A Guide to Random Data Analysis for

Computational Fluid Dyanamics

www.C AEAI.com
 A Guide to Random Data Analysis for Computational Fluid Dynamics

Copyright 2016
www.C AEAI.com

Published by Computer Aided Engineering Associates Inc.

All rights reserved. Except as permitted under U.S. Copyright Act of 1976, no part of
this publication may be reproduced, distributed, or transmitted in any form or by any
means, or stored in a database or retrieval system, without the prior written permission
of the publisher.

Visit our website at www.caeai.com.

A Guide to Finite Element Analysis of Par t-to -Par t Connections / 2

 Table of Contents

Introduction 4

Chapter 1
Techniques for Transient Flow Field Data 6

Chapter 2
Introduction to The Fast Fourier Transform 10

Chapter 3
Understanding The Fast Fourier Transform Inputs 14

Chapter 4
Application of The Fast Fourier Transform 17

Chapter 5
Accuracy of Your FFT and Parsevals Theorem 20
www.C AEAI.com

Chapter 6
The Short-Time Fourier Transform 22

Chapter 7
Cross-Coherence 25

Chapter 8
Auto and Cross Correlation 27

Chapter 9
Two-Point Statistics 30

Chapter 10
About CAE Associates 33
A Guide to Random Data Analysis for Computational Fluid Dynamics / 3
 Introduction

The growth of industrial computing capacity has allowed for increased turbulence-
resolving Computational Fluid Dynamics (CFD) approaches. As such, this has directly
led to the need to resolve finer scale turbulence which in turn requires decreasing time
steps. This increases run times and will result in larger data sets for post-processing.
Todays CFD engineers will increasingly find themselves in situations where they need
to integrate reliable statistical measures into their everyday post-processing routines. If
done properly, this will assure not only the quality of the post-processed data, but it will
also allow the engineer to extract more information from their data that may be other-
wise difficult to see.

At the beginning of a transient CFD analysis, the flow field is typically assigned an ar-
bitrary initial condition or steady state solution. This solution must be advanced to a
fully-developed statistically stationary state before meaningful sampling can begin.
This guide aims to discuss two primary topics. The first of which is often times misun-
www.C AEAI.com

derstood and is referred to as the initial transients in a CFD simulation and judging
whether or not the flow field has advanced beyond this stage. The length of this initial
transient is strongly case specific and is usually judged by intuitive inspection alone.
Often times, we simply look for repeating, oscillating signals at various points in the flow
fields which may not be adequate. Post-processing statistical laden results which are still
in this window may be misleading. The second topic relates to the post-processing of
transient data-sets and some of the more common techniques used for extracting the
most from your data.

A Guide to Random Data Analysis for Computational Fluid Dynamics / 4

 Introduction

As stated previously, the amount of useful data embedded in a transient signal can be
difficult to see for entry level and, often times, even advanced users. As such, this guide
is intended to serve as a point of reference to aid the everyday CFD engineer when
post-processing transient flow field data. It focuses on the introduction and application
of only a few of the commonly used mathematical approaches that may be useful for in-
terpreting transient data. The focus of this manual leans towards techniques which make
use of transforming a time series data-set into the frequency domain for further analysis.
www.C AEAI.com

A Guide to Random Data Analysis for Computational Fluid Dynamics / 5

 Chapter 1: Techniques for Transient
Flow Field Data

This chapter will discuss some of the post-proccessing techniques to consider when
your CFD analysis is transient in nature. We will first need to gain an understanding of
the concept of a stochastic process which is by definition a random process. A stochas-
tic process is said to be ergodic if its statistical properties can be derived from only a
subset of the entire dataset. This subset, or window of data, must represent the average
statistical properties of the entire process regardless of where the window was extract-
ed from in the process. In a nutshell, this implies that we expect that the statistics of the
signal would converge to a single value if given a large enough sampling window. With
Reynolds Averaged Navier-Stokes (RANS) CFD simulations, we most commonly focus
on the first three statistical moments: the mean, variance, and standard-deviation. For
advanced turbulence modeling with data acquired by the use of LES and DNS solvers,
we may also decide to include skew and kurtosis in our random data analysis.

For illustration purposes, lets consider a random pressure signal as shown in Figure 1.
www.C AEAI.com

The data was acquired at a single point in space and sampled at uniform frequency of
100,000 [hz].

