Abstract
Purpose
The purpose of this paper is to propose a seamless active interaction control method integrating electromyography (EMG)-triggered assistance and the adaptive impedance control scheme for parallel robot-assisted lower limb rehabilitation and training.
Design/methodology/approach
An active interaction control strategy based on EMG motion recognition and adaptive impedance model is implemented on a six-degrees of freedom parallel robot for lower limb rehabilitation. The autoregressive coefficients of EMG signals integrating with a support vector machine classifier are utilized to predict the movement intention and trigger the robot assistance. An adaptive impedance controller is adopted to influence the robot velocity during the exercise, and in the meantime, the user’s muscle activity level is evaluated online and the robot impedance is adapted in accordance with the recovery conditions.
Findings
Experiments on healthy subjects demonstrated that the proposed method was able to drive the robot according to the user’s intention, and the robot impedance can be updated with the muscle conditions. Within the movement sessions, there was a distinct increase in the muscle activity levels for all subjects with the active mode in comparison to the EMG-triggered mode.
Originality/value
Both users’ movement intention and voluntary participation are considered, not only triggering the robot when people attempt to move but also changing the robot movement in accordance with user’s efforts. The impedance model here responds directly to velocity changes, and thus allows the exercise along a physiological trajectory. Moreover, the muscle activity level depends on both the normalized EMG signals and the weight coefficients of involved muscles.
Keywords
Citation
Meng, W., Liu, Q., Zhou, Z. and Ai, Q. (2014), "Active interaction control applied to a lower limb rehabilitation robot by using EMG recognition and impedance model", Industrial Robot, Vol. 41 No. 5, pp. 465-479. https://doi.org/10.1108/IR-04-2014-0327
Publisher
:Emerald Group Publishing Limited
Copyright © 2014, Emerald Group Publishing Limited
1. Brief introduction
According to the official statistical data from the United Nations, the proportion of the world’s population aged over 60 years will be doubled from 11 to 22 per cent between 2000 and 2050. Meanwhile, there are about 650 million disabled people worldwide, accounting for around 10 per cent of the total world population. Currently, there is a considerable increase in the needs of healthcare and rehabilitation, especially among the elderly or disabled people (Zhou et al., 2013). Rehabilitation and medical robotics can not only liberate the therapists from heavy training missions but also help the patients perform scientific and repetitive training to provide a better motor function recovery. Because electromyography (EMG) signals contain much information of the muscle activity and can imply people’s movement action 30-100 ms in advance (Fan and Yin, 2013), they are widely used in clinical diagnosis, rehabilitation, prosthetic control and human–robot interaction. Benitez et al. (2013) gave an overview of different forms of therapy and explained the benefits of robotic-aided rehabilitation by monitoring and analyzing the bio-signals from patients. Krebs et al. (2003) described a concept of performance-based progressive robot therapy that utilized EMG thresholds to initiate the robot assistance. In the research, the EMG signals in 14 muscles of the upper limb were collected, and the assistance was triggered when the muscles’ activity increased above a threshold. This EMG-triggered assistance encourages self-initiated movement by patients, and this means that the patient has to apply some voluntary efforts to obtain a movement supported by the robot. However, this approach usually breaks the movement into two separate phases, an active phase driven by the patient and a passive phase driven by the robot (Marchal-Crespo and Reinkensmeyer, 2009). Furthermore, when the robot is driven to provide assistance in the second phase, typically passive controllers will be used to achieve the necessary movement, so the patient is not in a fully compliant environment when assistance is provided (Wolbrecht et al., 2008). In such EMG-triggered mode, the robot would operate with a predefined trajectory after being activated, which had no interaction with the human limbs during this period until the time allowed for the next trigger event (Song et al., 2008).
To increase patient’s interactivity during the rehabilitation, the patient-cooperative strategy should be combined with approaches that adapt the robot assistance according to the intention of subjects. DiCicco et al. (2004) presented an orthotic exoskeleton controlled by EMG signals with threshold approach, which means the robot can be only activated when the normalized EMG signals are above a certain threshold. However, as explained above, this method may divide the robot assistance into two independent sub-tasks. Therefore, the self-initiated movement support and patient-cooperative control strategies must be integrated into a robot-assisted rehabilitation system. Benitez et al. (2013) proposed a set of dynamic models to compute the robot triggering function and the assistance level people needed by using the muscular EMG signals and the sensor information from robot. The models can be integrated into an active 1-degree of freedom (DOF) elbow orthosis system, which provides a promising solution to such active rehabilitation problems. On the other hand, Song et al. (2013) proposed a continuous robot assistance method to provide robot-assistive torque proportional to the amplitude of EMG signals. However, the relationship between EMG and joint torque was only simplified as a linear model. Many researchers utilized EMG signals to estimate the joint torque and applied the estimated torque to the robot-assisted movement. Lenzi et al. (2012) studied a torque estimation-based EMG control method for powered exoskeletons, and the assistance was provided through a proportional EMG controller. These methods could provide effective support to the user. However, their usability was strongly limited by the application environment, and the estimation of user muscular torque was quite rough. The development of pattern classification can help patients perform rehabilitation with an EMG-controlled robotic system. A human-assisting manipulator operated by EMG signals and arm motions was proposed by Fukuda et al. (2003), where an active joint in the manipulator’s end-effector was controlled based on an EMG pattern discrimination. So, the robot can adapt itself to the changes of EMG patterns according to the differences among individuals. Similarly, Kiguchi and Hayashi (2012) proposed an EMG-based method to control an upper-limb robot according to user’s intention, in which, however, 16 channels of EMG signals were captured to estimate the upper limb joint torque vector. To have a better recognition accuracy, Ju et al. (2013) investigated nonlinear EMG features and classifiers, which made the robot system more reliable (Ouyang et al., 2014).
