Application of Newton/GMRES Method to Nonlinear Model Predictive Control of Functional Electrical Stimulation

Naser Mehrabi, Sadegh Tajeddin, Nasser L. Azad, John McPheeProceedings of the International Conference of Control, Dynamic Systems, and Robotics 2016 · DOI

Abstract

Recent studies have shown that functional electrical stimulation (FES) therapy can improve the motor recovery and range of the motion of stroke patients. The state-of-the-art functional electrical stimulators are open-loop control systems, i.e., the controller is unaware of the patient's posture and progress during the therapy. In this research, we have developed a closed-loop Newton/GMRES nonlinear model predictive controller that measures the patient posture and controls the muscle stimulation to reach a desired posture.

Introduction

Functional Electrical Stimulation (FES) is a critical rehabilitative technique that utilizes electrical currents to activate nerves, thereby restoring movement in patients suffering from stroke or spinal cord injuries. Despite its potential, most contemporary FES devices operate as "open-loop" systems. These devices deliver pre-set pulse sequences without feedback regarding the patient's actual limb position or muscle state, which limits their effectiveness in dynamic rehabilitation scenarios.

Developing effective closed-loop (feedback) controllers for FES presents several significant challenges. First, human musculoskeletal systems exhibit highly nonlinear dynamics that standard linear controllers struggle to manage. Second, while Nonlinear Model Predictive Control (NMPC) offers high accuracy by incorporating a "look-ahead" predictive capability, it is traditionally computationally prohibitive for real-time application in complex biological systems. Finally, many existing controllers fail to explicitly minimize electrical effort, leading to rapid muscle fatigue—a primary obstacle in prolonged FES therapy.

Methodology

To address these challenges, this research implements a fast optimization algorithm—the Newton/Generalized Minimal Residual (GMRES) method—to make NMPC viable for real-time FES. The system architecture is designed to measure patient posture and adjust muscle stimulation dynamically.

System architecture of the closed-loop FES controller, showing the integration of the controller, stimulator, and goniometer for joint angle feedback.

The researchers utilized a dual-model framework for development and testing:

  • High-Fidelity Model: Developed in MapleSim, this model incorporates complex muscle wrapping geometries and a modified Hill-type muscle model to simulate realistic knee extension.
  • Control-Oriented Model: A simplified version of the dynamics used within the NMPC algorithm to reduce computational load, approximating muscle parameters with 4th-order polynomials.
  • Biomechanical Parameters: Derived from the Ferrarin and Pedotti knee model, identified from clinical data of subjects with complete spinal cord injuries.

The core innovation lies in the Newton/GMRES algorithm. Unlike standard iterative solvers, this approach uses a forward-difference GMRES algorithm to determine the "Newton step" for the optimality problem. This allows the controller to solve the underlying differential equations only once per time step, drastically reducing computation time. The controller minimizes a cost function ($J$) that balances tracking accuracy (knee angle vs. target) and electrical effort (square of muscle activation, $a^2$).

Results and Discussion

The implementation of the Newton/GMRES NMPC yielded several key advantages over traditional control methods like PID.

Predictive Advantage and Effort Minimization: Because the NMPC incorporates the physics of the leg through its control-oriented model, it can anticipate gravitational and damping torques before they result in tracking errors. Furthermore, by mathematically minimizing the stimulation required for movement, the NMPC naturally mitigates muscle fatigue, potentially extending the duration of rehabilitation sessions.

Knee angle tracking performance comparing NMPC with varying tracking-to-effort ratios (alpha) against a traditional PID controller and the reference trajectory.

Performance Metrics: * Superior Tracking: As shown in the results, increasing the tracking-to-effort ratio (alpha) significantly improved the accuracy of knee extension compared to PID control. * Lower Current Consumption: The NMPC required less total electrical current to follow a trajectory than a manually tuned PID controller, confirming its efficiency in minimizing muscle activation. * Real-Time Feasibility: The Newton/GMRES method allowed the NMPC to maintain high performance with a sampling time of 50 ms, demonstrating its suitability for real-time clinical hardware.

Conclusion

This study demonstrates that a Newton/GMRES-based NMPC can provide efficient, real-time, closed-loop control for FES applications. While the current work focused on a 1-Degree of Freedom (DOF) knee model with a single muscle, it establishes a foundation for more complex systems. Future research will focus on scaling this approach to multi-joint models, such as a 2-DOF arm, where the computational efficiency of the Newton/GMRES method will be even more vital for simulating complex reaching movements in FES therapy.

Summary updated

Abstract

Recent studies have shown that functional electrical stimulation (FES) therapy can improve the motor recovery and range of the motion of stroke patients. The state-of-the-art functional electrical stimulators are open-loop control systems, i.e., the controller is unaware of the patient's posture and progress during the therapy. In this research, we have developed a closed-loop Newton/GMRES nonlinear model predictive controller that measures the patient posture and controls the muscle stimulation to reach a desired posture.
BibTeX
@inproceedings{mehrabi2016application,
  author = {Naser Mehrabi and Sadegh Tajeddin and Nasser L. Azad and John McPhee},
  title = {{Application of Newton/GMRES Method to Nonlinear Model Predictive Control of Functional Electrical Stimulation}},
  booktitle = {Proceedings of the International Conference of Control, Dynamic Systems, and Robotics},
  year = {2016},
  doi = {10.11159/cdsr16.121},
}
Figure