Lecture 2 v2 (Compatibility Mode)

ERASMUS Teaching (2008), Technische Universitt Berlin
Digital Signal Processing

Assoc. Prof. Dr. Pelin Gundes Bakir
gundesbakir@yahoo.com
References
Mitra S S.K., K Digital Digital Signal Processing Processing, A Computer based approach, approach Mc Mc-Graw Graw Hill Hill, 3rd Edition, 2006. Heylen W., Lammens S. And Sas P., Modal Analysis Theory and Testing, Katholieke Universiteit Leuven Leuven, 1997 1997. Keith Worden Signal Processing and Instrumentation, Lecture Notes, http://www.dynamics.group.shef.ac.uk/people/keith/mec409.htm "The Scientist and Engineer's Guide to Digital Signal Processing, copyright 19971998 by Steven W. Smith. For more information visit the book's website at: www.DSPguide.com" Boore, D. M. and J. J. Bommer (2005). Processing of strong-motion accelerograms: Needs, options and consequences, Soil Dynamics and Earthquake Engineering 25,93-115 Boore, D. M. (2005). On pads and filters: Processing strong-motion data, Bull. Seism. Soc. Am. 95,745-750.
P. Gundes Bakir,
Vibration based structural health monitoring
Signals
Definition
A signal is a function of independent variables such as time, distance, position, temperature and pressure. A signal carries information, and the objective of signal processing i is i t to extract t t useful f l information i f ti carried i db by th the signal. Signal processing is concerned with the mathematical representation p of the signal g and the algorithmic g operation p carried out on it to extract the information present.
P. Gundes Bakir,
Definition
For most purposes of description and analysis, a signal can be defined simply as a mathematical function,
y = f ( x)
where x is the independent variable which specifies the domain of the signal i l e.g.: y=sin(t) is a function of a variable in the time domain and is thus a time g ; signal; X()=1/(-m2+ic+k) is a frequency domain signal; An image I(x,y) is in the spatial domain.
P. Gundes Bakir,
Signal types
F a simple For i l pendulum d l as shown, h b basic i d definition fi i i i is:
where m is the peak amplitude of the motion and =l/g with l the length of the pendulum and g the acceleration due to gravity. As the system has a constant amplitude (we assume no damping for now), ), a constant frequency q y (dictated ( by y physics) p y ) and an initial condition (=0 when t=0), we know the value of (t) for all time..
P. Gundes Bakir,
Signal types
Also, two Al t identical id ti l pendula d l released l d from f = 0 at t t=0, 0 will ill have h the th same motions at all time. There is no place for uncertainty here. If we can uniquely specify the value of for all time, i.e., we know the underlying functional relationship between t and , the motion is deterministic or predictable predictable. In other words words, a signal that can be uniquely determined by a well defined process such as a mathematical expression or rule is called a deterministic signal. The opposite situation occurs if we know all the physics there is to know, but s bu still c cannot o s say yw what the es signal g w will be at the e next e time e instant-then s e the e signal is random or probabilistic. In other words, a signal that is generated in a random fashion and can not be predicted ahead of time is called a random signal.
P. Gundes Bakir, Vibration based structural health monitoring 7
Signal types
Typical examples to deterministic signals are sine chirp and digital stepped sine.
P. Gundes Bakir,
Signal types
Typical examples to random signals are random and burst random.
P. Gundes Bakir,
Signal types
Random signals are characterized by having many frequency components present over a wide range of frequencies. The amplitude p versus time appears pp to vary y rapidly p y and unsteadily y with time. The shhhh sound is a good example that is rather easy to observe using a microphone and oscillloscope oscillloscope. If the sound intensity is constant with time, time the random signal is stationary, while if the sound intensity varies with time the signal is nonstationary. One can easily see and hear this variation while making ki the h shhhh hhhh sound. d
P. Gundes Bakir,
10
Signal types
Random signals are characterized by analyzing the statistical characteristics across an ensemble of records. Then, if the process is ergodic, the time (temporal) statistical characteristics are the same as the ensemble statistical characteristics. The word temporal means that a time average definition is used in place of an ensemble statistical definition.
P. Gundes Bakir,
11
Signal types
Transient signals may be defined as signals that exist for a finite range of time as shown in the figure. Typical examples are hammer excitation of systems systems, explosion and shock loading etc etc.
P. Gundes Bakir,
12
Signal types
A signal i l with i h a time i varying i mean i is an aperiodic i di signal. i l
P. Gundes Bakir,
13
Signal types
It should I h ld b be noted d that h periodicity i di i d does not necessarily il mean a sinusoidal signal as shown in the figure.
For a simple pendulum as shown, if we define the period by , then for the pendulum pendulum,
and such signals are defined as periodic.
P. Gundes Bakir,
14
Signal types
A periodic i di signal i li is one that h repeats i itself lf i in time i and di is a reasonable model for many real processes, especially those associated with constant speed machinery. Stationary signals are those whose average properties do not change with time time. Stationary signals have constant parameters to describe their behaviour. Nonstationary signals have time dependent parameters. In an engine excited vibration where the engines speed varies with time; the fundamental period changes with time as well as with the corresponding dynamic loads that cause vibration.
P. Gundes Bakir,
15
Deterministic vs Random signals