Figure 1: Acquired random pressure signal

A Guide to Random Data Analysis for Computational Fluid Dynamics / 6

 Chapter 1

Equation 1 represents the calculation of the running mean and standard deviation for
given variable xi. To calculate the statistics as a function of sample size, we simply in-
crement from i=1 to i=N, where N represents the total sample size, Figure 2 plots the
running calculation for the average i and the standard deviation i as a function of
sample size.

The dotted red lines in Figure 2 represent the mean and standard deviation to within
plus or minus one percent of the calculated value when using the entire dataset. Upon
inspection of Figure 2, we observe two key points:

1) The mean converges much more rapidly than the standard-deviation (within ap-
proximately 500-600 samples). This is not a surprising result as the mean is utilized in the
calculation of the standard-deviation.

2) The block size required for convergence of the first 3 statistical moments (to within
one percent) is approximately Nc=20,000 (recall, the variance is nothing more than the
standard-deviation squared).
www.C AEAI.com

Figure 2: Running mean and standard deviation

A Guide to Random Data Analysis for Computational Fluid Dynamics / 7

 Chapter 1

This implies we would require approximately of simulation time per

block for our mean and variance to be fully converged to within one percent.

For transient analysis, which is expected to be stationary and ergodic, this simple anal-
ysis can be quite powerful. The only way to ensure the initial transients have settled in
your time marching simulation is by monitoring these statistics and ensuring they have
converged to within a reasonable tolerance. If one wishes to only post-process mean
flow field results from the transient data set - such as velocity and pressure quantities,
the required amount of simulation effort is far less demanding than acquiring statistics
related to the fluctuating components; (as they relate to the higher order statistical
moments) such as turbulent velocities, Reynolds stress, or block sizes used for Fast Fou-
rier Transform (FFT) analyses.

In the example above, we have only included the results from a single point in the
flow field. Using ANSYS CFX, you can monitor these results using the transient-statistics
tab and the Output Control menu which writes the running mean and standard
deviation for the entire computational domain as illustrated in Figure 3 and 4 for flow
over a cylinder. Using CFD-Post and its ability to compare-cases (say, at iteration i
compared to iteration i+1000), you could generate contour plots and monitor these
statistics for multiple planes. This would allow you to rapidly gain a qualitative assess-
ment of global transient-statistical convergence levels.
www.C AEAI.com

Figure 3: Instantaneous u-velocit y contours for f low over a c ylinder star ted from 0 velocit y

A Guide to Random Data Analysis for Computational Fluid Dynamics / 8

 Chapter 1

Figure 4: Mean u-velocit y contours for f low over a c ylinder star ted from 0 velocit y

This chapter illustrated how you could apply random data analysis statistical tech-
niques to monitor the convergence levels related to the initial transients for your time
marching analysis. As we have seen, when post processing transient CFD data, there
are multiple items to consider. We must consider how well the solver converges on an
iteration by iteration basis and how far the simulation has been carried out. We also
have found we need to monitor the flow field statistics to assess whether or not the
converged transient solution is completely reliable and producing meaningful results.
www.C AEAI.com

In Chapter 2, the Fast Fourier Transform and its implementation will be discussed.

A Guide to Random Data Analysis for Computational Fluid Dynamics / 9

 Chapter 2: Introduction to The Fast
Fourier Transform

In Chapter 1, the concept of a stochastic process and random data analysis were dis-
cussed. Specifically, the focus was on computing temporal statistics such as the mean
and standard deviation. In this chapter, the Fast Fourier Transform (FFT) and its imple-
mentation and use in data post-processing of CFD analyses will be discussed.

The French mathematician J. B. J. Fourier discovered that periodic waveforms can be

modeled as the sum of sine and cosine waves called a Fourier series. The mathematical
expression is given in Equation 1 for a generic signal with a known range. Application of
Equation 1 for a square wave with increasing values of n yields Figure 1, where it is ap-
parent that as n -> the square wave is more accurately reproduced.
www.C AEAI.com

Figure 1: Par tial sums of the Fourier series of a square wave

(ht tp: //w w w.reed.edu/physics/courses/Physics331.f 08/pdf/Fourier.pdf )