The potential problem with EMG-triggered passive robot training is that it does not consider patient’s voluntary participation or muscular efforts, thus reducing the capabilities of the system to exercise (Riener et al., 2006). The patient’s muscle activity level can be reflected by EMG signals, while the voluntary participation is closely related to the interaction force between the patient and the robot. Several control strategies have been developed to provide robotic assistance according to the patient’s disability level and the patient’s voluntary participation in the training process. In most literature, the subject’s muscle activation level is evaluated by using root mean square (RMS) of EMG signals collected from related muscles. However, the level of EMG readings is dependent on the skin impedance at the electrodes location, which may vary between different sessions (Mobasser et al., 2007). So to compensate this variability, the EMG data have to be normalized. Song et al. (2008) used the maximum voluntary contraction method to normalize EMG signals. In fact, using this method to calculate the muscle activity level is problematic, as there will be very different contributions of each muscle when conducting different motion patterns (Ji and Liu, 2010). Thus, the weights of constriction muscles of the limb have to be reconsidered when evaluating patient’s muscular strength. On the other hand, the human robot interaction control is mostly achieved by the use of impedance control. It is widely accepted that the behavior of the robot should be adjusted according to patient’s recovery conditions, which can be reflected from muscle activities (Ai et al., 2014). The basis of adaptive impedance assistance is to modify the robot motion in a way that is desired by the patient, which is believed to be most appropriate for rehabilitation. However, the issue of reference trajectory adaptation has some drawbacks, for example, the extent of trajectory adaptation can not be well determined, and the changes may result in an un-physiological motion pattern. To tackle this problem, the robot assistance speed can be adaptable (Duchaine and Gosselin, 2007). In this situation, the physiological trajectory can be followed, while the change in the motion velocity influenced by the patient can also be obtained in the meantime. Therefore, it may provide a better opportunity for the patient to actively contribute muscular efforts during the exercise compared to the trajectory control.
Interactive rehabilitation and training encourages a patient’s active participation and may aid in rapid motor function recovery (Liu, 2011). In this paper, an active interaction controller is built by the fusion of EMG signals (reflecting human movement intention) and the force sensor (providing human–robot interaction information). To obtain the surface EMG signals from lower limb muscles more conveniently and effectively, a wireless EMG acquisition device based on a Wi-Fi protocol is developed. An effective motion recognition controller based on autoregressive (AR) features and support vector machine (SVM) classifier is established to predict lower limb motion intention in advance and trigger the robot assistance. To allow the user to interact with the robot and take more efforts during the rehabilitation, an impedance controller is designed to make the robot movement speed adaptable to people’s participation efforts. Furthermore, the impedance controller takes into account the user’s muscle activity levels, which are evaluated from EMG signals by considering the contributions of each muscle in different motion patterns, to adapt the robot compliance accordingly. The proposed active interaction controller allows the patients to determine the trajectory pattern by recognizing EMG signals and influence the speed of their leg movements along a physiological trajectory during the whole practice.
2. A lower limb rehabilitation robot
The mechanical design of the robot is the basis of robot-assisted rehabilitation system. In general, there are two types of rehabilitation robots: end-effector robots and exoskeleton-type robots. The exoskeleton robots usually have to be worn by the patient, and such robots have drawbacks of inferior adaptability to different patients. On the contrary, the end-effector robots usually contact with the patient body at a certain point, making this type of robot easy to design and control. Although a number of robots show great prospect in the rehabilitation of the lower limbs, they have not yet been widely applied to the clinical rehabilitation (Yin et al., 2012). To avoid complex dynamics problems such as coupling effects caused by multi-DOFs, the human–machine interface by Yin et al. (2012) was applied to knee controlling with only 1-DOF joint extension. However, this structure cannot meet the requirement of lower limb movement in multi-DOFs. Recently, parallel robots have drawn a lot of interests in the robotic community due to their superiority over the classical serial structures in terms of stiffness, accuracy and high payload. It has been found that parallel robots are better candidates for lower limb rehabilitation (Jamwal et al., 2014). Hussain et al. (2011) had also proposed a 4-axes redundant parallel robot based on the modeling and kinematics of ankle anatomy.
The lower limb rehabilitation robot designed here is also basically a parallel mechanism with six transitional and rotational DOFs. The mechanical structure is composed of a fixed base, a moving platform and six actuated limbs with kinematic chains. The Stewart platform shown in Figure 1 was designed by the authors’ research group for the purpose of investigating lower limb rehabilitation. The designed robot can be adaptable to subjects of varying lower limb abilities. Specifically, the system mainly includes an industrial PC, six motion controllers based on digital signal processor (DSPs) (model: TMS320LF2407A) and Panasonic Minas servo drivers, as well as the Stewart platform. Linear position and velocity of each actuator are measured by incremental photoelectric encoders. The controller is coded in C++ language and runs under a Windows environment. The device is interfaced to a standard personal computer (PC) through a CAN BUS interface. Geometric diagram of the Stewart platform can be seen in Figure 1(b), where the radius of the upper platform and angle are defined as r b and Θ 2 , respectively, and the parameters of the fixed platform are defined as r a and Θ 1 , respectively. The radius of the upper moving plate is 180 mm, while the radius of the fixed base plate is 270 mm, and the angles of the platforms are 28° and 22°. The robot’s range of motion is able to fully meet the requirements of lower limb rehabilitation and training.
3. Wireless EMG acquisition device based on Wi-Fi protocol
It is important for the assistive rehabilitation robot to be controlled in accordance with the user’s motion intention, which can be directly reflected by the biological signals. However, it is difficult to combine the biofeedback with the mechanical robot control (Riener et al., 2005), as the traditional EMG acquisition devices are always wired, inconvenient and non-portable. The current wireless techniques for EMG equipment usually focus on Bluetooth, Zigbee or FSK (Hussain et al., 2013). The disadvantage is that the collected EMG data can only be used for local processing. As the Wi-Fi protocol has advantages of network access ability, high transmission bandwidth and low error rate over other wireless techniques, it can be used to complete the wireless transmission of EMG signals from the acquisition devices to the processing platforms. There are several commercial Wi-Fi-based EMG acquisition devices, such as the BTS FREEEMG from BTS Bioengineering Corp. and the devices from Shanghai NCC Medical Co., Ltd. However, these commercial equipments are not selected by the authors because they are very expensive, and this will increase the cost dramatically. The real-time performance is indeed an important issue when applying EMG devices to robot control. The self-developed device can also satisfy the real-time requirement by using the Wi-Fi protocol. Considering the system compatibility when integrating it with the robot platform in our experiment, we choose to develop such a device by ourselves. The wireless EMG acquisition device designed by the researcher’s group is presented in Figure 2, in which (a) is the system architecture and (b) is the picture of the hardware device. The central unit is a high-performance monolithic processor STM32F103RC6, and the communication unit uses a WM-MR-08 module as the 802.11 b/g baseband controller. In addition, the device is also integrated within a lithium battery and a power management system. Specifically, the sampling resolution is 12 bit, with a bandwidth of 20-500 Hz. The system amplification gain is set to 60 db, wherein the preamplifier gain is 20 db and the final gain is 40 db.
