Affiliations: Department of Mechanical Engineering, Texas A&M University, College Station, TX, United States
Correspondence:
[*]
Corresponding author. Kuang Nie, Department of Mechanical Engineering, Texas A&M University, College Station, TX, United States. E-mail: [email protected].
Abstract: Surface electromyography (sEMG) signals have great potential for predicting upper limb motion. Although prior investigations have explored diverse applications of sEMG signal analysis, but few studies have focused on real-time motion prediction within the context of upper limb configuration space. Additionally, previous research has not adequately considered individual variability in sEMG features. This study aims to accomplish two main objectives. Firstly, it seeks to examine the dissimilarities in signal distribution across different subjects when employing various features. Additionally, the study aims to establish a correlation between signal distribution patterns and the model’s predictive accuracy. Secondly, the study introduced a personalized standardization (PSD) technique, which will serve to normalize the shape of the signal distribution across different subjects, thereby addressing the inter-individual differences in sEMG features. A bi-directional long short-term memory (Bi-LSTM) network is employed to estimate the real-time moving intention of the upper limb after applying the PSD technique. The analysis of signal distribution involved nine combinations of features, encompassing six features, namely mean absolute value (MAV), wave length (WL), variance (VAR), root mean square (RMS), mean frequency (MNF) and median frequency (MDF). To assess predictive capabilities, several models were evaluated. Remarkably, the distribution analysis clearly demonstrated that the shape of the signal distribution notably influences the model’s performance. Accroding to results, the incorporation of the PSD technique resulted in a notable improvement in the accuracy of the Bi-LSTM model, which leds to an enhancement of up to 2.8 percentage points in predictive accuracy. Additionally, the Bi-LSTM model emerged as the highest-performing model among all the compared models during the analysis. These findings underscore the importance of considering individual variability in sEMG features when developing predictive models for upper limb motion and highlight the potential benefits of employing the PSD technique to enhance model performance.
Keywords: Surface electromyography, Real-time motion prediction, Deep Learning, Signal pre-processing
