This study aimed to highlight the demand for upper limb compound motion decoding to provide a more diversified and flexible operation for the electromyographic hand. In total, 60 compound motions were selected, which were combined with four gestures, five wrist angles, and three strength levels. Both deep learning methods and machine learning classifiers were compared to analyze the decoding performance. For deep learning, three structures and two ways of label encoding were assessed for their training processes and accuracies; for machine learning, 24 classifiers, seven features, and a combination of classifier chains were analyzed. Results show that for this relatively small sample multi-target surface electromyography (sEMG) classification, feature combination (mean absolute value, root mean square, variance, 4th-autoregressive coefficient, wavelength, zero crossings, and slope signal change) with Support Vector Machine (quadric kernel) outstood because of its high accuracy, short training process, less computation cost, and stability (p < 0.05). The decoding result achieved an average test accuracy of 98.42 ± 1.71% with 150 ms sEMG. The average accuracy for separate gestures, wrist angles, and strength levels were 99.35 ± 0.67%, 99.34 ± 0.88%, and 99.04 ± 1.16%. Among all 60 motions, 58 showed a test accuracy greater than 95%, and one part was equal to 100%.
Keywords: compound motion decoding; deep learning; machine learning; myoelectric prosthesis; surface EMG.
Copyright © 2022 Zhang, Lu, Fan, Wang, Zhang, Li and Tao.