Small-Data-Driven Temporal Convolutional Capsule Network for Locomotion Mode Recognition of Robotic Prostheses.

Area of Science:

• Robotics
• Biomedical Engineering
• Artificial Intelligence

Background:

• Precise control of robotic lower-limb prostheses is crucial for effective rehabilitation.
• Current methods often rely on large datasets, limiting their application in real-world scenarios.

Purpose of the Study:

• To propose a Temporal Convolutional Capsule Network (TCCN) for accurate locomotion mode recognition in robotic prostheses, particularly with small datasets.
• To enhance the control of lower-limb prostheses by improving the recognition of diverse walking conditions.

Main Methods:

• Developed a TCCN integrating spatial-temporal, dilated convolution, dynamic routing, and vector-based features.
• Compared TCCN performance against traditional machine learning (SVM) and deep learning models (CNN, RNN, TCN, CN) using 5-fold cross-validation.
• Evaluated performance on three- and five-locomotion mode recognition tasks.

Main Results:

• TCCN achieved 4.1% higher accuracy than CNN for three-locomotion modes and 5.2% higher for five-locomotion modes.
• The model demonstrated effectiveness in handling small datasets by balancing global and local information.
• Identified transition states as the primary source of confusion, suggesting areas for future refinement.

Conclusions:

• TCCN offers a promising approach for locomotion mode recognition in robotic prostheses, outperforming existing methods, especially with limited data.
• The network's ability to process vector information, retaining magnitude and direction, closely resembles human brain function.
• This advancement could lead to more intuitive and responsive prosthetic limb control.