CC BY
0ВЕСТНИК ТСГУ. 2023. № 2 (69)
YflK 004.62:615.47
Zhang Dong, Zheng Bangsheng, Zeng Xuecheng, Bai Duanyuan
BIONIC ROBOTIC ARM MOTION RECOGNITION BASED ON PSO-SVM ALGORITHM
Zhang Dong - School of Electronic and Information Engineering, Changchun University of Science and Technology, Changchun; Zheng Bangsheng - School of Electronic and Information Engineering, Changchun University of Science and Technology, Changchun; Zeng Xuecheng - School of Electronic and Information Engineering, Changchun University of Science and Technology, Changchun; Bai Duanyuan - Professor, School of Electronic and Information Engineering, Changchun University of Science and Technology, Changchun, email: 11237243@qq.com (China).
Surface Electromyography (sEMG) signals as a control source of the mechanical prosthetic hand has great advantages over other control sources, it contains a wealth of movement-related information, can be used to reflect the state and function of nerves and muscles. With the increasing development of science and technology, a large number of scholars have discussed sEMG gesture recognition. In this paper, a support vector machine (SVM) classification model based on particle swarm optimization algorithm (PSO) is proposed, which first extracts the active segment of the collected sEMG signals, and then extracts the time domain features and frequency domain features of the sEMG signals, and finally introduces the sEMG gesture recognition process based on PSO-SVM, and explains the parameters of PSO-SVM. Experimental results show that the average recognition rate of the proposed method for the four gestures is 95%, and the average recognition accuracy is improved by 5% compared with the traditional BP neural network. The proposed algorithm has good results in recognition effect.
Keywords: gesture recognition, PSO-SVM algorithm, electromyography signals, mechanical hand prosthesis.
Introduction
EMG signal is a bioelectric signal that accompanies muscle activity, which is the superposition of motor unit potentials in many muscle fibers, containing various information about muscle activity, which is the smallest contraction unit of muscle, consisting of a motor neurons, endplates and multiple muscle fibers [1, 2]. Since the sEMG signal can well reflect the movement intention of the human hand, by extract-
© Zhang Dong, Zheng Bangsheng, Zeng Xuecheng, Bai Duanyuan, 2023
$
BECTHHK TOry. 2023. № 2 (69)
ing the sEMG signal and using the action mode information contained in it, the purpose of precise prosthetic hand intelligent control can be achieved, and the sEMG-based motion intention recognition technology has become one of the research hotspots in the field of human-computer rehabilitation.
Nazarpour, Sharafat and others [3] extracted the high-order statistics of arm EMG signals as features, and input them into the clustering analysis classifier to realize four classifications of arm movements, with an accuracy rate of 91 %. Wang Jingfang et al. used MYO bracelet to collect 8-channel EMG signals, extracted 5 eigenvalues and input them into BP neural network for four classifications of finger movements, with an accuracy rate of 90.35 %. William et al. realized binary classification of gestures through BP neural network, and the accuracy rate was 94.6 %. Chen Yuru [4] et al. used MYO bracelet to collect 8-channel EMG to achieve 2 classifications, with an accuracy rate of over 90 %. Yu Bihong and others designed a wearable EMG acquisition device, and the accuracy of classifying four movements was 98.55 %.
Based on the mechanism of surface electromyography [5], POS-SVM algorithm is used to identify and classify the input test set.
1. Experimental data acquisition and feature extraction 1.1 Data acquisition
According to the four movements designed in this paper: holding, grasping, wrist rotation and rest, and then according to the relationship between the movements and muscle groups, the flexor carpi radialis is finally selected as the electrode placement position, which adopts the differential measurement method, in which two electrodes are placed at the flexor carpi radialis at a distance of 2 cm, and the other electrode is placed at the elbow as a reference electrode. As shown in Fig. 1. Four participants whose arms were healthy but differed in age, sex and weight were selected as the control group. According to the order of holding, grasping, wrist rotation and rest, the subjects did each of the four kinds of movements 100 times.
a) Holding, b) Grasping, c) Rotating Wrist, d) Resting, e) Placing Position of Electrode
Patch
Fig. 1. Schematic diagram of arm movement and electrode paste installation
1.2 Feature extraction
Before feature extraction, it is necessary to extract the active segment of EMG signal, which is the data from the beginning to the end of an action cycle [6]. We use moving average method to extract the active segment of EMG signal. There are many feature extraction methods in the process of pattern recognition. In this experiment, we choose time domain and frequency domain feature extraction methods [7].
