Chethan Parameswariah
Title: Understanding wavelet analysis and filters for engineering applications
Chethan Parameswariah
March 26, 2003
Electrical Engineering, Louisiana Tech University
Chethan Parameswariah
CONTENTS
Wavelet Basics Wavelet filter: pole zero locations and magnitude phase characteristics comparison between different families. Methods of identification and classification of inrush and fault current wavelet for protective relaying technique Energy Distribution of wavelet output decomposition Frequency Bandwidth characteristics of wavelet output bands Wavelet Seismic Event Detection Conclusion
March 26, 2003 Electrical Engineering, Louisiana Tech University 2
Chethan Parameswariah
WAVELET BASICS
Frequency Domain Fourier Transform
Time Series Input Signal
break
Frequency
Wavelet Transform Time Scale Domain. Multiple outputs with different frequency bands called scales (levels). X axis of each output in time units no time information lost. Y axis of each output in magnitude units.
NEXT PAGE
Time information lost
Other Method :
STFT Short Time
Fourier Transform
March 26, 2003
Electrical Engineering, Louisiana Tech University
Chethan Parameswariah Wavelet Transform Input Signal
a = approximation Fs = 1024 samples/sec # of samples = 1024 = 2^N
break
0 1.0 Hz 1.0 2.0 Hz
N = # of levels
d = details
2.0 4.0 Hz 4.0 8.0 Hz
Wavelet output decompositions
8.0 16.0 Hz 16.0 32.0 Hz 32.0 64.0 Hz 64.0 128.0 Hz 128.0 256.0 Hz
Break identified
256.0 512.0 Hz
March 26, 2003
Electrical Engineering, Louisiana Tech University
Chethan Parameswariah
Fourier Transform:
F ( w) = f (t )e j t dt
basis functions sines and cosines :
e j t = cos t + j sin t
Wavelet Transform:
CWT (b, a)= f (t )k b ,a (t )dt
basis functions compressed and translated functions of mother wavelet:
k b ,a (t ) =
t b w a a
a is the scale variable of the wavelet replaces the frequency variable in Fourier transform and b is the time shift variable.
March 26, 2003
Electrical Engineering, Louisiana Tech University
Chethan Parameswariah
Thanks to Mallats multiresolution wavelet analysis Fast wavelet transform sub-band codification using filter banks
Scaling function low pass filter output
(t ) = 2 h(k ) (2t k )
k =0
LPF and HPF are Quadrature Mirror Filters Wavelet function high pass filter output Frequency in octave bands f to 2f March 26, 2003 Electrical Engineering, Louisiana Tech University 6
g (k ) = (1) k h( N k )
w(t ) = 2 g ( k ) ( 2t k )
k =0
Chethan Parameswariah
Ideal : Filters should have perfect box function with sharp edges and no overlap In practice : it is very important to select wavelet coefficients that lead to good
filter characteristics.
Wavelet Research community presented several wavelet families - Daubechies, Symlets, Coiflets, Haar .. Each with different shape and lengths of mother wavelet leading to different wavelet filters with different properties.
Choosing the right wavelet for a specific application has been an open question due to lack of sufficient understanding.
It is a misunderstanding that any wavelet is suitable for any signal and any applications. Choosing or designing the right wavelet is crucial for a successful WT application [18] March 26, 2003 Electrical Engineering, Louisiana Tech University 7
Chethan Parameswariah Some Wavelet families and properties
Type Haar Daubechies Symlets Coiflets
Filter FIR FIR FIR FIR
Symmetry Symmetric Asymmetric Near Symmetric Near Symmetric
Orthogonality Orthogonal Orthogonal Orthogonal Orthogonal
March 26, 2003
Electrical Engineering, Louisiana Tech University
Chethan Parameswariah
WAVELET FILTERS CHARACTERISTICS.
