Kontrol Exoskeleton Robot Jari Tangan Berbasis EMG (Electromyograph) dengan Kombinasi Metode Transformer dan CNN (Convolutional Neural Network) untuk Rehabilitasi Pasien Stroke

Abstract

This research focuses on the development of an EMG (Electromyograph) based finger exoskeleton robot control system for the rehabilitation of stroke patients. The system uses a combination of Transformer and Convolutional Neural Network (CNN) methods to predict the angle of the fingers This research aims at developing a system for the rehabilitation of stroke patients. In an effort to help patients rehabilitate, scientists are developing robotics innovations. In stroke rehabilitation technology, there are various methods that can be used, one of which is using an exoskeleton type robot that can be used to assist movement in the body parts of stroke patients. The study began with data collection in the form of EMG signal recording and finger angle movements of respondents in healthy conditions and post-stroke conditions.Data collection is assisted by Myo Armband to record EMG signals as input and Flex Glove to record changes in hand finger angles as targets. Both types of data were taken by following a predefined procedure of repetition of grasping, opening, and resting movements. The raw data is processed through pre-processing consisting of filtering, windowing, and feature extraction with Root Mean Square (RMS) to then be trained on the CNN-Transformer method deep learning system.The best R2 results on training data show that Transformer gives the best R2 Score on healthy data (0.84), while CNN-Transformer has the lowest R2 Score (-3.59) on stroke data. On cross-validation, CNN showed better performance with R2 Score 0.56 on healthy data, compared to Transformer and CNN-Transformer. Overall, the results show that the combination of Transformer and CNN methods can provide more accurate finger angle predictions than the individual methods, especially on training data. However, the poor performance of CNN-Transformer in some cases could be due to the high model complexity and the mismatch between the CNN and Transformer architectures. These findings are expected to contribute to improving the effectiveness of stroke patient rehabilitation through more advanced and accurate robotic exoskeleton technology