结构动力学及其工程应用

Structural Dynamics and Its Applications

Spectral Analysis of Stationary

Stochastic Process

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: javayang@gxu.edu.cn
Dr. Jia-Hua Yang, Ph.D., Associate Professor
College of Civil Engineering and Architecture
Scientific Research Center of Engineering Mechanics
Guangxi University
Contents
 Correlation Function

 Power Spectral Density (PSD)

2
Correlation Function
A real-valued stochastic vector process with respect to (w.r.t.) time

Assume zero mean, a commonly seen case in Engineering

Correlation function is used to describe the stochastic process’ basic

statistical characterization

A matrix function of two arguments 𝑟, 𝑡.

The (𝑖, 𝑗)-entry of 𝐑(𝑠, 𝑡) is
𝐸[𝑥𝑖 𝑠 𝑥𝑗 (𝑡)]
3
The covariance matrix of 𝐱 𝑡

A strict-sense process requires that the joint probability distribution of

𝐱 𝑡 + 𝜏 and 𝐱 𝑡 is invariant to a time shift 𝑡, which is difficult to
model in practice. A wide-sense stationary process only requires that
the correlation function is invariant to a time shift 𝑡. The correlation
function depends only on the time lag 𝜏:

Generally, 𝑡 + 𝜏 and 𝑡 in the expectation cannot be swapped, i.e.,

. The correlation function need not be a
symmetric matrix.

4
For a stationary process it has the ‘transpose mirror property’ that

Proof

The second equality follows because the expectation is invariant to t

(stationary).
For two scalar processes, their cross-correlation function is defined as

If they are joint stationary, their cross correlation function only depends
on time lag

5
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
By definition, the (𝑖,𝑗)-entry (𝑖 ≠ 𝑗) of the correlation function of a
vector process is the cross correlation function between its 𝑖-th and 𝑗-th
component, i.e.,
𝐑 𝑖,𝑗 𝜏 = 𝐸[𝑥𝑖 𝑡 + 𝜏 𝑥𝑗 𝑡 ]

6
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Power Spectral Density
 In practice, we usually need to deal with stochastic vector
processes, i.e., 𝐱(𝑡) ∈ 𝑅𝑁𝑑 that denotes dynamic responses
measured at 𝑁𝑑 degrees of freedom.

 We cannot tell exactly what the value of 𝐱(𝑡) is at a particular

time instant, because it is stochastic and “unknown”.

 But we can model 𝐱(𝑡) as a random variable and characterize it

using its joint probability density function.

 For a stationary stochastic process 𝐱(𝑡), we use power spectral

density (PSD) matrices (or the counterpart correlation functions)
to characterize its statistical properties.
7
 A PSD matrix is used to characterize a stationary stochastic process in
an energy sense in the frequency domain:
𝐒 𝜔 = lim 𝐸[𝐗 𝑇 (𝜔)𝐗 ∗𝑇 (𝜔)]
𝑇→∞

The word “density” is used considering that this quantity

gives the energy at one frequency instant, i.e., power density,
and the sum (integration) over the frequency range gives the
energy of the signal.

Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Power Spectral Density
 𝐗 𝑇 (𝜔) is actually a scaled FT 𝐗 𝜔 that is defined for non-
periodic processes. 𝐗 𝜔 is not suitable for a stationary process
𝐱 𝑡 because when 𝑇 → ∞, 𝐱 𝑡 is non-decaying and 𝐗 𝜔 =
∞
ධ 𝐱 𝑡 𝑒 −i𝜔𝑡 𝑑𝑡 → ∞.
−∞

It’s 𝑇 but not 𝑇, because 𝑇 is related

 𝐗 𝜔 grows with 𝑇.
to the signal energy, e.g., PSD.
 A scaling factor is thus added to 𝐗 𝜔 , so that the FT for a
stationary process is finite even when 𝑇 → ∞.
∞
1
𝐗 𝑇 (𝜔) = න 𝐱 𝑡 𝑒 −i𝜔𝑡 𝑑𝑡 𝐗 𝑇 (𝜔) is the FT for a stationary
𝑇 −∞ process
9
 The Properties of PSD

It is Hermitian:

The transpose mirror property,

Positive semi-definite

Modulus of a complex number must be ≥ 0

10
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Operationally, it is more convenient to take the limit operator inside the
expectation, and interpret PSD directly as the product of the Fourier
transform of dynamic responses:
Compared to the original definition,
𝐒 𝜔 = lim 𝐸[𝐗 𝑇 (𝜔)𝐗 ∗𝑇 (𝜔)], it can be
𝑇→∞
seen that the limit operator is taken inside
the expectation. This is usually justified for
“well-behaved” ambient vibration data.