The signals can be further classified as monofrequency (sinusoidal) signals and multifrequency signals such as the square wave which has a functional form made up of an infinite i fi it superposition iti of f diff different t sine i waves with ith periods ,/2,/3, 1 D signals are a function of a single independent variable. The speech signal is an example of a 1 D signal where the independent variable is time time. 2D signals g are a function of two independent p variables. An image signal such as a photograph is an example of a 2D signal where the two independent variables are the p variables. two spatial
Classification of signals
The value of a signal at a specific value of the independent variable is called its amplitude. The variation of the amplitude as a function of the independent variable is called its waveform. For a 1 D signal, signal the independent variable is usually labelled as time. If the independent variable is continuous, the signal is called a continuous-time signal. A continuous time signal is defined at y instant of time. every If the independent variable is discrete, the signal is called a g . A discrete time signal g takes certain numerical discrete-time signal values at specified discrete instants of time, and between these specified instants of time, the signal is not defined. Hence, a discrete time signal is basically a sequence of numbers.
A continuous-time signal with a continuous amplitude is usually called an analog signal. A speech signal is an example of an analog signal.
A discrete time signal with discrete valued amplitudes represented p by y a finite number of digits is referred to as a digital signal.
P. Gundes Bakir,
18
A discrete di t ti time signal i l with ith continuous valued amplitudes is called a sampled-data signal. A digital signal is th thus saq quantized anti ed sampled-data signal.
A continuous-time signal with discrete valued amplitudes has been referred to as a quantized boxcar signal. This type of signal occurs in digital electronic circuits where the signal is kept at fixed level (usually one of two values) between two instants of clocking.
CLASSIFICATIONS OF SIGNALS
1 1D D signals signals 2D signals Stationary Non-Stationary
D t Deterministic i i ti
Random
Continuous
Transient
Periodic
Aperiodic
Monofrequency (sinuzoidal)
P. Gundes Bakir,
Multi Multifrequency
T Transient i t
I fi it aperiodic Infinite i di
20
Typical Signal Processing Operations p
Typical signal processing operations p

In the case of analog signals, most signal processing operations p are usually y carried out in the time domain. In the case of discrete time signals g , both time domain and frequency domain applications are employed. I In either i h case, the h desired d i d operations i are implemented i l d by a combination of some elementary operations such as:
Simple time domain operations Filtering Amplitude A lit d modulation d l ti
Simple Time Domain Operations

The three most basic time-domain signal operations are: Scaling Delay Addition Scaling is simply the multiplication of a signal by a positive or a negative constant. In the case of analog signals, this operation is usually called amplification if the magnitude of the multiplying constant constant, called gain, is greater than one. If the magnitude of the multiplying constant is less than one, the operation is called attenuation. Thus, if x(t) is an analog signal, the scaling operation generates a signal y(t)=x(t), x(t) where is the multiplying constant.
P. Gundes Bakir,
23

The three most basic time-domain signal operations are: Scaling Delay Addition Delay operation generates a signal that is delayed replica of the original signal. For an analog signal x(t), y(t)=x(t-t0) is the signal obtained by delaying x(t) by the amount t0, which is assumed to be a positive number. number If t0 is negative, negative then it is an advance operation.
P. Gundes Bakir,
24

