WAVELET METHODS FOR
TIME SERIES ANALYSIS
Curso: ANLISIS DE SISTEMAS DINMICOS NO LINEALES:
APLICACIN A LOS SISTEMAS NATURALES
1. Introduction
The first recorded mention of what we now call a wavelet seems to
be in 1909 in a thesis by Alfred Haar.
The concept of wavelets in its present theoretical form was first
proposed by Jean Morlet and the team at the Marseille Theoretical
Physics Center working under Alex Grossmann in France.
The concept of wavelet analysis have been developed mainly by Y.
Mayer and his colleagues.
The main algorithm dates back to the work of Stephane Mallat in
1988.
Such research is particularly active in the United States, where it is
spearheaded by the work of scientists such as Ingrid Daubechies,
Ronald Coifman, and Victor Wichkerhauser.
Wavelets are a relatively new way of analyzing time series or images.
Wavelet analysis is in some cases complementary to existing
analysis techniques (e.g., correlation and spectral analysis) and in
other cases capable of solving problems for which little progress had
been made prior to the introduction of wavelets.
2. Summary of Fourier Theory
3. Why Wavelet analysis?
The results of the Fourier transform are the Fourier coefficients A(f),
which when are multiplied by a sinusoid of frequency f yield the
constituent sinusoidal components of the original signal.
To study Non-stationary Signals, in which their frequency
components are changing over time, we cannot use the Fourier
Transform because time information is lost.
To correct this deficiency, Gabor (1946) adapted the Fourier
Transform to analyze only a small section of the signal at a time.
Gabors adaptation, called Short-Time Fourier Transform (STFT)
maps a signal into a two-dimensional function of time and
frequency.
The STFT represents a sort of compromise between the time- and
frequency-based views of a signal. It provides some information about
both when and at what frequencies a signal event occurs. However,
you can only obtain this information with limited precision, and that
precision is determined by the size of the window.
Real signals, in the nature, have their principal frequencies in the lowfrequency region. Therefore, it is totally necessary to have higher
accuracy in this region. Therefore, wavelet analysis represents the
next logical step: a windowing technique with variable-sized regions.
Wavelet analysis allows the use of long time intervals where we want
more precise low-frequency information, and shorter regions where we
want high-frequency information.
4. The essence of a wavelet
a wavelet is a small wave.
A wavelet grows and decays essentially in a limited
time period.
To quantify the notion of a wavelet, let us consider
a real-valued function () defined over the real axis
(-, +) and satisfying two basic properties:
The integral of () is zero:
The square of () integrates to unity:
Haar wavelet function (this is the oldest wavelet)
To form the next wavelets, we start with the Gaussian probability
density function for random variable with mean zero and variance 2
The first derivative of
()
is
If we renormalize the negative of the above to satisfy Equation 2, we obtain
the wavelet.
With proper renormalization again, the negative of the second derivative
of also yields a wavelet, usually referred to as the Mexican hat:
()
5. The Essence of Wavelet Analysis
let x() be a real-valued function of an
independent variable t, that we will refer to
as time. Let us consider the integral
Where (a,b) is called the average value of
x() over the interval [a, b]
Instead of regarding an average value (a,b) as a function of the end
points of the interval [a,b] over which the integration is carried out, we
can just as easily consider it to be a function of the length of the
interval, i.e. =b-a, and the center time of the interval, i.e. t=(a+b)/2.
We refer to as the scale associated with the average. Using and t,
we can define
We call A(,t) the average value of the signal x() over a scale of
centered about time t
As a specific example consider a series of measurements related to how well
a particular cesium beam atomic clock (clock 571) keeps time on a daily basis
as compared to a master clock.
The measurements are daily average fractional frequency deviates. They can
be regarded as averages of a signal x() over intervals of the form
[t 12 , t + 12 ]
When A(1,t)<0 then clock 571 has lost time.
A good indicator of the inherent ability of clock 571 to keep good time over a
scale of a day is thus not A(1,t) itself, but rather how much A(1,t) changes from
one time period to another.
Or
if |D(1,t)| is close to zero, the daily performance of the clock is good. More generally,
the timekeeping abilities of a clock over a scale can be judged by examining
(3)
scale
We are now in a position to connect wavelets to the scheme of looking at
changes in averages. Because the two integrals in Equation (3) involve
adjacent non-overlapping intervals, it is easy to combine then into a single
integral over the entire real axis to obtain
Specializing to the case =1 and t=0 yields
Haar wavelet
The Haar wavelet extracts information about how much difference there is
between the two unit scale averages of x() bordering on time t=0
To adjust the Haar wavelet to extract information about unit scale changes at
other values of t
To shift the location of H(.)
To extract similar information about other scales and times t
It is proportional to the difference
between two adjacent averages of scale
, the first average beginning at time t,
and the second, ending at time t.
The collection of variables
{W ( , t ) : > 0, < t < }
H
is known as the
HAAR CONTINUOUS WAVELET TRANSFORM (CWT) of x(.)
We now state a fundamental fact about the CWT that it preserves all the
information in x
if
satisfies the admissibility condition of
()
()
and
Then we can recover
x()
from its CWT via
6. Five easy steps to a CWT
The CWT produces wavelet coefficients that are a function of scale and
position of the wavelet. Multiplying each coefficient by the appropriately
scale and shifted wavelet yields the constituent wavelet of the original
signal.