Using this interpretation of PSD brings convenience for theoretical

derivation, because now for dynamic analysis in the frequency domain,
we just take the FT symbolically for dynamic responses, and then
calculate the product the FT, 𝐗(𝜔), and its conjugate transpose, 𝐗 ∗ (𝜔),
to get PSD.

11
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 Cross PSD and Coherence

By definition, the (𝑖,𝑗)-entry (𝑖 ≠ 𝑗) of the PSD matrix of a vector

process is the cross PSD between its 𝑖-th and 𝑗-th component
The ‘coherence’ is a dimensionless quantity that is conveniently used in
analysis (in a normalized way)

12
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Summary on Fourier Series, Fourier Transform
and PSD
 Consider that 𝐱(𝑡) is a zero-mean stationary stochastic process.

 It does not have a FS because it is not periodic.

 Its FT does not exist either because it extends over (0, ∞).

The magnitude of grows with 𝑇 (not 𝑇 because 𝑇 is

related to energy). The FT involves this integral, so it tends to infinity.
 A proper scaling 1/ 𝑇 is needed for a stationary process so that it has
converging and bounded statistics.

13
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
14
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Sample Correlation Function
 The definition requires multiple samples of 𝐱(𝑡) for a given 𝑡. This
could imply a very tedious data collection process; imagine
measuring the temperature at the same place and same time everyday.
For a stationary process, it is possible to estimate using only a single
rather than multiple time histories.

For a given 𝜏, the term is averaged over different 𝑡’s. This

works when the process is stationary and ‘ergodic’.

15
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 ෡ 𝑇 (𝜏)
The averaging factor is 𝑇 rather than 𝑇 − |𝜏|. This is so that 𝐑
satisfies the Wiener-Khinchin formula (will talk about it later).

 Transpose Mirror Property

Verify. For 𝜏 > 0, changing integration variable from 𝑡 to 𝑠 = 𝑡 − 𝜏

16
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 The sample correlation function can be written in a compact manner by
considering both positive and negative time lags.

Note that the response is a windowed process, i.e., it only has values
within a certain time period 0, 𝑇 , so we also know:

Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 Sample Power Spectral Density. Using only one sample process

The sample PSD is also Hermitian and it has the transpose mirror
property:

18
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Wiener-Khinchin Formula
 One important result in the  The sample counterpart of
theory of stationary process is Wiener-Khinchin formula that
that the correlation function uses only one set of response:
and PSD are Fourier
Transform pairs:

19
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 Proof.
∞ ∞ ∞
1
෡𝑇 𝜏
න𝐑 𝑒 −𝐢𝜔𝜏 𝑑𝜏 = න න 𝐱 𝑇 𝑡 + 𝜏 𝐱 𝑇 𝑡 𝑇 𝑒 −𝐢𝜔𝜏 𝑑𝑡𝑑𝜏
𝑇
−∞ −∞ −∞
∞ ∞
Let 𝑠 = 𝑡 + 𝜏 and change the
1
= න න 𝐱 𝑇 𝑠 𝐱 𝑇 𝑡 𝑇 𝑒 −𝐢𝜔 𝑠−𝑡 𝑑𝑡𝑑𝑠 change integration variable
𝑇 from (𝜏, 𝑡) to (𝑠, 𝑡)
−∞ −∞
∞ ∞
1
= න න 𝐱 𝑇 𝑠 𝑒 −𝐢𝜔𝑠 𝐱 𝑇 𝑡 𝑇 𝑒 𝐢𝜔𝑡 𝑑𝑡𝑑𝑠 Operator * denotes
𝑇 conjugate transpose
−∞ −∞
∞ ∞ ∗
1 1
= න 𝐱𝑇 𝑠 𝑒 −𝐢𝜔𝑠 𝑑𝑠 න 𝐱 𝑇 𝑡 𝑒 −𝐢𝜔𝑡 𝑑𝑡
𝑇 𝑇
−∞ −∞

