Automated Power-Precision Grasp Recognition from sEMG: A Multivariate Signal Decomposition-based Approach

This is the authors' abstract. We don't add key points for this paper.

Grasp recognition through surface electromyogra-phy (sEMG) has emerged as a significant element in the fields of automated control systems, human-computer interfaces, rehabil-itation systems, virtual environments, and related applications. Grasps can be categorized into different types based on their origination and force level. In power grasps, the object and hand have a rigid relationship, necessitating that all object movements originate from the arm. In contrast, precision handling enables the hand to perform intrinsic movements on the object with-out requiring arm movement. This necessitates development of methodology for recognizing these grasps from sEM G signals. This paper proposes a novel method for automated power-precision grasp recognition from sEMG. Overlapped windowing is utilized for sEMG signal segmentation. Each window frame is decomposed into its constituent modes using simplified successive multivariate variational mode decomposition (SSMVMD). Time domain features are then extracted from each mode for classification process. The KNN classifier has been utilized for binary (power and precision) and 14-class classification for NinaPro-DB4 dataset. The average subject-wise binary classification accuracy, precision, recall, and F1-score of 89.98%, 90.51 %, 89.30%, and 89.87 %, respectively are obtained. The cross-subject binary classification accuracy, precision, recall, and F1-score of 88.75 %, 89.78%, 87.40%, and 88.57%, respectively are achieved. The average subject-wise 14-class classification accuracy of 77.62% has been obtained. Further, the accuracy of 75.95% has been achieved for cross-subject 14-class classification.