So, discrete wavelet transform is equivalent to filtering it by a bank of filters of non-overlapping bandwidths which differ by an octave.
coefficients of these filter banks determined by the mother wavelet design important to know the behavior of these filters with these wavelet coefficients How are they different from the other filters ? Daubechies D4 scaling coefficients c(1) = 0.3415, c(2) = 0.5915, c(3) = 0.1585, c(4) = -0.0915 Daubechies D4 low pass filter coefficients
reverse(sqrt(2)*(c(1), c(2),..., c(n))
c(1) = 0.4830, c(2) = 0.8365, c(3) = 0.2241, c(4) = -0.1294
March 26, 2003 Electrical Engineering, Louisiana Tech University 9
Chethan Parameswariah
Block Diagram of wavelet filter bank
High Pass Filter Coefficients are obtained by reversing the low pass filter coefficients and then negating the odd coefficients.
Filter coefficients Daubechies D4 and D20
March 26, 2003
Electrical Engineering, Louisiana Tech University
10
Chethan Parameswariah zeros at pi (z = -1 in the figure) for the low pass decomposition filter are at the heart of wavelet theory. For the filter to behave well, they must have an extra property not built into earlier designs - sufficient number of zeros at pi [22]. Daubechies D4 wavelet with 4 coefficients has 2 zeros at pi and Daubechies D20 with 20 coefficients has 10 zeros at pi. Low Pass filter: pole zero map High Pass filter: pole zero map
zeros at pi
March 26, 2003
Electrical Engineering, Louisiana Tech University
11
Chethan Parameswariah
Comparing to equiripple filters:
equiripple filters have smallest ripple in both passband and stopband equiripple filter with 4 coefficients has one less zero at pi than Daubechies D4 equiripple filters not optimal for iteration decimation by 2 operators mix up the frequency bands
For every pth zero at pi for wavelet low pass filter: there exists p vanishing moments in the high pass filter for orthogonal wavelets. The decay towards low frequencies of high pass filter governed by p. smoothness (s) of the wavelet function is less than the order of the polynomial p and corresponds to the decay at high frequencies of low pass filter
March 26, 2003 Electrical Engineering, Louisiana Tech University 12
Chethan Parameswariah Scaling function Symlets wavelets Coefficients
more symmetric than Daubechies
pole zero maps
Zeros at pi
March 26, 2003
Electrical Engineering, Louisiana Tech University
13
Chethan Parameswariah Scaling function Coiflets wavelets Coefficients
more symmetric than both Daubechies and Symlets
pole zero maps
March 26, 2003
Electrical Engineering, Louisiana Tech University
14
Chethan Parameswariah
Magnitude Characteristics
LOW PASS WAVELET FILTERS
Some filters have exactly same magnitude characteristics but different phase characteristics Phase Characteristics
Magnitude Characteristics in dB
March 26, 2003
Electrical Engineering, Louisiana Tech University
15
Chethan Parameswariah
Magnitude Characteristics
HIGH PASS WAVELET FILTERS
Some filters have exactly same magnitude characteristics but different phase characteristics Phase Characteristics
Magnitude Characteristics in dB
March 26, 2003
Electrical Engineering, Louisiana Tech University
16
Chethan Parameswariah
Observations:
For use of wavelets in signal processing, knowledge of the magnitude and phase response of filters is important. information obtained - one of the properties that might be useful in choosing the right wavelet. Two wavelet filters - same magnitude characteristics but different phase characteristics. For applications with importance of time, the wavelet filter with lesser slope is chosen since it has a smaller time delay.
So lets see an application of wavelets to fast detection of fault current in a transformer !!!
March 26, 2003 Electrical Engineering, Louisiana Tech University 17
Chethan Parameswariah
METHODS OF INRUSH AND FAULT CURRENT IDENTIFICATION/PROTECTION USING WAVELETS Inrush current transient current which occurs when transformer turned on Amplitude much higher than rated full load current causes false tripping of differential relays used for protection. Fault current short circuit of output or high current at output. transient at the time of occurrence
March 26, 2003 Electrical Engineering, Louisiana Tech University 18
Chethan Parameswariah
Inrush Current :
inrush current has intermittent peaks with valleys in between them
Fault Current :
fault current on the other hand is perfectly symmetric with both positive and negative peaks March 26, 2003 Electrical Engineering, Louisiana Tech University 19
Chethan Parameswariah
Applying Daubechies D4 and D20 wavelet transform to inrush current:
D4 detail d4
large amplitude peaks corresponding to the peaks of the inrush current and smaller amplitude signal almost closer to a value of zero corresponding to the valleys (null period) in between the peaks
D20 detail d4
Higher number of coefficients smoothens the output decomposition so peaks and valleys indistinguishable.
d4
Shown only relevant wavelet output March 26, 2003 Electrical Engineering, Louisiana Tech University 20
Chethan Parameswariah
Applying Daubechies D4 and D20 wavelet transform to fault current:
D4 detail d4
d4
do not show any distinctive behavior except for a few large spikes at the state of transition from the normal current to the fault current.
