Development of a sEMG-based torque estimation control strategy for a soft elbow exoskeleton

Accepted Manuscript Development of a sEMG-based torque estimation control strategy for a soft elbow exoskeleton Longhai Lu, Qingcong Wu, Xi Chen, Ziyan Shao, Bai Chen, Hongtao Wu

PII: DOI: Reference:

S0921-8890(18)30512-8 https://doi.org/10.1016/j.robot.2018.10.017 ROBOT 3105

To appear in:

Robotics and Autonomous Systems

Received date : 17 June 2018 Revised date : 21 September 2018 Accepted date : 30 October 2018 Please cite this article as: L. Lu, Q. Wu, X. Chen et al., Development of a sEMG-based torque estimation control strategy for a soft elbow exoskeleton, Robotics and Autonomous Systems (2018), https://doi.org/10.1016/j.robot.2018.10.017 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

Development of a sEMG-based Torque Estimation Control Strategy for a Soft Elbow Exoskeleton Longhai Lu, Qingcong Wu, Xi Chen, Ziyan Shao, Bai Chen, and Hongtao Wu

Abstract—Motor dysfunction has become a serious threat to the health of older people and the patients with neuromuscular impairment. The application of exoskeleton to motion assistance has received increasing attention due to its promising prospects. The major contribution of this paper is to develop a joint torque estimation control strategy for a soft elbow exoskeleton to provide effective power assistance. The surface electromyography signal (sEMG) from biceps is utilized to estimate the motion intension of wearer and map into the real-time elbow joint torque. Moreover, the control strategy fusing the estimated joint torque, estimated joint angle from inertial measurement unit and encoder feedback signal is proposed to improve motion assistance performance. Finally, further experimental investigations are carried out to compare the control effectiveness of the proposed intention-based control strategy to that of the proportional control strategy. The experimental results indicate that the proposed control strategy provides better performance in elbow assistance with different loads, and the average efficiency of assistance with heavy load is about 42.66%. Index Terms—Torque estimation control strategy, soft exoskeleton, sEMG, motion intention, power assistance.

I. I NTRODUCTION In recent years, the rapid increase of the population of older people, physically handicapped patients, and stroke patients with motor dysfunction has brought heavy economic and medical burden to the society. There has been growing interest in the research of exoskeleton technology to assist the motion of subjects with motor disorder and rehabilitation problems [1]–[4]. However, most of existing exoskeletons developed for rehabilitation training and motion assistance have heavy and rigid mechanical structures [5]. Besides, they use complicated anthropomorphic mechanism to match the human anatomy and enable natural human-robot interaction for comfort and safety [6]. Compared to these rigid exoskeletons, the soft exoskeletons utilize human joint to rotate so that it can avoid human-robot interaction problems caused by the bias between This work was supported in part by the National Natural Science Foundation of China (Grant No. 51705240), the Natural Science Foundation of Jiangsu Province of China (Grant No. BK20170783), the Aeronautical Science Foundation of China (Grant No. 2017ZC52037), the State Key Laboratory of Robotics and System (HIT). Longhai Lu, Qingcong Wu, Xi Chen, Ziyan Shao, Bai Chen and Hongtao Wu are with laboratory of Robotics and System, Nanjing University of Aeronautics and Astronautics, nanjing210016, jiangsu, China). Corresponding author: Qingcong Wu. E-mail:[email protected]).

robot joint center and human joint center [7]. Furthermore, the soft exoskeleton robots are more lightweight, compliant, comfortable, convenient and harmonious [8]. Therefore, the soft exoskeleton robots have drawn an increasing attention for the numerous potential applications in military and medical fields [7], [9], [10]. Many researchers have developed various kinds of soft exoskeleton robots, which are wearable, low-cost and compliant [11]–[13]. One of the major challenges in this research field is the development power-assist control strategy, which should be suitable for the characteristics of wearer and soft exoskeleton. In order to realize efficient and comfortable assistance from exoskeleton robot, it is crucial to recognize the human motion intention from the interpretation of central nervous system signals. Human motion intention can be divided into three expressions according to the transmission steps of intention, including neural signal, such as electroencephalogram (EEG) [14], physiological signal, such as electromyography signal (EMG) [15], and physical force signal, such as human-robot interaction force signal [16], [17]. Among these signals, the surface electromyography signal (sEMG) have been widely applied in exoskeleton, as it can reflect the activation level of human muscles and non-invasive properties [18]. Compared to human motion intention recognition based on interaction force signals, the recognition strategy proposed based on sEMG is more suitable for characteristics of soft exoskeleton in aspects of comfort and safety [18]. Moreover, because the muscular activation arises before the actual muscle contraction about 20∼100 milliseconds, the control strategies have enough time to deal with real-time sEMG signals and generate commands for appropriate assistance [19]. The traditional control strategies of sEMG-based assistance system include machine learning algorithms, continuous linear and non-linear control strategies. The machine learning algorithms are capable of classifying the movement patterns according to sEMG signals [20], [21]. The linear discriminant analysis, support vector machine and artificial neural networks are the most common algorithms applied to recognize motion patterns based on sEMG signals [15]. In [22], Xin Li et al proposed a novel control strategy fusing angles and sEMG signals by time-delay recurrent neural network to estimate the motion intention of hand and wrist, and control the movement of prosthetic hand. Although using offline approaches, the accuracy of motion pattern classification can be higher than 90% by applying various classifiers. However, the online real-time

analysis is not robust and the algorithms, which need modeling and training, are complex and time-consuming. Furthermore, the motion patterns classified by the above algorithms are limited, resulting in discontinued response of exoskeleton. On the other hand, the continuous linear and non-linear control strategies generating continuous control signals have drawn plenty of attentions [23], [24]. In [25], Rong Song et al developed a linear proportional method to provide continuous signals based on sEMG signals for elbow exoskeleton robot. Other more complicated non-linear control strategies have also been developed to evaluate human motion intention and predict joint torque. In [15], Ao Di et al proposed an EMG-driven Hill-type neuromusculoskeletal model to estimate the realtime ankle joint torque for an ankle power-assist exoskeleton system. However, the performance of this continuous linear control strategy is not satisfying, and the Hill-type-based nonlinear control strategy requiring joint kinematics parameters are too complicated. In this paper, a continuous sEMG-based torque estimation control strategy (STECS) fusing signals from sEMG sensor, inertial measurement unit (IMU) and encoder is developed for a soft elbow exoskeleton suit to provide effective power assistance. The proposed strategy estimates the joint torque and position of elbow by processing sEMG signals and IMU signals. Different from the traditional sEMG-based strategies, the proposed control strategy is more efficient and simple to implement, the average efficiency of assistance is as high as 42.66% with heavy load. The rest of the paper is organized as follows. The second section describes the proposed strategies in detail. The overall components of the power-assist soft exoskeleton system and the control strategy estimating joint torque from sEMG signals are introduced and analyzed. Besides, the experimental procedures and evaluation strategies are discussed. In the third section, the experimental researches are carried out, and the results of under various conditions are displayed and discussed. Furthermore, the future applications and limitations are introduced. Finally, the summary of this paper is proposed in the last section. II. M ETHODS A. Power-assist Soft Exoskeleton Suit As shown in Fig. 1, a soft exoskeleton suit with one degree of freedom (DOF) is developed to assist the elbow movement of wearer. The whole power-assist soft exoskeleton system (Fig. 2) mainly consists of a customized actuated mechanical module, a sEMG sensor (Myoware sensor), two high sensitive IMUs (MPU9250 module), a high precision AC servo motor (YASKAWA SGM7J-01A7C6S) with a servo driver (YASKAWA SGD7S-R90A00A002), an IMU data processing device (ALIENTEK STM32F407), a 16-bit data acquisition card (Advantech PCL-818L), a 6-channels analog output ISA card (Advantech PCL-726), and Matlab/xPC real-time system which include a host PC (personal computer) and a target PC (Advantech IPC-610L). In the soft exoskeleton suit, the servo motor is connected to the actuated mechanical module to generate driving torque. The driving power can be transmitted to elbow of wearer via

IMU Tendon Ͳsheath sEMG sensor Wire rope Sheath support Driver system

Sheath support

Fig. 1. The overview of the power-assist soft elbow exoskeleton.