1.2. 1 Time domain feature extraction
Time domain feature extraction is the most traditional and direct sEMG feature extraction method. This paper chooses the following two time domain features:
First, suppose that an action cycle signal is and N is the data length, then there are:{x(j)|i = 1,2,..., N}
(1) Integral EMG (iEMG)
Integral EMG mainly describes the degree of signal concentration, and the formula is shown in formula
¿£MG=1lf=1|x(t)| . (1)
(2) Root mean square (RMS)
The root mean square value reflects the contribution of each muscle group in the process of completing limb movements, and it is an EMG characteristic parameter that can effectively distinguish different limb movements. The root mean square value of any sEMG can be calculated by equation
RMS = J^p2 . (2)
1.2. 2 Feature extraction in frequency domain
The following two frequency domain features are selected here:
(1) Median frequency (MF)
Median frequency can be very sensitive to detect changes in physiological parameters, although there is a certain amount of noise in the signal, but it has a good anti-aliasing ability, so that the impact of noise can be ignored. The calculation formula is shown in equations (1-3) (PSD (f) represents power spectral density).
MF = 1j;J PSD(f)df. (3)
(2) Average power frequency (MPF)
Average power frequency is a commonly used index to detect muscle fatigue in clinical and other related fields, and its calculation formula is shown in formula
MPF = £££££» <4)
J0 psd(f)df
2. Classification experiment
In this paper, support vector machine (SVM) classifier optimized by particle swarm optimization (PSO) is used to identify and classify EMG signals. SVM classifier distinguishes positive and negative samples through a specific classification
3Tl
plane, and the interval between these two samples needs to be maximized. This method is a machine learning method to classify data according to structural risk minimization [8]. When using SVM method to classify data or eigenvalues, we only need to select the appropriate kernel function and set its parameters. Secondly, set the penalty coefficient C, which is used to control the ratio parameter between empirical risk and confidence range, and to improve the generalization efficiency of machine learning. In order to make the parameters achieve satisfactory classification recognition rate, this paper will use particle swarm optimization method, using SVM for action classification pattern recognition.
SVM is optimized by particle swarm optimization method as follows [9, 10]
(1) Firstly, the data should be preprocessed, that is, the values of the training set and the test set extracted from the sEMG signal should be between [0, 1] by using the normalization function;
(2) Initialize the parameters of PSO, including sizepop and maxgen. Set the weight factor and termination condition of the algorithm, and also need to encode the initial particles;
(3) The individual extremum of each particle is set to the current position, and the fitness value of each particle is calculated by using the "fitness function", and the individual extremum corresponding to the particle with good fitness is taken as the initial global extremum;
(4) Carry out iterative calculation according to the updating formula of particle position and velocity, update the position and velocity of particles, and then calculate the fitness value of each particle after each iteration according to the "fitness function";
(5) Compare the fitness value of each particle with the fitness value of its individual extremum. If it is better, update the individual extremum, otherwise keep the original value. The updated individual extreme value of each particle is compared with the global extreme value. If it is better, the global extreme value is updated, otherwise the original value is retained;
(6) Judge whether the termination condition is met at the same time of iteration, and terminate the iteration if the maximum iteration times are reached or the obtained solution converges or the obtained solution has achieved the expected effect, otherwise return (4);
(7) Using penalty parameter c and RBF parameter g to train SVM classification;
(8) Input the training set into the SVM model for training, and then input the test set for classification, and get the classification result and recognition rate. The complete flow of this process is shown in Fig. 2.
The experimental environment is Inter (R) Core (TM) i7-11800H @ 2.30 GHz, and the classification model is programmed by Matlab. In PSO-SVM model, the maximum number of population evolution is 80 times and the maximum number of population is 30. RBF kernel function is selected, and the penalty parameter C and kernel function range are 0.1-50 and 0.1-100 respectively.
Zhang Dong, Zheng Bangsheng,
BECI™K TOry. 2023 № 2 (69) Zeng Xuecheng, Bai Duanyuan
Fig. 2. PSO-SVM algorithm flow chart
3. Experimental results
The classification effect obtained after input test set into PSO-SVM and BP neural network for classification is shown in Fig. 3 and 4. It can be seen that the prediction of PSO-SVM is better than that of BP neural network, and the prediction result of PSO-SVM is not much different from the actual one, while the BP neural network cannot achieve the same classification accuracy as PSO-SVM.
20 40 60 80 100 120 140 160 180 200
Test set sai^iLe Fig. 3. Classification effect of BP neural network
$
BECTHHK TOry. 2023. № 2 (69)
40 60 SO 100 120 140 160 180 200 Test set sample
Fig. 4. SVM classification effect
The recognition results of four gestures by BP neural network and PSO-SVM classifiers are shown in Table 1 and Table 2 respectively.
