结构动力学及其工程应用

Structural Dynamics and Its Applications

Fourier Series and Fourier Transform

阳佳桦，Ph.D.，副教授
广西大学土木建筑工程学院，工程力学研究中心
Email: [email protected]
Dr. Jia-Hua Yang, Ph.D., Associate Professor
College of Civil Engineering and Architecture
Scientific Research Center of Engineering Mechanics
Guangxi University
Contents
 Fourier Series (FS)

 Fourier Transform (FT)

 FS → FT → Discrete-time FT → Fast FT (FFT)

 Power Spectral Density (PSD)

2
Fourier Series
A periodic process
𝑥 𝑡+𝑇 =𝑥 𝑡

According to the Fourier theorem, it can be written as a Fourier series

(FS) Multiple of the
signal frequency
∞ ∞
2𝜋𝑘
𝑥 𝑡 = 𝑎0 + ෍ 𝑎𝑘 cos𝜔𝑘 𝑡 + ෍ 𝑏𝑘 sin𝜔𝑘 𝑡 , 𝜔𝑘 =
𝑇
𝑘=1 𝑘=1

The Fourier theorem says that by including an infinite number of them

with systematically increasing frequencies it is possible to represent ANY
periodic process.

3
𝑎𝑘 and 𝑏𝑘 are the (real) Fourier series coefficients. 𝑎0 accounts for the
non-oscillatory level of the process.
𝑇
1 2
𝑎0 = න 𝑥 𝑡 𝑑𝑡
𝑇 −𝑇
2
𝑇
2 2
𝑎𝑘 = න 𝑥 𝑡 cos𝜔𝑘 𝑡 𝑑𝑡
𝑇 −𝑇
2
𝑇
2 2
𝑏𝑘 = න 𝑥 𝑡 sin𝜔𝑘 𝑡 𝑑𝑡, 𝑘 = 1, 2, …
𝑇 −𝑇
2
𝑇 𝑇
Note that the integration domain need not necessarily be − , . It can
2 2
be any continuous interval of the signal period 𝑇.

4
Proof.
𝑇 𝑇
To derive 𝑎0 , integrating both sides of the Fourier series in − ,
2 2
gives
𝑇 𝑇 ∞ 𝑇 ∞ 𝑇
2 2 2 2
න 𝑥 𝑡 𝑑𝑡 = 𝑎0 න 𝑑𝑡 + ෍ 𝑎𝑘 න cos𝜔𝑘 𝑡𝑑𝑡 + ෍ 𝑏𝑘 න sin𝜔𝑘 𝑡𝑑𝑡
𝑇 𝑇 𝑇 𝑇
− − 𝑘=1 − 𝑘=1 −
2 2 2 2

=0 =0
𝑇
1 2
⟹ 𝑎0 = න 𝑥 𝑡 𝑑𝑡
𝑇 −𝑇
2
To derive 𝑎𝑘 and 𝑏𝑘 , note the following equations

5
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
6
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Multiplying both sides of the Fourier series with and
integrating gives

Only one term

is left, i.e., 𝑗 = 𝑘

𝑎𝑘 is then obtained.

7
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Multiplying both sides of the Fourier series with sin 𝜔𝑘 𝑡 and integrating
gives
𝑇
2
න 𝑥 𝑡 sin 𝜔𝑘 𝑡 𝑑𝑡
𝑇
−2
𝑇 ∞ 𝑇
2 2
= 𝑎0 න sin 𝜔𝑘 𝑡 𝑑𝑡 + ෍ 𝑎𝑘 න cos𝜔𝑗 𝑡 sin 𝜔𝑘 𝑡 𝑑𝑡
𝑇 𝑇
−2 𝑗=1 −2

=0 =0
∞ 𝑇
2
+ ෍ 𝑏𝑘 න sin𝜔𝑗 𝑡 sin 𝜔𝑘 𝑡 𝑑𝑡
𝑇
𝑗=1 −2

𝑇
= if 𝑗=𝑘
2

The above equation gives 𝑏𝑘 .

Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Fourier Series:
Complex Exponential Form
 The complex exponential form will connect a Fourier series to the
Fourier Transform.

Euler’s formula
𝑒 i𝜃 = cos𝜃 + i sin𝜃,
𝑒 −i𝜃 = cos𝜃 − i sin𝜃, i2 = −1

 cos𝜃 and sin𝜃 can be written in terms of 𝑒 ±i𝜃 using Euler’s formula

𝑒 i𝜃 − 𝑒 −i𝜃
sin𝜃 =
2i
𝑒 i𝜃 + 𝑒 −i𝜃
cos𝜃 =
2
9
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Substituting these two cosine and sine identities into the original Fourier
series gives
∞ ∞
𝑎𝑘 i𝜔 𝑡 −i𝜔 𝑡
i𝑏𝑘 i𝜔 𝑡
𝑥 𝑡 = 𝑎0 + ෍ 𝑒 𝑘 +𝑒 𝑘 +෍− 𝑒 𝑘 − 𝑒 −i𝜔𝑘 𝑡
2 2
𝑘=1 𝑘=1
∞ ∞
1 i𝜔𝑘 𝑡
1
=𝑎
ด0 + ෍ 𝑎𝑘 − i𝑏𝑘 𝑒 + ෍ 𝑎𝑘 + i𝑏𝑘 𝑒 −i𝜔𝑘 𝑡
2 2
𝑘=1 𝑘=1
𝑐0
𝑐𝑘 𝑐−𝑘

If we define the complex Fourier series coefficients,

1
𝑐0 = 𝑎0 , 𝑐𝑘 = 𝑎𝑘 − i𝑏𝑘
2
Then
We will first prove this
1
𝑐−𝑘 = 𝑎𝑘 + i𝑏𝑘
2
10
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Substituting −𝑘 into the formulation of 𝑎𝑘 and 𝑏𝑘 gives, where 𝑘 is a
positive integer
𝑇 𝑇
2 2 2 2𝜋 −𝑘 2
𝑎−𝑘 = න 𝑥 𝑡 cos𝜔−𝑘 𝑡 𝑑𝑡 = න 𝑥 𝑡 cos 𝑡 𝑑𝑡
𝑇 −𝑇 𝑇 −𝑇 𝑇
2 2
𝑇 𝑇
2 2 2𝜋𝑘 2 2𝜋𝑘 2
= න 𝑥 𝑡 cos − 𝑡 𝑑𝑡 = න 𝑥 𝑡 cos 𝑡
𝑇 −𝑇 𝑇 𝑇 −𝑇 𝑇
2 2
= 𝑎𝑘
𝑇 𝑇
2 2 2 2𝜋𝑘 2
𝑏−𝑘 = න 𝑥 𝑡 sin𝜔−𝑘 𝑡 𝑑𝑡 = න 𝑥 𝑡 sin − 𝑡 𝑑𝑡
𝑇 −𝑇 𝑇 − 𝑇 𝑇
2 2
𝑇
2 2𝜋𝑘
2
= − න 𝑥 𝑡 sin 𝑡 𝑑𝑡
𝑇 − 𝑇 𝑇
2
= −𝑏𝑘
11
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
With these two relationships of 𝑎−𝑘 and 𝑏−𝑘 , we have
1
𝑐−𝑘 = 𝑎−𝑘 − i𝑏−𝑘
2
1
= 𝑎𝑘 − i(−𝑏𝑘 )
2
1
= 𝑎𝑘 + i𝑏𝑘
2
The Fourier series can then be written in a unified way using the
complex Fourier series coefficients as
∞ It gives a glimpse of Fourier
𝑥 𝑡 = ෍ 𝑐𝑘 𝑒 i𝜔𝑘 𝑡 Transform. Next, let’s look
𝑘=−∞ at a non-periodic process.
𝑐𝑘 is obtained by substituting 𝑎𝑘 and 𝑏𝑘 :
𝑇
1 −i𝜔
2
𝑡
2𝜋𝑘
𝑐𝑘 = න 𝑥 𝑡 𝑒 𝑘 𝑑𝑡 , 𝜔𝑘 = , 𝑘 = 0, ±1, ±2, …
𝑇 −𝑇 𝑇
2
12
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 Parseval Equality

The ‘energy’ of a process when viewed in the time or frequency domain

can be equated via the ‘Parseval equality’, which appears in different
forms depending on the context. For a periodic process with finite
energy

The leftmost expression may be interpreted as the average energy per

unit time. The Parseval equality says that this energy can be viewed as
the sum of contributions from harmonics at different frequencies.

