Efficient Identification of Multiple-Muscle Functional Electrical Stimulation Systems

This thesis develops a method to identify the dynamics of a human arm controlled by functional electrical stimulation and uses the identified model of the arm to control force and motion of the hand. Functional electrical stimulation is a means to restore basic daily functions that require reaching to people with high spinal cord injuries. Previous FES controllers focus on fixed stimulation patterns or on single-joint movements which do not provide the flexibility to achieve arbitrary reaching tasks. The model developed in this thesis accounts for coupling of joints of the arm as well as the kinematic and muscular redundancy that make the human musculoskeletal system flexible to different tasks. The model also allows the use of centralized control strategies which are well-studied in robotics.;The model identification technique involves stimulating muscles at different configurations of the arm while measuring the joint configuration and velocity along with interaction forces at the hand. A model for static force output at a subject's wrist was identified and used for control. Over a large space of 3D endpoint forces the controller could predict a force at the wrist produced by stimulation of the muscles within 11% of the maximum force produced by muscle stimulation. A model for the shoulder and elbow torques produced by muscles while the hand moves along smooth reaching trajectories was identified and used for control. The model's predictions of shoulder and elbow torques were on average less than 20% of the maximum torque produced by muscle stimulation. The model was used for a demonstration of motion control of the hand.;Finally, the model's ability to predict torque for new areas of the model's input space was assessed. A purely black-box model will not make accurate torque predictions when presented with inputs unlike those used for training the model. A semiparametric model that incorporates knowledge of the arm dynamics into a Gaussian process model has much smaller expected errors in shoulder and elbow torque predictions than the purely black box model for inputs unlike those used for training the model.