Detectores de novidades e classificadores especializados em sistemas de sonar passivo

• Ano de defesa: 2019
• Autor(a) principal: Muniz, Victor Hugo da Silva.
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Rio de Janeiro; Brasil; Instituto Alberto Luiz Coimbra de Pós-Graduação e Pesquisa de Engenharia; Programa de Pós-Graduação em Engenharia Elétrica; UFRJ
• Palavras-chave em Português: Sonar; Navegação autônoma; Sistema de apoio à decisão; Aprendizado computacional; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA::ELETRONICA INDUSTRIAL, SISTEMAS E CONTROLES ELETRONICOS
• Link de acesso: http://hdl.handle.net/11422/12291

In submarines, sonar operators have the task of identifying and classifying passive sonar contacts, so that possible threats are detected. The automation of this process is extremely relevant, since it facilitates the work of the professional of this area, requiring less physical and mental efforts during the surveillance. The proposal of this study is to investigate the efficiency of specialized models in the constitution of such a system, aiming to derive a mechanism that effectively detects unknown ships, as well as correctly identifies the labels of those already known. Three levels of specialization were considered: non-specialized, specialized in classes, and specialized in ships, assuming the following techniques for the construction of the system: Principal Component Analysis (PCA), Kernel Principal Component Analysis (KPCA), One-Class Support-Vector Machines (OCSVM), Gaussian Mixture Models (GMM), k-Nearest Neighbors (kNN), sparse k-Nearest Neighbors (s-kNN) and Local Outlier Factor (LOF). Experiments conducted with real data acquired on an acoustic lane showed a better performance of the models specialized in ships, which reached a novelty detection rate of 83.4%, conjugated with an average recognition rate of known classes of 90.5%. Regarding specifically the task of classifying the known classes, 98.7% are correctly labeled.