FusionTCN-Attention: A Causality-Preserving Temporal Model for Unilateral IMU-Based Gait Prediction and Cooperative...

Area of Science:

• Biomechanics
• Robotics
• Machine Learning

Background:

• Human gait displays consistent contralateral coupling, where healthy limb motion predicts affected limb movement.
• Accurate prediction of limb kinematics is crucial for developing effective assistive technologies.

Purpose of the Study:

• To develop a causality-preserving temporal model, FusionTCN-Attention, for predicting contralateral hip and knee trajectories using unilateral IMU data.
• To establish a methodological basis for unilateral prediction in hemiparetic exoskeleton applications.

Main Methods:

• Implemented FusionTCN-Attention, integrating dilated temporal convolutions and a lightweight attention mechanism.
• Ensured strict real-time causality for low-latency predictions.
• Evaluated the model on data from twenty-one subjects.

Main Results:

• Achieved low root-mean-square errors (RMSEs) for hip (5.71°) and knee (7.43°) predictions.
• Obtained correlation coefficients exceeding 0.9 and a phase lag of 14.56 ms.
• FusionTCN-Attention outperformed conventional sequence models like Seq2Seq and causal Transformers.

Conclusions:

• Unilateral IMU sensing enables stable, low-latency prediction of limb kinematics.
• The developed model provides a foundation for unilateral prediction in advanced exoskeleton control systems.
• This research is a prerequisite for future hemiparetic exoskeleton applications.