A Novel Active Learning Framework for Cross-Subject Human Activity Recognition from Surface Electromyography.

Area of Science:

• Biomedical Engineering
• Robotics
• Machine Learning

Background:

• Wearable sensor-based human activity recognition (HAR) is crucial for advanced exoskeleton control.
• Existing HAR methods struggle with data quality and adapting to new users (cross-subject adaptation).

Purpose of the Study:

• To develop an active learning framework for robust HAR in exoskeleton systems.
• To address challenges in data quality and cross-subject adaptation for wearable sensor data.

Main Methods:

• Integrated a relation network architecture with data sampling techniques.
• Utilized auxiliary classifiers for subject-specific boundary establishment.
• Employed classifier discrepancy for data significance assessment and partitioning into sample/template sets.
• Applied category clustering for parameter tuning and template data optimization.

Main Results:

• The proposed framework demonstrated superior performance across all statistical metrics on public and self-constructed datasets.
• Ablation experiments confirmed the critical role of data screening in the adaptation process.
• Achieved enhanced accuracy and generalizability in human activity recognition.

Conclusions:

• The active learning framework effectively adapts HAR models to target subjects, improving accuracy and generalizability.
• The method addresses key limitations in data quality and cross-subject adaptation for wearable sensor-based HAR.
• This work validates the practical implementation of user-independent HAR for exoskeleton control systems.