Documente Academic
Documente Profesional
Documente Cultură
Abstract - In this paper, robot hand design aiming at Amplitude and frequency are typical information
application to a myoelectric prosthetic hand and recognition of extracted from the EMG signal, which can be used for the
finger operation via surface EMG are discussed. The robot hand identification of operation. Ito at el.[8] presumed muscle
with two fingers, thumb and index finger, was built to implement tension from the EMG signal, and tried to control the forearm
fundamental motions required in daily life, such as grasping and
type myoelectric prosthetic arm driven by ultrasonic motor.
holding. The fingers are driven by the wire actuation system to
imitate the human's tendon mechanism. In order to control each Farry et al.[9] has proposed a technique of teleoperating the
finger of the developed myoelectric prosthetic hand robot hand by identifying the frequency spectrum pattern of
independently, using the neural network, identifier which the EMG signal.
classifies four finger operations, namely flexion and extension of On the other hand, as points for the improvement of the
the thumb in metacarpophalangeal (MP) joint and flexion and myoelectric prosthetic arms/hands, enhancement of the
extension of the index finger in MP joint, was constructed. The operativity, simplification of the structure, weight saving of
recognition of each finger operation was performed using the the artificial arms/hands, etc. are mentioned. At present, most
four patterns of the neural network based identifiers in which the of the myoelectric prosthetic arms/hands are restricted to
recognition method of finger operations differs each other, and
realization of the specified operations such as palmar seizure,
the results were compared.
flexion-extension of a wrist, and inward-outward rotation of a
Index Terms - Neural network, EMG, Myoelectric prosthetic wrist. Probably, a present cutting edge practical myoelectric
hand, Gesture recognition prosthetic hand is the "i-LIMB Hand" produced by Touch
Bionics Inc.[10]. However, the myoelectric prosthetic
I. INTRODUCTION arms/hands which imitate human's hand and in which various
Recently, myoelectric prosthetic arms/hands, in which an operation is possible, have hardly spread to general public yet,
operation is identified by the electromyogram (EMG) and the since they are expensive due to the requirement of
artificial arm/hand is controlled based on the identification, measurement of highly precise EMG signals and use of plural
has been studied. The EMG has attracted an attention of number of actuators.
researchers as an interface signal of an electric actuated arm Therefore, to spread the myoelectric prosthetic hands to
for many years, and many of studies on the identification of the public, the purpose of this study is development of the
the EMG signal have been executed. Nowadays, it can be said inexpensive myoelectric prosthetic hand which can realize
that the EMG is the most powerful source of control signal to fundamental operation, such as holding and grasping, required
develop the myoelectric prosthetic arms/hands. in daily life. In order to make it budget prices, an
In early stage, the practical myoelectric prosthetic arms underactuated robotic hand which realizes flexion and
such as "Boston Arm" by Massachusetts Institute of extension of fingers by tendon mechanism, is introduced. In
Technology[1], "WIME Hand" by Waseda University[2], and addition, the "fit grasp mechanism" in which the fingers can fit
"Utah Artificial Arm" by Utah State University[3] were the shape of the object when the hand grasps the object, is
developed. These arms are attached to an upper arm of the proposed. The "fit grasp mechanism" makes it possible for the
person who lost his/her forearm. Then, extension and flexion robotic hand to grasp a small object, a cylindrical object, a
of the arm are performed based on the EMG signals measured distorted object, etc.. In this study, operation of three fingers
by electrodes attached to flexor and extensor of the upper arm other than a thumb and an index finger is regarded as
part. equivalent to the operation of the index finger. Therefore,
From the 1970s to the 1980s, elementary pattern robotic hand with thumb and index finger was designed and
recognition technique such as linear discriminant analysis, was built as a prototype.
used for the identification of EMG signals[4],[5]. In the 1990s, As for the identification of operations, independent
research on learning the nonlinear map between EMG pattern operation of each finger by simple measurement and control
and operation using a neural network has been performed[6],[7]. method is targeted. To the best of our knowledge, in the
Kelly et al.[6] identified four kinds of operations of forearm by previous studies, there is little research that tries to identify
combining Hopfield-type neural network and back operation of plural fingers and to actuate each finger
propagation neural network. independently. In this study, using the neural network,
identifier which classifies four operations, namely flexion and
274
III. MEASUREMENT AND SIGNAL PROCESSING OF EMG Therefore, the amplitude value in 100Hz which is
considered as the influence of AC power source noise being
A. Measurement part of EMG
minimum is used for recognition of the finger motions.