The three most basic time-domain signal operations are: Scaling Delay Addition Addition operation generates a new signal by the addition of signals. For instance, y(t)=x1(t)+x2(t)-x3(t) is the signal generated by the addition of the three analog signals x1(t), x2(t) and x3(t) .
P. Gundes Bakir,
25
Fourier transforms
This chapter focuses on Fourier-series expansion, the discrete Fourier transform, properties of Fourier Transforms and Fast Fourier Transform
Fourier transforms
Fourier Fo rier analysis anal sis is a family famil of mathematical techniques, techniq es all based on decomposing signals into sinusoids. The discrete Fourier transform (DFT) is the family member used with digitized signals. Why are sinusoids used? A sinusoidal input to a system is guaranteed to produce a sinusoidal output. Only the amplitude and phase of the signal can change; the frequency and wave shape must remain the same. Sinusoids are y waveform that have this useful p property. p y the only The general term Fourier transform can be broken into four categories, g from the four basic types yp of signals g that can be encountered. resulting
P. Gundes Bakir,
27
Categories of Fourier Transforms
P. Gundes Bakir,
28
Fourier transforms
These four Th f classes l of f signals i l all ll extend t d to t positive iti and d negative ti infinity. i fi it What Wh t if you only have a finite number of samples stored in your computer, say a signal formed from 1024 points? There isnt a version of the Fourier transform that uses finite length signals. Sine and cosine waves are defined as extending from negative infinity to positive infinity. You cannot use a group of infinitely long signals to synthesize something finite in length. The way around this dilemma is to make the finite data look like an infinite length signal. This is done by imagining that the signal has an infinite number of samples on the left and right of the actual points. If all these imagined samples l have h a value l of f zero, the th signal i l looks l k discrete di and d aperiodic, i di and d the th discrete time Fourier transform applies. As an alternative, A lt ti the th imagined i i d samples l can be b a duplication d li ti of f the th actual t l 1024 points. In this case, the signal looks discrete and periodic, with a period of 1024 samples. This calls for the discrete Fourier transform to be used.
P. Gundes Bakir,
29
Time and frequency domains

As shown in the figure, the Discrete Fourier transform changes an N point input p signal g into two N/2 + 1 p point output p signals. g The input p signal g contains the signal being decomposed, while the two output signals contain the amplitudes of the component sine and cosine waves. The input signal is said to be in the time domain. This is because the most common type of signal entering the DFT is composed of samples taken at regular intervals of time. The term time domain in Fourier analysis, may actually refer to samples taken over time. The term frequency domain is used to describe the amplitudes of the sine and cosine waves.
P. Gundes Bakir,
30

The frequency domain contains exactly the same information as the time domain, just in a different form. If you know one domain, you can calculate the other. Given the time domain signal, the process of calculating the frequency domain is called decomposition, analysis, the forward DFT, or simply, the DFT. If y you know the frequency q y domain, , calculation of the time domain is called synthesis, or the inverse DFT. Both synthesis and analysis can be represented in equation form and computer algorithms. The number of samples in the time domain is usually represented by the variable N. While N can be any positive integer, a power of two is usually chosen, i.e., 128, 256, 512, 1024, etc. There are two reasons for this. First, digital data storage uses bi binary addressing, dd i making ki powers of ft two a natural t l signal i l length. l th Second, S d the th most t efficient algorithm for calculating the DFT, the Fast Fourier Transform (FFT), usually operates with N that is a power of two. Typically, N is selected between 32 and 4096. In most cases, the samples run from 0 to N-1 1 , rather than 1 to N.

