Short-Time Fourier Transform

4/7/2014
An Introduction to
Short-Time Fourier Transform (STFT)
Advanced Structural Dynamics
M Ahmadizadeh, PhD, PE
Contents
• Scope and Goals
• Fourier Transform Review

– Advantages of Fourier Transform
– Limitations of Fourier Transform
• Short Time Fourier Transform

– Concept
– Formulation
– Examples
• Use of MATLAB
– Applications
1
4/7/2014
Scope and Goals
• To expand the capabilities of Fourier transform for time-

varying signals
• In addition to showing the frequency content of the signals,

it is also desirable to have an idea of when each frequency
content is dominant
• Describing complicated signals with fewer number of

frequencies
Fourier Transform
• Fourier analysis expands signals or functions in terms

of sinusoids (or complex exponentials).
• It reveals all frequency components present in a
signal.
Inverse DFT
DFT
2
4/7/2014
Fourier Transform
• Example: 5 Hz Signal
5 Hz
1
0.5
Signal
-0.5
-1
0 0.2 0.4 0.6 0.8 1
Time (s) Output
Amplitude 20
10
0
0 20 40 60 80 100
5
Phase
-5
0 20 40 60 80 100
Frequency (Hz) 5
Fourier Transform
15 Hz
1
0.5
Signal
-0.5
-1
0 0.2 0.4 0.6 0.8 1
Time (s) Output
20
Amplitude
10
0
0 20 40 60 80 100
10
Phase
0
0 20 40 60 80 100
Frequency (Hz) 6
3
4/7/2014
Fourier Transform
30 Hz
1
0.5
Signal
-0.5
-1
0 0.2 0.4 0.6 0.8 1
Time (s) Output
Amplitude 20
10
0
0 20 40 60 80 100
10
Phase
0
0 20 40 60 80 100
Frequency (Hz) 7
Fourier Transform
• Example: Signal Containing 5, 15 and 30 Hz Components

3 Frequencies
3
1
Signal
-1
-2
-3
0 0.2 0.4 0.6 0.8 1 Output
Time (s) 15
Amplitude
10
0
0 20 40 60 80 100
10
Phase
0
0 20 40 60 80 100
Frequency (Hz)
8
4
4/7/2014
Fourier Transform - Limitations
• Cannot provide information about both time and

frequency – i.e. cannot provide simultaneous time and
frequency localization
–  Not suitable for analyzing time-variant, non-

stationary signals
• As an example, it cannot be used to analyze the signals

obtained from nonlinear or damaged structures, since they
may contain varying frequency content at different times
• Signals: Stationary vs Non-stationary
– Stationary signals have properties (e.g frequency

content) that does not change over time
3 Frequencies
3
Sum of 3 signals, present at all

1
Signal
times -1
-2
-3
0 0.2 0.4 0.6 0.8 1
Time (s)
– Non-stationary signals may have different frequencies at

different times 1
3 Frequencies
0.5
Combination of 3 signals
Signal
occurring at different times -0.5
-1
0 0.2 0.4 0.6 0.8 1
Time (s)
10
5
4/7/2014
• 5, 15 and 30 Hz Components at Different Times

3 Frequencies
1
0.5
Signal
-0.5
-1
0 0.2 0.4 0.6 0.8 1
Output
Time (s)
4
Amplitude
2
Excellent frequency 0
0 20 40 60 80 100
localization (shows all 50
existing frequencies), but 0
no time localization
Phase
-50
-100
0 20 40 60 80 100
Frequency (Hz)
11
• Not appropriate for representing discontinuities or

sharp corners
–  Requires a large number of Fourier components to

represent discontinuities).
12
6
4/7/2014
• El Centro Earthquake Record and its Fourier transform

El Centro
3
1
Acceleration
-1
Output
-2 3
Amplitude
-3 2
-4 1
0 5 10 15 20 25 30 35
Time 0
0 5 10 15 20 25 30
500
Phase
-500
0 5 10 15 20 25 30
Frequency (Hz) 13

El Centro
3
1
Acceleration
-1 -3 El Centro - Components: 1
x 10
1
-2
-3
0.5
-4
0 5 10 15 20 25 30 35
Acceleration
Time
0
Included Signals: 1 -0.5
-1
0 5 10 15 20 25 30 35
Time
14
7
4/7/2014

El Centro
3
1
Acceleration
-1 El Centro - Components: 10
0.15
-2
0.1
-3
0.05
-4
0 5 10 15 20 25 30 35
Acceleration
Time 0
-0.05
Included Signals: 10 -0.1
-0.15
-0.2
0 5 10 15 20 25 30 35
Time
15