A Guide to Random Data Analysis for Computational Fluid Dynamics / 10

 Chapter 2

The Fourier transform is used to represent a periodic function by a discrete sum of com-
plex exponentials. Instead of Equation 1 above, we could equally write the summation
in complex form:

If we define a new variable k=2n we can define the Fourier integral (after some
math)

If we want a result for F(k), which is called the Fourier transform of f(x), we apply
equation 2 in the limit as approaches infinity with the Fourier series coefficients

The Fourier transform is closely related to a Fourier series and is used to represent a
general function by a continuous superposition (or integral) of complex exponentials.
The Fourier transform decomposes an arbitrary waveform into its sine components,
thus revealing its frequency content that might else be difficult to detect. A Discrete-
Time Fourier transform (DTFT) is a form of the Fourier transform that is applicable to
www.C AEAI.com

uniformly sampled continuous functions. The Fast Fourier Transform (FFT) takes ad-
vantage of powers of two and is a numerically efficient implementation of computing
DTFTs. Simply stated, the FFT is an efficient numerical algorithm that allows for the
transformation of a time dependent signal into the frequency domain (or vice-versa).

For illustration purposes, lets consider a random fluctuating pressure signal with the
mean removed as shown in Figure 2. The data is acquired at a single point in space
and sampled at a uniform frequency of 100 [kHz]. It is difficult to extract meaningful
information from this signal outside of typical temporal statics like mean and variance
that was illustrated in Chapter 1.

A Guide to Random Data Analysis for Computational Fluid Dynamics / 11

 Chapter 2

Figure 2: Acquired Random Pressure Signal

In order to overcome this shortfall, we pass this data series through an FFT algorithm.
The resulting transform results in what is called a Power-Spectral (PS) plot, and is illustrat-
ed in Figure 3. The y-axis has been non-dimensionalized using the free stream dynamic
pressure. Bear in mind that during the FFT, the resulting units will be based on the signal
passed. In this case, since pressure was used, the resulting units are [Pa^2] and results
in a power spectra. If scaling were applied using the sampling frequency a Power
Spectral Density (PSD) plot could be obtained resulting in units [Pa^2/Hz] as given in
Equation 4. The PSD normalizes the power spectrum to the spectrum measured by a 1
Hz-wide square filter. This convention is most often used for noise-level measurements.
www.C AEAI.com

The PS plot describes the distribution of the mean square power over a frequency
range related to the Nyquist frequency. Examining this figure, we can now clearly see
at least 5 discrete peaks occur as tabulated in Table 1.

Table 1. Frequenc y content from FF T

A Guide to Random Data Analysis for Computational Fluid Dynamics / 12

 Chapter 2

Figure 2: Acquired Random Pressure Signal

Figure 3: FF T of the Pressure Signal

Further, it is clear that the third tone is dominant and exhibits over three times the
peak power compared to the other tones. The total power could be determined by
obtaining the area under the curve with specified frequency intervals if desired such as
www.C AEAI.com

done with Octave analysis.

In Chapter 3, Fast Fourier Transform Inputs will be explored in greater detail.

A Guide to Random Data Analysis for Computational Fluid Dynamics / 13

 Chapter 3: Understanding The
Fast Fourier Transform Inputs

This chapter will take a more in-depth look at the inputs required for the FFT and will
help the CFD engineer understand how to utilize it properly. To do this, the focus will be
on a built in function in Python (Anaconda) to generate the FFTs using Welchs method.
Below is the function call used and the inputs will be discussed systematically.

Pxx,f = mpl.psd(pfluc,nfft,fs,window=mlab.window_hanning,scale_by_freq=False,side
s=onesided,noverlap=nlap)

The FFT Command Line and Welchs Method:

Welchs method is based on the concept of decomposing the acquired signal into
segments, performing the FFT on individual blocks, and taking the resulting ensem-
ble-average of the constructed periodograms. Welchs method may be considered
an improvement on the standard spectrum methods, as it reduces the noise in the
www.C AEAI.com

estimated power spectra. The drawback is that it does this by reducing the frequency
resolution (as nfft is smaller than N).

