Deep learning and feature engineering techniques applied to the myoelectric signal for accurate prediction of movements

• Ano de defesa: 2021
• Autor(a) principal: SOUZA, Gabriel Cirac Mendes
• Orientador(a): MORENO, Robson Luiz
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de Itajubá
• Programa de Pós-Graduação: Programa de Pós-Graduação: Doutorado - Engenharia Elétrica
• Departamento: IESTI - Instituto de Engenharia de Sistemas e Tecnologia da Informação
• País: Brasil
• Palavras-chave em Português: Bio-sinais; BioPatRec; CNN; Engenharia de feature; Engenharia de reabilitação; LSTM
• Área do conhecimento CNPq: CNPQ::ENGENHARIAS::ENGENHARIA ELÉTRICA
• Link de acesso: https://repositorio.unifei.edu.br/jspui/handle/123456789/2477

Pattern recognition techniques in the Myoelectric Signal (EMG) are employed in the development of robotic prostheses, and for that, they adopt several approaches of Artificial Intelligence (AI). This Thesis proposes to solve the problem of recognition of EMG standards through the adoption of profound learning techniques in an optimized way. The research developed an approach that extracts the characteristic a priori to feed the classifiers that supposedly do not need this step. The study integrated the BioPatRec platform (advanced prosthesis study and development) to two classification algorithms (Convolutional Neural Network and Long Short-Term Memory) in a hybrid way, where the input provided to the network already has characteristics that describe the movement (level of muscle activation, magnitude, amplitude, power, and others). Thus, the signal is tracked as a time series instead of an image, which allows us to eliminate a set of points irrelevant to the classifier, making the information expressive. In the sequence, the methodology developed software that implements the concept introduced using a Graphical Processing Unit (GPU) in parallel this increment allowed the classification model to combine high precision with a training time of less than 1 second. The parallel model was called BioPatRec-Py and employed some Engineering techniques of Features that managed to make the network entry more homogeneous, reducing variability, noise, and standardizing distribution. The research obtained satisfactory results and surpassed the other classification algorithms in most of the evaluated experiments. The work performed a statistical analysis of the outcomes and fine-tuned the hyperparameters of each of the networks. Ultimately, BioPatRec-Py provided a generic model. The network was trained globally between individuals, allowing the creation of a standardized approach, with an average accuracy of 97.83%.