Advantech STM32Board PCLͲ818LDAQ MPU9250

TargetPC

Advantech PCLͲ726Board

HostPC

Human&Soft Exoskeleton

YaskawaServo DriverSystem Driver

Myoware sEMGSensor

Encoder

Fig. 2. The flow of soft exoskeleton assistance system and introduction of system components.

the flexible tendon-sheath transmission system [26]. The inner wire rope goes through the tendon-sheath support attaching at the upper arm and, after that, connects to another sheath support attaching at the forearm. To ensure the compliance of soft exoskeleton and the comfort of user experience, the tendon-sheath support is sewn onto resilient strip. The position of resilient strip is adjustable according to the length and size of human arm. The sEMG electrodes are attached to biceps brachii, which dominates the main muscular contraction of elbow movement. The sEMG feedback signals collected and amplified by the sEMG sensor signals are inputted into the target computer through a 16-bit data acquisition card for joint torque estimation. The IMUs, which are attached to the upper arm and forearm of wearer, are connected to the STM32F407 board to fuses the data from gyroscope, accelerometer and magnetometer and, moreover, calculate arms configuration. Then a sEMG-based torque estimation control strategy developed based on the Matlab/xPC real-time system is proposed for the soft exoskeleton suit to provide assistance. The Matlab/xPC real-time system has three functions: 1) running the sEMG-based torque estimation control strategy

and controlling the servo motor with analog output card; 2) providing real-time signals about sEMG and joint angle to guide the experiments and preventing from the dangerous situation; 3) storing the sEMG, estimated joint torque, and estimated joint angle signals for offline analysis. B. Joint Torque Estimation Strategy In order to estimate joint torque for motion intention estimation, in this study, the sEMG signals representing the muscle activation level are collected and processed. The traditional approaches estimating joint torque from the sEMG signals are proposed to approximate the active state for a Hill-type muscle model [15]. In addition, motion capture system is needed to acquire the accurate joint kinematics parameters. In this paper, the developed approach is capable of estimating joint torque without complex modeling processes and joint kinematics parameters for the Hill-type model. Furthermore, this approach can be easily implemented for the developed soft exoskeleton suit and suitable for different users without plenty of training time and classification processes. 10Ͳ500Hz BandͲpass Filter

Kalman Filter

50HzNotch Filter

NonͲLinearly Normalization

410HzHighͲ passFilter

FullWave Rectifier

Linearly Normalization

1HzLow Filter

Fig. 3. The procedure of sEMG-based joint torque estimation strategy.

The diagram of the joint torque estimation strategy is illustrated in Fig. 3. During the data acquisition, the sEMG signals from biceps are collected by the sEMG sensor and amplified by a signal amplification circuit. In order to acquire the optimal signals, the sEMG sensor are put on the center of muscle and it should be parallel to the direction of muscle fiber [27]. The raw sEMG signals are collected at a 1000 Hz sampling frequency. However, due to the undesirable noise and disturbances, these signals cannot be used to estimate joint torque. For this reason, a 10-500 Hz second order Butterworth band-pass filter and a 50 Hz notch filter are utilized to filter noise and 50 Hz power frequency disturbance. The 410 Hz first order Butterworth high-pass filter is used to remove 99% of the sEMG signal power due to the muscle fatigue and tissue filtering properties [28], [29]. Then the signal envelope is obtained through the full wave rectifier and 1 Hz first order Butterworth low-pass filter. The linearly normalization makes signal envelope adjusted in the range of 0-100 for reducing the variability of different subjects. The nonlinearity between sEMG signals and joint torque is also taken into account [30], which can be described as: e(−EM GL ξ) − 1 (1) e(−100ξ−1) where EM GL is the normalization of sEMG signal, EM GN denotes the nonlinearity normalization of sEMG signal, ξ is the parameter of defined exponential curvature. Then the EM GN = 100

Kalman filter is used to remove the noise and smooth the estimated joint torque. In this approach, as the signal is a one dimensional signal, the state-transition matrix and observation matrix are simply equal to 1. Hence the following state equations are established to implement the Kalman filter: m(t) = m(t − 1) + Q

EM GN (t) = m(t) + R

(2) (3)

where m(t) is the t moment true state, i.e., the desired smoothing joint torque, m(t − 1) represents the previous state, Q and R denote the covariance of process noise and measurement noise, which are selected according to Zhijun Li et al [29], EM GN (t) shows the t moment measurement value. Then the prediction and update equation of Kalman filter are established as follows: M (t|t − 1) = M (t − 1|t − 1)

P (t|t − 1) = P (t − 1|t − 1) + Q

K(t) = P (t|t − 1)(P (t|t − 1) + R)−1

(4) (5) (6)

M (t|t) = M (t|t − 1) + K(t)(EM GN (t)

(7)

P (t|t) = (1 − K(t))P (t|t − 1)

(8)

− M (t|t − 1))

where M (t−1|t−1), M (t|t−1) and M (t|t) are the previous, priori and posteriori state estimate. P (t − 1|t − 1), P (t|t − 1) and P (t|t) represent the previous, priori and posteriori error covariance. K(t) denotes the kalman filter gain. According to linearity assumption, the current moment estimated state is just equal to the previous one, and the measurement value is used to update the estimation. With the iterations of prediction and update equation, the current smoothing joint torque can be estimated in real time. After calibrating by measured joint torque, the estimated joint torque can represent the joint torque within a certain margin of error. C. Joint Angle Estimation Strategy The conventional approaches estimating joint angle are mainly measured directly by the different angle sensors, such as rotational potentiometer, encoder, Hall sensor, and so on. However, the inherent non-rigid structure characteristic of soft exoskeleton suit makes these conventional angle sensors unable to be mounted on the exoskeleton structure. Obtaining the joint angle indirectly through IMU becomes an alternative solution [31]. In [32], ZC Ong et al attached two IMUs to upper and lower limb. The joint angle could be calculated according to the length of limb, the position and orientation of IMU. And in [33], RV Vitali et al also proposed a new IMUbased method to estimate 3D knee joint rotation. However, there are some limitations in these traditional IMU-based methods, such as the location of IMUs, parameters of limbs’ kinematics and so on. This study proposes a method using IMUs to estimate joint angle in real time. Compared to the conventional methods, this method is insensitive to the bound location and orientation of IMUs and it can calculate joint angle without parameters of limbs’ kinematics.

q te qos ( qts )

e

J Fig. 4. The orientation transformation from initial moment to t moment during elbow movement. q0s , q0e are beginning orientation and ending orientation at the initial moment, qts , qte are beginning orientation and ending orientation at the t moment, r represents the rotary axis of elbow, γ denotes the rotation angle of elbow during movement.

