1 Sungyoon Lee, 1 Jaesung Oh, 1 Youngwon Kim, 1 Minsuk Kwon * Jaehyo Kim 1 Department of mechanical & control engineering, Handong University, qlfhlxhl@nate.com * Department of mechanical & control engineering, Handong University, jhkim@handong.edu Abstract This paper aims to estimate physical human intention during lifting movement from EMG signals. Subjects conducted periodic target following up/down movement which the target moves in sinusoidal speed profile on sagittal plane. The movement conditions vary with target speed, 0.2Hz, 0.3Hz, 0.4Hz, and 0.5Hz, and with weight, 1.5kg, 3kg, 4.5kg, and no load. The difference between loaded and unloaded lifting movement in muscle contraction level is investigated and a simple mathematical method to recognize the motor intention and load condition is suggested. Through the logistic regression model, movement direction was estimated with over 80% accuracy. In addition, it was found that the pattern of TCL is similar to the joint angle in unloaded condition and to the angular speed in loaded condition. 1. Introduction Keywords: EMG, Movement Characteristic Estimation, Human Motor Planning Conventional robot systems exclude the involvement of human and assume that the entire operational circumstances are known. Therefore accuracy and speed were the main design issues. However, in recent years, robots working in the human environment have been developed and consequently the paradigm of robot has changed; flexibility to handle unexpected obstacles, safety to reduce the risk of human, and dependability to support human-in-the-loop condition [1]. The most important operation in Human-Robot Interface (HRI) is recognizing the human intention and reaction, and this could be divided into two level. One is at cognitive level to recognize the nonverbal cues in order to infer the human intention and to offer an intuitive, human-like way of interaction [2]. The other one is at body level, which indicates physical interaction. In this level, we consider the situations such as follows; humans and robots share the same workspace, have contact with each other, exchange forces, and cooperate when doing actions on environment [1]. The key issue of physical HRI is deciding which signal to measure and how to measure and process the signal to estimate the human motor intention. In a contact task, even if the same extrinsic (articular) movements are observed, it may represent a different intrinsic (muscular) coordinate called the stiffness, which is a spring-like property of the human musculoskeletal system. There have been many efforts to measure the stiffness during posture and movement such as using a manipulator and the perturbation method, for example [4, 5, 6, 7]. However, these methods are based on the assumption that the stiffness is constantly high in the extrinsic movement and low in the intrinsic movement, which makes the estimation not natural. In the contact task, the conditions of the environment and the load changes with time and therefore the stiffness should be actively change during movement, not constant [12]. In this context, EMG signal is the most appropriate since it has the advantage of estimating the stiffness level that can be used as the input parameter not only in the position control but also in the force control. Many researches have been carried out to estimate motion characteristics using EMG signals. It is known that the muscle force is needed to accelerate an arm from a resting state and to decelerate it back to a resting state when the designed position is accomplished [3]. Studies have shown that joint torques and trajectory can be estimated from quasi-tension [4, 5]. It also has been shown that joint the stiffness corresponds to the movement speed [6, 7, 8]. Based on these relationships, we have International Journal of Engineering and Industries(IJEI) Volume2, Number4, December 2011 doi : 10.4156/ijei.vol2.issue4.11 97
developed an estimation method for human motor intention during the contact task using EMG signals and adopted the lifting movement as the contact task. In the lifting task on the sagittal plane, subjects stood in position and only used upper limb except shoulders. Elbow angle and the EMG signals from muscles were continuously measured during the target following task. The angle was represented as a tracer on the screen with the target. The trajectory of the target was pre-programmed and the various loads of weights were used. This paper proposes a simple method to estimate 1DOF elbow movement in a lifting task from EMG signal. A mathematical model is proposed that estimates the coefficient for joint torque from the difference between agonist and antagonist muscle tension for the elbow joints and estimates the stiffness of the agonist and antagonist muscle co-activation from the sum of the agonist and antagonist muscle tension. Using this model, movement intention and object weights are estimated. 2. Materials and methods 2.1. Muscle tension and joint stiffness 2.1.1. EMG measurement: For the flexion and extension of the elbow joint, the biceps long head (BILH, flexor) and the triceps lateral head (TRIA, extensor) were measured. The wrist mono-articular muscles included the flexor carpi ulnaris (FCU, flexor) and the extensor carpi ulnaris (ECU, extensor). Total Contraction Level (TCL) was calculated from EMG signals of these muscles. Figure 1. Selected muscles and electrodes attachment. These four muscles are activated during the periodic target tracking up/down movement of upper limb. 2.1.1. EMG signal processing: The amplitude of EMG signal reflects the muscle tension. Raw EMG signals from the CNS contains high frequency contains, which needs rectification and low-pass filtering to get the amplitude. To process the EMG signal, we used a 2nd order low-pass filter with natural frequency of 13.36 rad/s and damping ratio of 1.002 [9]. (1) (2) 98
