Date of Award

Winter 1-1-2020

Degree Type

Thesis

Degree Name

Master of Science In Mechanical Engineering Degree

Department

Mechanical Engineering

First Advisor

Bogert, Antonie Van Den

Second Advisor

Dr. Eric Schearer

Third Advisor

Dr. Majid Rashidi

Abstract

Collegiate athletes rely on their muscles to compete in their respective sports. However, one injury requiring extended time out of competition could lead to muscle atrophy. As a result, athletes may learn to compensate for weakened muscle groups with stronger muscle groups; a change that may be almost undetectable. Consequently, compensating can add unnecessary stress to the musculoskeletal system, leading to reinjury. One way to combat this is by measuring muscle force. However, there are currently no methods to directly measure muscle force, so it must be solved for indirectly. This research aims to explore state estimation with trajectory optimization and a convolutional neural network. Both methods will be used to estimate the trajectories of the state variables and muscle force associated with forearm flexion. To serve as an input to both solution methods, artificial data was generated. This data contained measured trajectories for forearm position, angular velocity, muscle fiber length, muscle activation, and muscle force. In addition, the generated data included artificial sensor signals comprised of an electromyography (EMG) and inertial measurement unit (IMU). For testing, different signal to noise ratios were added to the generated sensor data. The trajectory optimization method was tested using different weight ratios. The results from this simulation study confirm that the tuning parameter should be chosen based on the noise levels present within the data. Moreover, this method of state estimation can iv accurately and precisely predict state variable trajectories at all noise levels. However, it struggles to predict muscle force when there is noise added to the data. A similar process was conducted to test the neural network; however, the batch size, was the tuning parameter selected for this method. The results from this portion of the simulation study conclude that the convolutional neural network was able to estimate the state variables precisely and accurately at all noise levels. Moreover, it was not as susceptible to estimation errors when utilizing an intermediate batch size. However, the network was unable to estimate muscle force precisely or accurately when there was noise in the data.

Download
COinS