𝐗 𝑇 (𝜔) 𝐗𝑇 𝜔
= 𝐒෠ 𝑇 (𝜔) We thus prove that the correlation function and PSD of a sample
process are a Fourier Transform pairs. Taking limit 𝑇 → ∞ gives
another FT pairs for the theoretical correlation function and PSD.
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Parseval Equality For Correlation Function
and PSD
For a stationary process, the Parseval equality explains the variance as a
sum of PSD contributions from different frequencies. It results directly
from the Wiener-Khinchin formula at zero time lag. Setting 𝜏 = 0

For a scalar process, the

Parseval equality implies that
the area under the PSD in
the frequency domain (w.r.t.
The formula for the sample response 𝜔/2𝜋, in Hz) gives the variance
of the response.

21
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 White noise
A white noise has a constant PSD

Using the Wiener-Khinchin formula the correlation function

where is the Dirac Delta function.

Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 Example. First order process driven by white noise

We calculate the correlation function first. For , write the above

equation at 𝑡 + 𝜏 and multiply by 𝑥 𝑡 . Taking expectation

This is 0 because 𝑥(𝑡) depends

When taking the derivative for the expectation, the only on the past, which is
derivative is taken for the quantity with 𝜏, because independent of the future 𝑤(𝑡 + 𝜏)
the variable of the correlation function is 𝜏

Solving gives
We need to determine this.

23
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Multiplying the governing equation by 𝑥 𝑡 and taking the expectation
gives

The derivative of the variance is

zero, or the change of variance
with time is zero, because the
process is stationary.
Therefore we have

24
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
To evaluate , express the response using Duhamel’s integral

Substitute this in the

previous equation
The Wiener-Khinchin formula for the Gaussian
process is used, and there is a ½ factor because
the integral only covers the right half of the
Delta function, i.e., positive time only.

Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Next, we calculate response power spectral density (PSD) of the response
using the definition:

Taking Fourier transform for the governing equation

Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 Verify the Weiner-Khichin formula

Separate the integral into two, one positive time lag and one for negative
time lag.

This is the PSD we

obtained from definition.

Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Discrete-Time Sample Process
The only difference is that the sample process is now in discrete-time:

Discrete-time sample correlation function

Note that averaging
factor is 𝑁 instead of 𝑁 − 𝑟 so that Positive time-lag index. The
the correlation function satisfies the final 𝑟 samples are left,
Wiener-Khinchin formula. because the time lag is 𝑟.
But when the data length is long
enough, using either factor is ok.

Negative time-lag index. Start

from the 𝑟-th samples are left,
because the time lag is 𝑟.
29
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 Transpose mirror property

Verify:

Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
 Structure of sample correlation function
Consider a discrete time process with 4 samples

Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Power Spectral Density Using Discrete-Time
Processes
෠ 𝑇 𝜔 using data 𝐱 𝑡
 In practice, we approximate 𝐒 𝜔 with 𝐒
with a finite duration 𝑇:
𝐒෠ 𝑇 𝜔 = 𝐗 𝑇 (𝜔)𝐗 ∗𝑇 (𝜔)
Note the integration limit.
1 𝑇
where 𝐗 𝑇 (𝜔) = ‫𝐱 ׬‬ 𝑡 𝑒 −i𝜔𝑡 𝑑𝑡
𝑇 0

 To get the sample PSD matrix using discrete-time processes, we

first re-define the FFT (with a scaling factor) as an approximation
of 𝐗 𝑇 (𝜔) 𝑇
𝑁𝑗 −1 2𝜋𝑘 𝑁𝑗 −1
1 −i 𝑗∆𝑡 ∆𝑡
෡𝑘 =
𝐗 ෍ 𝐱𝑗 𝑒 𝑁𝑗 ∆𝑡
∆𝑡 = ෍ 𝐱𝑗 𝑒 −i2𝜋𝑗𝑘/𝑁𝑗
𝑁𝑗 ∆𝑡 𝑁𝑗
33 𝑗=0 𝑗=0
 Wiener-Khichin formula for discrete-time processes

Note that a factor 2 is added at both numerator

denominator, because the correlation function is a
2𝑁 sequence (see the summation). The definition of
FFT requires that the denominator equals to the
number of time series sequence, in this case 2𝑁.