4. Motion recognition and muscle activity evaluation
4.1 EMG features extraction and motion recognition
EMG signals are often utilized to control the power-assist robot, as they can directly reflect the user’s movement intention and the muscle activity levels (Kiguchi et al., 2004). In recent years, a number of methods have been proposed to extract useful information from EMG signals (Ai et al., 2013). However, it is difficult to estimate the intended motions accurately from multichannel EMGs by using a fixed classifier. The existing methods tend to be complicated or require a huge amount of signal samples; most of them for lower limb EMG signal are less than ideal and can not meet the real-time requirement. Because the coefficients of AR model contain sufficient features of EMG signals, they are selected as the features to model the motion patterns in this paper. The transfer function of AR model is defined as: Equation 1 where, p is the order of the used AR model, a i is the AR coefficients (i = 1, 2, ⋅⋅⋅, p), x(n)is the EMG sample and w(n)is the residual value of white noise.
In this paper, four-channel EMG signals are collected and utilized for lower limb’s motion recognition, so a four-order AR model is established in this context. That is: Equation 2. The canonical equation of AR model is the Yule–Walker style, which builds the relationship between the AR coefficients a i and the self-correlation matrix of x(n): Equation 3. The equations above can be solved by the Levinson–Durbin-based iterative algorithm as follows: Equation 4 Equation 5 Equation 6 Equation 7. With equations (4)–(7), the AR coefficients a i can be obtained, and they are used as the inputs to a classifier for motion recognition.
SVM has gained wide acceptance in pattern recognition fields. It is shown that SVM is superior to other machine learning methods such as back-propagation neural network, as SVM is able to gain better generalization ability for unseen data (Liu et al., 2007). In this study, a method based on the AR features of EMG in combination with an SVM classifier is proposed, which is quite effective for solving nonlinear classification problems and reducing computation time. Supposing the samples are{x i , d i }, where x i is the sample to be classified, and d i is the category of samplex i . So the optimal hyper-plane w T x i + b = 0 should meet the formula, and we need to find the minimum‖ w ‖2/2, which can be rewritten as: Equation 8. The objective function above is to find the optimal solution under inequality constraints, and this can be resolved by Lagrange functions: Equation 9 where, λ is the Lagrange multiplier, and, finally, we can obtain the classification function as: Equation 10 where, x i is the support vector, Inline Equation 1 is the Lagrange multiplier with respect to the support vector and b*is a constant. For a nonlinear classification system, the slack variable ξ ≥ 0should be introduced, and the objective function becomes: Equation 11 where, C > 0is the penalty factor which is used to find the optimal hyper-plane under nonlinear conditions by transforming them to the samples suitable for linear classification.
Considering the nonlinear samples, the kernel function is utilized in nonlinear transformation to map the nonlinear variables to the high-dimensional linear space. Based on the features of EMG signals, a radial basis function is selected as the kernel function here: Equation 12
4.2 Evaluation of muscle activity level
To provide the robot assistance based on patient’s recovery conditions, progressive rehabilitation should be considered by monitoring the patient’s muscle condition and updating the robot controller in real time. As muscle activity levels are closely related to the coordination of multiple muscles, a number of researchers have used EMG signals to evaluate patient performance during the training or exercise sessions (Akdogan et al., 2012). However, the level of EMG readings may vary between different sessions. So to compensate this variability, the EMG signals have to be normalized. Moreover, there are different contributions of each muscle when conducting different motion patterns. Therefore, the weight coefficients of involved muscles of the limb have to be considered when evaluating the patient’s muscle activity level.
The SVM classifier is utilized to predict movement intention in this paper, so the process can be divided into two stages: the training stage and the active testing stage. In the first stage, the subject is asked to keep a certain motion for a period of time, so that EMG signals can be sampled and used to train the SVM classifier. In the second stage, the subject is asked to move the limb freely and the motion can be identified by the trained classifier. The EMG signals in the first stage can be regarded as the reference criterion to identify the movement intention, and thus, it is also reasonable for that to be used for normalizing the muscle activity levels. The scheme of motion recognition and muscle activity evaluation based on EMG signals is illustrated in Figure 3. Specifically, in the training stage, for each motion pattern, the weight coefficients of four selected muscles are calculated based on the total RMS values of four-channel EMG signals: Equation 13
In the second stage, subjects are asked to perform an active movement task, and the EMG signals are used to identify the motion pattern and evaluate the muscular activity level for the immediate pattern. Because different motions activate different limb muscles, the patient’s muscle activity level is calculated by integrating the normalized RMS values and the weight coefficients (Yin et al., 2012). In other words, when subjects intend to move, the limb’s motion will be recognized by EMG signals, and as the weight coefficients of muscles contributing w i have been obtained in the training stage, so it is clear how much a muscle is involved in such a certain motion. Therefore, it is more reasonable to evaluate the muscle activity level by integrating the normalized RMS values and the weight coefficients of activated muscles: Equation 14 where, w i is the weight coefficient of channel i for the current motion pattern, and R M S i (n) presents the RMS value of the nth sampled signals of channel i and i n i t(R M S i )is the initial RMS value of channel i, which is equal to the average RMS of channel i ( Inline Equation 2) computed above. The average of RMS values of EMG signals and muscle activity levels are computed for each movement session, according to the total EMG values during this session and the period of time.
