7/28/2019 Wnn Slides
1/60
7/28/2019 Wnn Slides
2/60
Wavelet Network
Wavelet + Neural Network
Wavelet decomposition as universal
approximation- like neural network
An alternative to neural network- having poor
convergence
Wavelet Transform as a neural network
7/28/2019 Wnn Slides
3/60
Fourier Analysis-
Which breaks down a signal into constituent sinusoids of
different frequencies
A mathematical technique for transforming our view of the
signal from time-based to frequency-based
in transforming to the frequency domain, time
information is lost.
From Fourier transform of a signal-impossible to tellwhen a particular event took place !
7/28/2019 Wnn Slides
4/60
Signals contain different non-stationary or transitory
characteristics: abrupt changes-often the most
important part of the signal
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
Time (s)
V
-And Fourier analysis is not suitable
7/28/2019 Wnn Slides
5/60
Wavelet Analysis-
Wavelet- a small wave on the surface of a liquid
Performs local analysis
--to analyze a localized area of a larger signal.
-capable of revealing aspects like trends, breakdown points,
discontinuities in higher derivatives of data set
7/28/2019 Wnn Slides
6/60
A wavelet -- a waveform of effectively limited duration that
has an average value of zero.
Sinusoids do not have limited duration they extend from
minus to plus infinity. Sinusoids are smooth and predictable,
wavelets tend to be irregular and asymmetric.
Wavelet Analysis-
7/28/2019 Wnn Slides
7/60
Fourier analysis - sine waves of various frequencies.
Wavelet analysis --breaks a signal into shifted and scaled
versions of the original (ormother) wavelet
just as some foods are better handled with a fork than a spoon
It also makes sense that local features can be described better
with wavelets that have local extent
7/28/2019 Wnn Slides
8/60
Mathematically, the process of Fourier analysis is represented by theFourier transform:
( ) ( )j t
F f t e dt
=
7/28/2019 Wnn Slides
9/60
= ''2'' )]()([),( dtettgtxftSTFT ftjx
Gabors : Short Time Fourier Transform
7/28/2019 Wnn Slides
10/60
7/28/2019 Wnn Slides
11/60
Fourier Wavelet
7/28/2019 Wnn Slides
12/60
7/28/2019 Wnn Slides
13/60
Continuous Wavelet Transform (CWT)
Multiplying each coefficient by the appropriately scaled and shifted wavelet
yields the constituent wavelets of the original signal:
( ) ( )C scale,position f t (scale,position,t)dt
=
7/28/2019 Wnn Slides
14/60
Scaling a wavelet -stretching (or compressing) it.
Scaling
7/28/2019 Wnn Slides
15/60
Shifting means---delaying (or hastening) its onset.
delaying a function f(t) by k ; f(t-k)
Shifting
7/28/2019 Wnn Slides
16/60
Take a wavelet and compare it to a section at the start of the original signal.For a C value (scale)
Shift the wavelet to the right and repeat steps 1 until youve coveredthe whole signal.
Scale (stretch) the wavelet and repeat above steps
Repeat all above steps for all scales.
Processing
7/28/2019 Wnn Slides
17/60
the CWT can operate at every scale, from that of the original signal
up to some maximum scale that you determine by trading off your
need for detailed analysis with available computational capability
The CWT is also continuous in terms of shifting: during
computation, the analyzing wavelet is shifted smoothly over the full
domain of the analyzed
function.
Continuous Wavelet Transform?
7/28/2019 Wnn Slides
18/60
Calculation at a subset of scales and positions at which to
make our calculations.
if we choose scales and positions based on power of two
so-called dyadic scales and positions then our
analysis will be much more efficient and just as accurate.
Discrete Wavelet Transform
7/28/2019 Wnn Slides
19/60
For many signals, the low-frequency content is the most
important part. It is what gives the signal its identity. The high-
frequency content, on the other hand, imparts flavor.
Human voice- remove the high-frequency components, the voice
sounds different, but we can still tell whats being said.
remove enough of the low-frequency components, we hear
gibberish.
In wavelet analysis, we often speak of:
approximations :the high-scale, low-frequency components of thesignal
details are the low-scale, high-frequency components.
Approximations and Details
7/28/2019 Wnn Slides
20/60
The filtering process, at its most basic level, looks like this:
7/28/2019 Wnn Slides
21/60
By looking carefully at the computation, we may keep only one
point out of two in each of the two 2000-length samples to get the
complete information. This is the notion ofdownsampling.