Figure 1: Welchs Method for FF T with No O verlapping

A Guide to Random Data Analysis for Computational Fluid Dynamics / 14

 Chapter 3
Frequency Resolution:
IFrequency resolution is illustrated by Equation 1. In order to increase the frequency
resolution (i.e. reduce f ), the ensemble period (T) must be increased. This is achieved
by either reducing the sampling frequency (fs), or increasing the number of sampling
points per ensemble (nfft).

pfluc:
This is the signal passed to the FFT that is assumed to be discretely sampled with a
uniform sampling frequency. The mean has been removed from the signal resulting in
just the fluctuating component of the variable.

nfft:
For FFTs, the signal will be decomposed into powers of nfft = 2m where m is speci-
fied by the user. Trailing data points after the last block is created are simply truncated.

Window:
Applying a Window to an FFT in essence modulates the input signal so that the spec-
tral leakage is evened out. If we take a sine wave, you would expect to see just one
sample at the frequency of the signal. Spectral leakage occurs due to the fact we are
binning the data and sometimes the frequency calculated falls right at the edge of
our bins. The power from the sample may leak out to the surrounding frequency bins,
www.C AEAI.com

and is thus referred to as spectral leakage. Windowing reduces the amplitude of the
samples at the beginning and end of the window and alters this leakage. Windowing
is implemented by multiplying the input signal with a windowing function. There are
many common windowing functions, but for CFD and turbulent flows, the Hanning
window is the most widely adopted. The Hanning window is given by:

A Guide to Random Data Analysis for Computational Fluid Dynamics / 15

 Chapter 3
Scaling:
Scaling by frequency specifies whether the resulting spectral density values should be
scaled by the scaling frequency, which gives density in units of This allows for
integration over the returned frequency values. Commonly, PSDs are reported in terms
of decibels in which case the PSD must be scaled using:

where THH represents the Threshold of Human Hearing typically taken as 2e-05 [Pa].

One or Two-Sided:
Specifies which side (or both) of the spectrum to be returned.

Overlapping:
Overlapping your data is a method to ensure you get the highest number of Nblocks
increasing your averaging accuracy and frequency resolution by using repetitive data.
As shown in Figure 2, we are in essence sliding our nfft segment back to the left reuti-
lizing a defined percentage of overlap points.
www.C AEAI.com

Figure 2: Illustration of 50% O verlapping

Now that we have a background on the FFT implementation in codes like Python, we
can exercise the code and discuss the results. This will be the focus of Chapter 4.

A Guide to Random Data Analysis for Computational Fluid Dynamics / 16

 Chapter 4: Application of The Fast Fourier
Transform

With a firm understanding of the required inputs for performing a FFT analysis, we are
ready to execute the algorithm for a representative data set and discuss the effects of
the chosen parameters. This chapter will focus on the built in function in Python (Ana-
conda) to generate the FFTs using Welchs method as discussed in the previous chap-
ters and repeated below for clarity.

Pxx,f=mpl.psd(pfluc,nfft,fs,window=mlab.window_hanning,scale_by_freq=False,sides
=onesided,noverlap=nlap).

Selection of nfft:
Based on the frequency resolution given by Equation 1, it is easy to see that the larger
the size of your nfft, the more accurate your frequency resolution becomes for a fixed
sampling rate.
www.C AEAI.com

Unfortunately, by increasing your nfft you lower the number of available periodo-
grams used in the averaging of the final FFT, as each window has now been extended.
To mitigate this, you can use overlapping. Typically, the optimal overlap percentage is
between 50 and 75% of the nfft size. Ultimately, the choice of nfft becomes a balanc-
ing act between the desired accuracy of the frequency resolution vs. the magnitude
of the spectra.

Lets assume we have a a scenario where the lowest known resonate frequency is near
1 [kHz] . Lets further assume we have the capability for infinite sampling frequency.

A Guide to Random Data Analysis for Computational Fluid Dynamics / 17

 Chapter 4
In this scenario, the lowest frequencies are the most demanding to capture because
they require they longest sampling times. For this scenario, lets take the sampling fre-
quency to be fixed at 1e-05 [s] or (100 [kHz]) with N=219 total points. We desire at least
20 cycles per nfft period to ensure statistical accuracy (the more the better, of course).
This means we would need at least 0.05 [s] of simulation time and 5,000 points per
nfft. We have now given ourselves a reasonable estimate for our selection of nfft that
should be between 212 and 213. The trade-off here is with the accuracy of the resulting
frequency resolution as seen in Table 1. But, we also want to average multiple periodo-
grams (per Welchs method) to ensure accuracy. Thus, we may want on the order of 30
ensembles which requires at least a total signal length of N=150,000 points. Remember,
we are using powers of 2 so we really need the next power of 2, which would be 218.

