Shoulder muscle activation pattern recognition based on sEMG and machine learning algorithms

Area of Science:

• Robotics
• Biomedical Engineering
• Machine Learning

Background:

• Surface electromyography (sEMG) is crucial for volitional control in robotic rehabilitation, particularly for prostheses and exoskeletons.
• Perfecting upper limb exoskeleton control using sEMG remains a challenge.
• This research aims to process shoulder muscle bio-electrical signals for advanced robotic assistive device motion control.

Purpose of the Study:

• To assess machine learning algorithms for recognizing shoulder motion patterns from sEMG signals.
• To investigate how motion speed, individual differences, EMG device, and dataset size affect recognition accuracy.

Main Methods:

• A novel convolutional neural network (CNN) was developed to process sEMG from 12 muscles for upper arm motion pattern recognition (resting, drinking, forward-backward, abduction).
• Statistical analyses (ANOVA, GLM Univariate, Chi-square) were used to evaluate CNN model accuracy across different conditions.
• The impact of EMG dataset size on recognition accuracy was examined by incrementally increasing data.

Main Results:

• CNN models achieved high accuracy in motion pattern recognition: 97.57% for normal speed and 97.07% for fast speed motions.
• Cross-subject and cross-device model accuracies were 79.64% and 88.93% (normal speed), respectively.
• A statistically significant difference in pattern recognition accuracy was observed among different CNN models.

Conclusions:

• CNN algorithms effectively process shoulder and upper arm sEMG signals for recognizing distinct upper limb motions.
• CNN models trained on specific motion speeds demonstrated superior accuracy for those speeds compared to mixed-speed models.
• Augmenting EMG datasets for CNN training significantly enhances pattern recognition accuracy.