Unsupervised layer-wise feature extraction algorithm for surface electromyography based on information theory.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Surface electromyography (SEMG) signal processing heavily relies on feature extraction.
• Deep learning methods achieve high performance but supervised learning is costly due to label requirements.
• Unsupervised methods are increasingly important for SEMG analysis.

Purpose of the Study:

• To develop an unsupervised method for learning disentangled feature representations of SEMG signals.
• To understand the attribute information within SEMG data more effectively.
• Introduce the Layer-wise Feature Extraction Algorithm (LFEA).

Main Methods:

• Proposed an information-based method for unsupervised feature learning.
• Designed a layer-wise network structure to accommodate different attribute abstraction levels.
• Utilized TC score and MIG metric to evaluate disentanglement performance.

Main Results:

• LFEA achieved superior performance in disentanglement metrics (TC score and MIG).
• The algorithm demonstrated a significant accuracy lead of at least 5.8% in motion classification compared to other models.
• Experimental results validated the effectiveness of the layer-wise network design.

Conclusions:

• The Layer-wise Feature Extraction Algorithm (LFEA) offers an effective unsupervised approach for SEMG feature representation.
• LFEA enhances the understanding of SEMG signal attributes and improves classification accuracy.
• This method addresses the limitations of supervised learning in SEMG analysis.