Table 1 illustrates our corresponding frequency resolution (bin width), number of aver-
aged periodograms, and the number of shedding cycles per period for selected values
of nfft.
www.C AEAI.com

The following figures were generated using the FFT call presented at the beginning of
this chapter with a Hanning window and no overlapping. Figure 1 illustrates the effect
of the selection of nfft on your FFT.

Figure 1: FF T Based on Selec ted Sizes

for nf f t

A Guide to Random Data Analysis for Computational Fluid Dynamics / 18

 Chapter 4
Figure 2 illustrates the effect of the selection of frequency resolution on your resulting
FFT. The original signal length was decimated by a factor of 2 and 4 to emulate larger
sampling frequencies. Similar to Figure 1, the peak values decrease and the broadening
of the peaks is still evident as the frequency resolution increases.

This brings to light that when comparing spectral plots with data obtained from ex-
perimental or numerical methods, care must be taken to understand exactly how the
plots were constructed. (Not just the physics of the problem!) It is important that the
spectral plots be constructed in the same fashion. If complete information is not sup-
plied on how the FFT was performed, one may end up with misleading discrepancies
when comparing back to back figures!
www.C AEAI.com

Figure 2: FF T Based on Frequenc y Resolution Acquired by

Decimating the Original Signal

Now that the background, required inputs, and execution of a typical FFT algorithm
have been presented, we need to address the accuracy of the computation. This will
be the focus of Chapter 5.

A Guide to Random Data Analysis for Computational Fluid Dynamics / 19

 Chapter 5: Accuracy of Your FFT and
Parsevals Theorem

The previous chapter discussed some of the most common variables utilized when
performing an FFT analysis. This chapter will discuss some key points when evaluating
the accuracy of a computed FFT by discussing Nyquists criterion and Parsevals theo-
rem.

Nyquist Criterion:
For a signal to be accurately reconstructed from a discrete set of sampled points, the
sampling rate should be at minimum twice the maximum frequency of interest. Thus,
the Nyquist frequency is defined as half of the sampling frequency of a discrete signal.
The Nyquist rate is the minimum sampling rate that satisfies the Nyquist sampling
criterion. If you wish to extract the frequency content from a 1 [kHz] signal, then the
Nyquist criterion tells us that our minimum required sample rate is 500 [Hz]. The most
straightforward way to avoid aliasing is to ensure your sampling frequency is high
www.C AEAI.com

enough to capture the spatial and or temporal variations in a given signal.

Parsevals Theorem:
Parsevals Theorem is a simple statement of the conservation of energy. The theorem
states that the total energy computed in the time domain must equal the total en-
ergy computed in the frequency domain. In other words, no loss of energy (or power)
should occur during the transformation from one space into another. For a continuous
integrable function, this is given by Equation 1below.

A Guide to Random Data Analysis for Computational Fluid Dynamics / 20

 Chapter 5
For discretely sampled data series, the relationship is simply the sums given by Equa-
tion 2.

The following section focuses on the built in function(s) in Python (Anaconda) and
the implementation of Parsevals Theorem. The snippet of code is well commented so
adaption to another engine of choice should be somewhat seamless.

m=13
## Do FFT
# pfluc is the acquired signal from a uniformly sampled time-history
nfft = int(2**m) # size of each record to ensemble average
nlap = int(0) # No. of pts to overlap
fr = fs / float(nfft) # Calculate frequency resolution
k = int((N-nlap)/(nfft-nlap)) # No. of periodograms to average
# Get 2-sided FFT with appropiate scaling that yields [units^2/Hz]
Pxx,f = mpl.psd(pfluc,nfft,fs,window=mlab.window_hanning,scale_by_
freq=True,noverlap=nlap)
parseval_side2 = (np.trapz(Pxx,f )/fs) # Num. int. PSD to get units [Pa^2]. Scale by fs for
units matching
parseval_side1 = np.var(pfluc) # Get variance of time domain to get [Pa^2]
Error = (np.abs(parseval_side1-parseval_side2)/parseval_side1)*100 # Compute error

Chapter 6 will discuss some of the shortcomings of the FFT and discuss other potential
www.C AEAI.com

data that can be extracted from your time series.

