Prediction of orientation-aware sEMG gesture classification by a Hybrid Phase–Frequency Dynamical Map Framework with the FORS-EMG dataset.

As an emerging research field, the recognition of gestures using surface electromyography (sEMG) has drawn significant attention in diverse applications such as assistive robotics, wearable human–computer interaction (wHCI), rehabilitation engineering, and prosthetic control. Nevertheless, accurate gesture classification is still difficult due to the high non-stationarity of the sEMG signals, and the inter-subject variability, forearm orientation, and electrode displacement. This paper introduces a new hybrid feature engineering approach for orientation-aware sEMG gesture classification using FORS-EMG dataset, which utilizes traditional time domain descriptors and a proposed Phase–Frequency Dynamical Map (PFDM) sEMG representation. The proposed method uses 160 features extracted from every signal epoch such as statistical time domain features and nonlinear features based on PFDM analysis including phase space and spectral descriptors. Four machine learning classifiers, Support Vector Machine (SVM), Random Forest (RF), K-Nearest Neighbors (KNN), and Naive Bayes (NB), are compared in regard to both subject-specific five-fold cross validation (5x5) and leave-one-subject-out (LOSO) validation. The experimental results show that the classification accuracy is high under the subject specific evaluation with SVM having the maximum mean accuracy of 96.32%. The proposed framework also shows the capacity to promise generalization in the cross subject validation. The results acquired suggest that the incorporation of nonlinear phase frequency dynamics towards conventional EMG descriptors could offer an understandable and a lightweight approach to robust wearable sEMG gesture recognition systems.