In order to estimate joint angle during elbow movement, the Mahony algorithm is utilized to fuse data from gyroscope, accelerometer and magnetometer and calculate the orientation of IMUs [34]. The thought of this algorithm is that the orientation of IMUs attached to arms can be corrected by the error between the directions of gravity and the accelerometer vector, and the error between the directions of the geomagnetic vector in object coordinate and the magnetometer vector. Since the quaternion can avoid Gimbal Lock [35], the orientation of IMU can be described by quaternion as follow:   q = q0 q1 q2 q3 h γ γ γ = cos 2 γx sin 2 γy sin 2

γ γz sin 2

(11)

γ = 2 arccos(q0t )

(12)

where qt , qts and qte denote rotation quaternion, transposition of beginning orientation, and ending orientation at the current moment, respectively. Then the elbow angle can be estimated as: −1

r

qo

qt = qts −1 qte · λ

i

(9)

where q is the quaternion of orientation of IMUs, q0 denotes the real number which contains the rotation angle information, q1 , q2 , q3 are the real part of the imaginary number i, j, k respectively. γ represents the angle of rotation about vector r, and rx , ry , rz consists of rotary axis r. Then the angle solution algorithm is proposed to estimate joint angle. Fig. 4 shows the orientation transformation from initial moment to t moment during elbow movement. IMUs are attached to upper arm and forearm. In this paper, the orientation of the IMU attached to upper arm is treated as the beginning orientation. Similarly, the orientation of the IMU attached to forearm is treated as ending orientation. Inverse treatment also works. The angle of elbow is considered as the angle rotated about axis r from beginning orientation to ending orientation. At the first moment, the beginning orientation and ending orientation need to be initialized to a uniform orientation. The process can be described as:

where q0t is real number of qt . Since the elbow angle can be calculated by orientation transformation between upper arm and forearm, the elbow angle is insensitive to bound location and initial orientation of IMUs, making this method suitable for the soft exoskeleton applications. D. sEMG-based Torque Estimation Control Strategy In this paper, a sEMG-based torque estimation control strategy fusing the signals from sEMG sensor, IMUs and encoder is developed to regulate the actuated mechanical module for elbow assistance. The conventional approaches mainly utilize machine learning algorithms such as neural network or support vector machine to classify the motion patterns, and they use the discrete results of motion patterns to drive the exoskeleton robots [15]. However, on the one hand, using the discontinuous results to control exoskeleton robot is not convenient. On the other hand, the supervised learning process consumes plenty of training time. This study proposes an efficient control strategy providing power-assist for the wearer according to human motion intention, which is simpler and more suitable to be applied in practical applications. Compared to the conventional approaches, this developed effective strategy can output the continuous signals without time-consuming classification of motion patterns. Estimatingjointtorque

Torque Estimation sEMG

Biceps

Human body

Estimatingjointangle

Kalman Filter M (t )

IMU Angle T (t ) Closeloop

Mapping +T (t ) + H (t ) & Ͳ Limition

PID u(t ) Soft Angle Exoskeleton Control Eblow Angle Mapping

Encoder Angle

(10)

Fig. 5. The diagram of sEMG-based torque estimation control strategy. M (t) is the estimated joint torque, ∆θ(t) denotes the incremental angle mapped from estimated joint torque, θ(t) represents the current joint angle estimated from IMUs, ε(t) is the expected signal inputted into PID controller, u(t) shows the output signal driving the soft exoskeleton suit.

where q0s , q0e are beginning orientation and ending orientation at the first moment. λ is the factor to initialize orientation of IMUs. After the initialization, the movement angle of elbow joint can be described as the rotation angle from beginning orientation to ending orientation around the elbow axis. The rotation quaternion can be represented as:

The diagram of proposed control strategy is described in Fig. 5. The sEMG signals are collected by sEMG sensor attached to biceps, while the joint torque is estimated by the aforementioned approach. The “Mapping and limitation block” shown in Fig. 5 has two basic functions. Firstly, the estimated joint torque can be mapped into the incremental angle of

q0s = q0e · q0e −1 q0s = q0e · λ

elbow (the effectiveness of this assumption will be verified in power-assist experiments) and used as one input of the proportion integration differentiation (PID) feedback loop to control elbow configuration. Secondly, the limitation block can be used to restrict the limitation of incremental angle to safety reasons. The limitations of incremental angle are set to 0◦ -30◦ by experiments. The higher limitation causes faster movement and more dangerous. Then the elbow angle including incremental angle and estimated joint angle from IMUs is set to 70◦ -180◦ to safety reasons. The linear relationship between estimated joint torque and incremental angle is determined by interactive experiments. Because the linear coefficient finally decides the incremental angle, it is relevant to the efficiency of assistance. In the interactive experiments, we use a priori value which is determined by half of the linear coefficient of previous subject. Since the sEMG signals can significantly vary from person to person, the linear coefficient is different for each subject. The selected control parameters of the PID controller are estimated via Ziegler and Nichols method [36] and carefully optimized by trial and error. The encoder signals are used as feedback signals to make a close-loop system. In the feedback loop, there is a relationship model between the length of wire rope and elbow angle, as illustrated in Fig. 6.

c d + arctan + (13) a bp 2cd − 2ab + 2βk (a + b)2 + (c − d)2 − (βk)2 p arccos 2 (a2 + c2 )(b2 + d2 ) where β is the encoder data representing motor motion, k is the factor depending on the actuated mechanical module. α = arctan

E. Experimental Setup and Signal Processing In the current study, five healthy subjects, aged 22-51, height 162-182 cm, weight 52-81 kg, two females and three males, had been involved in the experiments and provided informed consent. The implemented experimental approaches of this research have been approved by the Institutional Review Board of Nanjing University of Aeronautics and Astronautics.

TargetPC

PowerSupply

band

AdvantechPCLͲ 818LDAQ

c

sEMGSensor

TensionSensor

HostPC

Fig. 7. Introduction of components in the system which validates the efficiency of joint torque estimation strategy.

L a

d

D b

Fig. 6. The relationship model between elbow angle and length of wire rope.

Here, a, b are the distance from tendon-sheath support attached to forearm and upper arm to elbow rotation axis; c, d denote the height of tendon-sheath support; α means the current joint angle of elbow; L is the distance between two supports. There are some assumptions in this relationship model. Firstly, the wire rope is assumed to have no elastic deformation due to that force transferred by tendon-sheath system is not very large. Secondly, the resilient strip is assumed to have no large displacement and deformation. Since the motor directly drives the tendon-sheath system, encoder signals are related with the length of wire rope. Considering the actuated mechanical module and assumptions, the relationship between encoder signals and joint angle is described as:

Before the experiments, subjects had enough rest to avoid muscular fatigue. 70% alcohol was used to clean the skin of biceps brachii, the main contributing muscle for elbow contraction, and the conductive gel was used to improve the contract of the electrode with the skin. Then the sEMG sensor electrodes were putted on the center of biceps brachii with an interelectrode distance of 2 cm. Moreover, the electrodes were placed parallelly to the direction of muscle fiber to avoid the innervation zone of the muscles and acquire the best signals [27]. The experimental setup shown in Fig. 7 was constructed to verify the efficacy of joint torque estimation strategy. Subjects were required to stand beside experimental platform and pull the band vertically. The frequency and last time of actions are decided by subjects, five trials were performed, and each trial had 3 min rest time to relieve muscular fatigue. The sEMG signals together with force signals from tension sensor (JLBSMD) were collected by data acquisition card. The joint torque was estimated by the aforementioned approach. The measured joint torque was calculated by times the moment arm. The correlation coefficient between the estimated joint torque and measured joint torque was applied to quantitative analyze the efficacy of estimated joint torque, which could be determined as follows: ρTex ,Tr =

Cov(Tex , Tr ) σTex σTr

(14)

where Tex , Tr reflect the estimated joint torque and measured joint torque. Cov is the covariance. σTex , σTr are the standard deviation of estimated and measured joint torque. Moreover, the errors between estimated and measured joint torque, rootmean-square error (RMSE) and normalized root-mean-square error (NRMSE) were considered, which were described as:

RM SE =

v u T uX u (Trt − Text ) u t t=1

T RM SE N RM SE = Texmax − Texmin

estimated joint angle and potentiometer signals was calculated to prove the efficacy of angle solution algorithm. The RMSE as well as NRMSE between the estimated and measured joint angles were used to quantitative analyze the accuracy of angle solution algorithm. RMSE and NRMSE of the estimated joint angle in different locations with same orientation or different orientations with same location were applied to quantitative analyze the location and orientation insensitivity.

