Diseño e implementación de un sistema de entrenamiento para personas protetizadas

Abstract

CONDITION FOR PUBLICATION OF PROJECT. Upper-limb myoelectric prostheses require users to learn consistent muscle activation patterns to generate reliable control commands, which can hinder early adaptation and increase the risk of underuse or abandonment. This work presents DynaTrain, a physical and digital training bench that enables guided myoelectric control practice before prosthesis delivery. The system integrates sEMG acquisition using uMyo sensors, BLE communication to an embedded module, local processing on a Jetson Nano, gesture execution on a Dynamixel based robotic hand, and a virtual twin for interactive visualization through a web interface. Dynamic prosthetic profiles were designed to emulate configurations with different degrees of freedom. A data pipeline was implemented for signal cleaning and normalization, including label correction, removal of invalid segments, and per-session Z-score standardization. As an initial validation, a 1D-CNN classifier was trained using real sessions segmented into continuous blocks and time windows, achieving high overall performance on the test set while highlighting limitations related to class imbalance and gesture duration variability. The results support the feasibility of the bench as a training and objective logging platform and identify key improvements for segmentation robustness, real time synchronization, and clinical evaluation. Keywords: sEMG, myoelectric prosthesis, training platform, digital twin, rehabilitation