Deep Lagrangian Network Learning and Control of Robotic Exoskeleton Based on Multi-Sensor-Cyber Information Fusion

The overall performance improvement of model-based controller depends on the accurate plant model. However, many complicated plants exist unmodeled uncertainties caused by irregular structure, motion frictions, and external disturbances, which are difficult to obtain the mathematical model with high accuracy. In this work, a multi-sensor-cyber is constructed to sample the physics information about human-exoskeleton interaction and exoskeleton joint motion. Meanwhile, a model identification method based on Deep Lagrangian Network (DeLaN) is presented in robotic exoskeleton to realize multi-sensor information fusion and obtain the reasonable parameters of Lagrangian model. Then a human-exoskeleton cooperative motion control based on nonlinear extended state observer is proposed to guarantee that the exoskeleton tracks two joint demands in the case of tolerable human-exoskeleton interaction. Finally, the effectiveness of the proposed model identification and control scheme is verified by the experimental results in two-DOF exoskeleton platform.