کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

Fatigue-Resistant s-EMG Hand Gesture Recognition with Lightweight Recurrent–Attention Networks

نویسندگان :

Aliasghar Bagheri Tirtashi¹ Mehran Jahed²

1- دانشگاه صنعتی شریف 2- دانشگاه صنعتی شریف

کلمات کلیدی :

Surface Electromyography(s-EMG)،Hand Gesture Recognition،Muscle Fatigue،Deep learning،Gated Recurrent Unit (GRU)،Multi-Head Self-Attention (MHA)

چکیده :

Surface electromyography (s-EMG) is widely used for upper-limb prosthetic control and rehabilitation, but practical systems must remain reliable when muscle fatigue changes the amplitude and spectrum of EMG signals. This study investigates fatigue-aware hand gesture recognition in a realistic low-channel setting using compact deep sequence models with self-attention. s-EMG signals were recorded from eight healthy subjects using six bipolar forearm electrodes while they performed four hand/wrist gestures. Classical feature-based classifiers (SVM, LDA, KNN, logistic regression, random forest) were compared with four deep architectures that operate directly on 250 ms raw windows: a CNN, a GRU-based model, and their multi-head self-attention (MHA) variants. All models were trained on clean (non-fatigued) data and evaluated under two families of simulated fatigue applied at the signal level: (i) amplitude scaling with additive noise and (ii) frequency-domain spectral tilt that shifts power toward lower frequencies and reduces median frequency. On clean data, deep models outperformed classical baselines, with MHA+CNN reaching 98.23% accuracy. Under amplitude scaling, CNN accuracy dropped to 29.21% at the highest fatigue level, whereas MHA+GRU still achieved 65.19%. Under spectral tilt, CNN accuracy decreased to 53.81% at the strongest distortion, while MHA+GRU maintained 84.58%. Subject-wise analyses showed that GRU-based models degraded more gracefully than CNN-based ones. These results indicate that lightweight recurrent architectures with self-attention provide improved robustness to fatigue-like distortions in low-channel s-EMG configurations, supporting more reliable control of wearable prosthetic and rehabilitation devices.