CC BY
0APPLICATION OF SEMG SIGNAL IN REHABILITATION TRAINING OF LOWER LIMBS
BECIHHK TOry. 2022. № 2 (65)
UDK 004.5:61
Liye Ren, Chen Wang, Junyi Tian, Jianwei Fang
APPLICATION OF SEMG SIGNAL IN REHABILITATION TRAINING OF LOWER LIMBS
Liye Ren - PhD, Associate Professor, School of Electronic and Information Engineering, Changchun University, Changchun, email: renly@ccu.edu.cn (China); Chen Wang - PhD, Associate Professor, School of Electronic and Information Engineering, Changchun University, Changchun (China); Junyi Tian - PhD, Associate Professor, School of Electronic and Information Engineering, Changchun University, Changchun, (China); Jianwei Fang -PhD, Associate Professor, School of Electronic and Information Engineering, Changchun University, Changchun, (China)
Surface EMG signals have important research value in rehabilitation training and clinical medicine. In this paper, some important links such as EMG signal acquisition and preprocessing, feature extraction, pattern recognition and classification have been studied. The following issues have been considered: the selection of the appropriate filter to completely remove the noise in the EMG signal, use of the wavelet transform and the wavelet packet transform to extract the eigenvectors of the lower extremity muscle signal, and to apply the classifier with the neural network constructed by the support vector machine. The research methods of each link are analyzed in detail to lay the foundation for the follow-up research work on rehabilitation training.
Keywords: surface EMG signal, feature extraction, pattern recognition.
Introduction
Surface electromyography (sEMG) signal is one of many bioelectrical signals, and it is relatively easy to obtain a kind of electrophysiological signal. Therefore, surface EMG has important practical value in motion recognition, rehabilitation medicine and human-computer interaction. At present, the research of sEMG signal is more and more in-depth, which provides great progress for the exploration of rehabilitation training. In recent years, many scholars at home and abroad have done a lot of research on the recognition of sEMG signal, and various research results also provide experience for the development of rehabilitation training, laying a foundation for the application of rehabilitation training research. Research based on sEMG signal usually includes signal acquisition and pretreatment, feature extraction, pattern recognition and classification and other important links. In view of the above important links, this paper briefly expounds the research methods at home and abroad, combined with analysis, aiming to provide guarantee for the follow-up related work.
© Liye Ren, Chen Wang, Junyi Tian, Jianwei Fang, 2022
BECTHHK TOry. 2022. № 2(65)
Surface EMG acquisition and pretreatment
The sEMG signal is a very weak signal and has non-stationarity. Generally, the useful signal frequency components are in the range of 0 to 500 Hz, and the main energy is concentrated in the range of 50 to 150 Hz. Because the human sEMG signal is very weak, the degree of interference is large, and the measurement is difficult, whether the sEMG signal can be effectively collected is the basic condition for its follow-up research. The movement of the lower limbs of the human body is a periodic movement, and the sEMG signals collected from the muscles of the lower limbs are unconscious contraction signals, and the background noise is complex, and the collected signals need to be processed. At present, the sEMG signal acquisition equipment on the market will perform preliminary preprocessing on the EMG signal, but there will still be noise in the signal. Therefore, according to the characteristics of the surface EMG signal, reasonable design of the filter is the key technology to deal with the noise. Design a Butterworth or Che-byshev low-pass filter with an appropriate order to filter out the high-frequency noise introduced by the amplifier. For power frequency interference, an analog notch filter can also be designed to reduce noise.
In addition, the noise of sEMG signal can also be removed by using wavelet transform. According to the basic noise model, the wavelet transform is used to decompose the high and low frequency signals in the signal frequency domain. The noise model is as follows:
In the formula, f(n) represents the original signal, s(n) represents the signal disturbed by noise, e(n) represents the white noise signal, and a = 1. The ultimate purpose of wavelet transform is to gradually layer the signal s(n) carrying interference noise, denoise the noise in the high frequency signal through the threshold, leave the low frequency component, and finally reconstruct to obtain the EMG signal.
Feature extraction
Feature extraction is combined with signal segmentation and adopts moving window method to extract feature variables in specific small Windows. On the one hand, dimension reduction can be achieved; on the other hand, feature quantities can better reflect the features of the segment compared with original data. Feature extraction methods can be divided into five categories: time domain analysis method, frequency domain analysis method, time frequency domain analysis method, nonlinear analysis method and parameter model analysis method.
