Abstract
The purpose of this research was to assess the
performance of a myoelectric-assisted controller used in a tracking task. Using
the index finger, a subject tracked force profiles displayed on a
screen using visual feedback from two different controllers: 1) the actual
force and 2) the myoelectric controller. The myoelectric-assisted controller
performed significantly better than using feedback from the force measurements
(p = 0.00001). The myoelectric controller provided a 14% improvement over
the force feedback during dynamic finger movements and performed just as well
during static finger tasks. These findings provide evidence that combining myoelectric
signals with mechanical sensors improves man-machine interfaces for
biomechanical movements and force control.
1.
INTRODUCTION
Myoelectric
signals have been integrated in many applications over the years to provide control
of prosthetic limbs [1], for functional neuroprostheses [2], and for
teleoperation of robotic devices [3]. Many experimental techniques and
algorithms have been developed in an attempt to further improve the ability to
predict forces and torques from these myoelectric signals. Adaptive filters [4],
artificial neural networks [3] and frequency spectrum classifiers [5] (to name
a few) have been used to improve both the speed and the accuracy of these
predictions. However, it has been challenging developing strictly myoelectric based
controllers that respond quickly and can accurately provide stable and accurate
control. In lieu of relying entirely on myoelectric signals for control, this
research examines the feasibility of combining myoelectric signals with
mechanical sensors to assist virtual control. The purpose of this research is
to assess the performance of a myoelectric-assisted controller used to regulate
finger force.
2.
METHODS
The experiment required a subject to track
force profiles displayed on a screen by generating isometric finger movements
with the dominant index finger, thereby minimizing the difference between the
target and the actual response. To assess the feasibility of using the
myoelectric controller, the experiment was performed on a trained able-bodied
subject.
Figure
1: Experimental setup. Five
EMG electrodes are drawn. The shaded electrode is located on the inner arm. The
ground is the circular electrode.
2.1. Recordings
Finger force was measured using a load cell (Elane
Electronic) placed under the tip of the index finger of the dominant hand.
Prior to analogue-to-digital conversion, the data was filtered using a first
order Butterworth low-pass filter with a cut-off frequency of 400 Hz. In
real-time, the data was further filtered using a first order 33 point (16.5 ms
delay) Savitzky-Golay filter. The myoelectric activity involved in generating finger
force was measured using surface electromyography (EMG) techniques. Five EMG
electrodes were placed on accessible sites of the hand and arm to capture myoelectric
activity from the flexor digitorum superficialis, flexor digitorum profundus,
extensor indicis, extensor digitorum, and first dorsal interosseous (Fig. 1).
The signals were amplified and filtered with a 10-1000 Hz bandwidth using an
eight-channel EMG pre-amplifier (Bortec Biomedical Ltd.).
Figure
2: Schematic of the
predictive controller. The controller compares the force measurements with the
EMG-based predictions to adjust the gain element in each pathway.
Figure 3: EMG-based force predictions
during the experiment. The spikes in the prediction error reveal the effects of
phase separation during the rapid movements.
2.2. Predictive force model
Previous experiments in our lab have revealed that
isometric finger forces can be modelled from EMG activity using the Laguerre
Expansion Technique (LET). The Laguerre expansion involves using an orthogonal
set of Laguerre functions to develop finite-impulse digital filters (i.e. one
per muscle) which effectively filter the rectified EMG signals to predict the force.
As anticipated, the resulting model is less accurate than measuring finger force;
however, since muscular activity is a necessary precursor to motor movements,
the EMG-based model can be adjusted to respond faster than the actual finger
movements. The model was calibrated to predict finger force 50 ms before the
movement begins.