(15) (16)

where T is number of joint torque data. Text , Trt denote the estimated joint torque and measured joint torque at time t. Texmax , Texmin represent the maximum and minimum of Tex .

Target PC

Advantech PCLͲ818LDAQ

SoftExoskeleton TendonͲ DriverSystem sheathSystem 70°

AdvantechPCLͲ726 AnalogOutputCard

potentiometer STM32 DevelopmentBoard MPU1

TargetPC

RS232

Host PC

STM32 Yaskawa ServoDriver Development Board

RS232 MPU1 sEMG Sensor MPU2

180°

AdvantechPCLͲ726 AnalogOutputCard HostPC

AdvantechPCLͲ818L DAQ

YaskawaServo Driver

MPU2 TendonͲsheath System Yaskawa ServoMotor

Fig. 8. Introduction of components in the system which validates the efficiency of joint angle estimation strategy.

Then a series of experiments were conducted to evaluate the accuracy of joint angle which was estimated from mentioned algorithm. The evaluation system was shown in Fig. 8. The mechanical arm is driven by tendon-sheath actuation system to perform sinusoidal movement and imitates the motion of elbow. Two IMUs were bound on mechanical arms to measure orientation of corresponding arm, and they could be located in any position with any orientation. The potentiometer was installed on joint to measure actual joint angle. Nine trials were carried out in three different positions with three different orientations. In order to examine the performance of angle solution algorithm under different frequencies, in each trial, the frequencies of sinusoidal movement were variable, which were 1/13 Hz, 1/10 Hz, and 1/6 Hz orderly. The amplitudes of sinusoidal movement were from 30◦ to 180◦ without fixed. Taking the characteristic of insensitivity to bound location and orientation of IMUs into consideration, several trials were conducted to prove the feasibility. During the experiments, two IMUs were placed in three different locations with the same orientation to verify the location insensitivity. Similarly, two IMUs were placed in three orientations with the same location to validate the orientation insensitivity. The estimated joint angle from STM32F407 development board together with potentiometer signals are input into the target PC for analysis. The correlation coefficient between

Fig. 9. Introduction of components in the system which validates the performance of sEMG-based torque estimation control strategy.

Finally, a series of experiments were performed to examine the performance of proposed sEMG-based torque estimation control strategy. Besides, the proportional control strategy (PCS) in [29] were conducted to contrast the proposed control strategy. The experimental setup used to evaluate the efficiency of the proposed control strategy was shown in Fig. 9. During the experiments, the subjects wearing soft exoskeleton suit were instructed to perform a repetitive elbow movement that began at extension in 180◦ and ended at contraction in 70◦ . A joint rotation animation was displayed on the screen of host PC placed in front of subjects. During the experiments, the subjects were required to lift dumbbells whose plates weighted 2.5 kg, 5 kg and 7.5 kg with or without assistance and track the rotating line in movement animation as accurately as possible. The frequencies of rotating target trajectory were set to 1/12 Hz, 1/8 Hz, 1/6 Hz, 1/4 Hz and 1/2 Hz, respectively. Each subject was asked to perform three assistance trials in different loads and they had 3 min break between every two trials to relieve muscular fatigue. In each trial, three cycles of tracking movement were conducted to evaluate the effect of assistance control strategy. In order to analyze the performance of assistance control strategy, before each assistance trial, subjects should lift dumbbells without assistance to obtain contrastive data. The estimated joint torque, joint angle as well as the raw sEMG signals were input into the target PC for controlling and analysis. The contrast of estimated joint torque and raw sEMG signals under assistance and non-assistance conditions were utilized to validate the effectiveness of the proposed control strategy. Integral of the absolute value of raw sEMG signals were calculated to quantitative analyze the efficacy of the assistance control strategy, and it could be expressed as:

(17)

where Ur and Uex are the raw sEMG signals under nonassistance and assistance conditions, η denotes the efficiency of sEMG-based torque estimation control strategy. III. R ESULTS AND D ISCUSSIONS A. Estimated Joint Torque Performance The contrast of estimated and measured joint torque as well as errors between these two are shown in Fig. 10 and Fig. 11. The RMSE, NRMSE and correlation coefficient in each trial are shown in Table. I. The average correlation coefficient between estimated and measured joint torque, i.e., 98.04 (-0.9510∼+0.8260)%, shows that these signals have a good correlationship. Also in Fig. 10, the estimated joint torque based on sEMG is consistent with the trend of actual torque in different frequencies and amplitudes. Moreover, Fig. 10(a), (c) indicate that the estimated joint torque can identify more details like muscle shaking, last time and magnitude of joint torque in muscular contraction. The errors between estimated and measured joint torque is shown in Fig. 11. The average RMSE and NRMSE between these two signals are 1.898 (-0.5224∼+0.4006) Nm, 8.108 (2.293∼+1.298)% respectively. And the maximum error band is 11.42 Nm. As shown in Fig. 11, the errors concentrating on the range of ±5 Nm which is acceptable, and it suggests that the proposed strategy have good performance in joint torque estimation. In general, Fig. 10 and Fig. 11 show that the estimated joint torque based on sEMG signals representing the human motion intention is highly correlative with actual joint torque and it can be used to control soft exoskeleton system.

1 2 3 4 5

40 (a)

correlation coefficient (%)

RMSE (Nm)

NRMSE (%)

98.53 97.09 98.87 97.60 98.04

1.376 2.299 1.654 1.885 2.278

5.815 9.406 8.121 8.558 8.638

B. Estimated Joint Angle Performance The experimental results of joint angle estimation are shown in Fig. 12, and RMSE, NRMSE as well as correlation coefficient between estimated joint angle and measured joint angle are shown in table. II. Fig. 12(a) shows the contrast of estimated and measured joint angle in different orientations with same position, and Fig. 12(c) shows the contrast results in different positions with same orientation. From these two charts, the estimated joint angle tracks measured joint angle very well, the average

Tex

Tr

20 0 400

(b)

2

4

6 8 Time (s)

10

6 8 Time (s)

10

6 8 Time (s)

10

6 8 Time (s)

10

6 8 Time (s)

10

12 Tex

Tr

20 0 400 (c)

2

4

12 Tex

Tr

20 0 400

(d)

2

4

12 Tex

Tr

20 0 400

(e)

2

4

12 Tex

Tr

20 0

0

2

4

12

Fig. 10. The data of estimated and measured joint torque in five trials. The blue dash line Tex denotes joint torque estimated from raw sEMG signals in biceps, the red solid line Tr shows actual joint torque measured by tension sensor.

