Abstract:
Electromyography (EMG) is a method for determining how muscles and the nerve cells that control them operate. To contract and relax the muscles, electrical signals are generated by motor neurons. EMG signals may be utilized to create myoelectric control systems for assistive and rehabilitative devices, whether they are recorded invasively or noninvasively. The performance of the proposed myoelectric control is compromised by a variety of parameters, such as: the channel count, electrode location, noise contained in the EMG data, feature selection, and classifier selection. The goal of this research was to investigate the impact of sophisticated signal processing frameworks on the intrinsic characteristics of EMG signals to enhance the signal to noise ratio of the recorded data. Different EMG filters based on Empirical (EMD) and Variational Mode Decomposition (VMD) were developed and tested to denoise EMG signals. The EMG filter was designed using Clear Iterative Interval Thresholding (CIIT), Iterative Interval Thresholding (IIT), and Interval Thresholding (IT) methods, as well as SOFT, HARD, and SCAD operators. The denoised signals were then transferred to Pattern Recognition algorithms based on Convolutional Neural Networks (CNN) and Linear Discriminant Analysis (LDA). Additionally, disjoint and overlap segmentation methods were used to assess the effectiveness of the optimal windowing configurations. For both surface and intramuscular EMG data, statistical analysis revealed Iterative Interval Thresholding with VMD produces the greatest SNR despite the amount of noise in the signal. Whereas EMD-based filters do not retain the intrinsic properties of intramuscular EMG signals. For both LDA and CNN, statistical analysis revealed that the optimal segment size for disjoint windowing is between 250ms - 300ms. For overlap segmentation, the optimal time range for LDA is 250ms-300ms and for CNN is 275ms-300ms. The settings suggested here may be utilized to create a strong and reliable MEC.
