Date of Award

6-27-2022

Document Type

Thesis

Degree Name

Master of Science in Mechanical Engineering (MSME)

Department

Mechanical Engineering

First Advisor

Joel Canino

Second Advisor

Cheuk Ming Lui

Third Advisor

Rebecca Bercich

Abstract

According to the Center for Disease Control (CDC), 1 in 7 American adults are affected by disabilities that affect mobility [1]. Assistive devices, such as exoskeletons, may be able to assist affected individuals. Though these devices may be helpful, constant assistance can create dependence on the device. We hypothesize that an assist-as-needed control system can be developed to aid the wearer only when necessary. To develop an assist-as-needed control system, it needs to be determined when assistance is needed and the magnitude of this assistance. We hypothesize that an artificial neural network (ANN) can predict future normal torque outputs of joints given past joint angles, torques, and sEMG data of the governing muscles. Actual torque outputs can then be extrapolated from sensor data and the difference in these torque values can be determined to be the required assistance magnitude. To evaluate this method, a dataset of 10 subjects walking at 7 speeds was ascertained [3]. This dataset included pelvis, hip, knee, and ankle torque and angle about the 3 major axes across 3 strides at each walking speed. The surface electromyography data (sEMG) of the Tibialis Anterior (TA), Gastrocnemius Lateralis (GAL), Biceps Femoris (BF) and Vastus Lateralis (VL) were also collected for each subject during each stride. The first step in this investigation was to evaluate this method on a single and multi-joint system while considering the sEMG activity of the pertinent muscles. A multi-joint system could then be evaluated when not all governing sEMG data can be measured. Finally, a principal component analysis could be performed on this multi-joint system to evaluate the contribution of each feature in the training outcomes of the ANNs. This allowed the researcher to reduce the order of this system while still maintaining prediction accuracy. Features that contribute the most to prediction accuracy could be retained and the others could be removed to simplify the network, reducing training time, testing time, and hardware complexity. These results of this research will be used to inform future development of assist-as-needed devices.

Download

Share

COinS