2.3. Myoelectric-assisted
controller
A myoelectric controller was developed in an attempt
to improve force control (Fig. 2). The controller uses: 1) the measured forces
from the load cell and 2) the predicted forces calculated from the EMG-based
predictive model. This controller acts as a proportional gain controller by
calculating the difference between both inputs (force and predicted force) and
adjusting the two variable-gain elements accordingly. The controller works on
the principle that phase separation between two similar signals varies linearly
with frequency. A delay between two low frequency signals has negligible effect
on the phase separation. However, a delay, and hence, a phase separation
becomes significantly more apparent when comparing higher frequency signals. In
this study, as the difference between the measured force from the load cell and
the predicted force (signal 50 ms ahead) increases, this indicates that the
finger is moving quickly. Conversely, when the difference is small, the finger
is moving slowly. Together, the force measurements and the fast- acting
EMG-based predictions were used to help improve the overall tracking
performance. The controller’s two gain elements were adjusted using two
equations that varied according to the error between both signals:
where
x is the absolute error and y and z represent the gains for the measured force and the EMG-based
predictions, respectively. When the error x between target force and the
response is less than 1.5 N, the controller uses information uniquely from the
force sensor. As the error increases thereafter, the controller begins using a
combination of both the measured force and the EMG-based predictions.
2.4. Experimental Protocol
To avoid aliasing of the EMG signals, all experiments
were performed using real-time feedback working at 2000 Hz. The trained subject
tracked the different force profiles using the two different visual feedback
controllers: 1) the measured force and 2) the myoelectric controller. The force
profiles were made up of five steps of 1, 2, 3, 4, and 5 N each lasting 2.5 s,
arranged in pseudo-random order. A total of three different force profiles were
used in the experiment. Each trial consisted of tracking the three force
profiles using the two types of visually guided feedback. The subject performed
a total of twelve trials. To assess the performance of each trial, the error
was calculated between the target force and the response. The performances of
both types of controllers were assessed at a 5% significance level.
Figure
4: Performance error
boxplots.
Figure
5: Average response for both
controllers. The response time of the predictive controller is significantly
shorter than for the force feedback.
3.
RESULTS
For n = 12 trials,
and p = 0.00001, the one-tailed t-test indicates that the performance error
using the myoelectric controller was significantly lower than using force
feedback (Fig. 4). This result reveals that the myoelectric signals improve the
overall tracking response. The myoelectric controller improved the dynamic
responses during the time period 190 ms to 1 s after onset of the stimulus,
compared to using force feedback. The myoelectric controller provided 14%
performance enhancement (p = 0.000002, for a one-tailed t-test) compared to the
force feedback control. During static finger forces (1-2.5 s after stimulus
onset), there was no significant difference between both controllers (p = 0.78,
for a two-tailed t-test). This result suggests that both controllers performed
equally well during the stable holding phase of the tracking task.
4. DISCUSSION AND CONCLUSIONS
It was found that the myoelectric-assisted
controller enhances the ability to control forces in a visually-guided tracking
task. The predictive EMG-based model did not provide significantly more
overshoot in the response (Fig. 5). The
predictive model merely provides a faster response while being equally capable
of providing finer touch control. These results provide evidence that an
individual’s ability to control a device via a man-machine interface is
improved by combining myoelectric signals with mechanical sensors. We believe
that an
References
[1]
Zecca
M, Micera S, Carrozza MC, et al. Control of
multifunctional prosthetic hands by processing the electromyographic signal. Crit Rev Biomed
[2]
Hart RL, Kilgore KL, Peckham
PH, A comparison between control methods for implanted
[3] Fukuda O, Tsuji T, Makoto K, et al. A human manipulator teleoperated by
EMG signals and arm motion. IEEE Trans.
Robot. Autom. 19: 210-2003, 2003.
[4]
[5] Farry KA, Walker ID, Baraniuk RG.
Myoelectric teleoperation of a complex robotic hand. IEEE Trans. Rob. Aut. 12:775-787, 1996.
Acknowledgements
The authors would like to thank César Marquez for his valuable insight and technical expertise, and acknowledge the Natural Sciences and Engineering Research Council of Canada for financial support.