| Popis: |
A Thesis Submitted to the Faculty of Graduate Studies and Research In Partial Fulfillment of the Requirements for the Degree of Master of Applied Science in Industrial Systems Engineering, University of Regina. xiv, 96 p. In this thesis, a novel Deep Learning approach to classify the Surface Electromyography (SEMG) signals for hand movement recognition is presented and compared to the other approaches in the literature. EMG or muscle’s cells electrical activity are the electrical signals that are carried from the brain to the muscles through the spinal cord. These signals can be recorded and used to measure the activity of muscles. The concept is to measure the signals of different movements for healthy individuals, classify and train the signals, and use the trained model to predict the movements of individuals suffering from disability. The classifier used in this thesis is a Time Wrapped CNN+LSTM. The contribution of this research is to predict the hand movements (hand gestures) with a very high accuracy in order to enhance the efficiency of mechanical prosthetic hands and mimic the natural hand movement. To ensure the highest performance, the dataset is pre-processed using Smoothing (noise reduction), Filling Outliers (dimension reduction), Detrending (feature selection) and Normalizing. As a result, the novel method is called SFDN+DNN. There are many complex and developed prostheses in the market. However, the bottleneck to improve the capabilities of the prostheses to a quasi-real-hand level is still a big challenge. In this thesis a thorough literature review of advancements in EMG signal classification and the history of prosthetic hands are presented and some significant achievements are highlighted. Secondly the novel method to classify the SEMG signals is presented and evaluated using three benchmark hand SEMG datasets in the literature, which are the UCI Repository SEMG dataset, the Ninapro dataset and the Mendeley dataset. The datasets are also tested using some conventional machine learning algorithms such as KNN, SVM, Decision Tree, Random Forest, Naïve Bayes, Multi-nominal Logistic Regression, XGBoost and Multi-Layer-Perceptron, to illustrate the power of classification of our proposed method compared to these algorithms. Finally, the classification accuracy of the proposed method is compared to some of the most up-to-date techniques. The obtained results indicates that the DNN classifier can achieve a higher classification accuracy compared to the state-of-the-art methods in the literature. The classifier was able to reach the accuracies of 99.63%, 95.56% and 96.14%, for the UCI SEMG dataset, the NinaPro DB6 dataset and the Mendeley SEMG dataset respectively. While the best claimed accuracy for these datasets in the literature, to the best knowledge of the author, was 99.49%, 93.7% and 89.5%, for the UCI SEMG dataset, the NinaPro DB6 dataset and the Mendeley SEMG dataset respectively. The contributions of this research are as follows: Inferring the hand movements as sequential, and time series classification. Low computation cost and the ability for learning online (end-to-end system), which is a significant advantage compared to the offline conventional methods, and substantially, reaching one of the highest classification accuracies in the literature, with acceptable results for designing real prosthetic hands. Student yes |