El Centro
3
1
Acceleration
1.5
-2
-3 1
-4 0.5
0 5 10 15 20 25 30 35
Acceleration
Time
0
-0.5
Included Signals: 50
-1
-1.5
0 5 10 15 20 25 30 35
Time
16
8
4/7/2014

El Centro
3
1
Acceleration

3
-2
-3 2
-4 1
0 5 10 15 20 25 30 35
Acceleration
Time
0
-1
-2
-3
0 5 10 15 20 25 30 35
Time
17

El Centro
3
1
Acceleration

3
-2
2
-3
1
-4
0 5 10 15 20 25 30 35
Acceleration
Time 0
-1
Included Signals: 800 -2
-3
-4
0 5 10 15 20 25 30 35
Time
18
9
4/7/2014
Short-Time Fourier Transform
• Basic Concept:
– Break up the signal in time domain to a number of

signals of shorter duration, then transform each signal
to frequency domain
• Requires fewer number of harmonics to regenerate the

signal chunks
• Helps determine the time interval in which certain

frequencies occur
19
• Consider the first 3 seconds of the El Centro Record

El Centro
3
1
Acceleration
-1
-2
Output
4
-3
Amplitude
-4 2
0 0.5 1 1.5 2 2.5 3 3.5
Time
0
0 5 10 15 20 25 30
100
50
Phase
-50
0 5 10 15 20 25 30
Frequency (Hz)
20
10
4/7/2014

El Centro
3
1
Acceleration
-1
3
-3
2
-4
0 0.5 1 1.5 2 2.5 3 3.5
Time 1
Acceleration
0
-1
-2
-3
0 0.5 1 1.5 2 2.5 3 3.5
Time
21

El Centro
3
1
Acceleration
-1
3
-3
2
-4
0 0.5 1 1.5 2 2.5 3 3.5
Time 1
Acceleration
0
-1
-2
-3
0 0.5 1 1.5 2 2.5 3 3.5
Time
22
11
4/7/2014

El Centro
3
1
Acceleration
-1
3
-3
2
-4
0 0.5 1 1.5 2 2.5 3 3.5 1
Time
Acceleration
0
Included Signals: 80 -1
-2
-3
-4
0 0.5 1 1.5 2 2.5 3 3.5
Time
23
• Or consider the first chunk of non-stationary record

3 Frequencies
1
Knowing the
location of the
0.5
considered chunk
in time and its
frequency
Signal
0 content, we can
get an idea of the
time each
-0.5 frequency exists.
-1
0 0.2 0.4 0.6 0.8 1
Time (s)
24
12
4/7/2014
 Need a local analysis scheme for a time-frequency

representation (TFR).
 Windowed F.T. or Short Time F.T. (STFT)
 Segmenting the signal into narrow time intervals (i.e.,
narrow enough to be considered stationary).
 Take the Fourier transform of each segment.
Non-Stationary
1
0.5
Signal
-0.5
-1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time (s)
25
• Steps:
– Choose a window function of finite length
• A window function is a function that is multiplied by the signal to
keep a certain portion of it
– Place the window on top of the signal at t=0
– Truncate the signal using this window

26
13
4/7/2014
• Steps (continued)
– Compute the FT of the truncated signal, save results.
• For each time location where the window is centered, we
obtain a different FT
– Each FT provides the spectral information of a separate time-
slice of the signal, providing simultaneous time and frequency
information
– Incrementally slide the window to the right
– Repeat until window reaches the end of the signal
27
• What is being Fourier-transformed: Window Length: 0.2s

Hann Windowed Non-Stationary Signal Hann Windowed Non-Stationary Signal Hann Windowed Non-Stationary Signal
0.8 1 1
0.6
0.4 0.5 0.5
0.2
Signal
Signal
Signal
0 0 0
-0.2
-0.4 -0.5 -0.5
-0.6
-0.8 -1 -1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time (s) Time (s) Time (s)
  0.1s   0.5s   0.9s

Hann Windowed Non-Stationary Signal Hann Windowed Non-Stationary Signal
0.8 1
0.6
0.4 0.5
0.2
Signal
Signal
0 0
-0.2
-0.4 -0.5
-0.6
-0.8 -1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time (s) Time (s)
  0.3s   0.7 s
28
14
4/7/2014
• Fourier transforms of windowed signal:
Output Output Output

1.5 1.5 1.5
Amplitude
Amplitude
Amplitude
1 1 1
0.5 0.5 0.5
0 0 0
0 20 40 60 80 100
0 20 40 60 80 100
0 20 40 60 80 100
  0.1s   0.5s   0.9s

