A Transformer-Based Neural Network for Gait Prediction in Lower Limb Exoskeleton Robots Using Plantar Force

Area of Science:

• Robotics
• Biomedical Engineering
• Artificial Intelligence

Background:

• Lower limb exoskeletons require robust human-robot collaboration for safety and effectiveness.
• Accurate gait prediction is essential for compensating sensor delays and optimizing exoskeleton control.
• Plantar force provides rich, intrinsic gait pattern information crucial for prediction.

Purpose of the Study:

• To develop a novel deep learning model for predicting bilateral hip and knee joint angles using plantar pressure data.
• To enhance human-robot collaboration in lower limb exoskeleton applications through improved gait prediction.
• To investigate the efficacy of a transformer-based neural network combined with variational mode decomposition (VMD) for this task.

Main Methods:

• Developed a transformer-based neural network (TFSformer) incorporating 1D convolution and variational mode decomposition (VMD).
• The encoder uses 1D convolution to extract features from temporal and force-space dimensions of plantar pressure.
• The decoder employs a multi-channel attention mechanism and a deep multi-channel attention structure for efficient processing.

Main Results:

• TFSformer demonstrated significant reductions in Mean Absolute Error (MAE) and Mean Squared Error (MSE) for joint angle prediction.
• Compared to CNN, Transformer, and CNN-Transformer models, TFSformer achieved MAE reductions of up to 15.04% and MSE reductions of up to 29.90%.
• The model was validated on a custom dataset comprising data from 35 volunteers.

Conclusions:

• The proposed TFSformer model effectively predicts lower limb joint angles from plantar pressure, offering a promising approach for exoskeleton control.
• Integrating VMD and attention mechanisms in a transformer architecture improves gait prediction accuracy and efficiency.
• This advancement holds potential for enhancing the safety, efficacy, and user experience of lower limb exoskeleton robots.