13
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Proof of Parseval equality using complex Fourier series:
Writing

The order of infinite sum and

integration can be swapped,
when the process has finite
energy.

14
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Fourier Transform (FT): From FS to FT
If a process is not periodic, it CANNOT be written as a Fourier series,
but it can be represented as a sum of harmonics with infinitesimal
contributions at a continuum of frequencies. The Fourier Transform
(FT) pair:
1 ∞ i𝜔𝑡
The Inverse Fourier Transform
𝑥 𝑡 = න 𝑋 𝜔 𝑒 𝑑𝜔 of 𝑋 𝜔 .
2𝜋 −∞
∞
𝑋 𝜔 = ‫׬‬−∞ 𝑥 𝑡 𝑒 −i𝜔𝑡 𝑑𝑡 The Fourier Transform of 𝑥 𝑡 .

15
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 The idea is to consider one segment of 𝑥 𝑡 and treat it as a periodic
𝑇
process 𝑥𝑝 𝑡 with period 𝑇. That is, within the range − <𝑡<
2
𝑇
,𝑥 𝑡 = 𝑥 𝑡 , and 𝑥𝑝 𝑡 repeats this segment of 𝑥 𝑡 periodically
2 𝑝
𝑇 𝑇
outside − < 𝑡 < . See the following illustration.
2 2

𝑥 𝑡

𝑇 𝑇
−
2 2

𝑥𝑝 𝑡
16
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Fourier Transform: From FS to FT
 The FS coefficients of 𝑥𝑝 𝑡 are (because 𝑥𝑝 𝑡 is periodic):
𝑇 𝑇
1 2 1 2
𝑐𝑘 = න 𝑥𝑝 𝑡 𝑒 𝑘 𝑑𝑡 = න 𝑥 𝑡 𝑒 −i𝜔𝑘𝑡 𝑑𝑡
−i𝜔 𝑡
𝑇 −𝑇 𝑇 −𝑇
2 2

𝑇 𝑇
Because 𝑥𝑝 𝑡 = 𝑥 𝑡 within − and
2 2

 We want to show that FS is not suitable to study a non-periodic

process, and how to go from Fourier series (FS) to Fourier
transform (FT).

17
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Fourier Transform: From FS to FT
Using Cauchy-Schwartz inequality,
2
𝑁 𝑁 𝑁
2 2
෍ 𝑥𝑖 𝑦𝑖 ≤ ෍ 𝑥𝑖 ෍ 𝑦𝑖
𝑖=1 𝑖=1 𝑖=1

we have Treating integration as sum

𝑇 2

2
1 2
−i𝜔𝑘 𝑡 𝑒 −i𝜔𝑘 𝑡 2 = 1
𝑐𝑘 = 2 න 𝑥 𝑡 𝑒 𝑑𝑡
𝑇 −𝑇
2
𝑇 𝑇
∞
1 2 2 2 2 1
≤ 2 න 𝑥 𝑡 𝑑𝑡 × න 𝑒 −i𝜔𝑘 𝑡 𝑑𝑡 ≤ න 𝑥 2 𝑡 𝑑𝑡
𝑇 −𝑇 −
𝑇 𝑇 −∞
2 2

∞ 𝑇
18 ≤‫׬‬−∞ 𝑥 2 𝑡 𝑑𝑡
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Fourier Transform: From FS to FT
 This inequality indicates that when 𝑇 → ∞, 𝑐𝑘 → 0 no matter
what process 𝑥 𝑡 is (because signal energy is finite), so FS
coefficients are not suitable for non-periodic processes.
1
 To obtain FT, first take away the problematic factor from FS
𝑇
and define the following term:
𝑇
2
𝑋𝑝 𝜔𝑘 = න 𝑥 𝑡 𝑒 −i𝜔𝑘𝑡 𝑑𝑡
𝑇
−2

