Feature Extraction Using Wavelet Scattering Transform Coefficients for EMG Pattern Classification

Abstract

AbstractThe Electromyogram (EMG) signal collected from the human muscles has been utilised for a long time to aid in diagnosing several medical conditions and for the control of external devices, including powered exoskeletons and prosthetic devices. However, there are still many challenges in analysing this signal to translate the findings into clinical and engineering applications. One of the significant challenges is the knowledge extraction part, as represented by the feature extraction stage, which is considered a vital factor in attaining the ultimate performance in EMG-driven systems. Wavelet transforms analysis is one of the several methods utilised for feature extraction with biomedical signals in the time-frequency domain (TFD). Wavelet analysis-based feature extraction methods can be primarily categorised into three categories: wavelet transform (WT), wavelet packet transform (WPT), and the recently proposed deep wavelet scattering transform (WST). While many researchers utilised the first two methods to extract features from the EMG and other biomedical signals, the WST has not been appropriately investigated for feature extraction with EMG pattern recognition. This paper examines the potential benefits associated with the use of deep WST as a feature extraction method for the EMG signal and compares it with other wavelet methods. We used three well-known different EMG datasets collected with laboratory and wearable armbands hardware to provide a comprehensive performance evaluation under different settings. The new method demonstrates significant improvements in the myoelectric pattern recognition performance compared to WT and WPT, with accuracy reaching up to 96%.KeywordsElectromyogramWavelet scattering transformFeature extraction