Development of myoelectric interface based on pattern recognition and regression based models

Abstract

This paper proposes a combinatorial strategy for myoelectric control of robotic arm. Activation of main muscles responsible for 1 DOF of elbow joint is recorded. The goal was to create a mapping between muscles' Surface Electromyogram (sEMG) data and kinematics of the joints. The proposed strategy includes two main phases. In the first phase, Linear Discriminant Analysis (LDA) was utilized to classify several classes in user's arm motions. Due to fast training, simple implementation and robustness against long term effect of non-stationary characteristics of sEMG signals, LDA is a common classifier in myoelectric signal classification researches. In the second phase, two Time Delayed Artificial Neural Networks (TDANN) were trained to estimate proportional and continuous angle and velocity related to joint motion classes. Furthermore, two additional methods were used to enhance the prediction results accuracy. First, noise reduction of sEMG signals plays a key role in accurate joint kinematics prediction. Therefore, a new noise reduction approach is investigated based on classification results. Second, final predicted angles were achieved by data fusion of angles and angle difference rates, estimated by TDANN. Results show that, LDA classifies the motion classes with 95% accuracy and final estimated angular positions are significantly close to actual values. Therefore, proposed method is able to create a mapping between muscles' sEMG data and joint kinematics with acceptable error. Practical results confirm the performance of the proposed method.