Lower case letters represent time domain signals and upper case letters represent frequency domain signals. The figure shows an example DFT with N = 128. The time domain signal is contained in the array: x [0] to x [127]. Notice that 128 points in the time domain corresponds to 65 points in each of the frequency domain signals signals, with the frequency indexes running from 0 to 64. points in the time domain That is, N p corresponds to N/2 + 1 points in the frequency domain (not N/2 points). Forgetting about this extra point is a common bug g in DFT p programs. g
P. Gundes Bakir,
33
Fourier series expansion

Fourier series are infinite series designed to represent general periodic functions in terms of simple ones, namely cosines and sines. A function f(x) is called a periodic function if f(x) is defined for all real x and if there is a positive number p, called a period of f(x), such that
f ( x + p) = f ( x)
The graph of such a function is obtained by periodic repetition of its graph in any y interval of length g p p.
P. Gundes Bakir,
34

Familiar periodic functions are the cosine and sine functions. Examples of functions that are not periodic are:
x, x 2 , x 3 , e x , cosh x, ln x
If f(x) has period p, it also has the period 2p because the equation
f ( x + p) = f ( x)
implies that
f ( x + 2 p ) = f ([x + p ] + p ) = f ( x + p ) = f ( x) f ( x + np) = f ( x)
thus for any integer n=1,2,3,...
P. Gundes Bakir,
35

Furthermore if f(x) and g(x) have period p, then af(x)+bg(x) with any constants a and b also has the period p.Our problem in the first few slides will be the representation of various functions f(x) of period 2 in terms of the simple functions All these functions f have the period 2. They form f the so called trigonometric system. The figure shows these functions all have period 2 except for the constant 1, which is periodic with any period.
1, cos x, sin x, cos 2 x, sin 2 x, L, cos nx, sin nx
P. Gundes Bakir,
36

The series to be obtained will be a trigonometric series, that is a series of the form:
ao + a1 cos x + b1 sin x + a2 cos 2 x + b2 sin 2 x + L = a0 + (an cos nx + bn sin nx )

n=
Here ao , a1 , b1 , a2 , b2 , L are constants called the coefficients of the series. We see that each term has the period 2. Hence if the coefficients are such that the series converges, g its sum will be a function of p period 2 . Now suppose that f(x) is a given function of period 2 and is such that it can be represented by a series as above which converges and moreover has the sum f(x). ( ) Then using g the equality q y sign, g , we write:
f ( x) = a0 + (an cos nx + bn sin nx )

n=
P. Gundes Bakir,
37

The equation

n =1
is called the Fourier series of f(x). We shall prove that in this case, the coefficicents of the above equation are the so called Fourier coefficients of f(x) given by the Euler formulas.
1 a0 = 2 an = bn = 1
f ( x)dx
n = 1,2, L n = 1,2, L
f ( x) cos nxdx
f ( x) sin nxdx
P. Gundes Bakir,
38

Let f(x) be periodic with period 2 and piecewise continuous in the interval x . Furthermore, let f(x) have a left hand derivative and a right hand d i ti at derivative t each h point i t of f th that ti interval. t l Th Then th the F Fourier i series i of f

n =1
with coefficients
1 a0 = 2 an = bn = 1
f ( x)dx
n = 1,2, L n = 1,2, L
f ( x) cos nxdx
f ( x) sin nxdx
converges. Its sum is f(x) except at points xo where f(x) is discontinuous. There the sum of the series is the average of the left and right limits of f(x) at xo.

The left hand limit of f(x) at xo is defined as the limit of f(x) as x approaches xo from the left and is commonly denoted by f(xo-h). Thus,
f ( xo h) = lim h0 f ( xo h) as h 0 through positive values.

The right hand limit of f(x) at xo is defined as the limit of f(x) as x approaches xo from the right and is commonly denoted by f(xo+h). Thus,
f ( xo + h) = lim li h0 f ( xo + h) as h 0 through th h positive iti values. l
P. Gundes Bakir,
40
Example
Find Fi d the h Fourier F i coefficients ffi i of f the h periodic i di f function i f( f(x) )i in the h fi figure. The formula is:
k if < x < 0 f ( x) = 0< x < if k and f ( x + 2 ) = f ( x)
P. Gundes Bakir,
41
Example
Find Fi d the h Fourier F i coefficients ffi i of f the h periodic i di f function i f( f(x) )i in the h fi figure. The formula is:
k if < x < 0 f ( x) = 0< x < if k and f ( x + 2 ) = f ( x)
1 a0 = 2 1 a0 = 2 =
f ( x)dx
1 ( k ) dx + (k )dx 2 0
0
1 f ( x ) dx = 2
1 1 1 1 0 (kx) + (kx) 0 = k + k = 0 2 2 2 2
The above can also be seen without integration, since the area under d th the curve of f f( f(x) ) between b t - and d is i zero.
Example
From
an = 1
f ( x) cos nxdx
0 1 an = f ( x) cos nxdx = (k ) cos nxdx + (k ) cos nxdx 0
0 1 sin nx sin nx ) + (k ) =0 = ( k n n 0 because sinnx = 0 at - ,0, and for all n = 1,2,....Similarly :
bn =
f ( x) sin nxdx
0 1 bn = f ( x) sin nxdx = ( k ) sin nxdx + (k ) sin nxdx 0
0 1 cos nx cos nx = k k n n 0
Since cos(- ) = cos and cos0 = 1, this yields :