5 4 20
0 2
Phase
Phase
Phase
10
-5 0
Output Output
-10 1 -2 1 0
0 20 40 60 80 0
100 20 40 60 80 100
0 20 40 60 80 100
Frequency (Hz) Frequency (Hz) Frequency (Hz)
Amplitude
Amplitude
0.5 0.5
0 0
0 20 40 60 80 100 0 20 40 60 80 100
  0.3s   0.7 s
0 300
-100 200
Phase
Phase
-200 100 29
-300 0
0 20 40 60 80 100 0 20 40 60 80 100
Frequency (Hz) Frequency (Hz)
• Now we can plot the spectra next to each other to

generate a surface:
Amplitude (Window) Time
Frequency
30
15
4/7/2014
• Formulation:
– Continuous STFT

STFT  x(t ) ,    X  ,     x(t ) w  t    e
 it
dt
of the phase result of the STFT

axis, τ, and frequency axis, ω, to suppress any jump discontinuity
Often phase unwrapping is employed along either or both the time

x(t ) Time-domain signal to be transformed
 Time (slow time; lower resolution than t)

 Frequency
w t  Window function, commonly a Hann window or

Gaussian window bell centered around zero
X  ,   A complex function representing the phase and

magnitude of the signal over time and frequency (this
is essentially the Fourier Transform of x (t ) w t   ) 
31
• Formulation:
Time discretized, frequency continuous; but if
– Discrete STFT FFT is used, they both will be discrete.

STFT  xn  m,    X  m,    xw
n 
n nm e  itn
xn Sequence of discretized time-domain signal to

be transformed
m Time index
 Frequency
wn Sequence of discretized window function
X  m,   STFT of the time-domain sequence
32
16
4/7/2014
• Inverse STFT
– The original signal can be recovered from the transform
by the Inverse STFT. The most widely accepted way of
inverting the STFT is by using the overlap-add (OLA)
method, which also allows for modifications to the STFT
complex spectrum.
• Calculating the Inverse STFT:

– First, it is required that the window function must be
scaled such that the area underneath the window
function is unity:
33
• Calculating the Inverse STFT (Continued):

– It follows that:
– Hence:
– The continuous Fourier transform is:
– Substituting from above:
34
17
4/7/2014

– Swap the integration order:
– Since the inverse Fourier transform is:
– Then the time-domain signal can be recovered:
35
– Or:
– Alternatively, one can obtain windowed grain or wavelet

of x(t) as:
36
18
4/7/2014
• Time and Frequency Resolution:

– The STFT has a fixed resolution
• The width of the windowing function relates to how the
signal is represented
• It determines whether there is good frequency resolution

(frequency components close together can be separated)
or good time resolution (the time at which frequencies
change).
• A wide window (wideband transform) gives better

frequency resolution but poor time resolution.
• A narrower window (narrowband transform) gives good

time resolution but poor frequency resolution.
37
• Time and Frequency Resolution (Continued):

– That is, we cannot have good resolutions in both time
and frequency.
Good time resolution, poor Good frequency resolution,

frequency resolution poor time resolution
38
19
4/7/2014

– To explain this limitation, note that in Fourier transform:
• To increase the frequency resolution of the window the
frequency spacing of the coefficients (sequence in
frequency domain) needs to be reduced.
– Decreasing Nyquist (max) frequency (and keeping N constant)

will cause the window size to increase — since there are now
fewer samples per unit time.
– The other alternative is to increase N, but this again causes

the window size to increase.
– So any attempt to increase the frequency resolution causes a

larger window size and therefore a reduction in time
resolution—and vice-versa.
39

– Best simultaneous resolution of both is reached with a
Gaussian window function. This STFT with some
modifications for multi-resolution becomes the Morlet
wavelet transform.
– Window should be narrow enough to make sure that the

portion of the signal falling within the window is
stationary.
– Very narrow windows do not offer good localization in

the frequency domain.
40
20
4/7/2014

– Windowing Function infinitely long: w(t )  1
• STFT turns into FT, providing excellent frequency
localization, but no time information.

X  ,     x(t )e
 it
dt  X  

– Windowing Function infinitely short: w(t )   (t )

• gives the time signal back, with a phase factor, providing
excellent time localization but no frequency information.

