Time frequency analysis

LMS Test.Lab

16A

Copyright Siemens Industry Software NV

 Table of Contents

Chapter 1 Introduction ................................................................................................... 5

Chapter 2 Linear time-frequency representations ....................................................... 7

Section 2.1 The Short Time Fourier Transform (STFT) ......................................... 7
Section 2.2 Wavelet analysis .................................................................................. 8

Chapter 3 Quadratic time-frequency representations ............................................... 11

Section 3.1 The Wigner-Ville distribution............................................................ 12
Section 3.2 Generalization .................................................................................... 13
Section 3.2.1 Spectrogram ....................................................................................... 14
Section 3.2.2 Smoothed Pseudo-Wigner Distribution (SPWD) ............................... 14
Section 3.2.3 Pseudo-Wigner Distribution (PWD) .................................................. 14
Section 3.2.4 Choi-Williams Distribution (CWD)................................................... 14

Chapter 4 References ................................................................................................... 17

16A 3
 Chapter 1 Introduction

Chapter 1 Introduction
A great deal of physical signals are non-stationary. Fourier analysis establishes
a one-to-one relationship between the time and the frequency domain, but
provides no time localization of a signal's frequency components. Whilst an
overall representation of all frequencies that appeared during the observation
period is presented, there is no indication as to exactly at what time which
frequencies were present.

Time-frequency analysis methods describe a signal jointly in terms of both time

and frequency. The aim is to find a distribution that determines the portion of
the signal's energy which lies in a particular time and/or frequency range. In
addition these distributions might or might not satisfy some other interesting
mathematical properties, such as the "marginal equations".

The instantaneous power of a signal at time t is given by

|st|2 = Energy or intensinty per unit time at time t

The intensity per unit frequency is given by the squared modulus of the Fourier
transform S(w)

|Sω|2 = Energy or intensity per unit frequency at frequency w

The joint function P(w,t) should represent the energy per unit time and per unit
frequency

Pωt = Energy or intensity per unit frequency (at frequency w) per unit time (at
time t )

Ideally summing this energy distribution over all frequencies should give the
instantaneous power

and summing over all time should give the energy density spectrum.

Equations 3-1 and 3-2 are known as the 'marginal' equations and in addition the
total energy, E

should be equal to the total energy in the signal while satisfying the marginals.

16A 5
 Chapter 1 Introduction

There are a number of distributions which satisfy equations 3-1 and 3-2 but
which demonstrate very dissimilar behavior.

In general there are two main classes of time-frequency analysis methods -

 linear techniques discussed in section 9.2
 quadratic techniques discussed in section 9.3.

6 LMS Test.Lab Time frequency analysis

 Chapter 2 Linear time-frequency representations

Chapter 2 Linear time-frequency representations

In This Chapter
The Short Time Fourier Transform (STFT) .......................7
Wavelet analysis .................................................................8
These are representations that satisfy the linearity principle. If x1, and x2 are
signals, then T(t,f) is a linear time-frequency representation if -

Two linear techniques are discussed -

 The Short Time Fourier Transform
 Wavelet analysis

Section 2.1 The Short Time Fourier Transform (STFT)

A standard method used to investigate time-varying signals is the so-called

Short Time Fourier Transform (STFT). This involves selecting a relatively
narrow observation period, applying a time window and then computing the
frequencies in that range. The observation window then slides along the entire
time signal to obtain a series of spectra shown as vertical bands in Figure 3-1.

16A 7
 Chapter 2 Linear time-frequency representations

Figure 3-1 The Short Time Fourier Transform

For a time signal s(t) multiplied by a window function g(t), the Short Time
Fourier Transform located at time t is given by

This is a useful technique if it is possible to select the observation period so that

the signal can be regarded as being stationary within that period. There are a
whole range of signals however where the frequency contents change so rapidly
that the time period required would be unacceptably small.

This technique suffers from a further disadvantage in that the same time
window is used throughout the analysis and it is this that determines the
frequency resolution (Df= 1/T). This fixed relationship means that there has to
be a trade off between frequency resolution and time resolution. So, if you have
a signal composed of short bursts interspersed with long quasi stationary
periods, then each type of signal component can be analyzed with either good
time resolution or good frequency resolution but not both.

An alternative view of the STFT is gained if it is expressed in terms of the

Fourier transforms of the signal S(w) and the window function G(w). Equation
3-4 then becomes

By analogy with the previous discussion this reflects the behavior around the
frequency w ''for all times'' as illustrated by the horizontal bands in Figure 3-1.
These bands can be regarded as a bank of bandpass filters which have impulse
responses corresponding to the window function.

Section 2.2 Wavelet analysis

A method that provides an alternative for the analysis of non-stationary signals,

where it becomes difficult to find the right compromise between time and
frequency resolution for the analysis window of the STFT is the Wavelet
analysis.

