Predicting the time to localized muscle fatigue using ANN and evolved sEMG feature

Abstract

Surface Electromyography (sEMG) activity of the biceps muscle was recorded from nine subjects. Data were recorded while subjects performed dynamic contraction until fatigue. The signals were initially segmented into two parts (Non-Fatigue and Transition-to-Fatigue) to enable the evolutionary process. A novel feature was evolved by selecting then using a combination of the eleven sEMG muscle fatigue features and six mathematical operators. The evolutionary program used the DB index in its fitness function to derive the best feature that best separate the two segments (Non-Fatigue and Transition-to-Fatigue), for both Maximum Dynamic Strength (MDS) percentage of 40 and 70 MDS. Using the evolved feature we enabled an ANN to predict the time to fatigue by using only twenty percent of the total sEMG signal with an average prediction error of 9.22%.