D20 detail d4
d4
Shown only relevant wavelet output March 26, 2003 Electrical Engineering, Louisiana Tech University 21
Chethan Parameswariah
D4 Daubechies wavelet :
clearly distinguishes the inrush current from fault current better compared to Daubechies D20 for this application uses the detection principle based on the null period that exists in the inrush current better than the existing second harmonic component method used [28], [29].
Protective relating techniques:
Identification of the fault current in the power system and shutting down the power detection and classification - reliably conclude, in a very short time (1 2 cycles), whether and which type of fault occurs under a variety of time-changing operating conditions [51] important characteristics of a good relaying system is its short time response.
Wavelets have very interesting characteristics that can provide answers to the time response of the relay algorithm
March 26, 2003 Electrical Engineering, Louisiana Tech University 22
Chethan Parameswariah
Fast detection of fault current and response:
Daubechies D4
Daubechies D20
Daubechies D4 with fewer coefficients has smaller phase slope, hence smaller delay and faster response. Promising tool for further study towards applying them to the area of power protection systems. March 26, 2003 Electrical Engineering, Louisiana Tech University 23
Chethan Parameswariah
ENERGY DISTRIBUTION OF WAVELET DECOMPOSITIONS
Previous fault current example: Can we identify the frequency composition of the fault signal using wavelets ? Daubechies D4: o the output amplitude is maximum in the detail d6 (level 5) after fault o detail d5 (level 6) having approximately half the maximum amplitude Daubechies D20: Frequency composition of signal easily identified. o the output amplitude is maximum in the detail d6 (level 5) after fault o detail d5 (level 6) having very small (1/5th the maximum) amplitude
Energy of the output wavelet decomposition using D4 is more spread out between two frequency bands compared to output wavelet decomposition using D20
March 26, 2003
Electrical Engineering, Louisiana Tech University
24
Chethan Parameswariah
Wavelet decomposition output of transformer fault current signal
Daubechies D4 Daubechies D20
d6
d6
Frequency band: 40 80 Hz
Frequency band: 40 80 Hz
d5
d5
Frequency band: 80 160 Hz
Frequency band: 80 160 Hz
March 26, 2003
Electrical Engineering, Louisiana Tech University
25
Chethan Parameswariah
To further study:
Consider a simulated sinusoidal signal with 60 Hz and 420 Hz
f (t ) = 2 sin( 2 60t ) + 0.5 2 sin( 2 420t )
Three cycles of the signal at 60 Hz with 512 data points and sampled at a 10,240 Hz
rms ( f (t )) = 1.25 = 1.118
March 26, 2003
Energy of the signal = (rms( f (t ))) 2 = 1.25
26
Electrical Engineering, Louisiana Tech University
Chethan Parameswariah
Perform wavelet transform on the signal Calculate the rms and energy values for each output frequency bands.