Table 1
Accuracy of classification of four actions by BP neural network
Action Grip Grasp Rotate the wrist Rest Average
Number of test 50 50 50 50 200
groups
Correct number of 46 43 44 48 181
groups
Accuracy 92% 86% 88% 96% 90.5%
Table 2
Accuracy of SVM Classifier for Four Kinds of Motion Classification_
Action Grip Grasp Rotate the wrist Rest Average
Number of test 50 50 50 50 200
groups
Correct number of 48 47 46 50 19
groups
Accuracy 96% 94% 92% 100% 95.5%
Compared with BP neural network classification algorithm, the proposed algorithm is higher than BP neural network in recognition accuracy of four kinds of gestures, among which action grip is 4 % higher, action grasp is 8 % higher, action wrist rotation is 4 % higher, rest action is 4 % higher, and the average recognition rate is improved by 5 %, which shows the superiority of the proposed algorithm.
Conclusion
Comprehensive analysis of the above results, PSO algorithm optimized SVM has better recognition accuracy than BP neural network, the average recognition accuracy is as high as 95.5 %, it can be seen that PSO-SVM algorithm has greater advantages. However, there are still some shortcomings in this study. First of all, in the process of signal acquisition, sEMG signal is easily affected by the outside world because of its weak and unstable characteristics. In addition, the subjects do not cover people of all ages, the number of subjects is limited, and the number of samples is not balanced enough. Finally, because the parameters of SVM parameter optimization are different for each tester, special training is needed to adapt to the user and improve the accuracy before the prosthetic hand control. In the follow-up work, we should introduce other algorithms to make SVM get similar classification accuracy for different subjects.
References
1. Duchene J., Hogrel J.Y. A model of EMG generation // IEEE Trans Biomed Eng. 2000. № 47. 192-201.
2. Basmajian J.V, Deluca C.J. Electromyography-physiology, engineering and noninvasive applications. New York : IEEE Press, 2004.
3. Nazarpour K., Sharafat A.R., Firoozabadi S. M.P. Surface EMG Signal Classification Using a Selective Mix of Higher Order Statistics // Proceedings of the IEEE Engineeringin Medicine and Biology 27th Annual Conference. 2005. pp. 1060-1066.
4. Real-time control system based on sEMG signal / Chen Yuru, Wang Xuan, Zhou Penghui, et al. // Internet of Things Technology. 2019.
5. A case study with SymbiHand: an sEMG-controlled electro-hydraulic hand orthosis for indiduals with Duchenne musical dystrophy / R. A. Bos, K. Nizamis, B.F. Koopman B.F., et al. // IEEE transactions on neural systems and rehabilitation engineering. 2019. Vol. 28. pp. 258-266.
6. Mokhlesabadifarahani B., Gunjan V.K. EMG signals character in three states of contract by fuzzy network and feature extraction. Springer. 2015.
7. Niu Jiabao. Research on human gesture recognition and strength prediction technology based on sEMG. Tianjin Polytechnic University. 2019.
8. Jaber H.A, Rashid M.T. HD-sEMG gestures recognition by SVM classification for controlling prosthesis // Iraq Journal of Computers, Communications, Control and System Engineering (IJCCCE). 2019. Vol.19.
9. Pan Binjie, Zhu Jianfeng, Xu Riqing. Optimal ratio of composite curing agent for soft soil based on PSO-SVM algorithm // Journal of Jiangsu University (Natural Science Edition). 2021. № 42. pp. 339-345.
ВЕСТНИК ТОГУ. 2023. № 2 (69)
10. Real estate risk measurement and early warning based on PSO-SVM / Zhou W., Chen M., Yang Z., et al. // Social-Economic Planning Sciences. 2020. Vol. 5.
11. Research progress of surface EMG signal classifier based on artificial neural network / Zhou Xiaobo, Zou Renling, Lu Xuhua, et al. // Electronic Science and Technology. 2021. Vol. 34. pp. 62-67.
Заглавие: Распознавание движений бионической роботизированной руки на основе алгоритма PSO-SVM
Авторы:
Чжан Дун - Чанчуньский научно-технический университет (КНР) Чжэн Баншэн - Чанчуньский научно-технический университет (КНР) Цзэн Сюэчэн - Чанчуньский научно-технический университет (КНР) Бай Дуаньюань - Чанчуньский научно-технический университет (КНР)
Аннотация: Сигналы поверхностной электромиографии (сЭМГ) в качестве источника управления механическим протезом кисти имеют большие преимущества перед другими источниками управления, они содержат огромное количество информации, связанной с движением и могут использоваться для отражения состояния и функции нервов и мышц. В статье предложена модель классификации 8УМ, основанная на оптимизации алгоритма роя частиц (Р80), который сначала извлекает активный сегмент регистрируемых сигналов БЕМв, а затем определяет параметры и характеристики во временной и частотной областях сигналов и, наконец, выполняется процесс распознавания основанный на алгоритме Р80-8УМ. Результаты эксперимента показывают, что средняя скорость распознавания предложенного метода для четырех жестов составляет 95 %, а средняя точность распознавания улучшена на 5 % по сравнению с традиционной нейронной сетью ВР.
Ключевые слова: распознавание жестов, Р80-8У алгоритм, сигнал электромиографии, механический протез руки.
