ElectroMyoGram Pattern Recognition for Real-Time Control of Upper Limb Prosthesis

The paper's goal is to utilize correctly classified surface EMG (sElectroMyoGram) signals to assist movement of upper limb prosthesis. The work involves four steps; first step consists of acquiring 260 EMG signals from the upper arm muscles (biceps, triceps and deltoid) for three different movements (elbow flexion, elbow extension, shoulder abduction). Next step involves extracting nine time-domain features from sEMG. Third step is the classification of EMG signals according to the extracted features. Twenty one classification methods have been implemented and compared. The accuracy of classification is the major criterion of comparison. The most accurate classification algorithm is then used to determine the contribution percentage of every feature. The Lazy/Locally weighted learning (LWL) method is found to be the most accurate among the compared algorithms (94.23%-97.11%). It indicates that the most contributing features are wave length (biceps), variance (deltoid), root mean square (deltoid), wave length (deltoid) and mean absolute value (deltoid). In the final step, the movement classes found out by the Lazy/LWL method are used as commands to control a real time designed virtual upper limb movement. LWL is a local approximation method that has many advantages compared to currently used methods for EMG classification. It is fast, non-parametric and efficient.