Rapid detection approach for electronic nose systems using deep learning models

• Ano de defesa: 2020
• Autor(a) principal: GAMBOA, Juan Carlos Rodriguez
• Orientador(a): OLIVEIRA JUNIOR, Wilson Rosa de
• Banca de defesa: MELLO, Rafael Ferreira Leite de, SILVA, Antonio Samuel Alves da, PAULA NETO, Fernando Maciano de, LUDERMIR, Teresa Bernarda
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal Rural de Pernambuco
• Programa de Pós-Graduação: Programa de Pós-Graduação em Biometria e Estatística Aplicada
• Departamento: Departamento de Estatística e Informática
• País: Brasil
• Palavras-chave em Português: Nariz artificial; Odor; Detecção; Processamento de dados
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::PROBABILIDADE E ESTATISTICA
• Link de acesso: http://www.tede2.ufrpe.br:8080/tede2/handle/tede2/8754

The present thesis work focused on proposing a method to accelerate the identification of the specimens in electronic nose systems. Since the conventional data processing approach used in E-Nose is based on an initial stage of signal preprocessing applying techniques to perform the feature extraction (dynamic and static characteristics) for obtaining the odor fingerprint. Besides, in some cases, it is required to implement a feature selection method to choose the best attributes before the classification tasks using pattern recognition methods. For example, a Support Vector Machine (SVM) is one of the most common processing methods for odor recognition by the electronic olfactory systems. Therefore, the use of the traditional approach needs the whole measurement to obtain the main odorant parameters involving preprocessing techniques, which represents a challenge when aiming to perform real-time odors recognition. Thus, in this work is presented an approach for electronic olfactory systems data processing focused on the treatment of raw data based on a rising-window protocol to find an early portion of the sensor signals with the best recognition performance. We compared the proposed approach against a traditional method (using the entire response curves, applying preprocessing techniques to extract the features and later processing them using an SVM algorithm) in a real application with measures acquired with our developed system. Further, to validate the use of the proposed approach at different settings of electronic olfactory systems, we conducted more tests with several datasets and using deep learning techniques like convolutional neural network CNN. The results showed outperformance accuracy compared with the traditional approach with the advantage of using an early portion of the responses of the sensors, reducing the necessary time to make forecasts.