Assessing the Resilience of sEMG Classifiers to Sensor Malfunction and Signal Saturation

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Surface electromyography (sEMG) is crucial for gesture recognition.
• Current sEMG pipelines lack quantified robustness against real-world signal degradations like amplitude saturation and channel dropout.

Purpose of the Study:

• To systematically map the robustness of conventional sEMG feature-classifier pipelines under controlled clipping and single-channel failure.
• To quantify the impact of signal degradations on different feature combinations, classifiers, and subjects.

Main Methods:

• Evaluated four time-domain sEMG features (RMS, Variance, Zero Crossing, Waveform Length) and their pairwise fusions.
• Tested three classifiers (SVM, LDA, Random Forest) under varying symmetric saturation thresholds and single-channel dropout scenarios.
• Analyzed subject-wise performance dispersion instead of aggregate scores.

Main Results:

• Lightweight feature pairs (e.g., RMS + Waveform Length) with Random Forest demonstrated consistent robustness.
• Performance recovered as clipping weakened and remained resilient to single-channel dropout.
• The conventional pipeline exhibited significantly faster training times compared to deep learning baselines.

Conclusions:

• A combination of specific lightweight features and Random Forest provides a robust operating point for sEMG gesture recognition.
• The findings support the use of conventional sEMG pipelines for real-time applications requiring recalibration and robustness.
• This systematic analysis offers a valuable reference for designing resilient sEMG-based systems.