5. Adaptive impedance control for interactive rehabilitation
The motion recognition system determines the rehabilitation trajectory and parameters based on real-time evaluation and motion decoding. To provide a complement environment after the robot being triggered, flexible assistance should be provided by monitoring the interaction force between the user and the robot. Therefore, a position-based impedance model is established in this context. The robot can be controlled in admittance with the reference trajectory of the robot when the patient attempts to move. By measuring the interaction force applied by users, it is possible to compute the reference position/velocity required to render certain mass, stiffness and damping parameters (Saglia et al., 2013). The control objective of this impedance controller is to modify the robot assistance movement in each direction of the task space with user’s active interaction, so the dynamic relationship between the robot end-effector position and the interaction force needs to be established (Lopes and Almeida, 2008). To fulfill such task-space requirements, a second-order, end-effector impedance model is chosen and expressed in equation (15): Equation 15 where, M d , B d and K d are the diagonal matrices representing the desired inertia, damping and stiffness parameters, respectively; x and x d are the actual and the desired end-effector positions vectors in task space (Lopes and Almeida, 2008), respectively. F e is the interaction force the user exerts upon the robot.
In this parallel robot-aided rehabilitation program, the end-effector’s acceleration changes very slowly, so the impact of the acceleration change x¨ − x¨ d may be ignored (Xu et al., 2011). Considering the actions of desired stiffness and damping, equation (15) can be simplified as: Equation 16. The traditional impedance controller imposes fixed parameters on the patient and cannot adapt the impedance to different users. However, it is widely accepted that the rehabilitation should consider the patient’s recovery condition and update the impedance parameters in real time. Because impedance parameters of the human limb will be changed for different muscle activity levels, the robot impedance should be changed depending on the activity levels (Kiguchi and Hayashi, 2012). In the proposed controller, impedance parameters of the robot are adaptable in accordance with the active ratio of lower limb muscles to realize the natural and adaptive assist compliance. As mentioned above, to guarantee the assistive trajectory to be physiological, this impedance controller works with a constant reference trajectory and only adapts the robot velocities. The desired damping is the most important factor when an impedance model following force and velocity control is applied, and it is also one of the impedance parameters acts as a velocity feedback gain (Nagata et al., 2009). On the other hand, in the force-controlled workspace, if only a simple force feedback is applied, the response can be under-damped for an environment with high stiffness under rigid conditions (Patel and Shadpey, 2005). So we introduce a method for tuning the desired damping parameters here, and similar strategies applied in robot compliance control can also be found in related works of Cho and Park (2005), Hu et al. (2012) or Saglia et al. (2013). Note that the inertia has been set to zero, and in the case of pure damping simulation, a proper stiffness is selected that is necessary to ensure system stability. This guarantees that the user interacts with the external environment with a velocity that can be properly adjusted by the impact forces. Therefore, the viscous damping of impedance controller is the parameter to be adjusted, while the stiffness is selected properly, depending on the muscle activity levels of the lower limb. The trajectories will not be modified, and the damping parameter B d is adjusted as follows: Equation 17 where, B 0 is the initial viscous damping matrix for the initial position, C E M G is the proportional coefficient matrix of EMG effects and m a l(n) is the muscle active level evaluated by equation (14). sat(.) is a saturation function. Thus, the amount of damping parameter B d is increased when the activity level of related lower limb muscles simultaneously increased. In addition, velocities of the robot are set to be proportional to active interaction force based on the damping parameter.
The controller architecture for damping-active training is shown in Figure 4. The outer loop is an impedance control loop with a damping parameter, and the inner loop is a position/velocity controller. The damping coefficient is changed for different training resistance according to the muscle activity level, which is identified by EMG signals considering the weight coefficients of involved muscles in different motion patterns. Furthermore, the adaptive controller is proposed to adjust the impedance parameters and influence the robot speed in accordance with the patient’s interactive efforts. Specifically, a fuzzy proportional-integral-derivative (PID) adaptive controller was implemented in a joint space as a basic position/velocity controller of the 6-DOF parallel robot, and the detailed implementation of such a controller can be found in our previous work (Zhou et al., 2013).
6. Experiments and results
6.1 Experimental setup
To evaluate the ability of the integrated parallel rehabilitation robot to provide interactive rehabilitation for lower limbs, the previously described robot and controller were implemented in a preliminary experiment. Information including EMG signals, robot position, velocity and the interaction force is collected and analyzed. The surface EMG signals are acquired by the self-developed wireless EMG acquisition device based on Wi-Fi, with a 60-dB amplifier, a 20-500-Hz band pass filter, a 50-Hz notch filter, and an A/D convertor. A Futek force sensor (MTA400 tri-axial load cell, Advanced Sensor Technology, INC., USA) is mounted between the moving platform and the footplate to sense the interaction force acting between human and robot. The force signals are amplified before being sent to the PC by the Futek amplifier modules, and a low pass filter is used to tune the system to minimize the effect of noises.
The experiments were carried out with nine subjects aged from 20 to 42 years, including both male and female subjects. Six motions of lower limb that are frequently used, namely, dorsiflexion, plantarflexion, inversion, aversion, adduction and abduction (Jamwal et al., 2014) were adopted. Four-channel EMG signals from the subject’s gastrocnemius medialis (GM), tibialis anterior (TA), flexor digitorum longus (FDL) and soleus (SL) muscles were captured. The location of electrodes is shown in Figure 5(b). The surface EMG signals of these muscles were acquired in real time, and the force sensor was located to understand the human robot interaction. The subjects sat on a chair with his/her right foot constrained to the robotic device, as illustrated in Figure 5(c). Considering the safety issue, this preliminary test was performed with healthy subjects. This experiment was designed to evaluate if the proposed control scheme could trigger the robot assistance and modify the robot movement based on subject’s EMG signals and active efforts. Because the participation of healthy subjects and patients are both indicated by the data from force sensors and EMG, it is reasonable to verify the effectiveness of the proposed method by experiments on healthy subjects, and this is also accepted by many works such as Hussain et al. (2013) and Kiguchi et al. (2012).
The experimental process of the proposed method contains four main steps (Figure 6):
EMG signals acquisition and processing, with filtering and amplification: the four-channel surface EMG signals from lower limb were acquired by using the self-developed wireless EMG acquisition device, and the raw signals were transmitted to the PC for further processing via the Wi-Fi protocol. Specifically, we used Ag/AgCl bipolar electrodes with a conductive diameter of 16 mm, and the inter-electrode distance was set to 20 mm. The EMG signals were digitized by an A/D converter (sampling frequency, 1 kHz; quantization, 12 bit) after being amplified (60 dB) and filtered through a band pass filter (20-500Hz) and a notch filter (50 Hz).