In effect, the Fourier transform decomposes the signal using a set of basis
functions, which in this case are sine waves. The Wavelet transform also
decomposes the signal, but it uses another set of basis functions, called
wavelets. These basis functions are concentrated in time, which results in a
higher time localization of the signal's energy. One prototype basis function is
defined, and a scaling factor is then used to dilate or contract this prototype

8 LMS Test.Lab Time frequency analysis

 Chapter 2 Linear time-frequency representations

function to arrive at the series of basis functions needed for the analysis.

This brings us to the definition of the Continuous Wavelet transform. If h(t) is

the prototype function (basic wavelet) localized in time t0 and frequency w0
then the scaled versions (wavelets) are given by

where a is the scale factor given by w0/w

The Continuous Wavelet Transform CWT is given by

where t is the time localization.

A disadvantage of the STFT is that it uses a single analysis window of constant

width. The result is that there is a fixed relationship between the frequency and
time resolutions. Improving one could only be achieved at the cost of the other.
Mapping this onto the time/frequency plane results in a fixed grid as shown on
Figure 3-2(a).

The use of the scaling factor to dilate or contract the basic wavelet results in an
analysis window that is narrow at high frequencies and wide at low frequencies.
Figure 3-1 likens the STFT to a series of constant width bandpass filters.
Using this concept again, the wavelet transform can be considered as a bank of
constant relative Bandwidth filters.

where c is a constant. This is illustrated in Figure 3-2(b) where by allowing both

the frequency and time resolutions to vary, a multi-resolution analysis is
possible.

16A 9
 Chapter 2 Linear time-frequency representations

Figure 3-2 Mapping of the time/frequency plane

This is in fact a very natural way to analyse a signal. Low frequencies are
phenomena that change slowly with time so requiring a low resolution in this
domain. In this situation, a good time resolution can be sacrificed for a high
frequency resolution. High frequency phenomena vary rapidly with time
which then becomes the important dimension, so under these conditions wavelet
analysis increases the time resolution at the cost of frequency. This type of
analysis is also very closely related to the human hearing process, since the
human ear seems to analyse sounds in terms of octave bands.

10 LMS Test.Lab Time frequency analysis

 Chapter 3 Quadratic time-frequency representations

Chapter 3 Quadratic time-frequency representations

In This Chapter
The Wigner-Ville distribution ............................................12
Generalization ....................................................................13
Whilst linearity is a desirable property, in many cases, it is more interesting to
interpret a time- frequency representation as a time-frequency energy
distribution which is a quadratic signal representation. This type of
time-frequency representations will exhibit many desirable mathematical
properties, but it is important to investigate the consequences of the bilinearity
principle.

The first two terms in this result can be seen as "signal terms", and the last two
terms as the "interference terms". These interference terms are necessary to
satisfy mathematically desirable properties like the "marginal equations", but
they often make interpretation of the results difficult.

The interference terms can be recognized by their oscillatory nature, and

different so -called "smoothing" techniques can be used to reduce their effect.
This, however, leads us to a new tradeoff; that of a reduction of interference
terms against time-frequency localization. The spectral smearing effect of the
smoothing windows will disperse the signal's energy in the time-frequency
plane, thereby reducing the time-frequency localization of all signal
components.

Two examples of quadratic time-frequency representations are the spectrogram

and the scalogram.

spectrogram = |STFT|2

scalogram = |WT|2

which are the two energy-counterparts of the Short-Time Fourier Transform

(STFT) and the Wavelet Transform (WT) respectively. The interference terms
for these representations only exist where different signal components overlap.
Hence if the signal components are sufficiently far apart in the time-frequency
plane, the interference terms will be essentially zero. While neither of these
representations satisfies the marginal equations, this is not of great concern for a
qualitative energy localization assessment.

For an adequate interpretation of time-frequency analysis results, it is often

16A 11
 Chapter 3 Quadratic time-frequency representations

good practice to use several techniques (STFT or WT together with a quadratic

method), which makes it possible to distinguish the "signal components" from
the "interference terms".

Section 3.1 The Wigner-Ville distribution

The Wigner-Ville distribution is

where t is the local time. In terms of the spectrum it is

where q is the local frequency.

This distribution satisfies the marginals and is real. In addition, time and
frequency shifts in the signal cause corresponding shifts in the distribution.

Many of its characteristics can be understood by considering the fact that in

equation 3-9 at any point (t) a section of data prior to this point is being
multiplied with a section following this point and the results summed. This can
be visualized by imagining that the segment to the left is folded over on top of
the segment to the right. Where there is an overlap there will be a product and
therefor a value for the distribution.

