Beruflich Dokumente
Kultur Dokumente
esd <0.2, k
h ( n) is the IMF component. x ( n ) = Ck ( n ) + r ( n ) (7)
i +1
k =1
k
Let Ck ( n) = hi +1 ( n) , end up solution process of k th Ck (n) is the k th IMF component, k is
Where
IMF component. And iterate on the residual signal decomposition number, r ( n) is the residual signal.
with the equation (4):
Then we use wavelet adaptive thresholding algo-
h1k +1 ( n) = h1k ( n) Ck ( n) (4)
rithm to remove noise in Ck ( n) obtained by EMD
k +1
If h1 (n) is constant or monotone function, end up algorithm in the frontal analysis. We choose one of
and let r ( n) = h
k +1
( n) . threshold value function for the de-noising
1
computation with the adaptive threshold estimated
Although EMD de-noising algorithm can simply value of each IMF component to obtain each new IMF
remove one or more components for ECG de-noising,
component[6]. Thresholding are obtained by
it also removes useful corresponding signal compo-
nents and leads to serious signal distortion. equation (8) as shown below which are calculated with
process in the frontal chapter:
3. Wavelet adaptive thresholding 1 N 1
=2 (d i di ) 2
N i =0
(8)
Wavelet adaptive thresholding algorithm is a techn- Lastly, complete signal reconstruction with new
ique used to remove random noise and outliers from IMF components (see equation 9):
the signal before reconstruction [4-5]. Here, we utilize n
(9)
an adaptive soft thresholding strategy, as shown in
x ' (t ) = c ' (t ) + r ' (t )
i =1
i n
d + , components.
i d i In order to effectively enhance the frequency band
Where is the indicator value, is the selectivity of different channels and reduce signal
thresholding, and i is wavelet coefficients. cross-talk in different frequencies, we choose the
sym(N) wavelet as the mother wavelet after many
The choice of a thresholding value is a experiments. On the other hand, according to the order
fundamental issue, once the thresholded factorization of decomposition, the frequencies of IMF components
f must stay close to f . The error between f and show a descending trend. With the serial number N
reduced, time domain performance of the sym(N)
its thresholded approximation f is given by: wavelet gets worse and frequency domain performance
1 N 1 2
gets better. With N increased, the effects are contrary.
f i fi
R( f , f ) =
N i =0
(6)
We hope there is a good resolution in frequency
2
domain. Therefore we choose wavelet function whose
Although wavelet transforms are suitable for noise
N is bigger.
reduction, its pre-divided frequency feature limits its
In addition, we restrict the decomposition number
ability to decompose the signal into different
with three IMF components and a remaining signal to
frequencies according to the inherent characteristics of
reduce calculation time and ensure EMD effects.
ECG signals.
5. Result of simulation
4. Improved EMD-Wavelet
Because the ECG energy mainly concentrates in the In this section, we compared the three algorithms
0~100Hz scope, the smaller Sd (standard deviation) from simulation figure, SNR and MSE with MAT-
value is, the better linearity and the stability of the LAB7.0 as simulation platform.
basic modality function are, the more basic modality First we compare the performance of three
functions can decompose, we takes Sd=0.2 here. After algorithms in ECG de-noising with sym7 as wavelet
decompose ECG signal x (t ) using EMD algorithm function as shown in Figure 1. ECG signal data was
gathered from signal gathering system and the
we can express it as shown in equation (7):
826
simulation was carried out in MATLAB 7.0 algorithm is the smallest, wavelet adaptive
environment. thresholding algorithm takes second place and EMD
algorithm is the worst. MSE reflects the similarity of
pre and post de-noising signals. From the data we can
see that the similarity of both wavelet adaptive
thresholding and improved EMD-wavelet algorithm
are well.
Table 2 SNR comparison data of three algorithms
W avel et
Dat a EMD Adapt i ve EMD- wavel et
t hr eshol di ng
1 0. 1079 0. 0126 0. 0132
2 0. 1069 0. 0126 0. 0126
Figure 1. The performance of three algorithms 3 0. 1077 0. 0128 0. 0126
Fig1 shows that the signal with wavelet-EMD de- 0. 1101
4 0. 0126 0. 0125
noising algorithm is smoother than signal with merely
5 0. 1085 0. 0157 0. 0155
EMD or wavelet adaptive thresholding algorithm. This
6 0. 1055 0. 0122 0. 0119
de-noising algorithm also improved the edge problem
of EMD de-noising algorithm with small distortion.
Then, we take SNR and MSE (see equation (10) 6. Discussion
and equation (11)) as measurement index for a further
comparison of effects. Whether de-noising effect is In this paper, we proved the superiority of the imp-
good or not, its measure standard is that the bigger roved EMD-wavelet algorithm for the actual gathering
SNR is, the smaller the MSE is. ECG signal de-noising combining with EMD and
wavelet adaptive thresholding algorithms and provided
N
y2
( ( xi iyi )2 )
SNR = 10*log10 i =1 (10) a new means for ECG signal de-noising.
1 N
There are still some disadvantages in this algorithm.
MSE =
N
(x y )
i =1
i i
2 (11)
For example, during process of the signal border in
ECG algorithm, it may appear divergence phenomenon
x
Where i is the original signal and i is the de-
y and it is hard to avoid. Our work need to be perfected
noised signal. The SNR comparison data is shown in and further optimized.
Table 1.
Table 1. SNR comparison data of three algorithms
References
W avel et
Dat a EMD Adapt i ve EMD- wavel et [1] Piotr Augustyniak "Time-frequency modelling and dis-
t hr eshol di ng crimination of noise in the electrocardiogram" Physiol. Meas.
1 87. 3842 78. 8829 82. 2937 24(3), 2003, pp. 753-767
2 88. 3813 82. 5637 86. 3011 [2] D.L.Donoho, Denoising by soft thresholding. IEEE
3 87. 8577 80. 0125 82. 5862 Transactions on Information Theory,1995,41(3):617-627.
4 84. 5721
[3] E.Huang,et al. The empirical mode decomposition and
87. 9964 81. 3625
the Hilbert spectrum for nonstasionary time series analysis
5 85. 6628 77. 2689 81. 6593
[J]. Proc.R.Lond.A,1998, 454:903-905.
6 84. 9562 75. 2665 79. 6128 [4] D. L. Donoho, De-noising by soft thresholding, IEEE
Transactions on Information Theory, vol. 41, no. 3, pp. 613
Table1 appears that SNR of EMD algorithm is the 627, 1995.
largest, EMD-wavelet algorithm takes second place [5] F. Abramovich, T. Sepatinas and B. Silverman, Wavelet
and wavelet adaptive thresholding algorithm is the thresholding via a Bayesian approach. Journal of the Royal
least. However, combining with Fig 1 we can see that Statistical Society B,1998,60(4):725-749.
the effect of EMD is the worst. Why its SNR is the [6] M. Jansen, Noise reduction by wavelet thresholding (New
largest just proved this result. From TABLE I we can York, Springer Verlag, 2000).
[7] H.Y.Gao and A.G. Bruce, Waveshrink with firm
draw a conclusion that the effect of EMD-wavelet
shrinkage, Statistica Sinica, vol. 7, no. 4, pp. 855874,
algorithm is the best. 1997.
The SNR comparison data of three algorithms is
showed in Table 2.
Table 2 appears that according to a same wavelet
function, the MSE of improved EMD-wavelet
827