Date of Award

Fall 2023

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Biomedical Engineering

First Advisor

Scott Beardsley

Abstract

This thesis aimed to create a predictive model for controlling lower limb prostheses, using electrical activity from muscles in the residual limb to predict ankle angle and moment across mobility tasks including level ground walking, stair ascent, and descent (LW, SA, and SD). In aim 1, kinematic and kinetic ankle data, were used together with electromyographic (EMG) signals obtained from six healthy ambulators to train a nonlinear autoregressive (NARX) model to predict ankle angle and moment. Networks were trained with a delay of 58.33 ms, using a 120 ms sampling window and tested for their ability to predict ankle angle and moment across mobility tasks using a 10-fold cross validation. The average root-mean-squared error (RMSE) of networks across tasks was 2.93°±0.70° and 0.12±0.03 Nm/Kg when shank velocity was provided as input, 4.92◦±1.39° and 0.15±0.04 Nm/Kg when EMG activity was provided as input, and 2.37°±0.58° and 0.084±0.026 Nm/Kg when inputs were combined. In Aim 2, the best performing networks were tested in a modified version of a model of the Marquette powered prosthetic foot. Three cases were considered: open-loop NARX prediction with internal feedback and closed-loop NARX prediction that used either ankle angle feedback, or ankle angle and moment feedback from the prosthetic model. The internal feedback model accurately predicted its target without incorporating information from prosthetic foot performance. Simulations that incorporated modelled moment feedback increased error, causing instability over time in some SA and SD trials from a mismatch in desired and achieved ankle moment. The NARX networks in Aim 1 accurately predicted ankle angle and moment across subjects and mobility tasks and could be used to provide kinematic/kinetic control input for a powered prosthesis. In Aim 2, the subsequent control simulation performed well with internal feedback, but full closed-loop validation requires an updated model that incorporates loading and ground reaction forces. The system's potential for real-time operation without conscious input is promising. Future work should explore human-in-the-loop control performance and testing with amputee data, to characterize the impact of individual sources of variability associated with changes in residual muscle signals, adaptation over time, and diverse movement patterns.

COinS