Representação de sinais de atividade eletrodérmica usando decomposições atômicas com aceleração via redes neurais artificiais

• Autor(a) principal: Lima, Paulo Victor Martins Rosa Huguenin de
• Data de Publicação: 2023
• Outros Autores: [e-mail]
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da UERJ
• Texto Completo: http://www.bdtd.uerj.br/handle/1/20419
• topic: Electronic engineering; Neural networks (Computer science); Natural language processing (Computer science); Algorithms; Engenharia eletrônica; Redes neurais (Computação); Processamento de linguagem natural (Computação); Algorítmos; ENGENHARIAS::ENGENHARIA ELETRICA::TELECOMUNICACOES

The profusion of devices incorporating sensors capable of monitoring and capturing biomedical signals in real-time, including electrodermal activity (EDA), demands techniques capable of efficiently representing the EDA signal and, at the same time, allowing analysis to provide useful information for interpretation. Greedy adaptive decomposition algorithms, such as Matching Pursuit (MP), have been used to obtain coherent approximations of signals, i.e., those related to the phenomena represented in the signals. The MP obtains a weighted sum of a few waveforms (atoms) selected from a redundant dictionary iteratively to represent the signal. The set of candidate atoms is the dictionary from which the ones to use in the approximation are selected. When it contains multiple prototype waveforms, from which the atoms are defined by parameters, we can say that the dictionary is composed of subdictionaries. In this dissertation, we investigate the use of a composite dictionary for the representation of EDA signals via MP, motivated by the fact that such a dictionary can provide more sparse decompositions of EDA signals. Appropriate parameterizations for different subdictionaries containing atoms similar to the phenomena represented in EDA signals are investigated. On the other hand, a composite dictionary entails greater computational complexity to obtain the representation. We propose a subdictionary selection method that acts in the loop of the MP algorithm using an artificial neural network (ANN) to reduce the computational cost of the algorithm without impairing the quality of the representation obtained. ANN models capable of predicting with good accuracy the subdictionary to be used in the MP decomposition loop are investigated. The results indicate that this approach allows dividing the computational complexity by the number of subdictionaries without strongly compromising the quality (measured as a function of the approximation error) as a function of the number of atoms (MP iterations) of the representation thus obtained.