Gesture Recognition via EMG and Neural Networks: Temporal Series and Key Feature Analysis

The advancement of prosthetic technology necessitates effective control systems to enhance user experience, particularly for individuals with limb loss. This study explores the classification of hand gestures through the analysis of electromyography (EMG) signals captured by a MyoWare 2.0 Muscle Sensor and processed by an ESP-32 from 91 participants. Focusing on three distinct gestures—fist, peace sign, and rock and roll sign—the research employs a robust methodology that encompasses data collection, feature extraction, and the training of a neural network. During the feature extraction phase, key metrics such as the root mean square (RMS) value, average value, and discrete waveform length are derived from the EMG signals, providing critical inputs for the model. The extraction of key metrics reduces the number of mathematical operations required by the neural network, thereby simplifying its overall architecture and decreasing its complexity. The trained neural network demonstrates a commendable training accuracy of 92.3% and a validation accuracy of $\mathbf{8 4. 2 1 \%}$, indicating its potential for realworld applications in prosthetic control.