Evolving Gaussian Process based Learning of Knee Angle and Velocity

Transparent human-exoskeleton interaction requires accurate human joint angle and velocity learning which are regarded as human intent detection to cope with the unspecific and irregular kinematics and dynamics of the system. This paper attempts to address the limitations and deficiencies encountered by traditional methods which make it challengeable to figure out the natural relationships among the strongly coupled multi-source information from each of the human-exoskeleton subsystems. Dependent Gaussian process (DGP) based data fusion algorithm is established and serves as the mathematics foundation to explore the deep layer correlation among the human joint angle, interactive force and processed sEMG achieving satisfactory prediction results of joint angle. Gradient estimation model is then performed to obtain the human joint velocity by differential of a GP model. The statistic nature of the proposed model offers superior flexibility and encouraging human motion prediction results. And the proposed model can achieve human joint angle and velocity learning simultaneously without accessional sensors which may incur marked cost or may be impossible. Experimental works on an in-house exoskeleton which is the key step to verify the superiority of the proposed algorithms are also presented.