Impact of Activity Pace and Arm Position on Classification of ADLs

This study focuses on EMAHA-DB7, a new set of multi-channel surface electromyography (sEMG) signals designed to assess daily activities in different conditions. The dataset comprises sEMG signals from 10 subjects engaging in 10 distinct ADLs. Each activity was executed at two different paces and in three different arm positions, ensuring a diverse set of measurement conditions. This diversity was introduced to investigate potential disparities in sEMG activity between rapid and normal contraction speeds during ADL across different upper limb positions. To the classification framework, features were extracted from the time domain, frequency domain, wavelet domain, and Eigenvalues. Subsequently, these features were employed in training and testing with six classical machine learning models. The study revealed that the proposed machine learning framework effectively classifies the ADL categories when trained and tested irrespective of context. However, the model test performance degrades when the model is trained on data from a subset of conditions and tested on left out data.