An upper limb movement estimation from electromyography by using BP neural network

https://doi.org/10.1016/j.bspc.2018.12.020Get rights and content

Abstract

The body electromyography (EMG) signals contain a large amount of information related to the movement of the human body. Identifying the patients’ movement intention from the EMG signals is the key to controlling the exoskeleton to assist their movement. In order to accurately extract the information about the patients’ movement intention from the EMG signals, we preprocessed the EMG signals including signals amplification, denoising, biasing and normalization. Then we extracted the features of EMG signals from the time domain, frequency domain, and time-frequency domain respectively. Based on the features obtained, we used the Matlab neural network toolbox to train BP neural network and tested the established continuous movement control model. The results suggested that the angles estimated by the continuous movement control model had smaller errors. In addition, instead of the traditional working mode that used the PC to process the EMG signals, we used the STM32 microcontroller to perform real-time control of the upper limb exoskeleton, which greatly reduced the size of the control equipment and provided convenience for the patients’ rehabilitation training.

Introduction

Health issues have gradually become a topic of concern for people. Some diseases may cause hemiplegia if handled improperly. According to the severity of hemiplegia, it can be divided into paresis, incomplete paraplegia, and paralysis. Hemiplegic patients often have difficulty in moving, and bring a heavy burden on their families and the society. In order to alleviate this problem, scholars have used exoskeleton robots in rehabilitation training for patients with hemiplegia, which has brought great convenience to patients to some extent [1]. Ettore E. Cavallaro et al. [2] designed a seven-degree-of-freedom exoskeleton, and its movement estimation uses an improved Hill muscle model. The 28-lead signals acquisition instrument was used to collect the electromyography (EMG) signals, and the pattern recognition was performed by using a genetic algorithm so as to estimate the angle changes of the upper limbs. Takamitsu Matsubar et al. [3] designed a control system to control the mechanical arm. The control system could recognize five kinds of movement, which adopted the bilinear model to realize the movement recognition independent of individuals. The results showed that the control system could well control the manipulator to open, relax, bend, grasp and extend. Yi long et al. [4] designed a power-driven exoskeleton robot for lower extremities, which used a force sensor to monitor the exoskeleton in real time and used the controlling torque to evaluate the performance of the exoskeleton. Ming Meng et al. [5] adopted Hidden Markov Model (HMM) to identify the gait phase in the electromyogram, and used the modified Baum-Welch model to estimate the HMM parameters. Arjan Gijsberts et al. [6] established a large-scale benchmark evaluation system and combined EMG and acceleration, which significantly improved the accuracy of the movement. Dapeng Yang et al. [7] designed a five-degree-of-freedom mechanical prosthetic hand, which used two-channel EMG signals combined with the DD-SVM model and the template matching model (DCTM) to control the mechanical prosthesis. Zhijun Li et al. [8] designed an impedance control model to improve the accuracy and adaptability of the movement control. However, most of the current rehabilitation and training equipment is bulky, slow in response, and inaccurate in exercise control, which is inconvenient to the rehabilitation training of hemiplegia patients.

In the present study, a three-degree-of-freedom manipulator is designed to evaluate the performance of the system through the accuracy of movement estimation. We use the BP neural network to have established a model of the relationship between elbow angles and EMG signals features, through which we estimate the angles of the elbow joint and achieve continuous motion control of the exoskeleton.

Section snippets

Physical design

We studied the degree of freedom at the elbow joint of the human upper limb, the rotational freedom of the forearm and the degree of freedom at the wrist joint, and designed the corresponding upper extremity exoskeleton. The designed exoskeleton manipulators had three degrees of freedom: the swing at the palm, the rotation of the forearm, and the rotation of the elbow. The main target of this exoskeleton was the patients who would undergo rehabilitation training. When they used the exoskeleton,

Experimental design

Experimental equipment used for the continuous control of exoskeleton: EMG sensors, STM32, exoskeleton manipulator, power source, photoelectric encoder, and six-axis gyro sensor JY61 etc. The control process is shown in Fig. 7.

EMG signals’ continuous control of the exoskeleton manipulator

Features of EMG signals characterized the activation degree of muscle fibers. Different degrees of rotation of the elbow joints had different degrees of activation of the muscle fibers, which in turn caused differences in the features of the muscle signals. On the basis

Discussion

The three-degree-of-freedom motion control system we designed had relatively small errors and was light in weight compared with equipment designed by Guoqing Cai [28]. Its weight is only 1626 g. In order to verify the performance of the control system, we conducted continuous motion control experiments for people with different physical indicators. The results suggest that the error of the continuous motion control experiment is small, and the error of the elbow joint angle estimation is

Conclusion

On the basis of predecessors, we used a BP neural network to establish the discrete motion control model and continuous motion control model of exoskeleton robotic arm. We first preprocessed the acquired EMG signals. Then the features of the signals were extracted from three aspects, and the final features were obtained by weighted summation. In order for the BP neural network to converge quickly, the input data was normalized and the output data was anti-normalized. We finally achieved

References (28)

  • A. Gijsberts et al.

    Movement error rate for evaluation of machine learning methods for sEMG-based hand movement classification

    Ieee Trans. Neur. Sys. Reh.

    (2014)
  • D. Yang et al.

    Experimental study of an EMG-controlled 5-DOF anthropomorphic prosthetic hand for motion restoration

    J. Intell. Robot. Syst.

    (2014)
  • Z. Li et al.

    Adaptive impedance control for an upper limb robotic exoskeleton using biological signals

    Ieee Trans. Ind. Electron.

    (2017)
  • F. Xiao et al.

    Continuous estimation of elbow joint angle by multiple features of surface electromyographic using grey features weighted support vector machine

    J. Med. Imag. Health

    (2017)
  • Cited by (0)

    View full text