7/28/2019 Wnn Slides
22/60
Multiple-Level Decomposition
7/28/2019 Wnn Slides
23/60
Assembled back into the original signal without loss of
information. This process is called reconstruction, or
synthesis.
The mathematical manipulation that effects
synthesis is called the inverse discrete wavelet transform
(IDWT).
7/28/2019 Wnn Slides
24/60
Reconstruction
7/28/2019 Wnn Slides
25/60
S=A1+D1
=A2+D2+D1
=A3+D3+D2+D1
Reconstructed
Signal
Components
7/28/2019 Wnn Slides
26/60
Daubechies family of wavelets (dbN)
Db1 is the Haar wavelet
Haar wavelet
7/28/2019 Wnn Slides
27/60
Morletwavelet no scaling function
7/28/2019 Wnn Slides
28/60
cA3 cD3
cD2 cD1
Signal Analysissignal
7/28/2019 Wnn Slides
29/60
Signal Reconstruction
7/28/2019 Wnn Slides
30/60
X= the input vector, T= the translation parameter, d= the dilation parameter,
= the wavelet function, w= weight, = the addition parameter in the
network to take care nonzero mean function
( ) ( )2
2 21X
X X e
=
( )X d X T =
translated by a vector T and then dilated by d (scaling)
Wavelet Network
=Mexican Hat
-TJ
d1 w1
d2 w2
dJ wJ
-T1
-T2X y
+
+
+
+ +
7/28/2019 Wnn Slides
31/60
x2
(dh, Th)
w1
v1
w2
wh
1
2
h
x1
v2
vd
Hidden layer
Linear output
xn
neuron
Input
(d1, T1)
+1
( )1
output vectorinput vector
the output layer weight vector
the wavelet function
the translation vector
the dialation vectorthe bias at the output
hn
j j jj
w d
=
= + + T
y x t v x
y
x
w
t
d
7/28/2019 Wnn Slides
32/60
The updating equation is,
n n-1 n nV V K e= +
where the parameter vector V for single output case consists of
[ ]1 J 1 J 1 JV ,w .......w ,T ...T ,d ...d
=
and n stands for iteration index.
The Kalman gain is obtained by1T
n n 1 n n n n 1 nK P h R h P h
= +
The covariance has the relation,
Tn n n n 1P I K h I QI = +
and the Jecobian hn
is given by,
1n n
n
V V
yhV
==
[ ]1 1 1 1 11, ... , ... , ...n J J J J Jh wF w F wG w F =
Kalman Filtering based WN training
7/28/2019 Wnn Slides
33/60
Wavelet Network for Signal Processing Applications
(i) channel equalization and (ii) system identification
Wavelet decomposition as universal approximation- like
neural networkAn alternative to neural network- having poor
convergence
7/28/2019 Wnn Slides
34/60
Channel Equalizer
Trainingalgorithm
Z-N
Input Output
e
_+
Noise
Basic channel equalization scheme
_-
++
Channel equalization--
Telephone linesMicrowave links
Satellite channels
Underwater acoustic channels
Intersymbol interference
7/28/2019 Wnn Slides
35/60
(i) A linear channel of the type
y (n) = 0.2602 x (n) + 0.9298 x (n-1) + 0.2602 x (n-2)
(ii) A non-linear channel
y (n) = x(n) + 0.1x2 (n) + 0.05 x3 (n)
7/28/2019 Wnn Slides
36/60
2 2
y(n) y(n-1)[y(n) 2.5]y((n 1) u(n)
1 y ( ) y ( -1)n n
+ ++ = +
+ +
y((n 1) (y(n),y(n-1),u(n))f+ =
with u(n) = sin (2n/25)
A nonlinear dynamic plant described as
7/28/2019 Wnn Slides
37/60
Mackey-Glass system described by
)-(n(y1
)-y(nby(n)a)-(11)y(n10 +
+=+
Where a = 0.1, b= 0.2, = 17 and y(0) = 12.The identification considers the models as:
y (n+6) = (y (n), (n-6), (n-12), (n-18))where is an unknown nonlinear function
The networks are trained with 1000 sets.
