A Deep CNN-Based Approach for 10-Class with Two-Channel EMG Signal Classification

Electromyography (EMG) signals have emerged as vital tools for prosthetic control, motor function assessment, and rehabilitation technology. Traditional methods, relying on manual feature extraction, struggle with high-dimensional EMG data. Deep learning, specifically Convolutional Neural Networks (CNNs), has shown promise in addressing these challenges. However, there is a need to explore CNNs’ effectiveness, optimal architecture, real-time processing, and data augmentation for EMG signal classification. This study aimed to introduce a CNN-based approach tailored to classify ten-class, two-channel EMG signals. Eight participants provided EMG data by performing ten distinct finger and hand movements. Data were collected using two-channel EMG sensors, amplified, sampled at 4000 Hz, and filtered. A CNN architecture with eight layers was designed, incorporating one-dimensional convolutional layers, max pooling, global average pooling, dropout layers, and a dense output layer. The CNN-based approach demonstrated promising results. It exhibited high accuracy in classifying EMG signals for most classes, with an overall accuracy of approximately 91%. It excelled in the thumb (THU) and thumb-ring (TH-R) classes. However, some classes, like thumb-middle (TH-M) and thumb-index (TH-I), showed room for improvement in terms of recall. This study introduced an innovative CNN-based approach to tackle the challenges of classifying ten-class, two-channel EMG signals. It addressed critical research gaps and significantly advanced EMG signal processing, particularly for prosthetic control and rehabilitation technology.