10 E1

5

E2

E3

E4

E5

0 -5 -10

TABLE I

CORRELATION COEFFICIENT, RMSE AND NRMSE BETWEEN ESTIMATED AND MEASURED JOINT TORQUE IN FIVE TRIALS

Trial Trial Trial Trial Trial

Torque (Nm) Torque (Nm) Torque (Nm) Torque (Nm) Torque (Nm)

Z |Ur | − |Uex | Z × 100% |Ur |

Torque Error (Nm)

η=

Z

0

2

4

6 8 Time (s)

10

12

Fig. 11. Errors between estimated and measured joint torque. The lines E1 , E2 , E3 , E4 and E5 represent the errors between estimated and measured joint torque from Fig. 10 (a), (b), (c), (d), (e) respectively.

correlation coefficient, i.e., 99.90 (-0.01∼+0.01)%, shows that estimated joint angle is highly consistent with measured joint angle, and the average RMSE and NRMSE between estimated and measured joint angle are about 3.647 (-0.1682∼+0.2105)◦ and 2.430 (-0.1340∼+0.1640)% respectively. However, there are still some errors when mechanical arm performs sinusoidal movement in variable frequencies and amplitudes, and corresponding errors between estimated and measured joint angle are displayed on Fig. 12(b) and 12(d). Furthermore, errors at peak and valley are largest, and errors rise with increasing frequencies. The average RMSE in 1/13 Hz, 1/10 Hz, and 1/6 Hz are 2.983◦ , 4.078◦ and 4.307◦ . The maximum error band is 15.10◦ . These errors mainly come from the Mahony algorithm. In this algorithm, the accelerometer is used to

measure the gravity vector. However, due to the acceleration or deceleration during the movement, the gravity vector measured by accelerometer is offset and the orientation solution errors resulting in differences between estimated and measured joint angle occur. Fig. 12(e) describes the relationship between θd , i.e., measured joint angle, and θex , i.e., estimated joint angle. It is almost a straight line, indicating that estimated joint angle shows good performance. In Fig. 12(a), there are many overlaps among lines θea1 , θea2 and θea3 , and the deviation of estimated joint angle in different orientations with same position is very small. More specifically, the average RMSE and NRMSE are 2.124 (-0.4402∼+0.4673)◦ , 1.402 (-0.2900∼+0.3090)%. As shown in Table. II, the RMSE and NRMSE between estimated and measured joint angle are almost the same. The experimental results indicate that the angle solution algorithm is insensitive to orientations of IMUs. Similarly, considering the overlaps in Fig. 12(c), the close data from row of Table. II and the

obtained average RMSE and NRMSE which are 1.258 (0.1621∼+0.2205)◦ , 0.8456 (-0.1087∼+0.1485)%, the angle solution algorithm is insensitive to positions of IMUs. In general, Fig. 12 reveals that the proposed algorithm can estimate joint angle very well, although there are still some errors when arms turn around. And its insensitivity to different positions and orientations is suitable for soft exoskeleton suit. TABLE II

CORRELATION COEFFICIENT, RMSE AND NRMSE BETWEEN ESTIMATED AND MEASURED JOINT ANGLE IN THREE DIFFERENT POSITIONS AND ORIENTATIONS

correlation coefficient (%) RMSE (deg) NRMSE (%) Orientation 1

Error (deg) Position (deg) Error (deg) Position (deg)

Orientation 2

300

3000

(a)

θd

200

θ ea1

θ ea2

θ ea3

100

Orientation 3

0

0 10 (b)

5

10

15 Time (s)

20

E ea1

25

E ea2

Position 1

Position 2

Position 3

99.89 3.659 2.458 99.90 4.128 2.713 99.89 3.726 2.496

99.90 3.848 2.567 99.91 3.502 2.320 99.89 3.865 2.534

99.91 3.456 2.324 99.91 3.675 2.418 99.89 3.696 2.491

E ea3

0

C. Power-assist Movement Performance

-10 5

(c)

10

200

15 Time (s) θ

20 d

θ ep1

25 θ ep2

θ ep3

100 0 0 10 (d)

5

10

15 Time (s)

20

5

10

15 Time (s)

20

80

100 120 θd (deg)

140

E ep1

25

E ep2

E ep3

0 -10 0

25

θex (deg)

150 (e) 100 50 20

40

60

160

180

Fig. 12. The performance of estimated joint angle strategy in different positions and orientations. The blue solid line θd denotes the actual angle measured by joint potentiometer, the lines θea1 , θea2 , θea3 are estimated joint angle from IMUs in three different orientations with same position, and the lines θep1 , θep2 , θep3 are estimated joint angle in three different positions with same orientation. The lines Eea1 , Eea2 , Eea3 show errors between measured and estimated joint angle in different orientations, and the lines Eep1 , Eep2 , Eep3 show errors in different locations. The axis θex denotes estimated joint angle base on IMUs. (a) The data of measured and estimated angle in three different orientations with same position. (b) Errors between two signals in different orientations. (c) The data of measured and estimated angle in three different positions with same orientation. (d) Errors between two signals in different positions. (f) The relationship between measured and estimated joint angle.

The results of sEMG-based torque estimation control strategy are shown in Figs. 13-15. The integral of sEMG signals and efficiency of assistance are shown in Table. III. Figs. 13(a), 14(a), 15(a) show the contrast of estimated joint torque in assistance and non-assistance conditions under different loads. The light load is 2.5 kg so that subjects can easily lift it, the acceptable heavy load is 5 kg, and the challenging heavy load is 7.5 kg. Thus the results can describe the efficiency of control strategy in different loads which represent different applications. When subjects lift up and down the dumbbells, there are two processes of muscle exerting force which can be observed from first chart of Figs. 13-15. The first half waveform is higher than the second one, and it suggests that muscular force at lift up stage is greater than the force at lift down stage. Moreover, the results show that the biceps make more contributions during elbow contraction. Whether at the lift up or down stage, the dash blue line is lower than solid red line, and it means there is always assistance in the whole process of lifting dumbbells. The estimated joint torque, which is continuous control signal, is directly used to drive soft exoskeleton system. There is an interaction between human and robot system during movement, and the estimated joint torque has a series of shaking due to exerting force intermittently in the whole process. Figs. 13(b), 14(b), 15(b) describe the contrast of raw sEMG signals in assistance and non-assistance conditions under different loads. The errors of estimated joint torque between assistance and non-assistance conditions are shown in the third chart of Figs. 13-15. As loads increase, joint torque errors under heavy loads are pretty larger

Torque (Nm) sEMG (V) Error (Nm)

50

(a)

0 50 (b)

Mr

Mex

5

10

15 20 Time (s)

25

30

5

10

15 20 Time (s)

25

30

35

5

10

15 20 Time (s)

25

30

35

Ur

Uex

35

0 -5 0 (c) 20 0 -20

0

Error (Nm)

sEMG (V)

Torque (Nm)

Fig. 13. The results of assistance experiment with sEMG-based torque estimation control strategy under 2.5 kg load condition. The lines Mex and Mr denote the estimated joint torque with and without assistance, the lines Uex and Ur are the raw sEMG signals. (a) Comparison of estimated joint torque with and without assistance in 2.5 kg load. (b) Comparison of sEMG signals with and without assistance. (c) Errors between estimated joint torque under assistance and nonassistance conditions.

