Optimizing Locomotor Task Sets for Training a Biological Joint Moment Estimator.

Area of Science:

• Biomechanics
• Robotics
• Machine Learning

Background:

• Accurate biological joint moment estimation from wearable sensors is crucial for advanced exoskeleton control in real-world locomotion.
• Current deep learning methods demand extensive in-lab data, hindering robust model development due to data acquisition challenges.

Purpose of the Study:

• To develop a locomotor task set optimization strategy to minimize data collection for wearable sensor-based joint moment estimation.
• To identify a minimal, representative set of tasks that preserves neural network performance for exoskeleton control applications.

Main Methods:

• Performed cluster analysis on dimensionally reduced biomechanical features from diverse cyclic and non-cyclic locomotor tasks.
• Identified minimal viable task clusters to train a neural network for hip joint moment estimation.
• Evaluated model performance using cross-validation across subjects.

Main Results:

• The optimized task set-based model achieved a root mean squared error of 0.29 ± 0.06 Nm/kg for hip joint moment estimation.
• Performance was significantly better than using only cyclic tasks (p<0.05) and comparable to using the full task set.
• Demonstrated significant reduction in data collection and model training costs without compromising accuracy.

Conclusions:

• The proposed task set optimization strategy effectively reduces data requirements for deep learning models in exoskeleton control.
• This approach enables maintaining high model accuracy while significantly lowering the burden of data collection and training.
• Future exoskeleton designers can utilize this strategy to minimize data needs for deep learning-based wearable robot control.