Machine learning for grasping recognition using wearable sensors

Abstract

Listen

Translate article

Introduction: The objective is to evaluate the traditional classifiers for the identification of the grasp while doing different jobs, in order to obtain information that can be used in the diagnostic of the physical work requirements and job design. Methodology: The analysis considered different combinations of the data acquired from inertial and force resistive sensors: a) acceleration and resistive force sensors, b) acceleration, angular velocity and resistive force sensors c) acceleration, angular velocity, magnetic fields, and resistive force sensors. Different combinations of window and step sizes were selected with two overlap options: 50% and greater than 50%. Traditional classification models were trained: support vector machines, ensembles, Naive-Bayes algorithm. Results: Results demonstrate that the window size that presented optimal performance in the present study was 3 seconds with an overlap greater than 50%, the window size is greater than that suggested in the literature, which ranges from 0.75 to 2.25 seconds. Conclusions: The accuracy and F-score metrics for the different window-step combinations are presented, both metrics indicate that the models trained through Support Vector Machine have the best performance (90 %) with the combination of acceleration, angular velocity, and resistive force sensor.