Motion prediction and muscle activity level evaluation by using EMG signals: the experiment consists of two parts. In the first part, the subject was asked to perform the six motions one by one and maintain it for 15 s so that the EMG dataset can be acquired to train the SVM. The SVM classifier was implemented by C++ language. The SVM training time was less than 100 ms and can be ignored compared to the subject’s movement duration. In the second part, the trained SVM was used to recognize the movement intention online and control the robot. In this part, the subject was asked to perform continuous movements. The robot assistance will be triggered once the subject’s movement intention is identified, and the muscle activity level was calculated simultaneously based on the normalized EMG signals and weight coefficients.
Updating impedance parameters and velocity commands: an adaptive impedance control law was proposed to relate the muscle activity level to the damping parameter, so that the robot assistance speed can be adjusted in accordance with the human interaction force and the recovery conditions. The concept behind this adaptive law is to set the robotic damping low if low muscle activity level is detected from the subject. This low impedance will increase the robot speed under the same interaction force to make the subject change the training speed more easily. Similarly, the damping of the robot is adapted to be high if the subject shows a higher muscle activity level. This high impedance allows the subject to participate more voluntary efforts to deviate the robot from the reference speed during exercise.
Following the predefined trajectory and generating the adaptive robot velocity based on inverse kinematics and fuzzy PID controller. The controller was implemented in joint space as a basic position/velocity controller of the 6-DOF parallel robot to guide the subject’s limb on reference trajectories. We refer the reader to Zhou et al. (2013) for the detailed implementation of such controller for parallel robotic manipulator. In this experiment, the position and velocity tracking results of the robot have showed satisfactory control performance for rehabilitation purpose. The average joint position tracking errors and velocity tracking errors were controlled to < 1 mm and 0.1 mm/s, respectively, and the plots can be found in the study by Zhou et al. (2013).
Experiments were carried out to demonstrate the comparison of conventional EMG-triggered control method and the proposed method. In the first trial, the subjects were asked not to exert any force and remain passive after the robot is being triggered by EMG signals. In the second trial, experiments on the proposed method were executed, where the impedance parameters are adjustable to allow the subjects to change the robot compliance by themselves. The subjects were instructed to perform voluntary movements and contribute their muscular forces in the training. The primary goal of the first trial was to determine if the EMG-based controller can identify the user’s movement intention and trigger the robot assistance, and the primary goal of the second trial was to determine whether the proposed active interaction controller can improve the user’s voluntary participation and increase the muscle activity levels. Although the motions can be classified continuously in any order, the subject in this test was asked to perform the motions from dorsiflexion to abduction one by one. Each motion took about 10 s. Each subject was asked to undergo six movement sessions, and each session had six continuous motions (dorsiflexion, etc.). Meanwhile, all experiments were conducted with the first and the second trials for comparison. In addition, the subject was asked to have a 3-minute rest between two movement sessions.
6.2 Results and discussion
First, we examined the EMG motion recognition ability. After training the SVM in the first part of the trial, the EMG signals of four channels were acquired and transmitted to the processing part, where we identified the user’s movement intention based on AR features extraction and an SVM classification. The experiment program was written by C++ language and run on a PC with dual 2.8 GHz CPU. The motion recognition results can be output 20 ms in advance before the subject is starting to move his/her leg. In other words, the subject’s movement intention can be identified 20 ms earlier to their actual movement. Once the movement intention was determined, the robot assistance will be triggered immediately. The communication between the EMG recognition part and the robot part was realized by the transmission control protocol (TCP) protocol. So, there was a time delay of about 50 ms before the robot was reacting. Actually, there will be around 30 ms delay when people attempt to drive the robot by their own efforts, this, however, may not be noticed by the subjects in actual experiment. In fact, the subjects found they can drive the robot immediately when they intended to move, and they can adapt the robot control during the whole movement.
In this experiment, the subject was suggested to perform the six continuous motions from dorsiflexion to abduction to facilitate the comparison between different trials, which also made it easier to verify the classifier’s ability to identify the movement types continuously. Figure 7 shows an example of the partial results for Subject2 (S2), Subject5 (S5) and Subject8 (S8). The captured four-channel EMG signals and classification results are demonstrated. To further evaluate the effectiveness of the classifier, more experiments were performed with a group of nine healthy subjects. In this part, each subject was instructed to undergo a total of 50 movement recognition trials. The recognition details of all subjects are illustrated in Table I. A statistical overview of motion recognition results, described by means of the average accuracy and standard deviation, was reported for each subject in Table I. In the case of S1 and S6, the average recognition accuracy is higher than others because these two subjects had been participating in our experiment for a long time, and they were very familiar with experimental protocol. In contrast, the results of S2, S7 and S9 are relatively unsatisfactory, partially due to the muscle activity variations with gender and age and partially because they did not perform the motions accurately. Regarding the other four subjects (S3, S4, S5 and S8), although their EMG signals of lower limb are very different with each other, the motion recognition results of them are all satisfied, which prove the reliability and adaptability of the proposed prediction methods. The results of almost all subjects suggest that the EMG feed-forward items and SVM method make it possible to precisely predict human motion intention in advance and the satisfactory motion recognition accuracy (the subjects’ average rate of success is 94.71 ± 2.83 per cent) can be obtained, especially for several motions such as dorsiflexion, plantarflexion and abduction, in which the success rate can reach above 95 per cent.
Second, the robot impedance tuning and robot control process in the proposed active mode were presented and analyzed. The objective is to demonstrate how the robot compliance can be tuned according to people’s muscle activity level and how the robot movement velocity can be influenced by the subject’s active interaction force. In this part, we just try to verify the robot impedance tuning and active control process by using the proposed method, while the comparison between the two modes will be discussed in next section. Although the muscle activity levels and interaction forces vary a lot to different subjects, the parameter tuning and robot control process are similar for all subjects. Without loss of generality and for simplicity’s sake, experimental results of three subjects are selected from nine. The experimental results of S2, S5 and S8 during the exercise with assist of robot controlled by the proposed mode are illustrated in Figure 8, where the muscle activity level, the shaped impedance parameter and the active interaction force, as well as the robot velocity tracking results, are demonstrated. From the results, one can see that the robot impedance (damping here) can be adjusted by the user’s muscle activity levels, and the robot movement velocity was determined by both the human–robot interaction force and the damping parameter. Specifically, the robot can be compliant (a lower damping) when the muscle activity level is low, while the robot’s damping will increase when the subject shows a higher muscle activity level. Meanwhile, the robot movement speed was greatly influenced by the subject’s active interaction force (a larger interaction force yielded a higher robot velocity), which encourages the users to contribute their own efforts during the exercise. In addition, the desired and the actual velocity tracking results are also shown in Figure 8, where the solid line is the desired velocity and the dotted green line is the actual one. It can be seen that the tracking accuracy of the controller is satisfied for such a parallel robot, and more details can be found in (Zhou et al., 2013).
