EMG-based Feature Engineering Approach for Load Classification

We demonstrate a machine learning-based Human Activity Recognition (HAR) model in terms of load classification. A self-collected dataset using surface electromyography (sEMG) and BIOPAC MP36 was employed for developing the proposed model. The sEMG signals were obtained while 12 volunteer subjects lifted weights of 1 kg, 3 kg, 5 kg, 7 kg, and 10 kg. In the first step, we utilized an IIR elliptic filter to de-noise the collected sEMG signals. Next, a region of interest extraction technique based on automated relative energy empirical mode decomposition (AREEMD) was used to remove the extraneous data. Then, feature extraction was carried out by employing several pairs of statistical, time domain, frequency domain, and cepstral features. A 10-fold cross-validation technique was used to test the validity of the chosen features on a variety of classifiers, including support vector machine (SVM), decision tree (DT), ensemble, neural network (NN), and K-nearest neighbor (KNN). SVM using a cubic kernel outperformed other classifiers in terms of accuracy, providing the highest classification accuracy of 98.5% for five-class load classification.