A Guide to Random Data Analysis for Computational Fluid Dynamics / 21

 Chapter 6: The Short-Time
Fourier Transform

There are many inherent shortcomings of an FFT analysis that the CFD engineer may
need to be aware of, along with the potential work arounds. When performing an FFT
analysis, you can switch to and from the frequency domain revealing frequency con-
tent and amplitude information with ease. One drawback is there is no information
provided for the time history of a given amplitude or peak. One of the most obvious
questions one may ask is: is my dominant amplitude apparent for all time or does it
reveal itself only intermittently? When the peak with the dominant amplitude switch-
es frequencies as a function of time, the signal is said to exhibit mode-switching.
Joint Time-Frequency Analysis (commonly referred to as Short-Time Fourier Trans-
forms, or STFT) using Fourier transforms, given by Equation 1, provides information
containing frequency, time, and amplitude allowing one to visualize the time-evolu-
tion of the frequency content.
www.C AEAI.com

The spectrograms for this approach are computed using a short time duration where
z(t) is an applied window function. The time dependent input signal is partitioned into
several disjointed or overlapped blocks by multiplying the signal with a specified win-
dow. The FFT is then applied to each block. The blocks occupy different time periods
so the resulting STFT computes the spectral content of the input signal at each period.
The STFT therefore represents the time-dependent power spectrum of the input sig-
nal. The steps for computing the STFT are as follows:

1. Choose a window function of finite length

2. Place the window on top of the signal at t = 0 [s]
3. Truncate the signal using this window
A Guide to Random Data Analysis for Computational Fluid Dynamics / 22
 Chapter 6
4. Compute the FFT of the truncated signal
5. Incrementally slide the window to the right
6. Repeat steps 3-5 until the end of the signal is reached

The resulting spectrogram (where the magnitude has been normalized) for the sig-
nal used in this series is provided in Figure 1(c) below. Bear in mind, that due to the
uncertainty principle discussed above, a wide window will result in good frequency
resolution, but poor time resolution. Conversely, a narrow window will yield better time
resolution at the expense of frequency resolution. Four discrete tones are still apparent
(where the 5th tone and beyond are masked by the color scheme as their magnitudes
are on the order of the broadband noise) and we can now see their magnitude as a
function of time. It is apparent that tone three is dominant for most of the time history
but there are periods in time where this signal drops out (between 0.15 and 0.20 [s] for
example) suggesting that mild mode switching may be apparent.
www.C AEAI.com

Figure 1: Illustration of the Uniform-

ly Sampled Raw Signal, the Corre -
sponding FF T and the S TF T.

A Guide to Random Data Analysis for Computational Fluid Dynamics / 23

 Chapter 6
A well-known property of the Fourier transform pair s(t) and s() is the uncertainty
principle which states the time duration (t) of s(t) and the frequency bandwidth ()
are related by t 1/2. This implies a longer time duration of the input signal will re-
sult in a smaller frequency bandwidth. Conversely, the larger the frequency bandwidth
of the transformed input signal, results in a shorter duration of the input signal. This
illustrates the tradeoff between the frequency resolution and time resolution when
using this method. The precision of the transform is determined by the size of the se-
lected window and is fixed for all frequencies.

A technique which allows the use of variable window sizes may be employed to over-
come the time and frequency resolution tradeoff inherent with the STFT. This analysis
is called Wavelet analysis and is a topic for future discussion. Chapter 7 will discuss
another topic related to the FFT called Cross-Coherence analysis.
www.C AEAI.com

A Guide to Random Data Analysis for Computational Fluid Dynamics / 24

 Chapter 7: Cross- Coherence

This chapter introduces the concept of Cross-Coherence and discusses some benefits
of performing this type of analysis.

So far, we have dealt with extracting information from a single time series. In this
chapter, we will discuss a method for relating two time-dependent signals as a func-
tion of frequency. This type of analysis is referred to as the level of Coherence be-
tween the signals. The Cross-Coherence of two-such signals is given by Equation 1
where yx (f) represents the cross-spectra of the two signals and is normalized by the
square root of each signals respective auto-spectra.

If the signals are identical, the correlation coefficient is unity, and if they are unrelated
www.C AEAI.com

the correlation coefficient is zero. If the signals are identical, but the phase is shifted by
exactly 180 degrees, then the correlation coefficient is -1.