The rms value of the individual wavelet level L is given by:
rms (levelL) = sqrt ( 1
512
sum( A( L, :) 2 )
where A(L,:) is the coefficients of the wavelet decomposition level L signal
Total energy of the signal is given by sum of energies of individual output frequency bands:
n =1
f ( n) = a j ( n) +
n =1 j =1
d
n =1
2 j
( n)
March 26, 2003
Electrical Engineering, Louisiana Tech University
27
Chethan Parameswariah
Daubechies D20 wavelet decomposition of the simulated signal
Wavelet Level Frequency band Center frequency Rms value Energy
0 (a8) 1 (d8) 2(d7) 3(d6) 4(d5) 5(d4) 6(d3) 7(d2) 8(d1)
0-20 Hz 20 Hz 40 Hz 40 Hz 80 Hz 80 Hz 160 Hz 160 Hz 320 Hz 320 Hz 640 Hz 640 Hz 1280 Hz 1280 Hz 2560 Hz 2560 Hz 5120 Hz
10 Hz 30 Hz 60 Hz 120 Hz 240 Hz 480 Hz 960 Hz 1920 Hz 3840 Hz
0.1981 0.2827 0.8779 0.1777 0.0775 0.4831 0.0596 0.0074 0.0021
0.0392 0.0799 0.7706 0.0316 0.0060 0.2334 0.0035 0.0001 0.0000
(0.1981) 2 + (0.2827) 2 + (0.8779) 2 + (0.1777) 2
(rmsvalue)
= + (0.0775) 2 + (0.4831) 2 + (0.0596) 2 + (0.0074) 2 + (0.0021) 2
rms ( f (t )) = 1.16 = 1.08
Energy = 1 .16
March 26, 2003
Electrical Engineering, Louisiana Tech University
28
Chethan Parameswariah
Daubechies D4 wavelet decomposition of the simulated signal
Wavelet Level Frequency band Center frequency Rms value Energy
0 (a8) 1 (d8) 2(d7) 3(d6) 4(d5) 5(d4) 6(d3) 7(d2) 8(d1)
0-20 Hz 20 Hz 40 Hz 40 Hz 80 Hz 80 Hz 160 Hz 160 Hz 320 Hz 320 Hz 640 Hz 640 Hz 1280 Hz 1280 Hz 2560 Hz 2560 Hz 5120 Hz
10 Hz 30 Hz 60 Hz 120 Hz 240 Hz 480 Hz 960 Hz 1920 Hz 3840 Hz
0.2704 0.2909 0.7603 0.4817 0.2342 0.4111 0.1890 0.0562 0.0156
0.0731 0.0846 0.5780 0.2320 0.0548 0.1690 0.0357 0.0032 0.0002
(0.2704) 2 + (0.2909) 2 + (0.7603) 2 + (0.4817) 2
(rmsvalue)
= + (0.2342) 2 + (0.4111) 2 + (0.1890) 2 + (0.0562) 2 + (0.0156) 2
rms ( f (t )) = 1.23 = 1.11
Energy = 1.23
March 26, 2003
Electrical Engineering, Louisiana Tech University
29
Chethan Parameswariah
Observations
Daubechies D20:
energy is concentrated in wavelet level 2 (d7: 40 Hz to 80 Hz ) and wavelet level 5 (d4: 320 Hz to 640 Hz ).
Daubechies D4:
energy is still concentrated in wavelet level 2 (d7: 40 Hz to 80 Hz ) and wavelet level 5 (d4: 320 Hz to 640 Hz ). the energy is also considerably high in level 3 (d6: 80 Hz to 160 Hz) and increased energy levels in adjacent bands compared to Daubechies D20. Energy distribution varies within wavelet levels for different wavelet families
So what is the frequency bandwidth characteristics of these wavelet levels?
March 26, 2003
Electrical Engineering, Louisiana Tech University
30
Chethan Parameswariah
Energy distribution in output wavelet decomposition of simulated signal
COIFLETS COIF2
SYMLETS SYM20
COIFLETS COIF5
March 26, 2003
Electrical Engineering, Louisiana Tech University
31
Chethan Parameswariah
COMPARISON
Signal with single frequency at the center of the whole frequency range
March 26, 2003
Electrical Engineering, Louisiana Tech University
32
Chethan Parameswariah
FREQUENCY BANDWIDTH CHARACTERISTICS
Select say, wavelet level 2 (d7: 40 Hz to 80 Hz), sweep the sinusoidal signals frequency from 20 Hz to 100 Hz calculate the rms value of the wavelet level at different frequency points along the sweep For output level 2 (d7), input signal passes through few low pass filters and one high pass filter with down-sampling in between each filter.
