Acionamento e controle de movimentos de uma prótese mioelétrica de mão

• Ano de defesa: 2024
• Autor(a) principal: Dellagostin, Jonatan
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• dARK ID: ark:/26339/0013000014t24
• Idioma: por
• Instituição de defesa: Universidade Federal de Santa Maria; Brasil; Engenharia Elétrica; UFSM; Programa de Pós-Graduação em Engenharia Elétrica; Centro de Tecnologia
• Palavras-chave em Português: Engenharia biomédica; Biomecânica; Prostética; Aprendizado de máquina; Redes neurais profundas; Biomedical engineering; Biomechanics; Prosthesis; Machine learning; Deep learning; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA
• Link de acesso: http://repositorio.ufsm.br/handle/1/31870

This thesis presents the steps for constructing a low-cost, open-source robotic prosthetic hand controlled by artificial intelligence (AI) and driven by servomotors, aimed at the social inclusion and labor market reintegration of patients who have undergone transradial amputation of the upper limb. The mechanical block provides a 3D printable anthropomorphic hand capable of emulating a set of selected useful hand grips. The electrical block is responsible for reading surface electromyographic signals (sEMG) from the patient’s forearm and the current of each servomotor, as well as activating the servomotors to perform the grips. The management block operates at two distinct times: training and inference. During the training phase, the block accesses sEMG readings from a patient’s forearm during a session where they are instructed to perform a set of predetermined grips, and uses them to train a machine learning classification model. Afterward, this management block reads the sEMG and servomotor currents in real-time, predicts the corresponding grip, and commands the electrical block to activate the servomotors to execute the grip. The grip intensity is controlled by a proportional-integral (PI) controller, indirectly using the servomotor current measurements. The result obtained in the training of the grip classification model, with an accuracy of 0.9708, proves the effectiveness of the chosen neural network architecture, even considering the use of only one sEMG sensor. The implementation of a control simulation environment proved to be a valuable resource for evaluating and debugging motor current control strategies, avoiding unnecessary wear on the physical system that can occur when using a strategy that has not been previously tuned and tested. The integration between all the blocks, however, in order to constitute a fully functional prosthesis, was not developed and remains as a challenge for future work.