Scaling a wavelet simply means stretching or compressing it. The scale
factor, , works how following: The smaller the scale factor, the more
compressed the wavelet.
t
It is clear that the scale factor is related inversely to the radian frequency of
the signal.
=1
=1/2
=1/4
Shitfting a wavelet simply means delaying or hastening its onset
Therefore, the CWT is
W ( , t ) = ,t (u )x(u )du where , t (u ) =
u t
(
Its a very simple process. In fact, here are the five steps of an easy recipe for creating a
CWT:
1) Take a wavelet and compare it to a section at the start of the original signal
2) Calculate a number, C, that represents how closely correlated the wavelet is with this
section of the signal. The higher C is, the more the similarity. More precisely, if the signal
energy and the wavelet energy are equal to one, C may be interpreted as a correlation
coefficient. Note that the results will depend on the shape of the wavelet you choose.
3) Shift the wavelet to the right and repeat step1 and 2 until you are covered the whole
signal.
4) Scale (stretch) the wavelet and repeat step1 through 3.
5) Repeat steps 1 through 4 for all scales.
When you are done, you will have the coefficients produced at different scales by different
sections of the signal. The coefficients constitute the results of a repression of the original
signal performed on the wavelets.
How to make sense of all theses coefficients? You could make a plot on which the x-axis
represents position along the signal (time), the y-axis represents scale, and the colour at
each x-y point represents the magnitude of the wavelet coefficient C.
7. The Discrete Wavelet Transform
The Discrete Wavelet Transform (DWT) of {Xt} is an orthonormal transform.
The series and its transform can be considered to be two representations of the
same mathematical entity.
An N x N real-valued matrix is said to be orthonormal if its transpose is its inverse, i.e. if
T = IN
where IN is the N x N identity matrix.
We can decompose a time series X by premultiplying it by to obtain
W= X
an N dimensional column vector containing the transform coefficients; likewise,
given O, we can reconstruct X since
X=T W
The orthonormality implies that ||W||2=||X||2. Hence W n2 represents the contribution to the energy attributable to the DWT
coefficient with index n.
WT
Let us now define exactly what the notion of scale means.
For a positive integer , let
Represents the average of the contiguous data values
note that
X t (1) = X t
, which we can regard as a single point average and
X N 1 (N ) = X
which is the sample average of all N values
Note that the first eight DWT coefficients W0,,W7 are proportional to
differences in adjacent averages of {Xt} on a unit scale; the next four W8,,W11
are differences in adjacent averages on a scale of two W12 and W13 are
proportional to differences on a scale of four; W14 is proportional to a difference
on a scale of eight; and the final coefficient W15 is proportional to the average of
all the data.
There are thus exactly N/(2i) coefficients in the DWT associated with changes on
scale i where i=2j-1 for j=1,,J (note 1=1 and J=N/2); additionally, there is one
coefficient WN-1 associated with an average at scale N. The N-1 coefficients that
are associated with changes on various scales are called wavelet coefficients,
whereas WN-1 is called a scaling coefficient.
Let us now decompose the elements of the vector W into J+1 subvectors. The
first J subvectors are denoted by Wj j=1,J and the jth such a subvector contains
al of the DWT coefficients for scale j. Note that Wj is a column vector with N/2j
elements. The final subvector is denoted as VJ and contains just the scaling
coefficient WN-1. We can write
When J=4, N=2J=16
Let us now consider the wavelet synthesis of X
S J = , JTwhich
VJ is a vector whose elements are all
Recall that we have defined
For
0 j J let
1
Details (Low-scale, Highfrequency components)
Sj is a smoothed version of X
J
Sj =
k = j +1
+ SJ
Approximations (High-scale,
Low-frequency components)
Wavelet transform algorithm
Wavelet filter
Scaling filter is defined in terms of the wavelet filter via the quadrature mirror relationship
The signal S analyzed at level j has the following structure:
Some exaples of Wavelets
Families
a) Haar.
b) Daubechies.
c) Biorthogonal.
Scale to frequency:
Pseudo-frequencies corresponding to the
scales
Wavelet Toolbox of Matlab
One-Dimensional Analysis Using the
Command Line of Wavelet Toolbox of
Matlab
Approximation coefficients at level 3 (cA3)
Details coefficients at level 3 (cD3)
Details coefficients at level 2 (cD2)
Details coefficients at level 1 (cD1)
The Continuous Fourier Transform &
Amperometric Record
3.0
1.0
2.5
0.5
2.0
0.0
1.5
1.0
-0.5
0.5
-1.0
0
10
20
30
40
-3
x10
50
60
50
seno Freq Response
Data Freq Response
60V
0
40
20
-100
dB
dB
-50
-150
-200
-20
-250
-40
200
400
Hz
600
800
0
500
1000
Hz
1500
2000
WT & Amperometric Record
S
1
0
d2
d2
d3
d3
d4
d4
d5
d5
d6
d6
0.5
d1
0.5
d1
a6
a6
0.0
-0.5
0.0
-0.5
10
Tiempo (s)
15
20
10
Tiempo (s)
15
20
Scale
Scale
WT & Amperometric Record
Shift (time)