In the built robot hand for myoelectric prosthetic hand, the
Three dimensional graph of the amplitude value data in
fingers which will be used for operation are thumb and index
100Hz when performing each motion in MP joint is shown in
finger. Eight motions of two fingers are considered, which are
Fig.7, in which each measurement position is taken as an axis
flexion and extension in MP joint, and flexion and extension
of coordinates. In addition, the distribution of Fig.7 was
in interlocked DIP and PIP joints, and inward rotation and
divided into the distribution along the thumb and the index
outward rotation are not taken into consideration. In addition,
finger respectively, which are shown in Fig.8 and Fig.9.
the thumb and the index finger are operated independently.
The measurement positions of EMG are shown in Fig.5. Those
are the following three positions, namely vicinity of a
musculus flexor carpi radialis / a musculus flexor digitorum
superficialis (ch1), vicinity of a musculus flexor digitorum
profundus (ch2) which are used when making the PIP joint of
each finger flexion, and vicinity of a musculus extensor
digitorum (ch3) which is used when making each finger
extension.
275
After the learning of the neural network was over,
simulation was carried out using the 30 data for each finger
motion, which differs from the data used for the learning, and
its identification rate was examined. Then, if the integral value
which is rounded off to the first decimal places of the output
value from the neural network is equal to the integral value
given by the teacher signal, the recognition of finger motions
is judged as success. The results are shown in Table 1.
From Table 1, the successful identification rate was only
20.8% on average for use of the identification method (x).
276
study, the following three patterns shown in Fig.8 were the thumb. Again, if the output of N.N.-2 is 0, N.N.-3
considered. identifies whether input signal is extension of the index finger.
(a) is the identification method in which recognition of Finally, input signal which was not identified any of above
finger motions is finally achieved by distinguishing each one motion, is identified as flexion of the thumb. The simulation
motion by one neural network and repeating this. In (a), the was carried out using the 30 data for each finger motion,
finger motion is recognized by distinguishing in order with the which differs from the data used for the learning, and
high successful identification rate such as flexion of the index recognition of each finger motion was executed. Then, if the
finger, extension of the thumb, and extension of the index output of the neural network is larger than 0.5, it is regarded as
finger. This order was determined through the simulation 1, and if the output of the neural network is smaller than 0.5, it
results. In the case where it was not identified as either of is regarded as 0. The results are shown in Table 3.
these three motions, it is finally recognized as flexion of the
thumb. This identification method has the drawback that TABLE 3 Simulation results of (a)
incorrectly-identified motions are regarded as flexion of the Motion Fl of Th Ex of Th Fl of If Ex of If Average
thumb.
Identification
rate [%] 66.7 30 90 43.3 57.5
Fl: flexion, Ex: extension, Th: thumb, If: index finger
277
TABLE 5 Simulation results of (b) V. CONCLUSION
Motion Fl of Th Ex of Th Fl of If Ex of If Average In this paper, in order to control each finger of the
Identification developed myoelectric prosthetic hand independently, using
rate [%] 66.7 33.3 83.3 63.3 61.65 the neural network, recognition of four finger motions, namely
Fl: flexion, Ex: extension, Th: thumb, If: index finger flexion and extension of the thumb and the index finger in MP
joint, was performed. First, identification of these four finger
In (c), at first N.N.-1, which was trained to output 1 for motions by one neural network was executed. Then, the
motion of the thumb and 0 for motion of the index finger, identification rate of only 20.8% on average was obtained. In
classifies the input signal to motion of the thumb or motion of order to improve the identification rate, the choices of the
the index finger. After that, N.N.-2 or N.N.-3 distinguishes the finger motions which are identified by one neural network
input signal whether it is flexion or extension. Thus, finally were divided, and the method in which the partial
the input signal is identified to one of the finger motions. As identification is repeated was employed. Three types of the
well as in the case of (b), before performing synthetic improved identification methods were constructed, and the
identification, in order to examine identification rate of each simulation was carried out using each identification method.
neural network, the simulation when each neural network is The simulation result for each improved identification method
used independently was performed. The results are shown in showed successful identification rate of more than 57% on
Table 6. average.
