-
Exoskeletons for human performance enhancement are wearable devices and
machines that can increase the speed, strength, and endurance of human being. The user
provides control signals for the exoskeleton, while the exoskeleton actuators provide
majority of the power necessary for performing the task. The human operator applies
a scaled down force compared with the load carried by the exoskeleton.
The earliest recorded work in the field of exoskeleton is that of Nicholas Yagn a
Russian army engineer. He was granted a US Patent for his work on “Apparatus for
facilitating walking, running, and jumping” [
1
]. Although Yagn's mechanism was designed
to augment running, there is no record that the device was ever built and successfully
demonstrated. Prominent earliest proposed exoskeletons were General Electric’s
“HARDIman” ( Human Augmentation Research and Development Investigation) and
Los Alamos National Laboratory’s“PITMAN”(Powered suit of armour for
infantrymen) [
3
].
The development in the field of exoskeleton picked pace in 2001 when the U.S.
Defense Advanced Research Projects Agency (DARPA) started a 50 million project
named “Exoskeletons for Human Performance Augmentation (EHPA)”. The main
goal of EHPA was “to increase the capabilities of ground soldiers beyond that of a
human”. The EHPA program made a number of institutes made research progress in
the technologies related to Exoskeletons.
Prominent exoskeleton developed under the DARPA project were that of
Massachusetts Institute of Technology (MIT) [
4
], Berkley University's “Berkley Lower
Extremity Exoskeleton (BLEEX)” [
5
] and Sarcos Research Corporation's “Wearable
Energetically Autonomous Robot (WEAR)” [
6
]. All these exoskeletons had one common
feature that human motion intent was measured from force sensor placed between the
human operator and exoskeleton. The exoskeleton was triggered either by the
kinematic position command or by the dynamic contact force command. This introduced a
delay in the system. This delay was overcome in the next generation of exoskeleton
based on biosensors.
Many of the new proposed exoskeleton predicted the motion intent from EMG
signals. Prominent among these exoskeletons was an exoskeleton named Hybrid
Assistive Limb (HAL) developed at the University of Tsukuba, Japan, targeted for both
performance augmenting and rehabilitative purposes [
7
]. The main advantage of
EMG based exoskeletons is the ability to estimate the forces that will be generated by
the muscles before the actual occurrence of mechanical contractions. This information
is fed into the exoskeleton system such that by the time the physiological muscles
contract the exoskeleton amplifies the joint moment by a preselected gain factor,
resulting in a decrease in the reaction time of the human/machine system.
This paper presents an algorithm for estimation of the intended knee joint angle from
sEMG signals acquired from four muscles of lower limb namely Vastus Lateralis
(VL), Semitendinosus (STEN), Biceps Femoris (BF) and Rectus Femoris (RF). The
proposed algorithm uses mean absolute value (MAV) and root mean square (RMS)
for feature extraction and a multi-layer back propagation neural network (BPN) for
predicting the knee angle. The algorithm was evaluated by experiments showing the
calculated intended motion while performing a simple daily life activity of sitting in a
squat position and standing from a squat position. The algorithm and the experimental
results are both presented. The predicted knee angle can be used to control a lower
limb exoskeleton.
Five subjects of mean age 31.08 ± 3.15 years, mean weight 81.42 ± 9.5 Kg and mean
height 1.72 ±0. 08 m respectively, were asked to perform simple daily life activity of
sitting in a squat position from standing position and vice versa. sEMG signals were
recorded from four muscles of lower limb namely Vastus Lateralis (VL),
Semitendinosus (STEN), Biceps Femoris (BF) and Rectus Femoris(RF). EMG signals were
recorded using BioTrace+ software for the NeXus-10 biofeedback system at a
sampling frequency of 2048Hz. SENIAM recommendations were followed for the skin
preparation and sensor locations [
8, 9
]. Reference electrode was placed on the tibia
bone. The sensor location for the four muscles are tabulated in Table 1. Block
diagram of the joint angle prediction is shown in Fig1. Knee angle was measured using a
linear potentiometer based goniometer using NI USB 6009 data acquisition system at
a sampling frequency of 2048 Hz. The raw sEMG signal was band pass filtered from
20Hz to 500Hz. The purpose of the lower frequency limit was to remove DC offset
and motion artifacts. The high frequency limit prevents the aliasing. This was
achieved by using a zero-lag sixth order recursive Butterworth filter. The sEMG
signal recorded from VL muscle along with knee angle for subject 1 are plotted in Fig 2.
The sEMG signal is not fed directly to a neural network, because of the
dimensionality and random characteristics of the signal. Instead features were extracted and fed to
the neural network. In this paper, two time domain features mean absolute value
(MAV) and root mean square value (RMS) had been used.
Where is thevoltage of ℎ sample, N is number of samples in a segment. The
number of samples was set to be 128 for the current study.
[− ∑ ( ( )− ( ))
2
]
2 ( )
2
∆ = 2
[− ∑ ( ( )− ( ))
2 ( )
] ( ( )− ( ))
2 ( )
2
ture of the BPN used is shown in Fig 3. The adaptive learning algorithm of the BPN
neural network used is explained in following equations:
1
( ) = 2 ( − )
2
( + 1) = ( ) + ∆
( + 1) = ( ) + ∆
( + 1) = ( ) + ∆
2
∆ = ( − )
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Where , are the learning constant, is the knee angle predicted and
is the actual knee angle recorded by goniometer.
3
Results
Subjects were asked to perform daily life activities like normal walking, stand to sit in
a squat and squat to stand operation, standing up and sitting on a chair and ascending
and descending of stairs. sEMG signals were recorded from VL, STEN, BF and
GMED muscles of lower limb. Knee angle was measured from the potentiometer
based goniometer and NI 6009.
The recorded data was processed and time domain features MAV and RMS were
obtained. The BPN was first trained using the 60% of the data and knee angle was
predicted. The predicted knee angle and actual knee angle for Subject 1 are plotted in
Fig 4. The accuracy of the predicted knee angle using the proposed algorithm is
92.4%. The predicted knee angle could be used to control an exoskeleton or other
orthotic or prosthetic device.