Electrode Shifts Estimation and Adaptive Correction for Improving Robustness of sEMG-Based Recognition

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Surface electromyography (sEMG) recognition systems suffer accuracy degradation due to electrode shifts during use.
• Existing static recognition models are not robust to these dynamic disturbances, limiting practical applications.

Purpose of the Study:

• To develop a transfer learning method to mitigate the impact of electrode shifts on sEMG-based gesture recognition.
• To introduce an adaptive transformation technique for correcting sEMG signals affected by electrode displacement.

Main Methods:

• A novel activation angle was used to establish a polar coordinate system for electrode localization.
• An adaptive transformation, based on estimated shifts, was applied to correct sEMG samples.
• Experiments involved ten subjects, eight gestures, and multiple arbitrary positions, with shifts recorded using a 3D-printed annular ruler.

Main Results:

• The proposed method achieved an average accuracy of 79.32% for eight-gesture recognition, a substantial improvement over non-adaptive models (35.72%) and iGLCM (60.99%).
• The error between recorded and estimated shifts was minimal (-0.017±0.13 radians).
• The method successfully adapted a pre-trained model using single-label samples to recognize gestures in arbitrary positions, demonstrating efficiency.

Conclusions:

• The proposed transfer learning method effectively corrects for electrode shifts in sEMG signals, significantly improving recognition accuracy.
• This adaptive approach reduces the burden of re-training for users in sEMG-based rehabilitation systems.
• The technique offers a robust solution for real-world sEMG applications where electrode placement can vary.