Is this project an undergraduate, graduate, or faculty project?
Undergraduate
Project Type
group
Campus
Daytona Beach
Authors' Class Standing
Katherine Clark Francis Genco Hope Lea, Senior Carolyn Ascha Richardson Tobiah Rosser
Lead Presenter's Name
Hope Lea
Lead Presenter's College
DB College of Engineering
Faculty Mentor Name
Dr. Fan Yang
Abstract
A Myoelectric Prosthesis Utilizing A Convolutional Neural Network and Signal Processing Commercially available myoelectric prostheses for transradial amputees utilize electromyography (EMG) sensors to observe electrical signals from superficial muscle contractions and control robotic fingers, yet often have flaws regarding adaptability and dexterity. These devices require users to learn awkward muscle patterns, as the EMG placement and gesture recognition does not consider how humans would normally think about moving their hand. Additionally, the number of available gestures is more limited, as the devices are threshold-based and do not sample enough muscles. This work seeks to address these issues by creating a low-cost myoelectric prosthetic device, utilizing a 1-D ResNet Convolutional Neural Network (CNN). This is a machine learning model which can learn to recognize patterns in the electrical signals generated naturally by the user and translate them into 3-4 hand gestures, thus the device adapts to the user’s muscles rather than vice versa. The CNN was trained offline on labeled segments of EMG signals, which were captured from the forearm as gestures such as open, close, and pinch were repeated, thus building a reference from which the network learned. Preliminary results with the CNN reached a classification accuracy of 97%. Future work includes compressing the CNN and ultimately embedding the neural network on a Raspberry Pi 5 with protocols to obtain real EMG signals and determine servo commands, thereby completing the goal of creating a system which will sense and recognize the user’s intention and control an adaptable, multi-gesture robotic hand.
Did this research project receive funding support (Spark, SURF, Research Abroad, Student Internal Grants, Collaborative, Climbing, or Ignite Grants) from the Office of Undergraduate Research?
No
Included in
A Myoelectric Prosthesis Utilizing A Convolutional Neural Network and Signal Processing
A Myoelectric Prosthesis Utilizing A Convolutional Neural Network and Signal Processing Commercially available myoelectric prostheses for transradial amputees utilize electromyography (EMG) sensors to observe electrical signals from superficial muscle contractions and control robotic fingers, yet often have flaws regarding adaptability and dexterity. These devices require users to learn awkward muscle patterns, as the EMG placement and gesture recognition does not consider how humans would normally think about moving their hand. Additionally, the number of available gestures is more limited, as the devices are threshold-based and do not sample enough muscles. This work seeks to address these issues by creating a low-cost myoelectric prosthetic device, utilizing a 1-D ResNet Convolutional Neural Network (CNN). This is a machine learning model which can learn to recognize patterns in the electrical signals generated naturally by the user and translate them into 3-4 hand gestures, thus the device adapts to the user’s muscles rather than vice versa. The CNN was trained offline on labeled segments of EMG signals, which were captured from the forearm as gestures such as open, close, and pinch were repeated, thus building a reference from which the network learned. Preliminary results with the CNN reached a classification accuracy of 97%. Future work includes compressing the CNN and ultimately embedding the neural network on a Raspberry Pi 5 with protocols to obtain real EMG signals and determine servo commands, thereby completing the goal of creating a system which will sense and recognize the user’s intention and control an adaptable, multi-gesture robotic hand.