Third, we analyzed the experimental results to determine whether the proposed active interaction controller could improve the user’s voluntary participation, i.e. whether the proposed approach increases muscle activity levels. To this end, we compared the differences in muscle activity levels between the proposed and traditional methods. The experimental results of S1-9) are shown in Figure 9. This figure shows that the both types of values of the active mode were higher than the traditional mode. Additionally, Table II shows the average of six sessions in the RMS values and muscle activity levels for each training mode and for each subject. Both Figure 9 and Table II show a distinct increase of both types of muscle activity values (muscle activity level and RMS value) for the active mode compared to the traditional mode. Taking S1 for example, the overall average muscle activity level and RMS value in EMG-triggered mode are 3.982 ± 0.509 and 1.445 ± 0.245, respectively, while the average values in active mode are 5.857 ± 0.594 and 3.076 ± 0.549, respectively, in which the rates of increase are 47.08 and 112.8 per cent when compared to the previous one. These results show that the proposed active control methods can be realized with a higher level of the people’s active participation rate. The similar results were also obtained with other subjects. Although different subjects yielded different rates of increase in muscle activity levels and RMS values, the improvements of their active participation and muscle activities are all obvious, indicating that the proposed method is applicable to any users. The experimental results show that the active interaction controller can increase the amount of the muscle activity levels for each movement session compared to the traditional mode. This would allow patients to contribute more efforts during the exercise and may increase the training effects.
During the EMG-triggered assistance trial, the subject kept passive after activating the robot by EMG action recognition, so the muscle activity is at a lower level during this movement except for when people trying to trigger the robot. Differently, in the active interaction assistance trial, the subject had to provide a certain effort to keep the robot move in agreement with his/her intention, so the selected muscles will be activated for a long period of time. In the proposed method, the robot compliance can be adjusted according to his/her muscle activity level, while the traditional method does not take the subject’s status into account. It is clear that there was a distinct increase in the muscle activity levels for all subjects with the active interaction mode compared to the EMG-triggered mode. The increase in RMS values indicated that the active mode had a positive effect in encouraging muscle activation. On the other hand, the muscle activity level evaluation calculated here takes into account the muscles’ weight coefficients in a certain motion. The weight coefficients were determined in the SVM training stage and were calculated by using equation (14). The obtained muscles’ weight coefficients of several subjects are illustrated in Table III. To simplify the table, three subjects are selected from nine. The objective of this table is to present the differences of muscles’ weight coefficients in different motion patterns. Other subjects’ data were not listed for simplicity’s sake, but all data were obtained by using equation (13). Consequently, the muscle activity level here not only depends on the normalized EMG signals but also considers the weight coefficients of each muscle in a certain motion pattern, so a more scientific model can be achieved, and this can be adaptable to different people.
EMG-triggered robot assistance and real-time adjustment of impedance during the training have been demonstrated previously by Emken et al. (2008) and Xu et al. (2011). However, the method proposed here differs in three main ways. First, it provides a seamless interaction control scheme for robot-assisted rehabilitation after activating the robot by motion recognition. Both subject’s movement intention and voluntary participation are considered, not only triggering the robot when the user intends to move but also changing the motion pattern in accordance with the user’s participation efforts during the movement. In contrast, the existing EMG-triggered approach is to provide assistance when the subject starts the movement but remains passive after the robot is triggered. Second, the impedance controller developed here is different, in that it responds to velocity changes rather than variations in trajectory, and thus allows only freedoms in the training speed and keeps the movement along a physiological trajectory. Moreover, the algorithm proposed here adapts the robot impedance as a function of the user’s muscle activity level. On the contrary, the existing algorithms always limited the robot speeds but changed the path severely, which may guide the impaired limb move along an un-physiological pattern. Third, as muscles have different contributions in conducting different motion patterns, the muscle activity evaluation here takes into account the muscles’ weight coefficients in a certain motion pattern (Table III) and thus a more scientific model can be achieved. By using the proposed method, the robot can be compliant if the patient has severe impairment with low muscle activity so that the patient can influence the robot movement more easily. On the other hand, the robot behavior can be made stiff if the patient shows high muscle activity level so that the patient can contribute more efforts to the training process. Consequently, the proposed active interaction control strategy can adapt the robot compliance by adjusting an impedance model from movement to movement based on the patient’s muscle activity levels and the recovery conditions.
7. Conclusion
In summary, this paper presents a seamless active interaction control method between the EMG-triggered assistance and the adaptive impedance control scheme for parallel robot-assisted lower limb motor exercise and training. People’s both movement intention and voluntary efforts are considered, not only triggering the robot assistance when user attempts to move but also changing the motion pattern in accordance with people’s efforts during the movement. First, a wireless EMG acquisition device based on a Wi-Fi protocol was developed and a motion recognition method integrating AR coefficients and an SVM classifier is proposed to identify the user’s movement intention. Second, an impedance control method in accordance with the people’s interactive efforts is implemented to make the robot compliant. Third, the proposed method adapts the robot impedance as a function of people’s muscle activity level, which considers the weight coefficients of each muscle in a certain motion pattern. Experimental results with several healthy subjects demonstrated that the robot was able to move coordinately with the user’s intention while the impedance can be updated with the muscle conditions. It is clear that there was a distinct increase in the muscle activity levels for all subjects with the active interaction control mode compared to the EMG-trigged mode. Such strategies will potentially increase the patient’s motivation because the muscle activity level can change the robot compliance, while the active efforts can be reflected in the training speed and thus can cause a consistent feeling of success.