Figure 1: Cross- Coherence of a Baseline Configuration amd a Geometrically Modified Scenario, Sampled at the
Same Points in Space and Frequenc y. T he FF T (at the same point) are Embedded in the Right Hand Corner

A Guide to Random Data Analysis for Computational Fluid Dynamics / 25

 Chapter 7
The coherence plot for the cases in the first subfrigure show a strong correlation with
coefficients exceeding 0.80 for most peak tones. Since the pressure signal is being cor-
related, this tells us that the pressure at the aft-sensing point is strongly correlated with
what is happening as you progress away from that point (as x/L decreases). Conversely,
for the altered case, the magnitudes of the coherence fall significantly, coinciding with
a broadening and lowering of the peaks presented on the embedded FFT spectra. This
suggests that the feedback mechanism thought to be responsible for generating these
peaks has been disturbed.

The use of Cross-Coherence plots is yet another method for extracting useful informa-
tion from your time series data, but their use requires at least two data sets with the
same sampling rate, equivalent length, and sampling period. Chapter 8 will explain
both auto and cross-correlation in detail.

www.C AEAI.com

A Guide to Random Data Analysis for Computational Fluid Dynamics / 26

 Chapter 8: Auto and Cross Correlation

The use of Short-Time Fourier to obtain amplitude, time, and frequency content was
discussed in Chapter 7. This chapter introduces the concept of Cross-Correlation and
we will discuss some benefits of performing this type of analysis.

Cross-Correlation analysis can be used to help determine how strongly two signals
in the time-domain are related. If the signals are identical the correlation coefficient
is unity and if they are unrelated the correlation coefficient is zero. If the signals are
identical except that the phase is shifted by exactly 180 degrees then the correlation
coefficient is -1. When two independent signals are compared in the time domain
the procedure is known as cross-correlation. Autocorrelation is a special case of cross-
correlation where the same signal is compared to itself. The autocorrelation coefficient
of a signal x(t) is given as:
www.C AEAI.com

The time lag, or separation between points, is given rt. The maximum number of lags
is given by r. The cross-correlation of two signals x(t) is and y(t) is then given by Equa-
tion 2.

A Guide to Random Data Analysis for Computational Fluid Dynamics / 27

 Chapter 8

The cross-correlation performed in the time domain is a computationally inefficient

approach which may be computed faster taking advantage of the FFT. The enhanced
correlation is written by convolving the first function, g(t), with the time-reverse of the
second function given by h(t) in Equation 3:

where the * is the complex conjugate and the x indicates the cross-correlation. This
may be implemented with the following routine using Python:

def xcorrNorm( y, y2, fs, maxlags):

import numpy as np
####### Cross Correlation via FFT ###############
fs=float(fs)
y = y - np.mean(y) # sig 1
y2 = y2 - np.mean(y2) # sig 2
N = len(y)
sp = np.fft.rfft(y) # compute FFT transform of 1st signal
sp2 = np.fft.rfft(y2) # compute FFT of 2nd sig
www.C AEAI.com

sp = np.conj(sp) # complex conjugate of 2nd sig

rho = np.fft.irfft( sp*sp2 ) # inverse transform of the convolution
rho = np.fft.fftshift(rho) # shift fft so 0 is centered
rho = rho / (np.std(y) * np.std(y2)) / N # normalize
tlag = np.int32(np.linspace(-len(rho)/2,len(rho)/2,len(rho))) / fs # Get lags
rho = rho[(N/2) - maxlags : (N/2) + maxlags] # extract maxlag data
tlag = tlag[(N/2) - maxlags : (N/2) + maxlags] # extract maxlag data
return tlag, rho

A Guide to Random Data Analysis for Computational Fluid Dynamics / 28

 Chapter 8

Figure 1: Auto and Cross Correlation for the T ime

Series Presented in Previous Chapters

Figure 1 represents the cross and auto correlation of given pressure signals discussed
in the previous blogs. Referring to Figure 1: since we know the offset from the leading
edge to x/L=0.25 we may estimate the upstream speed of the propagating wave by
looking at the offset of the corresponding peaks of these curves. In this example we
obtain (L=0.0762 [m]) and a velocity of
www.C AEAI.com

Also, analogous to comparing peaks for the cross-coherence in the previous chapter,
we see here that the cross-correlations exhibit well-correlated peaks indicating what is
sensed at the aft wall is also strongly correlated with the pressure signal sensed up-
stream. As expected, the black line exhibits a peak of unity at zero lag since this is the
auto-correlation. Chapter 9 will discuss the use of two point statistics using turbulent
velocity and fluctuating pressures.

