Aplicação de máquinas de vetores de suporte para geração de sinais de controle robustos para próteses mioelétricas

Abstract

The upper limbs have great importance while performing daily activities. In addition, they aid in communication through movements and gestures. This way loss of the upper limbs directly affects these abilities and can cause the individual to lose his independence, socialization capacity, self-esteem, etc. Numerous researches have been developed with the purpose of solving these problems, most of them focused on prosthesis design. Prosthetic control techniques that use EMG signals are on the rise, mainly from pattern recognition with the use of Machine Learning algorithms, capable of providing various degrees of freedom with high accuracy. However, there are several problems that still limits this form of control in clinical applications, so achieving a generic model has become fundamental because performing the control of prosthesis through EMG signals requires some physical and mental effort by the user. The electrodes used for EMG signal capture are usually positioned in the prosthesis socket and thereby can undergo a constant change of position by modifying the extracted features of the EMG signal. This causes an unstable system performance. Thus, this work aims to develop a classification system capable of generating a more robust and generic model, maintaining a good performance even if the position of the electrodes has been changed. The Support Vector Machines technique was chosen since several works in the literature have proved their effectiveness in relation to other techniques. Moreover, the Principal Component Analysis technique was used in order to significantly reduce the data dimensionality, maintaining a good performance. The validation of the results was done through the cross-validation technique using the k-Fold method. The construction of more generic models allowed a great performance, with the volunteers' average accuracy of 91%. In addition, the use of the Principal Component Analysis technique allowed the reduction of data dimensionality from 24 features to a maximum of 8 main components, maintaining 95% of the data variability.