About the authors
Wei Meng received his BEng and MS in Communication Engineering from Wuhan University of Technology, China, in 2010 and 2013, respectively. He is studying his PhD in Wuhan University of Technology. From 2009 to 2013, he served as a postgraduate with School of Information Engineering, where he participated in research projects in areas of rehabilitation robots and bio-signals processing. Wei Meng is the author of over ten academic journal and conference proceeding papers. In recent years, he has directed two research projects. His research interests include the human–robot interaction, rehabilitation and assistive robots control.
Quan Liu is a Professor and Dean of the School of Information Engineering, Wuhan University of Technology. She received BEng from the department of Electronic and Information, Huazhong Engineering College, in 1985, M.Eng from the department of automation, Wuhan University of Technology, in 1991, and PhD from the School of Mechanical and Electronic Engineering, Wuhan University of Technology, in 2004. Her research interests include signal processing, signal detection and embedded systems and applications, etc. She has extensive experience of participating in and coordinating the national-level projects, including the national high technology research and development program (863) of China, the key program, general program and international collaborative program of nature science foundation of China (NSFC). She obtained two national awards, as well as three provincial and ministerial awards. She has produced more than 90 research papers, including more than 45 papers that are indexed by SCI and EI. She was awarded as the “National Excellent Teacher” in 2007.
Zude Zhou received his BEng from the Department of Electric Power Engineering of Huazhong Institute of Technology in 1970. From 1984 to 1986, he continued his further studies and research at the Department of Mechanical Engineering, University of Birmingham, UK. He is the Chair Professor of School of Mechanical and Electronic Engineering, Wuhan University of Technology, and the Director of Hubei Digital Manufacturing Key Laboratory, China. He is the Council Member of the CIRP and the Senior Member of American Society of Mechanical Engineers. He has been awarded more than ten National, Ministerial or Provincial prizes, including four Second Prize of National Science and Technology Progress Awards and nine First Prize of Ministerial or Provincial Awards. Prof Zhou has authored or co-authored 11 academic books, and published more than 300 papers in academic journals as well as proceedings. He also has been granted more than 30 national invention patents. During the recent five years, he has directed more than 15 funded research projects. His main areas of expertise are digital manufacturing, numerical and networked control system.
Qingsong Ai received his MS and PhD from the Wuhan University of Technology, China, in 2006 and 2008, respectively. From 2006 to 2007, he was a visiting researcher at the Faculty of Engineering, University of Auckland, New Zealand, and worked on the project of medical robots. Since 2009, he has been the Associate Professor at the Wuhan University of Technology. Dr. Qingsong Ai is the author of over 40 technical publications, proceedings and editorials. In recent years, he has directed six research projects. His research interests include signal processing, medical robots and advanced manufacturing technology. Qingsong Ai is the corresponding author and can be contacted at: [email protected]
References
Ai, Q.S. , Liu, Q. , Yuan, T.T. and Lu, Y. (2013), “Gestures recognition based on wavelet and LLE,” Australasian Physical & Engineering Sciences in Medicine, Vol. 36 No. 2, pp. 167-176.
Ai, Q.S. , Chen, L. , Liu, Q. and Zou, L. (2014), “Rehabilitation assessment for lower limb disability based on multi-disciplinary approaches”, Australasian Physical & Engineering Sciences in Medicine, Vol. 37 No. 2, pp. 355-365.
Akdogan, E. , Shima, K. , Kataoka, H. , Hasegawa, M. , Otsuka, A. and Tsuji, T. (2012), “The cybernetic rehabilitation aid: preliminary results for wrist and elbow motions in healthy subjects”, IEEE Transactions on Neural Systems and Rehabilitation Engineering, Vol. 20 No. 5, pp. 697-707.
Benitez, L.M.V. , Tabie, M. , Will, N. , Schmidt, S. , Jordan, M. and Kirchner, E.A. (2013), “Exoskeleton technology in rehabilitation: towards an emg-based orthosis system for upper limb neuromotor rehabilitation”, Journal of Robotics, Vol. 2013.
Cho, H.C. and Park, J.H. (2005), “Impedance control with variable damping for bilateral teleoperation under time delay”, JSME International Journal Series C Mechanical Systems, Machine Elements and Manufacturing, Vol. 48, No. 4, pp. 695-703.
DiCicco, M. , Lucas, L. and Matsuoka, Y. (2004), “Comparison of control strategies for an EMG controlled orthotic exoskeleton for the hand”, in Proceedings of the IEEE International Conference on Robotics and Automation (ICRA ‘04), pp. 1622-1627.
Duchaine, V. and Gosselin, C.M. (2007), “General model of human-robot cooperation using a novel velocity based variable impedance control”, in EuroHaptics Conference and Symposium on Haptic Interfaces for Virtual Environment and Teleoperator Systems. World Haptics, pp. 446-451.
Emken, J.L. , Harkema, S.J. , Beres-Jones, J.A. , Ferreira, C.K. and Reinkensmeyer, D.J. (2008), “Feasibility of manual teach-and-replay and continuous impedance shaping for robotic locomotor training following spinal cord injury”, IEEE Transactions on Biomedical Engeering, Vol. 55 No. 1, pp. 322-334.
Fan, Y. and Yin, Y. (2013), “Active and progressive exoskeleton rehabilitation using multi-source information fusion from sEMG and force & position-EPP”, IEEE Transactions on Biomedical Engineering, Vol. 60 No. 12, pp. 3314-3321.
Fukuda, O. , Tsuji, T. , Kaneko, M. and Otsuka, A. (2003), “A human-assisting manipulator teleoperated by EMG signals and arm motions.” IEEE Transactions on Robotics and Automation, Vol. 19 No. 2, pp. 210-222.
Hu, J. , Hou, Z. , Zhang, F. Chen, Y. and Li, P. (2012), “Training strategies for a lower limb rehabilitation robot based on impedance control”, in 34th Annual International Conference of the IEEE EMBS, pp. 6032-6035.
Hussain, S. , Xie, S.Q. and Liu, G. (2011), “Robot assisted treadmill training: mechanisms and training strategies”, Medical Engineering & Physics, Vol. 33 No. 5, pp. 527-533.
Hussain, S. , Xie, S.Q. and Jamwal, P.K. (2013), “Robust nonlinear control of an intrinsically compliant robotic gait training orthosis”, IEEE Transactions on Systems Man Cybernetics-Systems, Vol. 43 No. 3, pp. 655-665.