7/28/2019 Wnn Slides
38/60
Convergence characteristic of MLANN and WN based equalizers for linear channel at(a) 20dB and (b) 30dB NSR
0 200 400 600 800 1000
-35
-30
-25
-20
-15
-10
-5
0
MLANN
WN
iteration
MSEindB
0 200 400 600 800 1000-40
-35
-30
-25
-20
-15
-10
-5
0
MLANN
WN
iteration
MSEindB
(a) (b)
7/28/2019 Wnn Slides
39/60
Bit error rate characteristic of the MLANNand WN based channel equalizer for linear
channel
0 2 4 6 8 10
MLANN
WN
SNR in dB
BitErrorRate
10-2
10-3
10-1
100
7/28/2019 Wnn Slides
40/60
Convergence characteristic of the MLANN and WN based channelequalizer for non-linear channel at (a) 20 and (b) 30 dB NSR
0 200 400 600 800 1000-40
-35
-30
-25
-20
-15
-10
-5
0
MLANN
WN
iteration
MSEindB
0 200 400 600 800 1000-40
-35
-30
-25
-20
-15
-10
-5
0
MLANN
WN
iteration
MSEindB
7/28/2019 Wnn Slides
41/60
0 2 4 6 8 10
100
Bit error rate characteristic ofthe MLANN and WN based channel equalizer for non-
linear channel
MLANN
WN
SNR in dB
BitErrorRate
10-1
10-2
7/28/2019 Wnn Slides
42/60
0 20 0 4 00 60 0 8 00 1 0 000
0 .2
0 .4
0 .6
0 .8
1
1 .2
1 .4
Test result [original () and predicted( ) data]
7/28/2019 Wnn Slides
43/60
Comparison of the system responseand the network output for MLP
7/28/2019 Wnn Slides
44/60
Comparison of the system response and thenetwork output for the WN
7/28/2019 Wnn Slides
45/60
Error square of the WN for the system
7/28/2019 Wnn Slides
46/60
Convergence characteristic of the wavelet for the system
7/28/2019 Wnn Slides
47/60
Wavelet Network forPower Disturbance Classification
Power Quality (PQ)?
Electrical Disturbances
Shunt Capacitor switching
Wh PQb i t t?
7/28/2019 Wnn Slides
48/60
computerized equipments -highly sensitive
Semiconductor industry Drive systems or robots
Programmable logic controllers
Deregulation of power industry
Why PQ becomes important?
7/28/2019 Wnn Slides
49/60
Impact to Silicon Valley
One cycle interruption makes a silicon deviceworthless
Five minutes shut down of a chip fabricationplant causes delay from a day to a week
One second of power outage makes e-commerce sites lose millions of dollars worthof business
Why PQ becomes important?
7/28/2019 Wnn Slides
50/60
Power Quality Problems Different PQ problems
Need of Diagnosis Remedy
Classification of disturbances- essential
7/28/2019 Wnn Slides
51/60
Several typical PQ disturbances
Voltage SagVoltage Swell
Impulse
Oscillatory Transient Interruption
Notch
Voltage Sag
7/28/2019 Wnn Slides
52/60
g g
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
Voltage Swell
Impulse
7/28/2019 Wnn Slides
53/60
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
Oscillatory Transient
Interruption
7/28/2019 Wnn Slides
54/60
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
Notch
Interruption
7/28/2019 Wnn Slides
55/60
Conventional Disturbance Classifier
Wavelet Decomposition
Feature Extraction
Neural Network as Classifier
Signal
T
7/28/2019 Wnn Slides
56/60
-TJ
d1 w1
d2 w2
dJ wJ
-T1
-T2X y
+
+
+
+ +
-TJ
d1 w1
d2 w2
dJ wJ
-T1
-T2X y
+
+
+
+ +
-TJ
d1 w1
d2 w2
dJ wJ
-T1
-T2X y
+
+
+
+ +
Inputfeatu
revector
C2
C6
C1
7/28/2019 Wnn Slides
57/60
50 Hz System
1-cycle data-input vector (60 samples)Training sets 15 only for each module
Testing sets 25 for each moduleStructure 60x12x1- for module-1
Network Design
Testresults:
7/28/2019 Wnn Slides
58/60
97.0%Total
38/40Notch
39/40Oscillatory Transient
39/40Impulse
40/40Outage
39/40Voltage swell
38/40Voltage sag
Classification rate ofProposed method
Different power qualitydisturbance
Test results:
Classification rate for 6-class PQ disturbances
7/28/2019 Wnn Slides
59/60
Solution to PQ classification problem
9 Wavelet network
9 EKF training
9 Classification accuracy
9 On-line application
7/28/2019 Wnn Slides
60/60
RBF network the basis functions are not orthogonal
Which implies that the RBF NN representation is not
unique
In WN- the basis function can be non radial-symmetric