Non Parametric
Non Parametric
PRESENTED BY:
BHAVIKA JETHANI (2) SUPRIYA ASUTKAR (8)
1
Advanced Digital Signal Presentation
Topics to be covered :
Non-parametric methods of Power Spectrum
Estimation.
The Bartlett Method
The Welch Method
The Blackman and Tukey Method
Comparison of performance of Non-periodic
Power Spectral Estimation Methods
2
• The well known form of power density spectrum
estimate is called as periodogram
• Periodogram is not a consistent estimate of true
Power Density Spectrum.
• That means, it does not converge to the true power
density spectrum.
• So the emphasis of classical Non-parametric Methods
is on obtaining a consistent estimate of power
spectrum through some averaging and smoothing
operations performed directly on the periodogram or
directly on the autocorrelation.
3
NON-PARAMETERIC METHODS FOR POWER
SPECTRAL ESTIMATION
4
Since the estimates are based entirely on
a finite record of data ,the frequency resolution of
these methods is , at best , equal to the spectral width
of the rectangular window of length N, which is
approximately 1/N at the -3dB points. We shall be
more precise in specifying the frequency resolution of
the specific methods. All the estimation techniques
described in this presentation “Decrease the frequency
resolution in order to reduce the variance in the
spectral estimate”.
5
THE BARTLETT METHOD : AVERAGING PERIODOGRAMS
------ (1)
------ (2)
6
Finally , we average the periodograms for the K segments
to obtain the bartlett power spectrum estimate.
------ (3)
------ (4)
7
as :
------ (5)
------ (7)
8
In return for this reduction in resolution , we have reduced the
variance. The variance of the Bartlett estimate is
------ (8)
------ (9)
9
The Welch Method:
………(2)
……..(3) 11
…..(4)
……..(5)
………(6)
……..(7)
12
……(8)
……(9)
……(10)
…….(11)
……..(12) 13
BLACKMAN AND TUKEY METHOD:SMOOTHING THE PERIODOGRAM
•In this method , the sample auto correlation sequence is windowed first
and then Fourier transformed to yield the estimate of the power
spectrum.
----- (1)
14
CONT...
•Where w(n) is window function having length (2M-1) and is zero for m≥ M
----- (2)
15
CONT...
•However some of the window function do not satisfy this condition and may
result in negative spectrum estimates
----- (4)
16
CONT...
----- (5)
----- (6)
•The expected value of the Blackman –Tukey power spectrum estimate is-
----- (7)
----- (8)
----- (9)
----- (10)
18
CONT...
----- (11)
----- (12)
19
CONT...
----- (13)
20
CONT...
----- (14)
•The variance of Pxx becomes-
----- (15)
----- (16)
21
CONT...
•When w(f) is narrowed compare to true power spectrum further approximate as-
----- (17)
22
PERFORMANCE CHARACTERISTICS
OF NON-PARAMETRIC POWER SPECTRAL ESTIMATORS
In this section the Quality of three methods i.e. Bartlett, Welch and Blackman
and Tukey power spectral estimate is being compared.
----- (1)
----- (2)
23
----- (3)
The fact that Qp is fixed and independent of data length N is another indication of
poor quality of the estimate
24
1. Bartlett power spectrum estimate:
The mean and variance of Bartlett power spectrum estimate is
----- (4)
----- (5)
25
2.Welch power spectral estimate :
The mean and variance of power spectrum estimate is
----- (7)
----- (6)
26
3.BLACKMAN-TUKEY POWER SPECTRAL ESTIMATE :
The mean and variance of this estimate is
----- (8)
----- (9)
----- (10)
Put 8 , 9 , 10 in 1
27
SUMMARY OF QUALITY OF POWER SPECTRAL ESTIMATE
CONCLUSION:
1. Welch and Blackman-Tukey power spectrum estimate is somewhat better
than Bartlett
2. However the difference in their performance is relatively small.
3. The main point is that Quality factor increases as N length of data increases.
4. This characteristic behavior is not shared by periodogram.
5. Furthermore Quality factor depends on product of length N and freq
resolution ∆f
6.For desired level of quality ∆f can be decreased(freq resolution increased)by
increasing length N of data and vice versa.
28
COMPUTATIONAL REQUIREMENT OF POWER SPECTRAL ESTIMATE
29
2 N additional computation is required for windowing.
Blackman-Tukey Power Spectral Estimate
We cant use N point DFT for its computation because its maximum value limits
to 1024 point DFT for which we required 2M point DFT and one 2M point
IDFT hence we are using the FFT algorithm.
30
Conclusion
Recorded length(M)=0.9/Δf
=0.9/0.09
=10
No. of FFT’s=N/M
=100/10
=10
L = 1.1NΔf x M
= 99.9
32
2] Welch
Welch method
method :-
Quality factor (Qw)=0.78N ΔF (Non-overlapping)
= 0.78 X 100 X0.09
= 7.02
Recorded length(m)=1.28/Δf
=1.28/0.09
=14.22
No. of FFT’s=2N/M OR =1.56 N ΔF
=2 X100/14.22 =1.56 X 100 X 0.09
=14.06 =14.04
33
BLACKMAN
3]Blackman- TukeyTUKEY
method :-
METHOD
Recorded length(2M)=1.28/Δf
=1.28/0.09
=14.22
M=7.11
No. of FFT’s=N/M
=2 X100/14.22
=14.06
34
References:
1. Digital signal processing
Fourth edition by JOHN G. PROAKIS , DIMITRIS G. MANOLAKIS
35
36