solved have their investigation completed within the first year after the incident. One of the methods for generating evidence on those suspected of shooting a firearm is the identification of gunshot residue on their hands in the form of inorganic particles, generally composed of lead (Pb) and antimony (Sb). The standard method for identifying GSR (Gunshot Residues) is SEM-EDS which, despite being extremely accurate, requires a lot of work time to identify each sample. This work proposes a new GSR collection and analysis protocol as an alternative to SEM-EDS, based on techniques that combine LIBS spectroscopy with machine learning algorithms. With the proposed methodology, the model trained based on the SVM algorithm was able to predict with positive and negative samples, establishing probabilities for each prediction. Based on the probabilities, a prediction threshold was set to rule out somewhat reliable outliers. The final model was able to accurately predict the true samples of each class and eliminate contaminated samples and outliers (outside the expected standard), which demonstrated the potential of the new protocol to analyze numerous samples simultaneously and in field conditions, where no it is possible to control contaminating elements.