EMG-based estimation of shoulder kinematic using neural network and quadratic discriminant analysis

This paper proposes a two phase strategy in myoelectric control of upper limb which benefits from strength of two common myoelectric interface methods. The aim of this research is to create a mapping between electrical activities of muscles known as electromyogram (EMG) signals and kinematics of corresponding joint. Activations of main muscles responsible for 3 DOFs of shoulder were recorded. According to anatomical surveys, 4 classes of motion are defined in shoulder joint. First phase concerns with motion classification by quadratic discriminant analysis (QDA). In fact, classifier plays the role of high level controller that decides between several networks to estimate corresponding motion class angle. Also, majority voting (MV) is used as post processing algorithm to improve classification results. The second phase concerns with continuous angle estimation of different motion classes in shoulder joint. Four different time-delayed artificial neural networks (TDANN) were trained using data of corresponding motion class. Classification results indicated relative superiority of QDA over Linear Discriminant Analysis (LDA) to achieve higher overall accuracy of 98.3% against 93.7%. Also, TDANN estimation showed the capability of this model to predict shoulder joint angles, with high accuracy.