Predicting Continuous Locomotion Modes via Multidimensional Feature Learning From sEMG.

Area of Science:

• Robotics and Human-Computer Interaction
• Biomedical Engineering
• Machine Learning for Healthcare

Background:

• Adaptive control in walking-assistive devices necessitates advanced methods for predicting human locomotion modes.
• Early detection of transitions (e.g., level walking to stair ascent) is vital for enhancing robotic system intelligence and user interaction.
• Surface electromyography (sEMG) signals offer rich physiological data for inferring human movement intentions.

Purpose of the Study:

• To develop and validate Deep-STF, a unified deep learning model for integrated feature extraction from sEMG signals.
• To enable accurate and robust continuous prediction of multiple locomotion modes and transitions.
• To assess the model's performance across varying prediction time intervals and its adaptability to new environments.

Main Methods:

• Proposed Deep-STF, an end-to-end deep learning architecture for spatial, temporal, and frequency feature extraction from sEMG.
• Trained and evaluated the model for predicting nine locomotion modes and 15 transitions.
• Tested prediction accuracy at intervals from 100 ms to 500 ms and evaluated performance on new terrains with fine-tuning.

Main Results:

• Deep-STF achieved 96.60% accuracy predicting 100 ms ahead, outperforming seven benchmark models.
• Accuracy remained high (93.22%) even with a 500 ms prediction horizon.
• The model demonstrated strong adaptability, improving from 71.12% to 96.27% accuracy on new terrains with 15 calibration trials.

Conclusions:

• Deep-STF provides a powerful, unified approach for predicting human locomotion modes and transitions using sEMG.
• The model's high accuracy, robustness, and adaptability show significant potential for integration into future walking-assistive devices.
• Successful implementation can lead to smoother, more intuitive, and responsive user experiences with assistive robotics.