Figure 2. EMG signal processing procedure. To estimate motion characteristics from EMG signals, raw EMG signals should be first rectified and filtered through linear, 2nd order low-pass filter. The coefficient in Eq. (1) was acquired by digitally processing. The filtered signals are called quasi-tension since it shows high correlation with the actual muscle tension [4, 9]. Normalization was also needed because the levels of muscle development among the subjects were different and the muscle activation levels may have changed during the trials. Usually, signals are normalized with MVC(maximum voluntary contraction), but in this study, the average EMG signals during unloaded state with the arm position fixed to 90 degree were used to normalize the signals because it is neutral posture of the experiment. 2.2. Estimation of joint stiffness; Total Co-contraction Level. Joint stiffness, or TCL, is the summation of effective muscle stiffness, and can be expressed as follows. (4) Assuming that muscle activation is proportional to the EMG level, we can estimate the stiffness directly using Eq. (4). When the movement velocity increases in multi-joint movement, the joint stiffness also increases[4, 9]. 2.3. Estimation of joint torque and movement trajectory Joint torque is generated by the difference in flexor and extensor EMG signals, and can be determined by Eq. (3) using muscle i acting on joint j ( is the length of the moment arm.) (3) 99
A constrained optimization method was applied to estimate the parameter. Since the quasi-tension is similar to actual muscle tension, joint torque can be estimated using quasi-tension. Joint torque is generated during movements, and movement trajectory can be estimated from the joint torque using numerous mathematical models such as Least Square Fitting (LSF) and neural networks [9, 10]. 2.4. Classification of movement status by logistic regression model Neural network is the most frequently employed pattern recognition method, but this is complicated due to the need for a great deal of training data over a long period, as well as layers for higher accuracy [7]. Reliable results can be obtained by a simple linear model related with quasi-tensions when we are handling a low degree-of-freedom task [11]. We estimated the parameters using a logistic regression model. The model is equal to a single-layer neural network; therefore, it is possible to accept both a linear regression method and back-propagation. (5) 3. Experimental setup and Procedures 3.1. Experimental setup We conducted our experiment under two conditions: one is the unloaded condition, and the other is the loaded condition in different movement speed. 0.2Hz, 0.3Hz, 0.4Hz and 0.5Hz sinusoidal signals were used to vary speed. In the loaded condition, three weights, 1.5Kg, 3.0Kg, and 4.5Kg, were used by connecting several 1.5Kg objects. Subject held a rope connected to the objects by the wrist to remove a DOF of wrist joint. To measure the angle of the arm position, EBIMU-9DOF sensor (AHRS) was used with the resolution of 0.01. The sensor was attached to the end of the wrist. Subjects were asked to lift the objects of different weight varying speed. Visual information such as the heights of the object and the target were displayed on the monitor screen. 3.2. Experimental procedures Six healthy, right-handed male subjects ranging in age from 21 to 26 years participated in the experiment without any information on the specific purpose of the experiment. Subjects were asked to track the target trajectory five times while holding an object, each time with different weight and different movement speed. The target was displayed on the monitor and moved on sagittal plane periodically in a sine wave with 20cm amplitude. EMG signals and elbow angle of arm position were recorded during the experiment 100
Figure 3. Experimental setup. Subjects focus on the monitor screen and conduct target tracking periodic movement varying movement speed and object weight. Subjects stand upright and hold the object using the wrist to reduce the DOF. Figure 4. Data recorded during the target tracking periodic movement. TCL is similar to the actual angle with time delay. One was sampled under loaded condition, and the other was sampled under unloaded condition. EMG levels of unloaded condition including TCL are drawn 10 times amplified. 101
