Date of Award


Document Type

Thesis - SCU Access Only


Santa Clara : Santa Clara University, 2019


Bioengineering; Computer Science and Engineering; Electrical and Computer Engineering

First Advisor

Daniel Lewis

Second Advisor

Sally Wood

Third Advisor

Yuling Yan


There are many people throughout the world who have recently lost an arm and are struggling to learn how to do simple tasks without it. A typical solution to this is to use a prosthetic arm but there is a trade-off between functionality and cost. The most functional options involve expensive, invasive procedures. One solution is to reroute nerves from other places in the body, and another method is to build a socket into the arm, which allows the prosthetic to connect directly to nerves still existing in the arm. With these solutions, a person can simply think about what they want to do, and the prosthetic will do it. Our goal was to improve the existing solutions. By processing electromyography (EMG) signals through a neural network and controlling the prosthetic arm with the output, we hoped to enable a noninvasive prosthetic to become as intuitive and nearly as functional as the more expensive alternative.

A common type of noninvasive prosthetic is a myoelectric prosthetic, which is controlled using EMG signals. However, the way they are controlled is drastically different than the way anatomically intact arms are controlled. They use indirect motions, such as moving a body part to control the prosthetic. This makes them very complicated to use, especially for people who are still coping with the loss of an arm and trying to control the prosthetic in the same way. To make the transition to a prosthetic as easy as possible and to ensure that the prosthetic is intuitive to use, our control system uses EMG signals collected from the shoulder above the missing limb and natural movements.

Our control system focused on four motions: opening/closing a hand and extending/contracting an arm. The goal was to accurately categorize the motion in each sample at least 80% of the time. The first step was to collect data. We collected 250 samples of each motion, for a total of 1000 samples. Within each five second EMG recording, the motion began at the two second mark and lasted one full second. Once the data was collected, it was separated into good and bad data, based on the number of heartbeats that occurred. The heartbeat signal was significant compared to the rest of the EMG, and it had to be removed. This was the used to train and test a neural network. The network used, a Long Short-Term Memory Recurrent Neural Network (LSTM RNN), was designed specifically for data that changes over time. Most of the data was used to train the data and teach it what characteristics were associated with each motion, and then tested the remaining 20% of the data to find the final accuracy.

Overall, our project succeeded in many ways. Our control system costs slightly more than a quarter than most myoelectric prostheses and with an inexpensive arm, that cost would not be raised much. This arm is controlled with entirely natural movements, which is unlike most myoelectric prostheses and that would make it intuitive. The four motions were included to get successful results. The training accuracy, at 96%, is well over our goal of 80%. The testing accuracy, made smaller by the limiting data set, is at 78.4% and just shy of 80%. With the combination of those two scores, we consider our control system to have met all of our objectives for a control system of a prosthetic arm.

SCU Access Only

To access this paper, please log into or create an account in Scholar Commons using your email address.