The purpose of muscle activity onset detection is to distinguish occurrence of active muscle activity from baseline. The accurate muscle activity onset detection is an indispensible component of the surface electromyogram (EMG) systems.
Various techniques for the determination of onset time of muscle activity have been proposed, and most techniques utilize the waveform amplitude as the feature for muscle activity onset detection.
However, one drawback of amplitude based features is that they are very sensitive to background noise level changes. The performance of these amplitude-based methods degrades as the signal to noise ratio of the processed signal decreases. To overcome this difficulty, several methods have been proposed to improve the performance of muscle activity onset detection, particularly when the signal to noise ratio of the surface EMG is low.
Recently, researchers from USA and China propose a novel and robust muscle activity onset detection technique based on an unsupervised EMG learning framework, proving to be capable of being adaptive to a dynamic environment and also applicable to real-time detection.
In their research, the unsupervised learning framework was developed based on a sequential Gaussian mixture model. The sequential Gaussian mixture model was employed to discriminate between burst and non-burst distributions at each Mel-spaced frequency band, using the energy distribution as feature parameter.
The Mel-frequency warping method as used in this research emphasized low frequency information of EMG signals. Compared with those based on equal-partition bands, the method based on Mel-frequency bands had higher resolution in lower frequencies. Mel-frequency cepstral coefficients were employed in previous research for developing EMG-based speech recognition systems.
In the current application, Gaussian mixture model concerned not only the posterior distribution, but also the a priori distribution. The posterior probability, a priori probability and constraints to distributions were incorporated into an unsupervised learning framework.
The unsupervised learning framework used in this research considered a priori distributions of EMG burst and non-burst signals, as well as the distribution relationships between the signals. Compared with conventional statistical models, it had advantages in both the initialization process and the sequential process. The proposed model could be correctly initialized regardless of the presence or absence of EMG burst in the beginning of the process. In the updating process, the statistical models were updated in a soft manner, controlled by the presence probability of the EMG burst. And the sequential Gaussian mixture model based muscle activity onset detection not only provided EMG presence information in the frame level, but also more detailed EMG presence probability in each frequency component. Also, the non-burst information in the EMG frames could be employed to update models.
Compared with previously developed methods, the proposed method was characterized by several features. The method not only worked well in low signal to noise ratios, but also was adaptive to changing signal to noise ratios, thus it was capable of robust muscle activity onset detection in a dynamic environment. Besides, the proposed algorithm was able to run in an online manner, so the muscle activity onset detection could be applied to real-time systems. Moreover, the method could also correctly estimate the muscle activity onset without the assumption of non EMG burst in the initial state. Thus, this muscle activity onset detection scheme was more practical than conventional ones.
The proposed method was evaluated by simulated and real surface EMG signals, and the experimental results confirmed its superiority.
This research was supported by the National Institutes of Health under Grants R24HD050821 and R01NS080839, the Memorial Hermann Foundation, the National Natural Science Foundation of China under Grants 61271426 and 81271658, and the Strategic Priority Research Program of the Chinese Academy of Sciences under Grant XDA06030100.
References:
LIU Jie, YING Dongwen, William Z. Rymer, ZHOU Ping. Robust Muscle Activity Onset Detection Using an Unsupervised Electromyogram Learning Framework. PLoS ONE (Vol. 10, No.6: e0127990, June 3, 2015). DOI: 10.1371/journal.pone.0127990