Optimizing CNN Hyperparameters for Enhanced sEMG Signal Classification using D-Optimal Design

Abstract The effectiveness of Convolutional Neural Networks (CNNs) in classifying hand gestures from surface electromyography (sEMG) signals is significantly influenced by the meticulous selection of hyperparameters. However, the impact of these structural hyperparameters on classification accuracy remains underexplored. This study investigates five CNN hyperparameters through the implementation of a D-optimal experimental design. This methodology enables the systematic identification of efficient combinations while minimizing the number of experiments necessary for evaluation. Additionally, it allows for a thorough assessment of the individual contributions of each factor to accuracy and inference time, thereby highlighting the most influential parameters. Experimental trials were conducted using three distinct subsets from a public sEMG signal database (NinaPro). The findings indicate that the convolution type exerts the most significant influence on accuracy, closely followed by the number of parallel layers and, to a lesser extent, the number of sequential layers. Notably, one of the identified configurations attained an accuracy of 98.62% \(±\) 0.84 on data obtained from subjects with transradial amputations at distinct levels. Furthermore, the most optimized models demonstrated inference times of below 200 ms on a standard processor, demonstrating their potential applicability in real-time environments.