Download - Academy Publisher

influence of emotion on the activity of the nervous
system is effectively reflected in the physiological signals
employed. Unlike the case of speech recognition or facial
expression recognition, where knowledge of the correct
class label of a given data point is self-evident, the
acquisition of a high-quality physiological signal
database with confidence in the underlying emotional
status is an intricate task.
The physiological signal data of ECG is from the
MIT-BIH, Standard MIT-BIH Arrhythmia database are
collected by Beth Israel Hospital Arrhythmia Laboratory
from 1975 to 1979. There are more than 4000 data Holtel
sets [6]. Database has a total of 48 records, which are from
47 individuals (of which 201 and 202 are from the same
individual).
In MIT-BIH database, each record contains data on
two channels, which leads settings as follows: The first
channel is the use of calibration limb leads Ⅱ; The second
channel used correction V1 (occasionally have V2, V5
leads, there is another V4 leads). Data’s sampling
frequency is 360HZ and the sampling accuracy is 11 bits
(sample data range between 0 ~ 2047).
analysis figure, and this paper extracts the maximum
value as the feature vector of different pattern of ECG.
Wavelet Transform analyzed ECG parameters are
statistically classified as emotional pattern joy, anger,
sadness and pleasure. These parameters are then applied to
ANN as training and testing data. Also, these parameters
are considered as neurons in ANN. The neurons in a
feedforward neural network are organized as a layered
structure and connected in a strictly feedforward manner.
The structure of a basic feedforward neural network is
presented in Fig.1. The feedforward neural network is one
of the most widely used ANNs. A great number of
successful applications of this type of network have been
reported [7].
So, after the four-scale decomposition procedure, we
get wavelet coefficients of maximum value value of a
typical ECG pattern. There are only four representative
samples given. Each ECG classification pattern signal is
composed by five coefficients. Then decompose the fourscale
wavelet to get the 5-dimensional feature vector as
the input feature vector of ECG for pattern recognition
with method of BP neural network.
Beacuse each line has only one element of each
column (under the painted lines are) which is much larger
than other elements, which indicates that separation
algorithm is very satisfactory.
For describing the separation effect quantitatively, we
n
1 si
− yi
2
∑
MN s
i=
1 i 2
use the expression
to compute the
average relative error of the source signals and separated
signals, where N is the number of the source, M is the
sampling points). The error curve is showed in Figure 4.
We can know the system has large error when starting the
train but the error slowly getting smaller and smaller, soon
met the desired error and the convergence effect is also
obvious in Figure 4.
18
16
14
12
10
8
6
Figure 3 The ECG signals of sample 100 and wavelet transform in four
scales
This paper uses a quadratural, compact Daubechies
2 wavelet as the base function for four-scale
decomposition of the ECG physiological signal data each
day. And extract the maximum values composition vector
of each layer in wavelet decomposition as the feature
vector of the ECG signal vector, constituting a 4-
dimensional feature vector. The ECG signals Waveform
of sample 100 and wavelet transform coefficients Wf (a,
b) in different scales are shown in fig.3.
Then we extract wavelet coefficients of the signal’s
maximum value and minimum value. We classify and
save four kinds of models of the wavelet coefficients. The
features of the ECG can be extracted from the statistical
4
2
0
0 500 1000 1500 2000 2500 3000
Figure 4. The graph of the error
IV. CONCLUSINO
In this study, a new method concluding wavelet
transform and artificial neural network is proposed for
classification of ECG arrhythmias. This method includes
two phases, which are valued feature extraction and
classification. In feature extraction phase of the proposed
method, the feature values for each arrhythmia are
extracted using discrete wavelet transform (DWT). In
classification phase, obtained valued features are used as
248