Knee Joint Torque Prediction with Uncertainties by a Neuromusculoskeletal Solver-informed Gaussian Process Model.

Research interest in exoskeleton assistance strategies that incorporate the user's torque capacity is rapidly growing, yet uncertainty in predicted torque capacity can significantly impact the user-exoskeleton interface safety. In this paper, we estimated knee flexion/extension torques by using a neuromusculoskeletal (NMS) solver-informed Gaussian process (NMS-GP) model with muscle electromyography signals and joint kinematics as model inputs. The NMS-GP model combined the NMS and GP models by integrating valuable features from an NMS solver into a GP model. The NMSGP model was used to predict knee joint torque in daily activities with uncertainty quantification. The activities included slow walking, self-selected speed walking, fast walking, sit-to-stand, and stand-to-sit. Model performance, defined as low prediction error between the model's predicted torque and measured torques from inverse dynamics computations, of both the NMS-GP and NMS models was analyzed. We found that prediction error was significantly lower in NMS-GP models than in NMS models. We observed relatively high uncertainties in the predicted knee torque at the beginning of each movement, particularly in self-selected speed walking. High uncertainties were also found during terminal stance and swing in self-selected speed walking. Compared to other torque prediction methods, the proposed NMS-GP model not only provides an accurate joint torque prediction but also a measure of the uncertainty. Our study showed that the NMS-GP model has a large potential in control strategy design for rehabilitation exoskeletons and to enhance the overall user experience.