50

(a)

0 50(b)

Mr

Mex

5

10

15 20 Time (s)

25

30

0 (c)

5

10

15 20 Time (s)

25

30

35

0

5

10

15 20 Time (s)

25

30

35

Ur

Uex

35

0 -5 20 0 -20

sEMG (V)

Torque (Nm)

Fig. 14. The results of assistance experiment with sEMG-based torque estimation control strategy under 5 kg load condition.

Error (Nm)

than errors under light load. The average RMSE of estimated joint torque between assistance and non-assistance are 18.72 Nm, 32.43 Nm and 27.77 Nm respectively. It indicates that performance of assistance has a growing tendency with loads increasing. The comparison of efficiency in different loads is shown in Table. III. The performance under light load, i.e., 26.87%, is not very satisfying due to that muscular activation level is low when subjects lift light load, and the control strategy does not have enough input to supply the powerful assistance. The efficiency under acceptable heavy load, i.e., 42.66%, is slightly higher than the efficiency under challenging heavy load, i.e., 41.70%. The main reason is that the output torque and speed are limited into reasonable range for the sake of security, causing the control system to not perform fully when subjects lift heavy load. The proportional control strategy is also carried out to contrast, and results are shown in Figs. 16-18. Similarly, this strategy has good performance in assistance. The efficiency is shown in Table. III. The contrast of estimated joint torque with and without assistance under different loads are shown in the first chart of Figs. 16-18. Compared to the proportional control strategy, the proposed strategy is slightly worse under light load, and it is better under heavy loads. The efficiency of proportional control strategy is 29.81%, 35.01% and 33.18% orderly. Furthermore, from the Figs. 16(a), 17(a), 18(a), estimated joint torque at lift down stage sharply decline to a low level under different loads, causing better assistance results. The main reason is that the proportional control strategy does not have closed position loop, so instead of providing assistance for subjects according to muscular activation level, it supplies the constant assistance at lift down stage. Compared to the Figs. 16(a), 17(a), 18(a), the proposed control strategy has better performance at lift up stage and worse performance at lift down stage, and the efficiency of proposed control strategy is better than proportional control strategy. It indicates that the better assistance performance at lift up stage makes higher efficiency and better assistance experience for wearer. The comparison of raw sEMG signals in assistance and nonassistance conditions under different loads is displayed in Figs. 16(b), 17(b), 18(b), and the errors of estimated joint torque between assistance and non-assistance conditions are shown in the third chart of Figs. 16-18. The sEMG power is calculated by integral of absolute value of raw sEMG signals. The comparison of sEMG power in different control strategies is shown in Fig. 19. Compared with the data of the cases without assistance, the sEMG power is significantly reduced with the proposed control strategy. The average sEMG power decreased by 2.949 V·s for 2.5 kg, 7.673 V·s for 5 kg and 10.30 V·s for 7.5 kg. Compared to the proportional control strategy, in the light load application, the average sEMG power of proposed control strategy is slightly higher than proportional control strategy by 0.3110 V·s, and in acceptable and challenging heavy loads applications, the average sEMG power of proportional control strategy is lower by 1.376 V·s and 2.106 V·s. It indicates that the proposed control strategy has better performance in acceptable and challenging heavy loads applications. The comparison of assistance efficiency in different fre-

50

(a)

0 50 (b)

Mr

Mex

5

10

15 20 Time (s)

25

30

0 (c)

5

10

15 20 Time (s)

25

30

35

0

5

10

15 20 Time (s)

25

30

35

Ur

Uex

35

0 -5 20 0 -20

Fig. 15. The results of assistance experiment with sEMG-based torque estimation control strategy under 7.5 kg load condition.

(a)

0 50 (b)

Mr

Mex

5

10

15 20 Time (s)

25

30

0 (c)

5

10

15 20 Time (s)

25

30

35

0

5

10

15 20 Time (s)

25

30

35

Ur

uex

35

0 -5 20 0 -20

Error (Nm)

sEMG (V)

Torque (Nm)

Fig. 16. The results of assistance experiment with proportional control strategy under 2.5 kg load condition. The lines Mex and Mr denote the estimated joint torque with and without assistance, the lines Uex and Ur are the raw sEMG signals. (a) Comparison of estimated joint torque with and without assistance in 2.5 kg load. (b) Comparison of sEMG signals with and without assistance. (c) Errors between estimated joint torque under assistance and non-assistance conditions.

50

(a)

0 50 (b)

Mr

5

10

15 20 Time (s)

25

30

5

10

15 20 Time (s)

25

30

Ur

Mex

Uex

Torque (Nm)

Pr

30

P ex1

P ex2

20 10 0

2.5

5 Weight (KG)

7.5

35

35

AVERAGE S EMG POWER AND EFFICIENCY OF ASSISTANCE IN DIFFERENT STRATEGIES

0 TABLE III

(c) 0 (c) 20 0 0

5

10

15 20 Time (s)

25

30

35

Fig. 17. The results of assistance experiment with proportional control strategy under 5 kg load condition.

sEMG (V)

40

Fig. 19. Comparison of sEMG power under non-assistance and assistance conditions with different control strategies in different load. The bars Pr describes the sEMG power calculated by integral without assistance. Pex1 , Pex2 denote the sEMG power with assistance in sEMG-based torque estimated control strategy and proportional control strategy respectively.

-5

-20

Error (Nm)

quencies with 5 kg load is shown in Table. IV. From the Table, we can see that the proposed method and conventional method both were affected in dynamical situation. The average efficiency of the proposed method and conventional method all vary with the movement frequency. However, the performance of the proposed strategy is always better than that of the conventional method. The average efficiency of the proposed method is decreased by 30%, and conventional method is decreased by 45.47%. The main reason is that the proposed method uses position control to realize the faster tracking performance than force control. Therefore, the proposed method can provide more assistance. sEMG Power (V—s)

Torque (Nm) sEMG (V) Error (Nm)

50

50

(a)

Mr

Pr without assistance (V·s) Pe with STECS (V·s) Pe with PCS (V·s) η with STECS (%) η with PCS (%)

2.5 kg

5 kg

7.5 kg

10.94 7.988 7.677 26.87 29.81

17.99 10.31 11.69 42.66 35.01

24.71 14.42 16.51 41.70 33.18

Mex

TABLE IV

0 50(b)

THE EFFICIENCY IN DIFFERENT FREQUENCIES

5

10

15 20 Time (s)

25

30 Ur

Uex

35

Frequency

0 -5 0 20 (c)

5

10

15 20 Time (s)

25

30

35

5

10

15 20 Time (s)

25

30

35

0 -20

0

Fig. 18. The results of assistance experiment with proportional control strategy under 7.5 kg load condition.