The time-domain analysis method is the earliest traditional method. The independent variable of time-domain characteristics is time, and the signal is regarded as a curve that changes with the strength of the signal. Among them, root mean square (RMS) and integrated electromyogram (IEMG) are frequently used in studies and have achieved good results. Their formulas are as follows:
APPLICATION OF SEMG SIGNAL IN REHABILI-
TATION TRAINING OF LOWER LIMBS BtOHHtC TCTy 2022. № 2 (65)
Frequency domain analysis method can directly observe the distribution and change of sEMG frequency band, using Fourier transform to obtain the signal spectrum or power spectrum. Average power frequency (MPF) and median frequency (MF) are widely used spectrum characteristic parameters, and their formulas are as follows:
Due to the limitation of frequency domain features, frequency domain features are not widely used in the research of motion pattern recognition.
Time-frequency domain analysis methods have been widely used in recent years. Due to the time and frequency localization characteristics of wavelet transform, the time-varying spectrum analysis of the signal can be realized, so that it can analyze the signal in any detail and is not sensitive to noise. Therefore, the wavelet transform is a powerful tool for the analysis of sEMG signals. After the function y(t) is shifted on the time axis by b, the scale is stretched, and the inner product operation is performed with the signal to be analyzed under the scale a:
If"" t-b WTs(a, b)=— S{ty$* (——)dt = (s(t), i|/(t)>, n > 0
,
The wavelet transform can decompose the signal at multiple scales according to different wavelet functions, and can reconstruct the coefficients according to the obtained high-frequency and low-frequency coefficients of each scale, and the singular values of the decomposition coefficients can also form eigenvectors. The maximum value of the wavelet coefficient, the singular value of the wavelet coefficient and the energy value of the wavelet coefficient are the most important features, and the formula is as follows:
1. Maximum wavelet coefficients :
2. Wavelet Coefficient Singular Values :
3. Wavelet coefficient energy value:
1 Al |2
, E = — T\ sj
1 N ' I
j=1
The wavelet packet transform is a signal decomposition and reconstruction algorithm based on the complete binary tree structure, which is extended from the wavelet transform. It can analyze and extract features on any signal, and its main three features are the same as wavelet transform. Wang Kunpeng et al. used the feature extraction method of wavelet transform to select the average power of active segments in the wavelet subspace in the frequency domain distribution to form a eigenvector for pattern recognition, and achieved good results.
1. Maximum wavelet packet coefficient:
BEGTHHK TOry. 2022. № 2 (65)
2. Singular value of wavelet packet coefficient:
3. Wavelet packet coefficient energy:
fearure = {log10 {Eiri = 1, -,
Because EMG signals have nonlinear characteristics, nonlinear analysis methods are gradually applied to EMG signal feature extraction. Among them, the most used parameter is entropy. For example, Song Fangyu et al. used the spectral approximate entropy theory to regress and extract the jumped surface EMG signal when the sEMG signal generated nonlinear jumps to realize the characteristics of linear and nonlinear signal parts. extract. There are not many studies on nonlinear analysis methods in feature extraction at home and abroad, and the remaining nonlinear parameters can be used for further exploration.
The analysis method of the parametric model is to establish an AR model, analyze the relationship between the parameters of the AR model and the body movements determined by the corresponding muscle activity, therefore extract the sEMG signal based on the action pattern. The mathematical expression of the AR model is:
In the formula, x(n) represents the nth sampled value of the EMG signal, ®(n) represents the white noise residual, and p represents the order of the AR model. Ok represents the k-th coefficient of the AR model. When the AR model order is 4, the feature extraction effect of its pattern classification is the best. The parametric model method is mainly used to drive the prosthesis and realize the bionic technology.
Pattern recognition
The effect of pattern recognition is directly related to the selection of eigenvalues. In the face of different needs and conditions, it is very important to select an appropriate pattern recognition method. Among them, support vector machine and neural network algorithm are the most representative learning methods in EMG pattern recognition. At present, the use of sEMG signals to identify human movements mainly focuses on the movements of upper limbs, wrists and fingers, and there are relatively few studies on the identification of movements of human lower limbs.
Support vector machine (SVM) is a machine learning method based on statistical learning theory developed in the mid-1990s. Support vector machine can map input samples to high-dimensional feature space through the nonlinear relationship of its kernel function, and construct the corresponding maximum. The optimal classification hyperplane, thereby distinguishing samples, is one of the most important factors affecting the performance of SVM. There are four commonly used kernel functions: linear kernel function, polynomial kernel function, radial basis kernel function and Sigmoid nuclear function. The function formula is as follows:
APPLICATION OF SEMG SIGNAL IN REHABILI-
TATION TRAINING OF LOWER LIMBS BECfflHEC TOTY. 2022. № 2 (65)
1. The linear nuclear:
2. Polynomial kernel:
3. Radial Base kernel:
4. Sigmoid nuclear:
K(xirx) = axjx Ku^i) = (xjx+l)d K(a£,3c) = exp{- \\x- Xi\\2l&} = tanh (v(x ■ xt) + c)
The SVM algorithm has two important parameters in the sEMG motion pattern recognition, namely the kernel function parameter and the optimal penalty parameter. The optimization of the parameters is the main way for many researchers to improve the SVM algorithm. When SVM is extended to multi-classification, the selection of its classification method is also particularly important.
