Optimization Of Myoelectric Pattern Recognition System Of Multifunction Upper-limb Prosthesis

The robustness of myoelectric prosthesis is largely influenced by many factors, which play a prominent role to determine the average accuracy rate of the classified patterns. These factors become more important when multifunctional prosthesis with high degrees of freedom (DOF) is constructed.In this research work, we proposed different optimized pattern recognition (PR) systems in order to accurately classify different patterns of hand motions, based on the extracted feature vector (FV) of surface electromyographic (sEMG) signal.The experimental results demonstrate that the combination of the following extracted features achieved the highest classification rate of98%using the linear discriminant analysis (LDA) classification algorithm:sample entropy (SampEnt), root mean square (RMS), myopulse percentage rate (MYOP) and difference absolute standard deviation value (DASDV). In addition, it was found that the performance of the classifier was improved through the implementation of more than one feature.Furthermore, the performance of five wavelet families was tested to select the proper wavelet family that leads to highest classification rate. The results show that the highest average classification accuracy was97.41%achieved by implementing general neural network (GRNN) classification method based on energy of wavelet coefficients (dblO at sixth decomposition level). In addition, the results of our experiment demonstrated that the use of wavelet families at a high decomposition level increases the recognition rate of hand motions.In this study, we also investigated the performance of three kernels functions of support vector machine (SVM) classifier. The result shows that polynomial kernel function is the optimal choice in most cases. The highest achieved classification accuracy was93%using extracted wavelet coefficients. Moreover, we investigated the best value of K that should be used as an input parameter in the K-nearest neighbor (K-NN) algorithm. The result demonstrates that k=5is the optimal choice in most cases.In addition, we proposed two intelligent hybrid pattern recognition systems based on swarm intelligence and evolutionary algorithm. Artificial bee colony (ABC) was proposed as an alternative solution to overcome the weak point caused by back propagation (BP) algorithm, and to improve the learning algorithm of multilayer perceptron (MLP) neural network. An obvious improvement of the average accuracy rate was achieved by2%based on ABC-MLPNN. In addition, genetic algorithm (GA) was recruited to solve the problem of selecting the optimal parameters of support vector machine (LibSVM). Choosing parameters (c:constant, g:gamma) plays a major role in determination of the performance of the selected classifier. The experimental results reveal that our proposed pattern recognition method GA-LibSVM performs better (with5%increment) than the existing classical classification algorithm SVM.