1/2Hz 1/4Hz 1/6Hz 1/8Hz 1/12Hz

η with STECS (%)

η with PCS (%)

32.23 33.50 26.55 27.16 42.66

16.45 26.98 16.98 15.95 35.01

In general, Figs. 13-19 reveal that the sEMG-based torque estimation control strategy has good performance to supply assistance under different loads according to human motion intention. The comparison experiments also indicate that the

efficiency value of proposed control strategy is higher than that of the proportional control strategy under heavy load condition. Moreover, due to introducing position control close loop, the proposed strategy is more stable than the proportional control strategy which makes use of force control. Furthermore, with the position control strategy, the subjects can change the assistance or non-assistance statement freely by control the activation level of muscles at any position. D. Future Application and Limitations This study proposed a sEMG-based torque estimation control strategy fusing several signals from sEMG, IMU and encoder to provide power-assist for users according to human motion intention. Although the subjects in these experiments were physically healthy and energetic, the soft exoskeleton assistance system can be extended to rehabilitation system for stroke patients by providing various rehabilitation training strategies. Moreover, the method estimating joint torque can be used to evaluate the rehabilitation effect, and the method calculating elbow angle can be extended to estimate other joint angle. However, there are still some limitations in proposed methods. The sEMG signals are highly sensitive to interference. Although there are various filters to remove noise, the estimated joint torque is still sensitive to strong electromagnetic interference. Moreover, the dynamical performance of the proposed method is not as good as static performance. In addition, the linear relationship between the estimated joint torque and incremental angle is a simple estimation. In the future work, we will develop a higher-order approximation strategy for the real-time adaptive adjustment of coefficients. IV. C ONCLUSION In this paper, a continuous sEMG-based torque estimation control strategy fusing sEMG, IMUs and encoder signals was developed and applied to the elbow soft exoskeleton suit successfully. The methods and strategies involved were provided and discussed in detail, and the corresponding experiments were also conducted to verify the efficacy of methods proposed. The average efficiency of assistance provided by elbow soft exoskeleton suit is as high as 42.66% under heavy load. Experimental results suggest that the proposed power-assist control strategy has good performance in elbow assistance and it is suitable for elbow soft exoskeleton suit. Besides, this strategy has potential to be applied for the active rehabilitation training for stroke patients. R EFERENCES [1] F. Giovacchini, F. Vannetti, M. Fantozzi, M. Cempini, M. Cortese, A. Parri, T. Yan, D. Lefeber, and N. Vitiello, “A light-weight active orthosis for hip movement assistance,” Robotics And Autonomous Systems, vol. 73, pp. 123–134, 2015. [2] N. Aliman, R. Ramli, and S. M. Haris, “Design and development of lower limb exoskeletons: A survey,” Robotics And Autonomous Systems, vol. 95, pp. 102–116, 2017. [3] Q. Wu, X. Wang, B. Chen, and H. Wu, “Development of a minimalintervention-based admittance control strategy for upper extremity rehabilitation exoskeleton,” IEEE Transactions on Systems Man CyberneticsSystems, vol. 48, no. 6, pp. 1005–1016, 2018.

[4] Q. Wu, X. Wang, and F. Du, “Development and analysis of a gravitybalanced exoskeleton for active rehabilitation training of upper limb,” Proceedings Of the Institution Of Mechanical Engineers Part C-Journal Of Mechanical Engineering Science, vol. 230, no. 20, pp. 3777–3790, 2016. [5] B. S. Rupal, S. Rafique, A. Singla, E. Singla, M. Isaksson, and G. S. Virk, “Lower-limb exoskeletons: Research trends and regulatory guidelines in medical and non-medical applications,” International Journal Of Advanced Robotic Systems, vol. 14, no. 6, 2017. [6] L. Zhou, Y. Li, and S. Bai, “A human-centered design optimization approach for robotic exoskeletons through biomechanical simulation,” Robotics And Autonomous Systems, vol. 91, pp. 337–347, 2017. [7] S. Lee, S. Crea, P. Malcolm, I. Galiana, A. Asbeck, and C. Walsh, “Controlling negative and positive power at the ankle with a soft exosuit,” in 2016 IEEE International Conference on Robotics And Automation, pp. 3509–3515, 2016. [8] A. T. Asbeck, S. M. M. De Rossi, I. Galiana, Y. Ding, and C. J. Walsh, “Stronger, smarter, softer next-generation wearable robots,” IEEE Robotics AND Automation Magazine, vol. 21, no. 4, pp. 22–33, 2014. [9] Y. Ding, I. Galiana, A. T. Asbeck, S. Marco, M. De Rossi, J. Bae, T. R. Teles Santos, V. L. de Araujo, S. Lee, K. G. Holt, and C. Walsh, “Biomechanical and physiological evaluation of multi-joint assistance with soft exosuits,” IEEE Transactions on Neural Systems And Rehabilitation Engineering, vol. 25, no. 2, pp. 119–130, 2017. [10] A. T. Asbeck, S. M. M. De Rossi, K. G. Holt, and C. J. Walsh, “A biologically inspired soft exosuit for walking assistance,” International Journal Of Robotics Research, vol. 34, no. 6, pp. 744–762, 2015. [11] M. Wehner, B. Quinlivan, P. M. Aubin, E. Martinez-Villalpando, M. Baumann, L. Stirling, K. Holt, R. Wood, C. Walsh, and Ieee, “A lightweight soft exosuit for gait assistance,” in 2013 IEEE International Conference on Robotics And Automation, pp. 3362–3369, 2013. [12] A. Villoslada, A. Flores, D. Copaci, D. Blanco, and L. Moreno, “Highdisplacement flexible shape memory alloy actuator for soft wearable robots,” Robotics And Autonomous Systems, vol. 73, pp. 91–101, 2015. [13] S. Sridar, P. H. Nguyen, M. Zhu, Q. P. Lam, and P. Polygerinos, “Development of a soft-inflatable exosuit for knee rehabilitation,” in 2017 IEEE/RSJ International Conference on Intelligent Robots And Systems, pp. 3722–3727, 2017. [14] K. Lee, D. Liu, L. Perroud, R. Chavarriaga, and J. d. R. Millan, “A braincontrolled exoskeleton with cascaded event-related desynchronization classifiers,” Robotics And Autonomous Systems, vol. 90, pp. 15–23, 2017. [15] D. Ao, R. Song, and J. Gao, “Movement performance of human-robot cooperation control based on emg-driven hill-type and proportional models for an ankle power-assist exoskeleton robot,” Ieee Transactions on Neural Systems And Rehabilitation Engineering, vol. 25, no. 8, pp. 1125–1134, 2017. [16] V. Grosu, S. Grosu, B. Vanderborght, D. Lefeber, and C. RodriguezGuerrero, “Multi-axis force sensor for human-robot interaction sensing in a rehabilitation robotic device,” Sensors, vol. 17, no. 6, 2017. [17] Q. Wu, X. Wang, F. Du, and R. Xi, “Modeling and position control of a therapeutic exoskeleton targeting upper extremity rehabilitation,” Proceedings Of the Institution Of Mechanical Engineers Part C-Journal Of Mechanical Engineering Science, vol. 231, no. 23, pp. 4360–4373, 2017. [18] Q. Meng, Q. Meng, H. Yu, X. Wei, and Ieee, “A survey on semg control strategies of wearable hand exoskeleton for rehabilitation,” in 2017 2nd Asia-Pacific Conference on Intelligent Robot Systems, pp. 165– 169, 2017. [19] F. Hug, T. Gallot, S. Catheline, and A. Nordez, “Electromechanical delay in biceps brachii assessed by ultrafast ultrasonography,” Muscle AND Nerve, vol. 43, no. 3, pp. 441–443, 2011. [20] F. Duan, L. Dai, W. Chang, Z. Chen, C. Zhu, and W. Li, “semgbased identification of hand motion commands using wavelet neural network combined with discrete wavelet transform,” IEEE Transactions on Industrial Electronics, vol. 63, no. 3, pp. 1923–1934, 2016. [21] A. Lopez-Delis, D. Delisle-Rodriguez, A. C. Villa-Parra, T. BastosFilho, and Ieee, “Knee motion pattern classification from trunk muscle based on semg signals,” in 2015 37th Annual International Conference Of the Ieee Engineering In Medicine And Biology Society, pp. 2604– 2607, 2015. [22] X. Li, Q. Huang, J. Zhu, W. Sun, and H. She, “A novel proportional and simultaneous control method for prosthetic hand,” Journal Of Mechanics In Medicine And Biology, vol. 17, no. 8, 2017. [23] T. Lenzi, S. M. M. De Rossi, N. Vitiello, and M. C. Carrozza, “Intentionbased emg control for powered exoskeletons,” IEEE Transactions on Biomedical Engineering, vol. 59, no. 8, pp. 2180–2190, 2012.