Neural network is an algorithmic mathematical model that imitates the behavioral characteristics of animal neural network and performs distributed parallel information processing. Neural network can classify various actions. There are many kinds of neural networks used for pattern recognition, and many researchers will choose a combination of the two neural networks to optimize and improve the recognition rate. Jin Hua et al. combined a three-dimensional convolutional neural network with a long-term and short-term long-term memory network to recognize compound hand movements, and the recognition rate was stable at 90%.
In addition, classification algorithms such as Linear Discriminant Analysis (LDA), Hidden Markov Model (HMM), and Decision Tree have also been used in the study of action pattern recognition related to sEMG signals, all of which have achieved good results. Linear discriminant analysis (LDA) is a generalization of Fisher's linear discriminant method, which finds a linear combination of features of two classes of objects or events to characterize them. The resulting combination can be used as a linear classifier, which is the most widely used in the field of EMG control.
Conclusion
The feature extraction and analysis methods based on wavelet transform and wavelet packet transform are more suitable for the research of lower limb rehabilitation training, and more feature vectors can be extracted on the wavelet band. The internal structure of the neural network is complex, and it is not suitable for the case where the training set is too small, and it is prone to overfitting. Support vector machine and linear discriminant analysis are more suitable for the classification of lower extremity pattern recognition. Optimizing various parameters of support vector machine is a means to improve the recognition rate, but the optimization of support vector machine also needs to consider the structural risk, and it must have stability.
Acknowledgment
Grant Sponsor: Education Department of Jilin Province. Research on Key Techniques of Lower Limb Rehabilitation Training Based on Human Surface EMG Sig-nal.No:JJKH20210628KJ.No:2021LY505L16. Grant Sponsor: Changchun university. Research on lower limb motion pattern recognition based on EMG. No:ZKQ202007.No:ZKC202006.
ВЕСТНИК ТОГУ. 2022. № 2(65)
References
1. Hou Xiuli. The application of SVM in the recognition of hand movement patterns from surface electromyography signals // Journal of jiujiang University. 2021. P. 68-71.
2. Research on Surface EMG Signal Pattern Recognition in Rehabilitation Engineering / Chen Zhiqing, Zhang Yue, Wang Zhelu, Zheng Zexiang // Electronic Technology & Software Engineering. P. 74-75.
3. Feature extraction method of human leg surface EMG signal / Wang Kunpeng, Pang Jie, Shi Lei, Qu Jianfeng // Journal of Chongqing University. 2017.
4. Song Fangyu, Liu Yehui, Zhu Lihua. Research on feature extraction of muscle fatigue surface EMG signal based on wavelet transform // Journal of Biomedical Engineering Research. 2019. № 38(1). P. 86-89.
5. Recognition of Electromyographic Signal Time Series on Daily Hand Motions Based on Long Short-Term Memory Network / Jin Hua, Qinkun Xiao, Li Wang, Yixin Liu, Xuhui Ning // Traitement du Signal. 2021. Vol. 38, № 2. P. 387-394.
Заглавие: Применение ЭМГ сигналов в реабилитационной тренировке нижних конечностей
Авторы:
Рен Лие - Чанчуньский университет, Чанчунь, КНР Ванг Чен - Чанчуньский университет, Чанчунь, КНР Тиан Джуню - Чанчуньский университет, Чанчунь, КНР Фан Джанвей — Чанчуньский университет, Чанчунь, КНР
Аннотация: Поверхностные ЭМГ сигналы имеют важное исследовательское значение в реабилитационном обучении и клинической медицине. В работе были изучены некоторые важные связи, такие как сбор и предварительная обработка сигналов ЭМГ, извлечение признаков, распознавание и классификация. Рассматривается вопросы выбора фильтр для удаления шума в ЭМГ сигнале, использования вейвлет-преобразование и вейвлет-пакетного преобразования для извлечения собственных векторов сигнала мышц нижних конечностей и применения классификатора с нейронной сетью. Данные вопросы подробно анализируются, чтобы заложить основу для последующих исследований по реабилитационному обучению.
Ключевые слова: поверхностный ЭМГ сигнал, извлечение признаков, распознавание образов.
