A study on comparing method of motion classification using muscle bulging for control of powered prosthetic hand

Abstract

Aiming at the control of a powered prosthetic hand, this paper compares methods for the classification of intended hand motions using muscle bulging patterns caused by muscle contraction. Two sheets of Polyvinylidene Difluoride (PVDF) film were used as sensors to detect the muscle bulging on the forearm caused by intended hand motions. A neural network had been successfully trained for the classification of six types of hand motions using the muscle bulging pattern detected by the two PVDF sensors. In this paper, we further studied the motion classification methods of back propagation neural network (BPNN), k-nearest neighbor algorithm (k-NN), and support vector machine (SVM) to compare their classification performance. We found that all three methods had a similar classification rate of about 95% for six types of hand motions. Moreover, a regressive analysis comparison of the time for each classification method to converge to 95% of the total classification rate showed that SVM converged significantly earlier than BPNN and k-NN. The time it takes for SVM to converge the classification rate to 95% is less than 0.1 s, suggesting that real-time motion classification is possible by using SVM. In a similar manner, we found that SVM requires the least training data of the three methods at only nine trials for a type of motion. Furthermore, SVM had the highest classification rate at about 90% in practical experimental conditions. In conclusion, SVM was found to be the most practical of the three methods.