The diagram below demonstrates that for a signal only starting a time (tstart),
all points to its left have value zero resulting in a distribution with the same
value. The same will apply at the end point (tend),

12 LMS Test.Lab Time frequency analysis

 Chapter 3 Quadratic time-frequency representations

Thus one characteristic of the Wigner Ville distribution is that for a signal of
finite duration the distribution is zero up to the start and beyond the end. The
same can be said when considering the frequency version which means that for
a band limited signal, the Wigner Ville distribution will be zero outside of that
range. The same manoeuvre can be used to see why the reverse is true if at
some point the signal level drops to zero. Consider the situation illustrated
below.

At a point where the signal itself is zero (t0), multiplying the section to the left
by the section to the right results in a non-zero value. In general it can be said
that the Wigner distribution is not zero when the signal is. This unwelcome
characteristic makes it difficult to interpret, especially when analyzing signals
with many components.

The same mechanism accounts for noisiness that can be seen in the distribution
in places where it is not present in the signal as shown below.

When evaluating the distribution at point (t1) the overlapping sections will not
include the noise, but even at point (t2) where there is no noise in the signal, it
will already influence the distribution. Noise will be spread over a wider period
than occurred in the actual signal therefore.

The same reasoning can be used to explain the appearance of the interference
terms along the frequency axis. This is especially so when a signal contains
multiple frequency components at the same moment in time, which will result
in interference terms at a frequency mid way between the frequencies of the
different components. As mentioned above, these terms can easily be
recognized by their oscillatory nature and smoothing techniques can reduce
their effect. Some possible smoothing techniques are discussed below.

Section 3.2 Generalization

A generalization of the Wigner-Ville distribution leads to a whole class of

time-frequency representations, with as main desirable mathematical property

16A 13
 Chapter 3 Quadratic time-frequency representations

their invariance against operations like time shift, frequency shift or

time/frequency scaling. This means that a shift in time or frequency of the
signal leads to an equivalent shift of the time-frequency representation of that
signal, or that scaling the signal leads to a corresponding scaling of the
time-frequency representation.

This more general class of time/frequency representations are defined as follows

where Wx(t',f') is the Wigner-Ville distribution of the signal x(t), and where ΨT
is the "kernel function". It is the choice of this kernel function that determines
the basic properties of each specific time-frequency representation derived from
this general definition. The kernel function can also be seen as a smoothing
function applied to the Wigner-Ville distribution.

Typical examples of techniques that can be defined in this framework are

Section 3.2.1 Spectrogram

where the kernel = Wigner distribution of the analysis window.

Section 3.2.2 Smoothed Pseudo-Wigner Distribution (SPWD)

where the kernel =separable smoothing function with independent smoothing

spread in time- and frequency domain.

Section 3.2.3 Pseudo-Wigner Distribution (PWD)

the same as the SPWD, but with no smoothing along the frequency axis. This
can also be considered as "short-time Wigner distribution".

Section 3.2.4 Choi-Williams Distribution (CWD)

where the kernel = exponential smoothing function.

The class of shift-invariant representations (time- and frequency shifting) is also

called Cohen's class, and examples of representations belonging to that class are

14 LMS Test.Lab Time frequency analysis

 Chapter 3 Quadratic time-frequency representations

the spectrogram, the Wigner-Ville distribution, PWD, SPWD, ...

The class of time shift/time scale invariant representations is also known as the
Affine class, and examples of representations belonging to this class are the
scalogram, the Wigner-Ville distribution, CWD, ...

16A 15
 Chapter 4 References

Chapter 4 References
Books

Time-frequency analysis :

Leon Cohen - Prentice Hall - 1995 - 299 pp. - ISBN 0-13-594532-1

Papers

Linear and Quadratic Time-frequency Signal Representations :

F. Hlawatsch, G.F.Boudreaux-Bartels (IEEE SP Magazine, April 1992)

Time-frequency distributions - A review :

Leon Cohen (Proc. of IEEE, July 1989)

Wavelets and signal processing :

O. Rioul, M. Vetterli (IEEE SP Magazine, October 1991)

Time-frequency analysis applied to door slam sound quality problems. :

H. Van der Auweraer, K. Wyckaert, W. Hendrickx (Journal de physique IV,

May 1994)

16A 17
 Index

C
Choi-Williams Distribution (CWD) • 17
G
Generalization • 16
I
Introduction • 5
L
Linear time-frequency representations • 7
P
Pseudo-Wigner Distribution (PWD) • 17
Q
Quadratic time-frequency representations • 13
R
References • 19
S
Smoothed Pseudo-Wigner Distribution
(SPWD) • 17
Spectrogram • 17
T
The Short Time Fourier Transform (STFT) • 8
The Wigner-Ville distribution • 14
W
Wavelet analysis • 9

16A 19