4. Experimental results 4.1 Movement prediction using EMG signals Fig. 5 shows the estimation of movement direction of the elbow. Regression was used to predict the movement intention. Comparing the estimated movement intention with actual elbow angle, it is possible to estimate the movement intention of the elbow joint. The accuracy of estimated direction is 88.7% on average as shown in the table 1. Figure 5. Estimated movement intention using logistic regression: Up represented upward intention and Down represented downward intention estimated form EMG signals. Comparing with actual angle, estimated movement intention shows a high coincidence. Table 1. Movement direction prediction accuracy using logistic regression analysis during periodic target tracking movement Speed [Hz] 0.2 0.3 0.4 0.5 Weight [Kg] U a D b U D U D U D 0 0.803 0.771 0.898 0.901 0.919 0.830 0.898 0.818 1.5 0.830 0.879 0.923 0.910 0.916 0.928 0.941 0.944 3 0.933 0.932 0.936 0.807 0.948 0.878 0.881 0.932 4.5 0.820 0.886 0.873 0.882 0.919 0.814 0.940 0.905 a Upward, b Downward 4.2 Relationship between weight and TCL Fig. 6 shows the average TCL patterns of four weighed object under the same movement speed condition. It has 10 sections and each section is part of a sine wave. In most cases, TCL does not overlap on different weights, but 4.5Kg occasionally cross with 3Kg because of larger deviation, although, we can distinguish the weight based on the total average of one cycle of TCL from the result. 102
Figure 6. Classification of the object weight for each speed from average TCL in each section. Each section is one of the ten parts of a sine wave. TCL is distinguishable without overlap under different weight condition except for some cases. 4.3. Tendency of TCL when load condition change Fig. 7 shows the tendency of TCL under the unloaded condition and the loaded condition. As shown in the figure, the tendency of TCL is similar to the actual angle under the unloaded condition, but under the loaded condition, the tendency of TCL is similar to the angular velocity. Figure 7. General tendency of TCL under loaded condition and unloaded condition. TCL pattern is similar to the actual angle under the unloaded condition and angular velocity under the loaded condition. The tendency could be found even when varying normalized state. 103
Table 2 shows the correlation of TCL with the actual angle and speed. Under the unloaded condition, the correlation of TCL with the actual angle is 79.1% on average. However, if the movement speed and object weight increase, the correlation of TCL with the actual angle is less than the correlation of TCL with the angular velocity. Table 2. The correlation of TCL with the actual angle and the angular velocity Speed [Hz] 0.2 0.3 0.4 0.5 Weight [Kg] A a V b A V A V A V 0 0.826 0.387 0.796 0.645 0.791 0.625 0.749 0.611 1.5 0.703 0.546 0.611 0.717 0.559 0.805 0.598 0.803 3 0.463 0.794 0.422 0.850 0.447 0.872 0.411 0.875 4.5 0.542 0.746 0.194 0.726 0.332 0.839 0.258 0.837 a Actual angle, b Angular Velocity 4. Conclusion This study has shown that the elbow movement characteristics during lifting tasks can be estimated solely from EMG signals. During the lifting task, movement direction was estimated from the difference between flexor and extensor EMG signals using logistic regression and the object weight was distinguished using the average TCL. With slow speed and light weight, TCL was similar to the actual angle since the torque needed to maintain the position was larger than the torque needed to change the movement. In the other cases, TCL was similar to the angular speed of elbow movement since the torque needed to maintain the position was smaller than the torque needed to change the movement because of the increased mass and weight on the wrist. From the result, we confirmed that the human motor intention varies depending on the movement speed and object weight. Further studies are required to set exact criteria to distinguish the loaded and unloaded movement. 5. References [1] De Santis A., Siciliano B., Safety Issues for Human-Robot Cooperation in Manufacturing Systems, from http://www.phriends.eu/virtual_08.pdf [2] A. J. Schmid, O. C. Schrempf, U. D. Hanebeck, H. Woern, Towards Intuitive Human-Robot Cooperation, 2nd International Workshop on Human Centered Robotic Systems HCRS 2006, pp.7-12, 2006 [3] J. V. Basmajian and C. J. Luca,Muscles alive, Baltimore: Williams & Wilkins (1985) [4] Koike, Y., Kawato, M., Estimation of dynamic joint torques and trajectory formation from surface electromyography signals using a neural network model, Biological Cybernetics, vol.73, 291-300, 1995. [5] Koike, Y., Kawato, M., Trajectory formation from surface emg signals using a neural network mode,l In Proceedings of annual international conference of the IEEE engineering in medicine and biology society 15, pp. 1628-1629, 1993. [6] Gomi H., Kawato M., Equilibrium-point control hypothesis examined by measured arm stiffness during multijoint movement, Science, vol. 272, pp.117-120, 1996. [7] Gomi H., Kawato M., Human arm stiffness and equilibrium point trajectory during multi-joint movement. Biol Cybern, vol.76, pp.163-171, 1997. [8] Osu R., Gomi H., Multi-joint muscle regulation mechanisms examined by measured human arm stiffness and EMG signals, J. Neurophysiol, pp.1458-1468, 1999 [9] Koike, Y., Kawato, M., Estimation of arm posture in 3D-space from surface EMG signals using a neural network model, IEICE Transactions on Fundamentals, E77-D 4, pp.368-375, 1994 [10] D. Shin, J. Kim, Y. Koike, M. Sato, Estimation of time-varying stiffnes ellipse from EMG signals using a musculo-skeletal model, The 2 nd International Symposium on Measurement, Analysis and 104
Modeling of Human Functions, pp. 255-258, 2004 [11] R. Kevin, Wheeler, Device Control Using Gestures Sensed from EMG, IEEE International Conference on Soft Computing in Industrial Applications, 2003 [12] J. Kim, M. Sato and Y. Koike, Human arm posture control using the impedance control ability of the musculo-skeletal system against the alternation of the environment, Trans. On CASE, Vol.4, No.1, pp.43-48, 2002 105