Computational Intelligence in Analysis of Electromyography Signals - Computational Intelligence in Biomedical Engineering - page 253

Rectification
smoothing

EMG signal

Motion 1

Motion
recognition

Bandpass filter

Motion M

FIGURE 5.23
An example of a limb-motion detection system used to classify between the
various motions captured by the EMG waveform.

information. Best prediction accuracies were 95.1% but were reduced if the
limb motion was significantly different from that used in the training data.
Cheron et al. (1996) designed a recognition system based on a dynamic
recurrent neural network (DRNN) to recognize more complex arm movements.
The objective was to recognize the arm's trajectory during fast movements,
such as drawing imaginary figures in the air. Their classification task was more
complex because many muscles were involved and the system had to react
quickly. EMG information from the posterior, anterior, and median deltoid,
and pectoralis major superior, inferior, and latissimus dorsi were recorded.
The DRNN network was shown to be more robust to signal noise and could
extrapolate the muscle motion, for example, the system could correctly pro-
duce shapes such as circles and ellipses when the corresponding EMG signals
were passed to it. This outcome offered some validation in that the DRNN
network had captured some of the intrinsic muscle information related to the
final arm movements.

5.7.2 Limb Motion in Prosthetics
The 1990s saw renewed research work into the application of NNs and other
CI techniques in systems designed for limb-motion detection in myoelectric
prosthetics. The basic design (Figure 5.23) of a limb-motion detection sys-
tem involves first filtering, rectifying, and smoothing the EMG signal before
processing the features and applying pattern-recognition techniques to deter-
mine the motion implied by the EMG input. The goal here is to develop
systems that are highly accurate and robust because it is always a challenge
to reduce the false detection rates whereby the user moves the prosthetic
wrongly due to the system misclassifying the intention.
Work on predicting limb motion began earlier with investigations into mus-
cle changes due to the interaction with a prosthesis. Morris and Repperger
(1988) used linear-discriminant analysis to investigate changes in the antag-
onist tricep-bicep muscle pair due to wearing a prosthesis. Their work was
an attempt to develop some measure of performance in the use of prostheses
from earlier studies by Saridis and Stephanou (1977), Lee and Grimson (2002),
and Graupe et al. (1982). The EMG signals for four different motions were