[24] S. Kwon, Y. Kim, and J. Kim, “Movement stability analysis of surface electromyography-based elbow power assistance,” IEEE Transactions on Biomedical Engineering, vol. 61, no. 4, pp. 1134–1142, 2014. [25] R. Song, K.-y. Tong, X. Hu, and L. Li, “Assistive control system using continuous myoelectric signal in robot-aided arm training for patients after stroke,” IEEE Transactions on Neural Systems And Rehabilitation Engineering, vol. 16, no. 4, pp. 371–379, 2008. [26] Q. Wu, X. Wang, L. Chen, and F. Du, “Transmission model and compensation control of double-tendon-sheath actuation system,” IEEE Transactions on Industrial Electronics, vol. 62, no. 3, pp. 1599–1609, 2015. [27] Z. Tang, H. Yu, and S. Cang, “Impact of load variation on joint angle estimation from surface emg signals,” IEEE Transactions on Neural Systems And Rehabilitation Engineering, vol. 24, no. 12, pp. 1342–1350, 2016. [28] J. R. Potvin and S. H. M. Brown, “Less is more: high pass filtering, to remove up to 99muscle force estimates,” Journal Of Electromyography And Kinesiology, vol. 14, no. 3, pp. 389–399, 2004. [29] Z. Li, B. Wang, F. Sun, C. Yang, Q. Xie, and W. Zhang, “semg-based joint force control for an upper-limb power-assist exoskeleton robot,” IEEE Journal Of Biomedical And Health Informatics, vol. 18, no. 3, pp. 1043–1050, 2014. [30] B. Hudgins, P. Parker, and R. N. Scott, “A new strategy for multifunction myoelectric control,” IEEE transactions on bio-medical engineering, vol. 40, no. 1, pp. 82–94, 1993. [31] H.-T. Chang, L.-W. Cheng, and J.-Y. Chang, “Development of imu-based

[32]

[33]

[34]

[35]

[36]

angle measurement system for finger rehabilitation,” in Proceedings Of 2016 23rd International Conference on Mechatronics And Machine Vision In Practice, pp. 196–201, 2016. Z. C. Ong, Y. C. Seet, S. Y. Khoo, and S. Noroozi, “Development of an economic wireless human motion analysis device for quantitative assessment of human body joint,” Measurement, vol. 115, pp. 306–315, 2018. R. V. Vitali, S. M. Cain, R. S. McGinnis, A. M. Zaferiou, L. V. Ojeda, S. P. Davidson, and N. C. Perkins, “Method for estimating three-dimensional knee rotations using two inertial measurement units: Validation with a coordinate measurement machine,” Sensors, vol. 17, no. 9, 2017. P. Gui, L. Tang, S. Mukhopadhyay, and Ieee, “Mems based imu for tilting measurement: Comparison of complementary and kalman filter based data fusion,” in Proceedings Of the 2015 10th IEEE Conference on Industrial Electronics And Applications, pp. 1998–2003, 2015. T. Beravs, P. Rebersek, D. Novak, J. Podobnik, and M. Munih, “Development and validation of a wearable inertial measurement system for use with lower limb exoskeletons,” in 2011 11th IEEE-RAS International Conference on Humanoid Robots, pp. 212–17, 2011. J. G. Ziegler and N. B. Nichols, “Optimum settings for automatic controllers,” Transactions of the ASME. Journal of Dynamic Systems, Measurement and Control, vol. 115, no. 2B, pp. 220–2, 1993.

Lon nghai Lu recceived the B.S. B degree in mechatro onics engineeering from Soocchow Univerrsity, Suzhou,, China, in 20016. He is cu urrently a Ph..D. student undeer the superviision of Profeessor Bai Cheen in the College of Mech hanical and Elecctrical Engineeering, Nanjiing Universitty of Aeronaautics and Asstronautics, Nanjjing, China. His major research intterests inclu ude robotics, nonlinear conttrol, embeddeed system, sE EMG analysiss, and the app plication of exoskeleton to neeuromuscularr rehabilitatio on.

Qinggcong Wu received th he B.S. andd Ph.D. deg grees in meechatronics engiineering from m Southeast University, N Nanjing, China, in 2011 and 2016, respectively. He is currently y an assistaant professorr with the College C of Mecchanical and Electrical E Eng gineering, Naanjing Univerrsity of Aeron nautics and Astrronautics, Naanjing, Chinaa. His major research inteerests include robotics, nonllinear controll, tendon-sheaath transmiss ion theory, grravity balancing theory, and the applicatioon of exoskelleton to neuroomuscular reh habilitation.

Xi C Chen receiveed the B.S. deegree in mechhatronics eng gineering from Nanjing Univversity of Aeeronautics an nd Astronautiics, Nanjing, China, in 20 017. He is curreently a Ph.D. student und der the supervvision of Pro ofessor Bai Chen C in the Colllege of Mechanical and Electrical E Engineering, Nanjing Uniiversity of Aeroonautics and Astronauticss. His major research interests include robotics, brainn-computer innterface, and intelligent coontrol.

ZiYaan Shao receeived the B.S.. degree in m mechatronics engineering e frrom Yantai Univversity, Shanndong, China, in 2013 annd the M.S. degree from Yangzhou Univversity, Yanggzhou, China,, in 2016. Hee is currently y a Ph.D. student under the supervision of Professorr Bai Chen iin the Colleege of Mechanical and Elecctrical Engineeering, Nanjiing Universitty of Aeronaautics and Asstronautics, Nanjjing, China. His major reesearch intereests include robotics, tend don-sheath transsmission theoory, robot kin nematics, annd the applicaation of exosskeleton to neurromuscular reehabilitation.

Bai Chen receiveed the B.S. deegree and Ph .D. degree fro om Zhejiang University, Hangzhou, Chinna, in 2000 0 and 20055, respectiveely, all in mechanical engiineering. Currrently, he is a full professsor in the Colllege of Mech hanical and Elecctrical Engineeering at Nan njing Universsity of Aeronautics and Astronautics.

His current reseearch interestts include m minimally inv vasive neurossurgery roboot, virtual surgery systeem, force feedback contro ol, interventioonal therapy.

Hon ngtao Wu recceived the B.S S. degree from m Yanshan University, U Heebei, China, in 19982, and the M.S. M degree and a Ph.D. deegree from Tiaanjin Universsity, Tianjin, Chinna, in 1985 annd 1992, resp pectively, all iin mechanicaal engineering g. Currently, he iss a full profeessor in the College C of Meechanical and d Electrical Engineering E at N Nanjing University of Aerronautics andd Astronauticcs. His curreent research interrests include parallel p robott, robot kinem matics, multib body system dynamics. d

Highlights 

A soft elbow exoskeleton suit is developed to provide effective power assistance.



A sEMG-based elbow joint torque estimation control strategy is proposed.



The proposed control strategy is more efficient and stable than proportional strategy.



The average efficiency of power assistance with heavy load is about 42.66%.