We are now proceeding the control experiments for the
TABLE 6 Independent simulation results of (c)
built myoelectric prosthetic hand using the identification
N.N.-1 Thumb Index finger Average
method (b) which showed the highest successful identification
Identification rate rate. As a future work, recognition of 8 kinds of finger
80 55 67.5
[%]
motions including flexion and extension of the thumb and of
N.N.-2 Fl of Th Ex of Th Average
the index finger in interlocked DIP and PIP joints, is left.
Identification rate
80 100 90 Then, even if the number of finger motions which have to be
[%]
N.N.-3 Fl of If Ex of If Average identified increases, realization of the high identification rate
Identification rate is required.
83.3 90 86.65
[%]
Fl: flexion, Ex: extension, Th: thumb, If: index finger REFERENCES
[1] R. Mann, and D. Reimers, “Kinesthetic Sensing for the EMG Controlled
From Table 6, the identification rate of the first Boston Arm,” IEEE Transactions on Man-Machine Systems, pp. 110-115,
identification was only 67.5% on average. In addition, the March 1970.
[2] I. Kato, et al., “The Evaluation Method of Rehabilitation Devices-Field
simulation which performs synthetic identification was carried
Testing of Powered Forearm Prosthesis, WIME Hand-,” Proc. of 6th Int.
out using the 30 data for each finger motion, and the Symposium on External Control of Human Extermities, pp.141-184, 1978.
identification rate was examined. The results are shown in [3] S. Jacobsen, et al., “Development of Utah Artificial Arm,” IEEE
Table 7. Transactions on Biomedical Engineering, Vol.29, No.4, pp.249-269,
TABLE 7 Simulation results of (c) 1982.
[4] D. Graupe, J. Magnussen, and A. A. M. Beex, “A Microprocessor System
Motion Fl of Th Ex of Th Fl of If Ex of If Average for Multifunctional Control of Upper Limb Prostheses via Myoelectric
Signal Identification,” IEEE Transactions on Automatic Control, Vol.23,
Identification
rate [%] 70 70 76.7 13.3 57.5 No.4, pp.538-544, 1978.
[5] S. Lee, and G. N. Saridis, “The control of a prosthetic arm by EMG
Fl: flexion, Ex: extension, Th: thumb, If: index finger pattern recognition,” IEEE Transactions on Automatic Control, Vol.29,
No.4, pp.290-302, 1984.
From Table 7, the synthetic identification rate of 57.5% [6] M. F. Kelly, P. A. Parker, and R. N. Scott, “The Application of Neural
on average was obtained, which is almost same level as (a). Networks to Myoelectric Signal Analysis: A preliminary study,” IEEE
Transactions on Biomedical Engineering, Vol.37, No.3, pp.211-230,
From the above simulation results, it turned out that in all 1990.
(a), (b) and (c), the identification rate was improved compared [7] B. Hudgins, P. A. Parker, and R. N. Scott, “A new strategy for
with (x). As for (a), although incorrectly-identified motion multifunction myoelectric control,” IEEE Transactions on Biomedical
would be identified as flexion of the thumb, recognition of the Engineering, Vol.40, No.1, pp.82-94, 1993.
[8] K. Ito, T. Tsuji, A. Kato, and M. Ito, “An EMG Controlled Prosthetic
finger motions succeeded at the identification rate of 57.5% on Forearm in Three Degrees of Freedom Using Ultrasonic Motors,” Proc. of
average. As shown in (b) and (c), also in the case where the the Annual Conf. the IEEE Eng. In Medicine and Biology Society, Vol.14,
identification of each neural network was alternative and pp.1487-1488, 1992.
identification was performed by two steps, the identification [9] K. A. Farry, I. D. Walker, and R. G. Baraniuk, “Myoelectric
Teleoperation of a Complex Robotic Hand,” IEEE Transactions on
rate of more than 57% on average was obtained. On the other Robotics and Automation, Vol.12, No.5, pp.775-787, 1996.
hand, from the observation of N.N.-1 in (b), N.N.-2 in (c) and [10]http://www.touchbionics.com/
N.N.-3, it can be said that the difference between flexion and
extension is comparatively easy to identify.
278