Convolutional Feature Engineering for Cross-day Personal Identification using Wrist Electromyography.

Area of Science:

• Biometrics
• Artificial Intelligence
• Signal Processing

Background:

• Surface electromyography (EMG) is a viable biometric for person identification.
• Performance degradation over multiple days is a key challenge for EMG biometrics.
• Deep learning offers advanced feature extraction for biosignal analysis.

Purpose of the Study:

• To introduce MyoBM-Net, a novel convolutional neural network for EMG-based person identification.
• To evaluate the cross-day robustness of EMG biometrics using deep feature extraction.
• To compare MyoBM-Net against conventional feature extraction methods.

Main Methods:

• Utilized 1D and 2D convolutional layers for spatial and channel-specific feature extraction from EMG signals.
• Employed wrist EMG data from 43 participants across three separate days over one month.
• Conducted cross-day identification analysis, training and testing on different days.

Main Results:

• MyoBM-Net achieved a median rank-1 accuracy of 98.5% in cross-day identification.
• The proposed method outperformed conventional feature extraction techniques.
• A lower Distinctiveness-Based Index (DBI) of 1.73 was recorded, indicating strong discriminative power.

Conclusions:

• MyoBM-Net demonstrates significant potential for robust and accurate EMG-based personal identification, even across different days.
• The deep learning approach effectively addresses the challenge of performance variability in multi-day biometric systems.
• The method's ability to extract spatial and channel-specific information enhances its suitability for practical biometric applications.