Electromyography (EMG) Signal Classification by Artificial Neural Networks

EMG signal processing is one of the active fields of biomedical signal processing. One unanswered question is how to determine whether a muscle is fatigued by analyzing the EMG signal. Fatigue detection could be useful in several different practical situations. There are several studies which show there are differences between EMG signal features before fatigue and after fatigue. Generally studies are based on an analytical analysis of the EMG signal instead of a quantitative analysis. In all previous studies in EMG signal processing for fatigue/nonfatigue detection, the result is that there exist some differences between the EMG signal before muscle fatigue and after fatigue. Unfortunately there are no comprehensive studies to quantify these changes. In our study, we concentrated on analyzing the effect of fatigue on muscles to see which set of features of the EMG signal are appropriate for fatigue detection. A MatLab code is developed which generates a set of 15 different features of an EMG signal. Some of these features are features mentioned in the other studies and some are originally designed during our study. After generating the features a multilayer perceptron artificial neural network is used to classify features into two classes (fatigue/nonfatigue). The result of study is that correct classification rate with a small NN (with 5 to 7 nodes) for 1.5 seconds of EMG signal is better than 95%. The results of this study and the achieved classification rate is one of the first studies in this field.