Estratégia de caracterização de sinais eletromiográficos baseada em redes neurais artificiais para sistemas de controle de máquinas de movimento contínuo

• Ano de defesa: 2019
• Autor(a) principal: Sponchiado, Grégori Stefanello
• Orientador(a): Vargas, Fabian Luis
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Pontifícia Universidade Católica do Rio Grande do Sul
• Programa de Pós-Graduação: Programa de Pós-Graduação em Engenharia Elétrica
• Departamento: Escola Politécnica
• País: Brasil
• Palavras-chave em Português: Aprendizado de Máquina; Redes Neurais Artificiais; EMG; CPM; CAM
• Palavras-chave em Inglês: Machine Learning; Artificial Neural Networks; EMG; CPM; CAM
• Área do conhecimento CNPq: ENGENHARIAS
• Link de acesso: http://tede2.pucrs.br/tede2/handle/tede/9333

Human beings often suffer from lower limb injuries which are mostly related to aging and daily-motion. This impacts health and exposes human body to undesirable surgical interventions and therapies. In this scenario, the goal of this work is twofold: (a) use artificial neural network (ANN) to identify and classify muscle usage patterns based on electromyographic (EMG) signals, and (b) use the ANN’s output decision to control a Continuous Passive Motion (CPM) machine during a patient physiotherapy session. The strategy uses surface electromyography (sEMG) combined with a supervised learning method and artificial intelligence (AI) to create a feedback signal which allows these devices to function in Continuous Active Motion (CAM) mode. Methods: This work used 300 EMG signals collected from the vastus lateralis muscle of 10 healthy individuals to develop a strength classifier system. The core’s classifier is composed of a trained (backpropagation) feedforward neural network. The EMG signals are classified into predefined force levels, which in turn are used as inputs to control a CPM machine. Thus, there is a direct correspondence between each of the predefined force levels and the CPM machine linear displacement. Results: The trained ANN classifies, at real-time, EMG signals into force levels at 81 % accuracy with computational efficiency. After receiving the predefined force levels from the ANN’s output, the delay of the mechanical control system to adjust the CPM machine is less than 100 seconds. Conclusion: The AIbased assertiveness of the proposed strategy allows us to consider extending the use of single muscle EMG signals to pave the way for controlling another biomechanical machines in a near future.