Jamwal, P.K. , Xie, S.Q. , Hussain, S. and Parsons, J.G. (2014), “An adaptive wearable parallel robot for the treatment of ankle injuries”, IEEE/ASME Transactions on Mechatronics, Vol. 19 No. 1, pp. 64-75.
Ji, X.F. and Liu, H.H. (2010), “Advances in view-invariant human motion analysis: a review”, IEEE Transactions on Systems Man and Cybernetics Part C-Applications and Reviews, Vol. 40 No. 1, pp. 13-24.
Ju, Z. , Ouyang, G. , Wilamowska, K.M. and Liu, H. (2013), “Surface EMG based hand manipulation identification via nonlinear feature extraction and classification”, IEEE Sensor Journal, Vol. 13 No. 9, pp. 3302-3311.
Kiguchi, K. and Hayashi, Y. (2012), “An EMG-based control for an upper-limb power-assist exoskeleton robot”, IEEE Transactions on Systems Man and Cybernetics B-Cybernetics, Vol. 42 No. 4, pp. 1064-1071.
Kiguchi, K. , Tanaka, T. and Fukuda, T. (2004), “Neuro-fuzzy control of a robotic exoskeleton with EMG signals”, IEEE Transactions on Fuzzy Systems, Vol. 12 No. 4, pp. 481-490.
Krebs, H.I. , Palazzolo, J.J. , Dipietro, L. , Volpe, B.T. and Hogan, N. (2003), “Rehabilitation robotics: performance-based progressive robot-assisted therapy”, Autonomous Robots, Vol. 15 No. 1, pp. 7-20.
Lenzi, T. , De Rossi, S.M. , Vitiello, N. and Carrozza, M.C. (2012), “Intention-based EMG control for powered exoskeletons”, IEEE Transactions on Biomedical Engineering, Vol. 59 No. 8, pp. 2180-2190.
Liu, H.H. (2011), “Exploring human hand capabilities into embedded multi- fingered object manipulation”, IEEE Transactions on Industrial Informatics, Vol. 7 No. 3, pp. 389-398.
Liu, Y.-H. , Huang, H.-P. and Weng, C.-H. (2007), “Recognition of electromyographic signals using cascaded kernel learning machine”, IEEE-ASME Transactions on Mechatronics, Vol. 12 No. 3, pp. 253-264.
Lopes, A. and Almeida, F. (2008), “A force-impedance controlled industrial robot using an active robotic auxiliary device”, Robotics and Computer-Integrated Manufacturing, Vol. 24 No. 13, pp. 299-309.
Marchal-Crespo, L. and Reinkensmeyer, D.J. (2009), “Review of control strategies for robotic movement training after neurologic injury”, Journal of Neuroengineering and Rehabilitation, Vol. 6 No. 20, pp. 1-1.
Mobasser, F. , Elklund, J.M. and Hashtrudi-Zaad, K. (2007), “Estimation of elbow-induced wrist force with EMG signals using fast orthogonal search”, IEEE Transactions on Biomedical Engineering, Vol. 54 No. 4, pp. 683-693.
Nagata, F. , Hase, T. , Haga, Z. , Omoto, M. and Watanabe, K. (2009), “A desktop NC machine tool with a position/force controller using a fine-velocity pulse converter”, Mechatronics, Vol. 19 No. 5, pp. 671-679.
Ouyang, G. , Zhu, X. , Ju, Z. and Liu, H. (2014), “Dynamical characteristics of surface EMG signals of hand grasps via recurrence plot”, IEEE Journal of Biomedical and Health Informatics, Vol. 18 No. 1, pp. 257-266.
Patel, R.V. and Shadpey, F. (2005), “Contact force and compliant motion control”, in Control of Redundant Robot Manipulators, LNCIS Vol. 316, Springer-Verlag, Berlin Heidelberg, pp. 79-117.
Riener, R. , Lunenburger, L. , Jezernik, S. , Anderschitz, M. , Colombo, G. and Dietz, V. (2005), “Patient-cooperative strategies for robot-aided treadmill training: first experimental results”, IEEE Transactions on Neural Systems and Rehabilitation Engineering, Vol. 13 No. 3, pp. 380-394.
Riener, R. , Luenenberger, L. and Colombo, G. (2006), “Human-centered robotics applied to gait training and assessment”, Journal of Rehabilitation Research and Development, Vol. 43 No. 5, pp. 679-694.
Saglia, J.A. , Tsagarakis, N.G. , Dai, J.S. and Caldwell, D.G. (2013), “Control strategies for patient-assisted training using the ankle rehabilitation robot (ARBOT)”, IEEE/ASME Transactions on Mechatronics, Vol. 18 No. 6, pp. 1799-1808.
Song, R. , Tong, K.-Y. , Hu, X. and Li, L. (2008), “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.
Song, R. , Tong, K.Y. , Hu, X.L. and Zhou, W. (2013), “Myoelectrically controlled wrist robot for stroke rehabilitation”, Journal of Neuroengineering and Rehabilitation, Vol. 10 No. 25, pp. 1-1.
Wolbrecht, E.T. , Chan, V. , Reinkensmeyer, D.J. and Bobrow, J.E. (2008), “Optimizing compliant, model-based robotic assistance to promote neurorehabilitation”, IEEE Transactions on Neural Systems and Rehabilitation Engineering, Vol. 16 No. 3, pp. 286-297.
Xu, G.Z. , Song, A.G. and Li, H.J. (2011), “Control system design for an upper-limb rehabilitation robot”, Advanced Robotics, Vol. 25 Nos 1/2, pp. 229-251.
Yin, Y.H. , Fan, Y.J. and Xu, L.D. (2012), “EMG and EPP-integrated human-machine interface between the paralyzed and rehabilitation exoskeleton”, IEEE Transactions on Information Technology in Biomedicine, Vol. 16 No. 4, pp. 542-549.
Zhou, Z. , Meng, W. , Ai, Q. , Liu, Q. and Wu, X. (2013), “Practical velocity tracking control of a parallel robot based on fuzzy adaptive algorithm”, Advances in Mechanical Engineering, Vol. 2013.
Acknowledgements
This research is funded and supported by the National Natural Science Foundation of China (Grant No. 51475342) and the Fundamental Research Funds for the Central Universities (Grant Nos. 2013-YB-002 and 2013-IV-129).