Detección de la intención de movimiento de extremidades inferiores usando métodos de aprendizaje supervisado

Abstract

This work is part of the project Prototype of usable exoskeleton in the lower extremities, through the use of adaptive control algorithms. The aim of this project was to develop a capable algorithm of detecting the motion intention based on electromyograms (EMG) of subjects with pathologies in the lower limbs using artificial neural networks (ANN) with pattern recognition and the Levenberg-Marquardt method. Contemplated a stage (filtering, rectification and normalization) and the annotation of the motion intention for EMG pre-processing. Trained and validated the algorithm using an EMG database of normal subjects. Obtained an overall performance of 90.96% for a point-to-point evaluation and 94.88% in an evaluation by events. Publishing these results for ETCM-IEEE2018. Recorded a database of six patients (42.83 ± 10.51 years), containing 78 EMG signals, corresponding to 13 muscles. With the training parameters obtained in the first database, determined the motion intention in the subjects with pathologies, additionally, the values of signal-to-noise ratio (SNR) and mean frequency (MNF). Obtained an overall performance of 93.14% point-to-point and 91.19% by events, the delay time was 31.06 ± 18.89 ms, SNR of 17.28 ± 1.67 dB and the MNF values found they are lower than those the literature reported, suggesting lower torque in this population. These results allow contemplating the implementation of the algorithm in real time for an exoskeleton.