512 data points which provides us with 9 wavelet levels March 26, 2003 Electrical Engineering, Louisiana Tech University 33
Chethan Parameswariah
FREQUENCY BANDWIDTH CHARACTERISTICS DAUBECHIES D4 wavelet level d7
Freq in Hz 20 21 22 23 24 25 26 27 28 29 30 31 : : RMS value 0.2489 0.2609 0.2671 0.2709 0.2779 0.2939 0.3214 0.3583 0.3979 0.4324 0.4540 0.4570 : : decibel value (dB) -26.7633 -25.8221 -25.3493 -25.0644 -24.5555 -23.4407 -21.6459 -19.4763 -17.3782 -15.7179 -14.7393 -14.6103 : :
Freq Range: 40 80 Hz
Freq in Hz : : 90 91 92 93 94 95 96 97 98 99 100
RMS value : : 0.5698 0.5324 0.4903 0.4568 0.4422 0.4473 0.4638 0.4798 0.4859 0.4772 0.4533
decibel value (dB) : : -10.1972 -11.5540 -13.2042 -14.6163 -15.2676 -15.0380 -14.3151 -13.6351 -13.3815 -13.7455 -14.7734
March 26, 2003
Electrical Engineering, Louisiana Tech University
34
Chethan Parameswariah
FREQUENCY BANDWIDTH CHARACTERISTICS DAUBECHIES D20 wavelet level d7
Freq in Hz 20 21 22 23 24 25 26 27 28 29 30 31 : : RMS value 0.0659 0.0630 0.0594 0.0638 0.0833 0.1152 0.1529 0.1904 0.2225 0.2441 0.2510 0.2400 : : decibel value (dB) -52.4409 -53.3332 -54.5011 -53.0709 -47.7443 -41.2637 -35.5995 -31.2079 -28.0998 -26.2465 -25.6867 -26.5807 : :
Freq Range: 40 80 Hz
Freq in Hz : : 90 91 92 93 94 95 96 97 98 99 100
RMS value : : 0.5004 0.4602 0.4182 0.3831 0.3612 0.3528 0.3529 0.3543 0.3511 0.3403 0.3222
decibel value (dB) : : -11.8857 -13.5594 -15.4772 -17.2284 -18.4082 -18.8767 -18.8686 -18.7896 -18.9724 -19.5953 -20.6934
March 26, 2003
Electrical Engineering, Louisiana Tech University
35
Chethan Parameswariah
Observations: The passband and stop band regions are not flat. Daubechies D20 wavelet decomposition is more flat in the passband. The lower frequency side lobe is better in the Daubechies D20. The higher frequency side lobes are same in both.
Daubechies D20 with higher number of coefficients is better wavelet for quantifying the signal frequencies in the passband.
March 26, 2003
Electrical Engineering, Louisiana Tech University
36
Chethan Parameswariah
Energy Leakage into adjacent bands. D4 D4
Level d6
Level d8
D20
Frequency band: 20 - 40 Hz
D20
Frequency band: 80 160 Hz
March 26, 2003
Electrical Engineering, Louisiana Tech University
37
Chethan Parameswariah
FREQUENCY BANDWIDTH CHARACTERISTICS wavelet level d6 Freq Range: 80 160 Hz
Daubechies D4
Daubechies D20
March 26, 2003
Electrical Engineering, Louisiana Tech University
38
Chethan Parameswariah
FREQUENCY BANDWIDTH CHARACTERISTICS Symlets wavelet level d7 Freq Range: 40 - 80 Hz
wavelet d6
wavelet d8
March 26, 2003
Electrical Engineering, Louisiana Tech University
39
Chethan Parameswariah
FREQUENCY BANDWIDTH CHARACTERISTICS Coiflets wavelet level d7 Freq Range: 40 - 80 Hz
wavelet d6
wavelet d8
March 26, 2003
Electrical Engineering, Louisiana Tech University
40
Chethan Parameswariah
So from frequency bandwidth characteristics: wavelets do not exhibit the ideal non-overlapping bandwidths Daubechies D4, the bandwidth characteristics are poorer lot of leakage Daubechies D20 wavelet is better than D4 narrower transition regions ripples in both the pass band and the stop band D20, Symlets and Coiflets flat response in the middle of the band Careful selection of the number of data points and the sampling frequency to avoid the frequency of interest from falling on the edges of the band
March 26, 2003
Electrical Engineering, Louisiana Tech University
41
Chethan Parameswariah
Selection of sampling frequency
Frequency bands for sampling frequency 1024 Hz Frequency bands for sampling frequency 10240 Hz 5120 Hz to 2560 Hz 2560 Hz to 1280 Hz 1280 Hz to 640 Hz 640 Hz to 320 Hz 160 Hz to 80 Hz 80 Hz to 40 Hz 40 Hz to 20 Hz 20 Hz to 10 Hz DC to 10 Hz
Wavelet level
d1 d2 d3 d4 d5 d6 d7 d8 a8
512 Hz to 256 Hz 256 Hz to 128 Hz 128 Hz to 64 Hz 64 Hz to 32 Hz 32 Hz to 16 Hz 16 Hz to 8 Hz 8 Hz to 4 Hz 4 Hz to 2 Hz DC to 2 Hz
1024 Hz sampling frequency 60 Hz at the end region of level d4 10240 Hz sampling frequency 60 Hz in the middle of the level d6 For qualitative and quantitative signal analysis sampling frequency and number of data points important.
