Melhoramento do índice de detecções na espectrometria gama em amostras ambientais usando inteligência artificial
Carregando...
Arquivos
Data
2016-12-19
Autores
Arine, Bruno Burini Robles [UNESP]
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
A utilização de elementos radioativos se estende hoje a diversos ramos da ciência e tecnologia, exigindo maneiras eficientes, precisas e sustentáveis de monitoramento para a preservação ambiental e a saúde humana. Uma técnica eficiente e não-destrutiva de análise de elementos emissores de raios gama é a espectrometria gama. No entanto, os softwares atuais de análise se deparam com certos empecilhos quando tratam de amostras ambientais, cujas concentrações de radioatividade são próximas do limite de detecção, ou quando certos radionuclídeos de interesse estão sujeitos a interferências. Este trabalho desenvolveu algoritmos de análise através de ferramentas do campo da inteligência artificial, de forma a obter um desempenho superior do que os métodos empregados em softwares clássicos de análise radiométrica. Três classificadores do campo da inteligência artificial foram testados; árvores de decisão (AD), máquinas de vetores de suporte (MVS) e redes neurais artificiais (RNA). Em testes com uma fonte de chumbo-210 de baixa atividade, a MVS obteve acurácia de 0,93, enquanto os métodos tradicionais obtiveram acurácia máxima de 0,73. Em testes com amostras reais de nitrato de uranila, a acurácia de classificação da RNA foi de 0,91, enquanto a acurácia dos métodos tradicionais foi de 0,70. Observou-se que os modelos baseados em inteligência artificial tiveram desempenho superior aos métodos tradicionais em todos os experimentos. Entretanto, esta vantagem diminui à medida que a relação sinal-ruído dos espectros aumenta, se tornando negligível quando a relação sinal-ruído supera o limite de detecção.
Usage of radioactive elements is present in several areas of science and technology, requiring efficient, accurate, and sustainable ways of environmental conservation and human health monitoring. Gamma-ray spectrometry is an efficient and non-destructive analysis technique for identification of gamma-ray emitting elements. Nonetheless, current analysis softwares have certain limitations regarding environmental samples, whose radioactivity concentration is often close to the detection limit, or when certain radionuclides of interest are subject to interference. This work aims the development of analysis algorithms based on the field of artificial intelligence in order to achieve better performance than the methods in classical radiometric analysis software. Three artificial intelligence classifiers were tested; decision trees (DT), support vector machines (SVM) and artificial neural networks (ANN). In tests with a low-activity lead-210 source, SVM obtained an accuracy of 0.93, while the traditional methods obtained a maximum accuracy of 0.73. In tests with real samples of uranyl nitrate, the accuracy of ANN classification was 0.91, while the accuracy of the traditional methods was 0.70. It was observed that artificial intelligence-based models performed better than traditional methods in all experiments. However, this advantage decreases as the signal-to-noise ratio of the spectra increases, becoming negligible when the signal-to-noise ratio exceeds the detection limit.
Usage of radioactive elements is present in several areas of science and technology, requiring efficient, accurate, and sustainable ways of environmental conservation and human health monitoring. Gamma-ray spectrometry is an efficient and non-destructive analysis technique for identification of gamma-ray emitting elements. Nonetheless, current analysis softwares have certain limitations regarding environmental samples, whose radioactivity concentration is often close to the detection limit, or when certain radionuclides of interest are subject to interference. This work aims the development of analysis algorithms based on the field of artificial intelligence in order to achieve better performance than the methods in classical radiometric analysis software. Three artificial intelligence classifiers were tested; decision trees (DT), support vector machines (SVM) and artificial neural networks (ANN). In tests with a low-activity lead-210 source, SVM obtained an accuracy of 0.93, while the traditional methods obtained a maximum accuracy of 0.73. In tests with real samples of uranyl nitrate, the accuracy of ANN classification was 0.91, while the accuracy of the traditional methods was 0.70. It was observed that artificial intelligence-based models performed better than traditional methods in all experiments. However, this advantage decreases as the signal-to-noise ratio of the spectra increases, becoming negligible when the signal-to-noise ratio exceeds the detection limit.
Descrição
Palavras-chave
Radiação gama, Meio ambiente, Inteligência artificial, Aprendizado de máquina, Gamma radiation, Environment, Artificial intelligence, Machine learning