EMG-based hand gesture classification using Myo Armband with feedforward neural network

Abstrak

This paper presents the development of an electromyography (EMG)-based hand gesture identification system for remote-controlled applications. Even though the Myo Armband is no longer commercially supported, the research discusses its use in EMG data collecting. Open-source libraries were utilized to capture EMG data from this device to solve this problem. Using the developed data acquisition platform, data was collected from 30 participants who performed three (3) gestures - a fist, an open hand, and a pinch. The energy spectral density (ESD) and power ratio (pRatio) were extracted to describe gesture-specific patterns. A feedforward neural network (FFNN) was implemented for classification, initially configured with 10 hidden neurons and later optimized to 40 neurons to improve the performance. The box plot analysis showed channels CH1, CH4, CH5, and CH7 as the most significant for enhancing classification accuracy. The optimized FFNN achieved 80% and 70% for the training and testing accuracies, respectively. However, the results suggest that implementing a systematic protocol during data acquisition to reduce signal overlap between movements could improve classification accuracy. In conclusion, the study successfully developed an open-source EMG data acquisition platform for MYO Armband and demonstrated acceptable hand gesture recognition using an optimized FFNN.