Data driven modeling and MPC Based control for Pathological Tremors

Pathological tremor is a common neuromuscular disorder that significantly affects the quality of life for patients worldwide. With recent developments in robotics, rehabilitation exoskeletons serve as one of the solutions to alleviate these tremors. For improved performance of such devices, we need to solve a few problems, which include developing a model for pathological tremors, and a safe control system that can conveniently incorporate constraints on the wrist's range of motion and it's input force/torque. Accurate predictive modeling of tremor signals can be used to provide alleviation from these tremors via various currently available solutions like adaptive deep brain stimulation, electrical stimulation and rehabilitation orthoses. Existing methods are either too general or too simplistic to accurately predict these tremors in the long term, motivating us to explore better modeling of tremors for long-term predictions and analysis. We explore towards the prediction of tremors using artificial neural networks using EMG signals, leveraging the 20- 100 ms of Electromechanical Delay. The kinematics and EMG data of a publicly available Parkinson's tremor dataset is first analyzed, which confirms that the underlying EMGs have similar frequency composition as the actual tremor. 2 hybrid CNN-LSTM based deep learning architectures are then proposed to predict the tremor kinematics ahead of time using EMG signals and tremor kinematics history, and the results are compared with baseline models. This is then further extended by adding constraints-based losses in an attempt to further improve the predictions. Then, we explore the application of model-based predictive control (MPC) for the full wrist exoskeleton designed in our lab for the alleviation of tremors. The main motivation for using MPC here relies on its ability to incorporate state and input constraints, which are crucial for the user's safety. We employ a linear MPC methodology, in which the forearm-exoskeleton model is successively linearized at each time sample to obtain a linear state space model, which is then used to obtain the optimal input by minimizing a convex quadratic cost function. This is then integrated with the tremor model developed via BMFLC and neural networks to provide tremor suppression. Simulation studies are provided to demonstrate the effectiveness of the control schemes. The numerical simulations suggest that the MPC framework is capable of accurate trajectory tracking while providing better tremor suppression than a PD controller without using any tremor model, while the neural network model outperforms the frequency-based BMFLC model. The findings could set up for devising physics-based Neural networks for pathological tremor modeling and experimentally evaluate the performance of the developed framework.