A Deep Learning Model with a Self-Attention Mechanism for Leg Joint Angle Estimation across Varied Locomotion Modes

Area of Science:

• Robotics and Biomechanics
• Artificial Intelligence in Healthcare

Background:

• Traditional lower limb assistive devices use finite-state strategies for trajectory planning, limiting adaptability.
• Deep learning offers potential for adaptive gait pattern learning from diverse user data.
• Current methods struggle with seamless transitions between different locomotion tasks and personalization.

Purpose of the Study:

• To develop a temporal deep learning model with self-attention for continuous lower limb joint angle trajectory generation.
• To evaluate the model's performance against existing methods for ankle and knee angle estimation.
• To demonstrate the benefits of data diversity and transfer learning for assistive device personalization.

Main Methods:

• A temporal deep learning model incorporating a self-attention mechanism was developed.
• Fast Fourier Transform and paired t-tests were used for performance analysis.
• A 10-fold leave-one-out cross-validation scheme was employed for testing.

Main Results:

• The attention model achieved low Normalized Root Mean Square Errors (NRMSE) of 11.50% (±2.37%) for ankle and 9.31% (±1.56%) for knee angles.
• Statistical analysis confirmed the attention model's superiority over baseline models in reducing prediction error.
• The model generated smoother joint trajectories, enhancing safety and comfort, and transfer learning improved performance.

Conclusions:

• The proposed attention-based model offers a data-driven approach for generating adaptable lower limb joint trajectories.
• This method can seamlessly switch between locomotion tasks, overcoming limitations of finite-state strategies.
• The approach reduces the need for extensive manual personalization, making assistive devices more accessible.