Estimation Of Motion Intention And Torque Of Human Elbow Joint In Physical Human Robot Interaction

Currently, the topic of human-machine systems and physiologically-based human-machine interfaces is receiving a large amount of attention. For the development of advanced Human-Robot Interaction (HRI) systems which are able to measure or predict the intention of humans' movement and therefore adapt their behaviors accordingly is an essential and challenging problem.Surface electromyographic (sEMG) signal acquisition offers a noninvasive method for operating a robotic arm based on human's volition or intention. There have been various research attempts using humans' EMG signals to control the motion of machines such as rehabilitation robotic systems.In present dissertation, experimental studies of our investigation to extract human's elbow joint motion intention from sEMG measurement were carried out. EMG activity was recorded from six superficial elbow and shoulder muscles while a human subject's limb is kept still on a horizontal table but subject to certain force applied to the hand, by changing the direction of the force applied. In each experiment a cable was attached to the subject's wrist, and a constant force magnitude was applied in various directions with the use of a pulley system. Based on the measurement obtained, we study the relation between static sEMG signals and movement trends of the elbow in eight directions using cosine tuning functions. Genetic Algorithm (GA) is applied to determine multivariate nonlinear regression coefficient to improve the prediction accuracy. We find three-cosine functions often provided the best fit to the EMG data.There are nonlinear nature of relationships exists in exskeleton muscle at different force level and model can not be used in scale simply. Thus, the force levels should be classified before we use model to extract the motion direction. Neural network is illustrated to investigate and classify different level due to its capability to predict data which has nonlinear character such as EMG signals. In present dissertation, a four-layer, fully connected, feed forward artificial neural network (ANN) with tan-sigmoid transfer function in the hidden layers and a linear transfer function in the output layer is selected to distinguish the different force levels of elbow joint when subjects are holding static load at different directions. The input vectors that have been normalized in advance are divided into two groups: the first group is the AR coefficients when p=4, the second group is the advisable features values from EMG signals extracted by wave packet transform (WPT). The output nodes are five, corresponding to the five force levels. The predictive ability is quite satisfying (above 86.7%) when force is above 1 Kg.All these results provide rich information for understanding the mechanism of extraction of intention from sEMG and some correlative dynamical processes.This work has been supported in part through the National Science Foundation of China (CNSF Grant No. 60505012,60674060).