X  ,     x(t ) (t   )e
 it
dt  x( )e  i

41
• Heisenberg (Uncertainty) Principle:
1
t f 
4
Time resolution: How well Frequency resolution:

two spikes in time can be How well two spectral
separated from each other components can be
in the transform domain. separated from each other
in the transform domain.
𝜟𝒕 𝒂𝒏𝒅 𝜟𝒇 𝒄𝒂𝒏𝒏𝒐𝒕 𝒃𝒆 𝒎𝒂𝒅𝒆 𝒂𝒓𝒃𝒊𝒕𝒓𝒂𝒓𝒊𝒍𝒚 𝒔𝒎𝒂𝒍𝒍 !
42
21
4/7/2014
• Heisenberg (Uncertainty) Principle (Continued):
– One cannot know the exact time-frequency

representation of a signal.
– We cannot precisely know at what time instance a

frequency component is located.
– We can only know what interval of frequencies are

present in which time intervals.
43
• Example:
– Chirp Signal (linear frequency variation 10-200 Hz)
0.5
x(t)
-0.5
-1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Time (sec)
500 points spaced at 0.002 s, variable window sizes
44
22
4/7/2014
• Example (Continued):
– Chirp Signal – Ordinary Fourier Transform
Output
0.8
Amplitude 0.6
0.4
0.2
0
0 50 100 150 200 250
-100
-200
Phase
-300
-400
-500
0 50 100 150 200 250
Frequency (Hz)
45
– Chirp Signal – Window Size: 128 points
Window size = 128 Window type = Hann Window size = 128 Window type = Hann
250
20 200
Frequency (Hz)
15
150
10
5 100
0
250
50
200 1
150 0.8
100 0.6
0.4
50 0.2 0
0 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Frequency (Hz) Time (sec) Time (sec)
46
23
4/7/2014
250
20 200
Frequency (Hz)
15
150
10
5 100
0
250 50
200 1
150 0.8
100 0.6
0.4
50 0.2 0
0 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
47
250
20
200
15
Frequency (Hz)
150
10
5 100
0
250
50
200 1
150 0.8
100 0.6
0.4
50 0.2 0
0 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
48
24
4/7/2014
• Use of MATLAB:
– The MATLAB command to perform STFT is spectrogram:
[B, f, t] = spectrogram(x, window, noverlap, nfft, fs)
– Spectrogram inputs:
• x: input signal (signal to be transformed)
• window function (discrete window at the same input rate)
• window overlap (in terms of number of points)
• sampling points for FFT (number of points used in FFT)
• sampling frequency (rate of sampling for the input signal
and the window function)
– Spectrogram outputs: Transformed signal, frequencies

and times.
49
• Applications:
– Signal processing of any non-stationary signal (audio

signals, earthquake excitations, structural responses to
ambient vibrations, …)
– In structural dynamics, STFT can be used to:

• Determine the dominant modes of vibration (and their
shapes and frequencies) at any time interval
• Health monitoring and damage detection through the study
of dominant frequencies
– e.g. a reduction in frequency is generally indicative of
damages leading to softer structures
50
25
4/7/2014
• Applications:
– Damage Detection Example:
• 2-Story structure, natural frequencies 2.2 and 5.8 Hz
• Response simulated to El Centro record
• Linear Response
5
x 10
0.04 2
Displacement, m
0.02 1
Force, N
0 0
-0.02 -1
-0.04 -2
0 10 20 30 40 -0.02 -0.01 0 0.01 0.02
Time, s Displacement, m
4
x 10
2.5 1
51
Number of Iterations
2 0.8
Energy, N-m
1.5 0.6
Internal
1 Input 0.4
0.5 0.2
0 0
0 10 20 30 40 0 10 20 30 40
Time, s Time, s
• Applications:
– Damage Detection Example:
• 2-Story structure, natural frequencies 2.2 and 5.8 Hz
• Response simulated to El Centro record
• Nonlinear Response (by defining a yield displacement)
5
x 10
0.04 2
Displacement, m
0.02 1
Force, N
0 0
-0.02 -1
-0.04 -2
0 10 20 30 40 -0.02 -0.01 0 0.01 0.02
Time, s Displacement, m
4
x 10
2.5 30
52
Number of Iterations
2
Energy, N-m
20
1.5
Internal
1 Input
10
0.5
0 0
0 10 20 30 40 0 10 20 30 40
Time, s Time, s
26
4/7/2014
• Applications:
– Damage Detection Example: Linear Response
0.02
0.01
x(t)
-0.01
-0.02
0 5 10 15 20 25 30 35
Time (sec)
Window size = 2048 Window type = Chebyshev

10
Frequency (Hz)
0
0 5 10 15 20 25 30
Time (sec)
53
• Applications:
– Damage Detection Example: Nonlinear Response
0.02
0.01
x(t)
-0.01
-0.02
0 5 10 15 20 25 30 35
Time (sec)
Window size = 2048 Window type = Chebyshev

10
Frequency (Hz)
0
0 5 10 15 20 25 30
Time (sec)
54
27

Short-Time Fourier Transform

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Short-Time Fourier Transform

Uploaded by

Copyright:

Available Formats

4/7/2014

Advanced Structural Dynamics

• Scope and Goals

• Fourier Transform Review