Download - Academy Publisher
Download - Academy Publisher
Download - Academy Publisher
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
others through the selective emotional music, with a total<br />
of Record a 25-day EMG physiological signals whose<br />
signal sampling frequency is 32Hz.<br />
This paper uses a quadratural, compact Daubechies<br />
wavelet as the base function for six-scale decomposition<br />
of the EMG physiological signal data each day. And<br />
extract the maximum and minimum values composition<br />
vector of each layer in wavelet decomposition as the<br />
feature vector of the surface EMG signal vector,<br />
constituting a 14-dimensional feature vector. The EMG<br />
signals Waveform of joy and wavelet transform<br />
coefficients Wf (a, b) in different scales are shows as fig.3.<br />
For emotional joy pattern recognition next<br />
procedure, Table I shows the maximum value of 6 level<br />
wavelet decomposition in five experiments.<br />
TABLE I. the maximum of typical surface EMG joy signal<br />
joy<br />
a6 d6 d5 d4 d3 d2 d1<br />
1 137.9 29 11.99 14.38 16.67 13.76 7.097<br />
2 130.6 24.51 10.11 40.3 27 50.58 27.75<br />
3 209.8 20.83 40.21 18.28 18.53 17.68 18.77<br />
4 133.4 24.95 11.21 23.47 16.94 18.01 9.231<br />
5 91.27 21.91 9.811 13.92 8.421 8.224 8.942<br />
Then a three-layer BP neural network is used in this<br />
paper. Input nodes number is 14 and the output nodes<br />
number is 4, which represents four kinds of emotional<br />
states, respectively, joy(1000), anger(0100), sadness(0010)<br />
and pleasure(0001). For how number of the nodes<br />
selection in the middle hidden layer, experiments show<br />
that the hidden layer nodes have a significant impact for<br />
the performance of neural networks. If we select too few<br />
nodes, each category can not be separated by the network,<br />
and if too many nodes, the operation is too big, there<br />
maybe "over learning". Therefore system performance<br />
and efficiency must be taken into comprehensive<br />
consideration to determine the hidden layer nodes. In this<br />
study, after many experiments comparison, finally we<br />
select 14 as the hidden layer nodes and the effect is quite<br />
good. The training sample set is closely related to network<br />
performance. To design a good set of training samples, it<br />
is necessary to select the sample size and the quality of the<br />
sample is important. That is, the determination of the<br />
samples number and sample selection organization is<br />
important. In this paper, data of 19 days are selected as the<br />
training set. Through experimental selection and<br />
comparison, the remaining data of six days are selected as<br />
test set. Learning rate has a great impact on network<br />
performance too. The experiments show that when the<br />
learning rate is 0.01 and the precision control parameter<br />
will under 0.01.<br />
Then we get the result of emotion recognition by the<br />
use of BP neural network. After many experiment, the<br />
recognition rate can reach more than 83.33%. We can see<br />
that the emotion recognition using the BP neural network<br />
is feasible. BP neural network for solving linear equations<br />
used to achieve the classification makes the training time<br />
greatly reduced.<br />
IV. CONCLUTION<br />
Emotion recognition has a promising development<br />
prospect. Identifying the person's emotional state through<br />
the physiological signal has drawn increasing attention.<br />
This experiment introduce multi-scale decomposition<br />
wavelet of EMG signals by wavelet transform and extract<br />
the maximum and minimum of wavelet decomposition<br />
coefficients to construct signal feature vector to present<br />
the original EMG. Then enter it into the standard BP<br />
neural network classifier for emotion recognition. The<br />
kind of classify is able to detect and identify the surface<br />
EMG of four kinds of emotions, joy, anger, sadness and<br />
pleasure. Compared to the classical classifier, the emotion<br />
recognition has a better classification effect, higher<br />
recognition rate and better robustness. Experimental<br />
results show that the surface EMG feature extraction<br />
based on wavelet transform and emotional type<br />
recognition method using BP neural network as a<br />
classification tool is feasible and effective in application<br />
of emotion recognition.<br />
REFERENCES<br />
[1] Luo Senlin, Pan Limin. Affective Computing Theory and<br />
Technology [J]. Systems Engineering and Electronics,<br />
2003,25 (7) :905-909.<br />
[2] Picard R. Affective Computing [M]. England: MIT Press,<br />
1997:68.<br />
[3] Kim KH, Bang SW, Kim S R. Emotion recognition system<br />
using short-term monitoring of physiological signals [J].<br />
Medical Boilogical Engineering Computing, 2004, 42<br />
(3) ,419-427.<br />
[4] Wagner J, Kim J, AndréE. From physiological signals to<br />
emotions: implementing and comparing selected methods<br />
for feature extraction and classification [J]. IEEE<br />
International Conference on Multimedia & Expo.New<br />
York: IEEE Computer Society Press, 2005:940 -943.<br />
[5] Yang Rui request, Liu Guangyuan. BPSO four<br />
physiological signals based on the emotional state<br />
identification [J]. Computer Science, 2008,35(03),137-<br />
139.<br />
[6] Niu Xiao-wei, LIU Guang-yuan. Physiology signals based<br />
on genetic algorithm emotion recognition [J]. Southwest<br />
University (Natural Science Edition), 2007, (9) ,134-138.<br />
[7] Wu Fang-liang, Shi Zhong Kun, Yang Xiang-hui, et al.<br />
Based on BP neural network with LM algorithm submarine<br />
sonar self-noise prediction [J]. Chinese shipbuilding,<br />
2006,47 (3): 45 - 50.<br />
[8] Kun, Wang Zhizhong. Using LM algorithm is an improved<br />
BP network in Surface EMG Signal in the applied research<br />
[J]. R & D on the, 2005,29 (6) :399-401.<br />
229