March 26, 2003 Electrical Engineering, Louisiana Tech University 42
Chethan Parameswariah
WAVELET SEISMIC EVENT DETECTION
Current Methods: STA/LTA (Short Time Average/Long Time Average) phase picker method new event detector by Murdock and Hutt the Seismic Research Observatories (1983) Wavelet Method:
properties of wavelet transforms - promising potential for use idea of using wavelets has been floated around and has created interest
Typical seismic earthquake signal
March 26, 2003
Electrical Engineering, Louisiana Tech University
43
Chethan Parameswariah
BACKGROUND
Seismic signal :
s (t )= n(t ) + e(t )
i.e., noise + event
seismic signal can be represented in its wavelet bases as:
s (t ) = d j ,n ( s ) j ,n (t )
j ,n
where
for
j>0
and
nI
d j , n ( s ) = d j , n ( n ) + d j , n (e)
The noise n(t) or is modeled for each wavelet level as :
d j ,n (n) = N (0, 2 ) j
Gaussian noise with a mean of zero 2 and a standard deviation of j
March 26, 2003 Electrical Engineering, Louisiana Tech University 44
Chethan Parameswariah
multiscale threshold test : hypothesize that for any event to be detected in the seismic signal
d j ,n ( s ) > 4 j
the occurrence of an event in the signal is true if any of the wavelet level amplitude falls outside the boundary of the noise model
Earthquakes considered:
Earthquake data: available from the IRIS Consortium full sampling rate of 40 samples/sec Central Alaska Southern Indiana New York Washington South India 7.9 5.0 5.1 6.8 7.7 [Link] [Link] [Link] [Link] [Link] Nov 3, 2002 Jun 18, 2002 Apr 20, 2002 Feb 28, 2001 Jan 26, 2001 Earthquake Magnitude Time in UTC Date
March 26, 2003
Electrical Engineering, Louisiana Tech University
45
Chethan Parameswariah
New York Earthquake:
55 minutes of seismic data with 127050 data points Oxford, Mississippi seismic station
Wavelet transform window
March 26, 2003
Electrical Engineering, Louisiana Tech University
46
Chethan Parameswariah
For sampling rate : 40 samples/sec Select number of data points N: 1024 two wavelet levels 2 and 3, whose frequency ranges from 0.075 to 0.31 Hz
Level Wavelet Scale Frequency range (Hz) Center Frequency (Hz)
covers microseism peak 0.1 0.3 Hz
0 1 2 3 4 5 6 7 8 9
a9 d9 d8 d7 d6 d5 d4 d3 d2 d1
0 0.037 0.037 0.075 0.075 0.15 0.15 0.31 0.31 0.62 0.625 1.25 1.25 2.50 2.5 5.0 5 - 10 10 - 20
0.018 0.058 0.117 0.234 0.468 0.937 1.875 3.75 7.5 15
Fourier Transform of seismic signal 47
March 26, 2003
Electrical Engineering, Louisiana Tech University
Chethan Parameswariah First 1024 points rms values of wavelet transform Next 4 windows rms values
Before the arrival of P wave
Perform wavelet transform - Daubechies D20 Calculate the rms values for each window Plot March 26, 2003 Electrical Engineering, Louisiana Tech University 48
Chethan Parameswariah
Next 4 windows
Level 5 (d5) wavelet levels rms values crosses threshold 200
Corresponding time series window
March 26, 2003
Electrical Engineering, Louisiana Tech University
49
Chethan Parameswariah
New York Earthquake:
P phase was detected at the sixth window time of arrival of 153.6 seconds - 25.6 seconds/window x 6 good 150 seconds earlier detection than conventional method used. inherent delay of 25.6 seconds reduced by over lapping windows.
