The estimation of Knee Joint angle based on Generalized Regression Neural Network (GRNN)

Abstract

Capturing of the intended action of the patient and provide assistance as needed is required in the robotic rehabilitation device. The intended action data that can be extracted from surface Electromyography (sEMG) signal may include the intended posture, intended torque, intended knee joint angle and intended desired impedance of the patient. Utilizing such data to drive robotic assistive device like exoskeleton requires a multilayer control mechanism to achieve a smooth Human Machine Interaction force. It is very important that the controller for gait assistive device is able to extract as many information as possible from the patient muscle with impaired limb and predict different parameters associated with gait cycle. Joint kinematics and dynamics are important to be estimated as the Gait cycle of lower limb consists of flexion and extension postures at knee, hip and ankle joints respectively. This paper proposes a new classification and estimation technique of the lower limb joint kinematics and dynamics based on sEMG signal to predict specifically knee joint flexion and extension postures as well as Knee Joint angles of two postures. In the technique proposed, the feature data of raw sEMG data have been filtered with a second order digital filter and then input to train the Neural Network (NN) and to Generalized Regression Neural Network (GRNN) model to estimate the angle of flexion and extension. The GRNN and NN have been tested with RMS, LOG, MAV, IAV, Hjorth, VAR and MSWT features. GRNN with Multi scale Wavelet Transform (MSWT) feature has ensured 1.5704 Mean Square Error which is very promising accuracy. The SVM has been used to predict postures (flexion and extension). The SVM also has classified flexion and extension with accuracy over 95%.