Date of Award

2026

Type

Thesis

Degree Type

Robotics Engineering Bachelor of Science

Department

Earth & Space Science

First Advisor

Dr. Mohammad Jafari

Second Advisor

Dr. Abiye Seifu

Third Advisor

Dr. Kimberly Shaw

Abstract

This thesis develops and evaluates a deep learning-based prediction model capable of identifying intended limb movement from surfaced electromyography (sEMG) signals using sequence learning techniques. sEMG signals change over time due to multiple factors such as muscle fatigue or user variability. Traditional prosthetics control methods rely on static feature extraction, ignoring how signals change over time, thereby limiting their ability to capture the temporal changes of muscle activity. As a result, these approaches often lead to poor accuracy, robustness, and generalization. Limited experimental validation has been conducted on sequence-based machine learning approaches using temporal sEMG data from publicly available datasets such as Ninapro DB5, to improve prosthetic movement prediction accuracy.

To address these challenges, a Long Short-Term Memory (LSTM) based deep learning model was implemented to learn sequential patterns in time-series sEMG data from three selected exercises in the Ninapro DB5 dataset, and to predict intended limb movement. The Ninapro DB5 dataset provides captured sEMG signals for muscle activity along with their corresponding label for each movement. The model was trained and evaluated across the three exercises, each one with a different level of complexity, enabling performance evaluation under varying conditions. Model predictions were compared against ground truth labels using classification accuracy and confusion matrix analysis.

Experimental results demonstrate that the LSTM model effectively learns sequential patterns from the sEMG data with high accuracy. All exercises achieved at least 79.82% validation accuracy, and a F1-score of 79.70%. Confusion matrix analysis shows that misclas- sifications happen between gestures with similar muscle activation patterns. This suggests that further improved predictions for subtle changes in muscle activity could be achieved with more sophisticated techniques, such as CNN+LSTM. Additionally, this machine learn- ing approach enables the incorporation of multimodal data by training the model on a variety of sensory information alongside sEMG data. These findings confirm that temporal modeling can improve sEMG signal interpretation compared to traditional approaches. Overall, this thesis highlights the potential of an LSTM-based model to predict intended movements of biological systems reliably, supporting the potential of a prosthetic limb control system that feels natural to the user and improves their health both physically and emotionally.

Download

Share

COinS