Fig 1.
Schematic illustration of logarithmic energy distribution of a frequency band.
(a) Distribution of EMG signal contaminated by noise. (b) Distributions of EMG burst and non-burst. The shadow denotes the classification error.
Fig 2.
Process of EMG onset/offset detection using the sequential GMM.
Fig 3.
Simulated EMG bursts and the corresponding detection performance of the unsupervised learning framework.
The detected segments with muscle activity were highlighted by the rectangular envelope built on the basis of the onset/offset estimates provided by the unsupervised learning framework. (a) Simulated clean EMG trial without added noise, which is composed of five EMG bursts with different durations and amplitudes. The corresponding actual onsets and offsets are marked by vertical dashed lines. (b) Simulated EMG signals at SNR level of 2 dB and the muscle activity onset detection performance. (c) Simulated EMG signals at SNR level of 10 dB and the muscle activity onset detection performance. (d) Simulated EMG signals with time-varying SNR levels and the muscle activity onset detection performance (the dashed lines indicate different signal segments for calculating SNRs).
Fig 4.
EMG burst presence probability in the time-frequency domain estimated by the unsupervised learning framework.
(a) Magnitude spectrogram of the simulated EMG signal with time-varying SNR levels. (b) Spectrogram of EMG burst presence probability. The transition from white to black corresponds to probability changing from 0 to 1. (c) Simulated EMG signals with time-varying SNR levels and the muscle activity onset detection performance (the rectangular envelope); the corresponding actual onsets and offsets are marked by vertical dashed lines.
Fig 5.
Comparison of onset detection performance using different methods (mean ± standard error).
AMP: the conventional amplitude thresholding method; TKEO: the method based on TKE operation conditioning; Bonato: the double threshold algorithm developed by Bonato et al. [6]; SGMM: the sequential GMM based unsupervised learning method. For each SNR level, the mean latency was averaged over 60 trials of simulated surface EMG signals.
Fig 6.
Experimental surface EMG onset detection.
Examples of the experimental surface EMG signals with relatively (a) high and (b) low SNRs, and the muscle activity segments identified by the unsupervised learning framework.