A Guide to Random Data Analysis for Computational Fluid Dynamics / 29

 Chapter 9: Two - Point Statistics

In the previous chapter, the use of auto and cross-correlations were discussed. These
are commonly used to gain an understanding of the relationship between random
signals. The final chapter in this guide discusses the use of two-point statistics using
turbulent velocities or pressures.

Several statistical concepts can be used for the analysis of unsteady quantities of inter-
est and in this chapter, I will focus on unsteady or fluctuating pressure signals. The
fluctuating component of the pressure are found from a classical Reynolds decompo-
sition. Equation 1 gives the fluctuating pressure where the prime represents the fluc-
tuating component of the variable, the overbar represents the mean, and a variable
with no demarcation represents the instantaneous value.

(1)
www.C AEAI.com

The mean takes the traditional definition given in Equation 2 below where the num-
ber of independent observations is given by N, which is sufficiently large to ensure the
mean value has converged.

(2)

The variation of the measured unsteady pressure is studied using the root-mean-
square (rms) of the fluctuating variable and is commonly referred to as the standard
deviation in statistical literature.

A Guide to Random Data Analysis for Computational Fluid Dynamics / 30

 Chapter 9

(3)

The length scale variations in the shear layer are studied by employing a two point
spatial cross-correlation on the components of the fluctuating pressure with an origin
defined at some fixed point with a known separation . The two point spatial
tensor is then given by:

(4)

At zero separation, if using velocity instead of pressure, the diagonal elements represent
the normal turbulent velocities and the off diagonals would represent the turbulent
shear stresses. With the spatial separation represented in the correlation, one may look
at how the velocity at the defined origin is related to the velocity at another point which
inherently defines a length of how far apart the selected variables are related.

Figure 1 (next page) illustrates spatial correlation plots of the fluctuating surface pres-
sure as a function of separation with the origin located at the center point of the aft wall
www.C AEAI.com

for each configuration, as introduced in previous chapters. The baseline case exhibits a
much larger, highly correlated region that extends from just below the origin to the top
of the aft wall. The controlled configurations well correlated region is noticeably more
compact, which suggests that smaller scale structures have impacted the aft wall. This
is consistent with the argument that the control is breaking up the organized turbulent
features found in the upstream shear layer which ultimately impact the aft wall.

A Guide to Random Data Analysis for Computational Fluid Dynamics / 31

 Chapter 9

The controlled configurations well correlated region is noticeably more compact, which
suggests that smaller scale structures have impacted the aft wall. This is consistent with
the argument that the control is breaking up the organized turbulent features found in
the upstream shear layer which ultimately impact the aft wall.

Figure 1: Figure 1. Fluc tuating sur face pressure correlation on the

af t wall with the origin located at the center point.

This chapter wraps up the guide on random data analysis techniques commonly used
for CFD analysis.
www.C AEAI.com

A Guide to Random Data Analysis for Computational Fluid Dynamics / 32

 Chapter 10: About CAE Associates Inc.

Since 1981, CAE Associates been helping companies large and small maximize value
from engineering simulation through expert FEA consulting and CFD consulting ser-
vices.

As an ANSYS channel partner since 1985, we provide full service technical support to
over 100 companies in our region, including some of the worlds largest and sophisti-
cated users of simulation like General Electric and United Technologies.

We offer a range of consulting options to fit unique and specialized needs of our clients.
It could be as simple as getting to the bottom of a product failure, or as complex as
developing the infrastructure and process for simulation within the entire organization
using a combination of software, training, mentoring and automation.
www.C AEAI.com

Our team of senior technical specialists bring graduate-level educations and an average
of 15 years of real-world experience in a variety of industries including: aerospace, elec-
tronics, consumer product, turbomachinery, civil engineering, manufacturing, biomedi-
cal, energy, and nuclear power.

For questions and to find out more about how CAE Associates can help you solve
your engineering analysis challenges, please contact us at [email protected].

A Guide to Random Data Analysis for Computational Fluid Dynamics / 33