Note that 𝜔𝑘 is DISCRETE, and we’ll show how

the continuous frequency 𝜔 comes into the picture.
19
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Fourier Transform: From FS to FT
 We have the following relation for FS 𝑐𝑘 :
𝑋𝑝 𝜔𝑘
𝑐𝑘 =
𝑇 Substitute

 𝑥𝑝 𝑡 can then be written as a FS representation:

∞ ∞
1
𝑥𝑝 𝑡 = ෍ 𝑐𝑘 𝑒 i𝜔𝑘𝑡 = ෍ 𝑋𝑝 𝜔𝑘 𝑒 i𝜔𝑘𝑡
𝑇
𝑘=−∞ 𝑘=−∞
2𝜋𝑘
 Note that by construction 𝜔𝑘 = , the frequency interval is
𝑇
2𝜋 1 ∆𝜔
∆𝜔 = 𝜔𝑘+1 − 𝜔𝑘 = ⇒ =
𝑇 𝑇 2𝜋
20
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Fourier Transform: From FS to FT
∞
i𝜔𝑘 𝑡
∆𝜔
𝑥𝑝 𝑡 = ෍ 𝑋𝑝 𝜔𝑘 𝑒
2𝜋
𝑘=−∞

 Recall the illustration figure and we know that when 𝑇 → ∞,

𝑥𝑝 𝑡 → 𝑥 𝑡 , so we have
∞
∆𝜔
𝑥 𝑡 = lim 𝑥𝑝 𝑡 = lim ෍ 𝑋𝑝 𝜔𝑘 𝑒 i𝜔𝑘𝑡
𝑇→∞ 𝑇→∞ 2𝜋
𝑘=−∞

2𝜋
 Note that when 𝑇 → ∞, ∆𝜔 = → 𝑑𝜔 (infinitly small), and the
𝑇
discrete 𝜔𝑘 tends to the continuous 𝜔.
21
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Fourier Transform: From FS to FT
 The infinite sum becomes an integral
1 ∞ The Inverse Fourier
𝑥 𝑡 = i𝜔𝑡
න 𝑋(𝜔)𝑒 𝑑𝜔 Transform of 𝑋 𝜔 .
2𝜋 −∞
This factor looks strange, but it ensures that the unit is right. 𝑋(𝜔) is in
V/Hz (See the unit of the FT in the following). On the left 𝑥 𝑡 is in V,
where so we need Hz on the right to cancel with V/Hz. But 𝑑𝜔 is in rad/s, so
we need 1Τ2𝜋 to make 𝑑𝜔Τ2𝜋 = Hz
𝑇
2
𝑋 𝜔 = lim 𝑋𝑝 𝜔𝑘 = lim න 𝑥 𝑡 𝑒 −i𝜔𝑘𝑡 𝑑𝑡
𝑇→∞ 𝑇→∞ −𝑇
2 If 𝑥 𝑡 in V, and the unit of 𝑑𝑡 is s,
∞ then the unit of 𝑋 𝜔 is V ∙ s = V/Hz.
= ‫׬‬−∞ 𝑥 𝑡 𝑒 −i𝜔𝑡 𝑑𝑡

The Fourier Transform of 𝑥 𝑡 .

22
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 Parseval equality for a non-periodic process

It also says that signal energy in time and frequency domain is equal.

Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Fourier Transform:
Discrete-Time FT and FFT
 Data are recorded in discrete time in practice, so we consider the
discrete-time FT:
𝑁𝑗 −1
∞
𝑋 𝜔 = න 𝑥 𝑡 𝑒 −i𝜔𝑡 𝑑𝑡 ≈ 𝑋෠ 𝜔 = ෍ 𝑥𝑗 𝑒 −i𝜔𝑗∆𝑡 ∆𝑡
−∞ 𝑗=0

2𝜋𝑘
 We only calculate it at discrete frequencies 𝜔𝑘 = , because
𝑁𝑗 ∆𝑡
we have an efficient algorithm to do it:
2𝜋𝑘
𝑁𝑗 −1 −i𝑁 ∆𝑡𝑗∆𝑡 𝑁𝑗 −1
෠
𝑋 𝜔𝑘 = σ𝑗=0 𝑥𝑗 𝑒 𝑗 ∆𝑡 = σ𝑗=0 𝑥𝑗 𝑒 −i2𝜋𝑗𝑘Τ𝑁𝑗 ∆𝑡

24
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Fourier Transform:
Discrete-Time FT and FFT
 Lose ∆𝑡 and we define the fast Fourier Transform (FFT) used in
MATLAB:
𝑁𝑗 −1
𝑦𝑘 = σ𝑗=0 𝑥𝑗 𝑒 −i2𝜋𝑗𝑘Τ𝑁𝑗 , 𝑘 = 0, 1, … , 𝑁𝑗 − 1

By this form, it is not “FAST” but only a normal summation.

It’s the algorithm that makes it “FAST” Fourier Transform.

We’ll use another definition of FFT that is consistent

with non-decreasing stationary stochastic processes.
The only difference is that a scaling factor will be used.
25
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 The ‘inverse FFT’ of

We can show that the inverse FFT recovers the original discrete time
series that produces the FFT, i.e., 𝑥𝑗 = 𝑧𝑗 , 𝑗 = 1, ⋯ , 𝑁 − 1. Substituting
into the above inverse FFT:

This sum of exponential terms can be proved.

Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 𝑘
For the first case when 𝑘 is an integer multiple of 𝑁, 𝑝 = , 𝑒 i2𝜋𝑗𝑝 = 1,
𝑁
so the sum is 𝑁.
For the second case, using the geometric series summation formula, with
𝑎 = 𝑒 i2𝜋𝑘/𝑁 ,
𝑁−1
𝑁
1 − 𝑎
෍ 𝑎𝑗 =
1−𝑎
𝑗=0

Using the above result we see that 𝑒 i2𝜋(𝑗−𝑟)𝑘/𝑁 in the inverse FFT is
non-zero only when 𝑗 = 𝑟 because 𝑘/𝑁 is not an integer.
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Conjugate Mirror Property of FFT
 FFT has the conjugate mirror property:

𝑦−𝑘 = 𝑦ത𝑘
𝑦− 𝑁−𝑘 = 𝑦ത𝑁−𝑘
Proof:

28
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
About half of the FFT sequence carries redundant information, in the
sense that it can be produced as the complex conjugate of the other half.

𝑁𝑦𝑞 = the integer part of 𝑁/2, is the index at or just below the
‘Nyquist frequency’

29
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 Parseval equality for FFT

30
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Distortions of Fourier series (FS): Aliasing and
Leakage
 Aliasing
1
The approximation is valid only up to the ‘Nyquist frequency’ ( Hz).
2Δ𝑡

Contaminated with harmonics in the original process beyond the

Nyquist frequency.
High frequency variations are mistaken as low frequency ones.
 Leakage

Contamination from harmonics that do not have an integer multiple of

cycles within the measured time span, in a neighborhood of
the subject frequency (and aliased counterparts).

31
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 Aliasing is not repairable after discrete-time sampling. Should filter out
high-frequency contents beyond the Nyquist frequency. Leakage can be
suppressed by increasing data duration.
 Nyquist frequency
1 1
𝑓𝑁𝑦𝑞 = Hz (half of the sampling frequency)
2 Δ𝑡
𝜋
𝜔𝑁𝑦𝑞 = 2𝜋𝑓𝑁𝑦𝑞 = rad/s
Δ𝑡

According to the conjugate mirror property , only the first half

of the FS carries the information. It can be shown that (we skip the proof)

This equation indicates that the discrete-time approximation FS contains

not only 𝑐𝑘 , but also other terms 𝑐𝑘+𝑁 , 𝑐𝑘+2𝑁 , ⋯

Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
To transform the contributions to the ones at positive frequencies, use the
conjugate mirror property to separate the sum

is contaminated with contributions from frequencies (in Hz) 𝑓𝑠 ±

1
𝑓𝑘 , 2𝑓𝑠 ± 𝑓𝑘 , 3𝑓𝑠 ± 𝑓𝑘 , ⋯, where 𝑓𝑠 = is the sampling frequency and
Δ𝑡
𝑘
𝑓𝑘 = is the subject FFT frequency.
𝑁Δ𝑇

Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 Aliasing Example (Au 2018).

Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 2𝑓𝑠 − 𝑓𝑘 = 2 × 10 − 6 = 4 Hz

Mis-interpreted as low harmonics

𝑓𝑠 =10 Hz 2𝑓𝑠 − 𝑓𝑘 = 2 × 10 − 9 = 1 Hz

Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 Leakage

Leakage occurs in the FFT approximation of FS when the data duration

𝑁Δ𝑡 is not an integer multiple of the period T. In this case, a given FS
frequency need not be matched by a FFT frequency
. When this occurs, the FS harmonic will ‘leak out’ to other
FFT frequencies.
Leakage and aliasing can be explained by a formula expressing as the
convolution of the target FS and a kernel function:
Indicates that 𝑐𝑘Ƹ contains contributions from all FS
frequencies. The contributions are not directly from 𝑐𝑟 . It
depends how far 𝜔 ഥ𝑟 is from 𝜔𝑘 through the kernel function.

Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Schematic plot of 𝐷𝑁 𝑢 . It indicates that the 𝑐𝑘Ƹ has the constributions
from the following frequencies:

Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 Leakage example.

Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 Leakage
2
𝑓𝑘 = 1.2 = 1.67 Hz

10
𝑓𝑘 = 1.2 ≈ 8.33 Hz
Aliased leakage 𝑓𝑠 − 𝑓𝑘 = 10 − 8.33 = 1.67 Hz

7
𝑓𝑘 = 1.2 ≈ 5.83 Hz
leakage
5 Aliased leakage 𝑓𝑠 − 𝑓𝑘 = 10 − 5.8
𝑓𝑘 = = 4.17 Hz
1.2

Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Distortions of Fourier transform (FT): Aliasing
and Leakage
 The discrete-time Fourier transform (DTFT) approximation is
related to the target FT by a convolution integral, which can be
used to explain aliasing and leakage of FT (we again skip the proof)

 To see aliasing, has local maxima 1 at 𝑢 = 0, ±1, ±2, ⋯, i.e.,

the kernel as a function of 𝜔′ has local maxima at
. According to the conjugate mirror property of FT,
40
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
so receives significant contributions from the frequencies
. It means FTs at high frequencies will have
significant contributions at their mirror conjugates at low frequencies.
This aliasing.
 To see leakage, note that at frequency 𝜔 receives contribution
from the FT of the original process at frequency 𝜔′. The
contribution is attenuated by , which depends on how far
𝜔′ is from 𝜔. 𝜔′ need not be the subject frequency 𝜔 of FT, or the
aliased counterparts, so the contributions of are non-zero.
This is leakage. The leakage effect diminishes as the data duration
increases, because is negligible for . In
the long data duration case 𝑁Δ𝑡 is large, so the right-hand side is
small.

Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
FFT in MATLAB
Let be the FFT of . In MATLAB, we can use the built-
in function to calculate the FFT: y=fft(x). In MATLAB, the index
starts from 1, so the array y in MABLAB and the FFT in theory have
the following relationship

42
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Up to the Nyquist frequency
where Nq = floor(N/2)(integer part of N/2).

Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
FFT Algorithm
 The key idea is that for 𝑁 being some power of 2, i.e., ,a
FFT sequence of length 𝑁 can be obtained from two FFT sequences
of length 𝑁/2; and similarly each FFT sequence of length 𝑁/2 can
be obtained from another two FFT sequences of length 𝑁/4; and so
on.

44
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Writing the previous equation here again for comparison

46
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
47
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.