After adding factor 2, “𝑘” has become “2𝑘”, so

despite the appearance of this formula, 𝐒෠𝑘 is not the
FFT of 𝐑 ෡ 𝑟 𝑟−1 . Rather, 𝐒෠𝑘 is equal to the 2k-
Undefined, short one element, because 𝑟=− 𝑁−1
𝑁 − 1 − − 𝑁 − 1 + 1 = 2𝑁 − 1 th entry of the FFT of the 2N-sequence formed by
So add a for in first odd element? ෡ 𝑟 𝑁−1
padding a zero in the beginning of 𝐑 𝑟=− 𝑁−1

Evaluate at the FFT frequencies

𝑁−1 𝑁−1
2𝜋𝐢𝑟𝑘 2𝜋𝐢𝑟2𝑘 2𝑘-th entry
− −
𝐒෠𝑘 = Δ𝑡 ෍ ෡ 𝑟𝑒 𝑁
𝐑 = Δ𝑡 ෍ ෡ 𝑟 𝑒 2𝑁
𝐑
𝑟=− 𝑁−1 2𝑁 sequence 𝑟=− 𝑁−1
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Au SK, 2018. Operational Modal Analysis. Modeling, Bayesian Inference, Uncertainty Laws. Springer.
Power Spectral Density Using Discrete-Time
Processes
2𝜋𝑘
 The discrete PSD matrix at the discrete frequency ω𝑘 = can
𝑁𝑗 ∆𝑡
then be written:

𝐒෠𝑘 = 𝐗 𝑘 𝐗 ∗𝑘

 Auto-PSDs of each measurement channel can be quickly

calculated using MATLAB.

1. Calculate unscaled FFTs using MATLAB function “fft()”.

a. Put 𝐱𝑗 , which is at time instant 𝑗∆𝑡 and measured at 𝑁𝑑 DOFs, in

the 𝑗-th row of the data matrix 𝐃.
36
Power Spectral Density
b. Calculate the FFT matrix 𝐁 = fft(𝐃) . The 𝑘 − th row of 𝐁 is the
2𝜋𝑘
unscaled FFT at frequency instant :
𝑁𝑗 ∆𝑡
𝑁𝑗 −1

𝐘෠𝑘 = ෍ 𝐱𝑗 𝑒 −i2𝜋𝑗𝑘Τ𝑁𝑗
𝑗=0

Take the FFTs up to the Nyquist frequency, i.e., the frequency

𝑁𝑗
index 𝑁𝑞 = int + 1, so 𝑘 = 0, 1, … , 𝑁𝑞 . This is because the
2
FFTs after the Nyquist frequency is just the conjugate mirror image
of those before the Nyquist frequency, and no additional
information is available.
37
Power Spectral Density
∆𝑡
෡𝑘 =
2. Calculate the scaled FFTs 𝐗 𝐘෠𝑘
𝑁𝑗

3. The PSD matrices are 𝐒෠𝑘 = 𝐗

෡𝑘𝐗
෡ ∗𝑘 , and the auto-PSDs at each
measurement channel are:
𝐬ො𝑘 = diag(𝐒෠𝑘 )

where diag() is the operator that takes the diagonal elements of

a matrix.

38
Power Spectral Density: Example

39
Power Spectral Density: Example

T T T T

T T T T

T T T T

T T T T
𝐒 𝜔 = lim 𝐸[𝐗 𝑇 (𝜔)𝐗 ∗𝑇 (𝜔)]. The expectation can be approximated by the sample
𝑇→∞
1 𝑁𝑎
෡=
average 𝐏 ෌𝑎=1 ෡ 𝑘,𝑎 𝐗
𝐗 ෡ ∗𝑘,𝑎 to get smooth PSDs. It can be done by dividing data into
𝑁𝑎
40
𝑁𝑎 non-overlapping segments.
 PSD using raw data

Story 1
Story 2
-5 Story 3
10 Story 4
PSD (g2/Hz)

-10
10

-15
10

0 2 4 6 8 10 12
Frequency (Hz)

41
Power Spectral Density: Example
-2
10
Channel 1
Channel 2
Channel 3
Channel 4

-3
10
PSD (g/sqrt(Hz))

10 -4

10 -5
2 4 6 8 10 12 14 16
Frequency (Hz)

42