Session: 07-12-01 Control Theory and Applications I

Paper Number: 73687

Start Time: Tuesday, 02:00 PM

73687 - Deep Neural Network Real-Time Control of a Motorized Functional Electrical Stimulation Cycle With an Uncertain Time-Varying Electromechanical Delay 

            

Neurological disorders (NDs) include a wide range of conditions including cerebral palsy, stroke, spinal cord injuries, Parkinson’s disease, traumatic brain injuries, and muscular dystrophy. Some of the most prevalent symptoms of these disorders are progressive muscle weakness and loss of voluntary coordinated limb motion. These issues can lead to a sedentary lifestyle, which often cause a wide range of additional adverse secondary health effects such as obesity, high blood pressure, and cardiovascular disease. In an effort to combat muscle atrophy, increase bone density, and improve cardiovascular and mental health, rehabilitative strategies for people with NDs often involve the development of closed-loop functional electrical stimulation (FES) control methods that facilitate motor-assisted stationary cycling. The FES causes the rider to produce a muscle forces that induces positive crank motion while the motor assists the rider in kinematic deadzones and at any other time the rider cannot meet the cadence objective.

In addition to the inherent time-varying, nonlinear, and uncertain dynamics of the switched cycle-rider system, a challenge for this type of control design is an input delay called the electromechanical delay (EMD) that exists between the start of stimulation and the onset of muscle contraction, as well as the end of stimulation and the end of the corresponding muscle contraction. The EMD can cause an otherwise stable system to become unstable. A controller that considers this muscle-induced delay can improve cadence tracking. Furthermore, by decoupling the two sides of the bike, rider asymmetries (i.e., the differences in dynamics between the nondominant and dominant legs) can be evaluated and compensated for, which is especially important to consider for individuals with NDs. On this split-crank cycle, the impaired or non-dominant leg tracks a desired cadence, and the dominant leg tracks a desired position that is offset 180 degrees from the non-dominant leg, so that the cycling is done in phase.

A deep neural network (DNN)-based control architecture can be used to estimate the nonlinear, uncertain, and switched dynamics of each leg of the cycle-rider system. In the developed method, the DNN can approximate the system’s dynamics, which can then be used as a feedforward term in the controller, which, empirically, improves system performance. A DNN update law also updates the output layer weights in real time which can further improve performance. Since the DNN approximation can account for the differences in dynamics between riders, the FES can be applied more conservatively which can delay the start of muscle fatigue. This DNN-based controller will open the door for more personalized treatment between individuals.

Presenting Author: Hannah M. Sweatland University of Florida

Authors:

Hannah M. Sweatland University of Florida
Brendon C. Allen University of Florida
Max L. Greene University of Florida
Warren E. Dixon University of Florida