Recurrent Neural Network as Estimator for a Virtual sEMG Channel

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Surface Electromyography (sEMG) is crucial for human-computer interaction.
• Signal contamination and noise saturation significantly degrade sEMG-based classification accuracy.
• Accurate hand posture classification is essential for prosthetic control and rehabilitation.

Purpose of the Study:

• To estimate a virtual sEMG channel using Long Short-Term Memory (LSTM) recurrent neural networks.
• To enhance the classification of hand postures using the generated virtual sEMG channel.
• To evaluate the performance of the virtual channel against clean and contaminated sEMG data.

Main Methods:

• Utilized Long Short-Term Memory (LSTM) nodes within a Recurrent Neural Network (RNN) architecture to create a virtual sEMG channel.
• Employed the NinaPro database for hand posture classification.
• Implemented a multi-class, one-against-all Support Vector Machine (SVM) classifier with Root Mean Square (RMS) of the sEMG signal as the primary feature.

Main Results:

• The virtual sEMG channel achieved an average hit rate of 73.96% ± 3.02% on clean data.
• Classification accuracy on contaminated data significantly improved from 9.29% ± 4.42% to 66.48% ± 6.11% with the virtual channel.
• Demonstrated the efficacy of the virtual channel in mitigating noise saturation effects on sEMG signal classification.

Conclusions:

• The proposed LSTM-based virtual sEMG channel effectively improves hand posture classification accuracy.
• This method offers a robust solution for enhancing sEMG signal analysis in the presence of noise.
• The virtual channel approach shows significant potential for real-world applications requiring reliable sEMG interpretation.