STA LTA method
STA/LTA Method: misses the P phase wave LTA 45 secs
March 26, 2003
Electrical Engineering, Louisiana Tech University
50
Chethan Parameswariah
Continuing:
corresponding time window
S phase wave - Signal peaked at 17th window Time of arrival : 435.2 secs
difference S phase arrival time - P phase arrival time is: 435.2 (7.25 min) 153.6 (2.56 min) = 281.6 secs (4.69min) approx equal to the distance between the epicenter of the earthquake and Oxford, Mississippi = 1583 miles
March 26, 2003
Electrical Engineering, Louisiana Tech University
51
Chethan Parameswariah
New York Earthquake: Junction City, TX station
P phase detected 9th window arrival time of 230.4 seconds
Seismic signal recorded
March 26, 2003
Electrical Engineering, Louisiana Tech University
52
Chethan Parameswariah
New York Earthquake: Junction City, TX station
S phase is detected in the 25th wavelet decomposition window the time of arrival is 640 seconds
640 (10.67 min) 230.4 (3.84 min) = 409.6 seconds (6.83 min)
Distance to epicenter = 2545 miles
March 26, 2003
Electrical Engineering, Louisiana Tech University
53
Chethan Parameswariah
Improving accuracy with moving window
Plots of rms values for level 5 (d5) wavelet decomposition
Fixed non-overlapping window Zoomed in at the P phase
overlapping window 128 point shift
March 26, 2003 Electrical Engineering, Louisiana Tech University 54
Chethan Parameswariah
Earthquakes recorded at given stations
New York (20 Apr 2002 Plattsburg, NY) - Oxford, MS - Junction, TX Southern Indiana (18 June 2002 - Evansville, IN) - Oxford, MS - Junction, TX Washington (28 Feb 2001 Nisqually, WA) - Pine, OR - Longmire, WA South India (26 Jan 2001 Gujarat, India ) - Kislovodsk, Ru - Ala Archa, Ky Central Alaska (3 Nov 2002 Cantwell, AK) - Oxford, MS - Junction, TX 297.6 297.6 1562.4 1562.4 1264(12640) 1264(12640) 230.4 204.8 723.2 556.8 492.8 (4928) 352 (3520) 60.8 3.2 192 38.2 131.2 (1312) 35.2 (352) 153.6 192.0 220.8 428.8 67.2 (672) 236.8 (2368) 124.8 156.8 422.4 614.4 297.6 (2976) 457.6 (4576)
P Phase arrival time (seconds)
S Phase arrival time (seconds)
Estimated time difference seconds (distance in kms)
March 26, 2003
Electrical Engineering, Louisiana Tech University
55
Chethan Parameswariah
CONCLUSION
Summarizing the contributions of this research: Understanding pole zero and magnitude phase characteristics of wavelet filters Shown some wavelet filters have same magnitude characteristics, different phase response Presented - methods of identification, detection of inrush and fault currents Proposed a protection method for faults Research on energy distribution and leakage in output wavelet decompositions Study of frequency bandwidth characteristics of wavelet levels and comparison Importance of careful selection of sampling frequency and number of data points Proposing a wavelet method for detection of P-phase and S-phase waves in earthquakes References: Listed in the dissertation
March 26, 2003
Matlab programs: Listed in Appendix
56
Electrical Engineering, Louisiana Tech University
Chethan Parameswariah
Thanks for your patience Questions ?
March 26, 2003
Electrical Engineering, Louisiana Tech University
57
