Classification techniques for adaptive distributed networks and aeronautical structures.

• Ano de defesa: 2018
• Autor(a) principal: Feitosa, Allan Eduardo
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Acoustic emission; Acústica (Emissão); Adaptive networks; Classification; Detection; Discrimination; Distributed networks; EMBRAER; Estruturas (Monitoramento); Maximum likelihood; Sensor; Sensors; Sistemas distribuídos; Structural health monitoring
• Link de acesso: http://www.teses.usp.br/teses/disponiveis/3/3142/tde-05022019-104746/

This master thesis is the result of a collaborative work between EMBRAER and the Escola Politécnica da USP for the study of structural health monitoring (SHM) techniques using sensors applied to aircraft structures. The goal was to develop classification techniques to discriminate between different events arising in the aircraft structure during tests; in the short term, improving the current SHM system used by EMBRAER, based on acoustic emission and, in the long term, fostering the development of a fully distributed system. As a result of studying classification methods for immediate use, we developed two techniques: the Spectral Similarity and a Support Vector Machines (SVM) classifier. Both are unsupervised solutions, due to the unlabeled nature of the data provided. The two solutions were delivered as a final product to EMBRAER for prompt use in the existing SHM system. By studying distributed solutions for future implementations, we developed a detection algorithm based on adaptive techniques. The main result was a special initialization for a maximum likelihood (ML) detector that yields an exponential decay rate in the error probability to a nonzero steady state, using adaptive diffusion estimation in a distributed sensor network. The nodes that compose the network must decide, locally, between two concurrent hypotheses concerning the environment state where they are inserted, using local measurements and shared estimates coming from their neighbors. The exponential performance does not depend on the adaptation step size value, provided it is sufficiently small. The results concerning this distributed detector were published in the journal IEEE Signal Processing Letters.