Deep Learning-Based Mapping of Textile Stretch Sensors to Surface Electromyography Signals: Multilayer Perceptron, Convolutional Neural Network, and Residual Network Models

This study evaluates the mapping accuracy between textile stretch sensor data and surface electromyography (sEMG) signals using Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), and Residual Network (ResNet) models. Data from the forearm, biceps brachii, and triceps brachii were analyzed using Root Mean Square Error (RMSE) and R2 as performance metrics. ResNet achieved the lowest RMSE (e.g., 0.1285 for biceps brachii) and highest R2 (0.8372), outperforming CNN (RMSE: 0.1455; R2: 0.7639) and MLP (RMSE: 0.1789; R2: 0.6722). The residual learning framework of ResNet effectively handles nonlinear patterns and noise, enabling more accurate predictions even for low-variability datasets like the triceps brachii. CNN showed moderate improvement over MLP by learning temporal features but struggled with low-variability datasets. MLP, as the baseline model, demonstrated the highest RMSE and lowest R2, highlighting its limitations in capturing complex relationships. These results suggest the potential reliability of ResNet for mapping textile stretch sensor data to sEMG signals, showing promising performance within the scope of this study. Future research could explore broader applications across different sensor configurations and activities to further validate these findings.