부산대학교 도서관

Robotic assistive devices (orthoses and exoskeletons) usually feature non-intrusive sensors and intelligent algorithms. These enable the decoding of different locomotion modes (LMs), allowing the devices to tailor their assistance and support users in executing everyday walking tasks. Despite recent advances, most LM decoding tools (i) decode a small number of LMs simultaneously; (ii) have high recognition delays and low prediction times in advance; and (iii) do not consider the usually slow preferred speeds of neurologically impaired users. This study aims to address these shortcomings by presenting an automatic and user-independent kinematic-based LM decoding tool to classify, in real-time, 4 LMs (standing (St), level-ground walking (LGW), stair descent (SD), and stair ascent (SA)) when using a robotic assistive device at slow speeds. The proposed deep learning tool revealed an average F1-score, Matthew's Correlation Coefficient, and accuracy of 0.96 ± 0.014, 0.95 ± 0.018, and 0.98 ± 0.0072, respectively. Real-time experiments with the robotic device showed the ability to predict the upcoming LM with an average computational load and prediction time in advance of 2 ± 1 ms and 478 ± 239 ms, respectively, across six transitions (St-LGW, LGW-St, LGW-SD, SD-LGW, LGW-SA, and SA-LGW). These results suggest that the proposed LM decoding tool has the potential to be used in real-time to adapt the assistance of robotic assistive devices in advance.