Example
bn = k [cos 0 cos( (n ) cos n + cos 0] n k [2 2 cos n ] = n 2k (1 cos n ) = n Now cos( ( ) = 1 , cos2 = 1, , cos(3 ( ) = 1 etc in g general
- 1 for odd n cosn = 1 for even n f odd dd n 2 for and thus 1 cosn = 0 for even n
P. Gundes Bakir,
44
Example
Hence the Fourier coefficients b n of our function are : 4k 4k 4k , b4 = 0, b5 = ,L , b2 = 0, b3 = 5 3 Since the an are zero, zero the Fourier series of f(x) is : b1 = 4k 1 1 sin x + sin 3 x + sin 5 x + L 3 5 Th partial The ti l sums are : S1 = 4k sin x 4k 1 S2 = sin x + sin 3 x etc. 3
Their Th i graph h seems to t indicate i di t that th t the th series i i is convergent t and dh has the sum f(x), the given function.
P. Gundes Bakir,
45
Example
We notice W i that h at x=0 0 and x=, the points of discontinuity of f(x), all partial sums have the value zero, the arithmetic mean of the limits k and k of our function at these points. points
P. Gundes Bakir,
46
Derivation of the Euler formulas

Th key The k to the h E Euler l f formulas l i is the h orthogonality h li of f
1, cos x, sin x, cos 2 x, sin 2 x,L , cos nx, sin nx

a concept of basic importance as follows: THEOREM 1: The trigonometric system above is orthogonal on the interval x (hence also on 0 x 2 or any other interval of l length th 2 because b of f periodicity); i di it ) th that ti is th the i integral t l of f th the product d t of f any two functions in 1, cos x, sin x, cos 2 x, sin 2 x,L, cos nx, sin nx over that interval is zero, , so that for any y integers g n and m, ,
( a ) cos nx cos mxdx = 0 (n m)
(b) sin i nx sin i mxdx d = 0 ( n m)
(c) sin nx cos mxdx = 0 (n m or
n = m)
P. Gundes Bakir,
47
Fourier series expansion of any period p=T

The functions considered so far had period 2, for the simplicity of the formulas. However, we will mostly use the variable time t and work with functions x(t) with period T. We now show that the transition from 2 to period T is quite simple. The idea is simply to find and use a change of scale that gives from a function f(x) of period 2 to a function of period T. In the equations below we can write the change of scale as: x=kt with n such that the old period x=2 gives for the new variable t the new period t=T. Thus 2 =kT kT hence k=2 k 2 /T and x=kt= x kt 2t/T. This implies dx= dx 2dt/T which upon substitution into
1 a0 = 2 an = bn = 1
f ( x)dx
n = 1,2, L n = 1,2, L
1 a0 = x(t )dt T T/ 2 2 2nt an = x ( t ) cos dt T T/ 2 T

T /2
T /2
f ( x) cos nxdx
n = 1,2, L n = 1,2, L
48
f ( x) sin nxdx
2 2nt bn = x ( t ) sin dt T T/ 2 T
T /2
P. Gundes Bakir,

Since we will mostly use the variable time t and in the frequency domain 2n/T, the equation

n =1
can be written as follows:
2nt 2nt x(t ) = a0 + an cos + bn sin T T n =1

1 a0 = x(t )dx T T/ 2 n = 1,2, L n = 1,2, L 2 2nt an = x ( t ) cos dt T T/ 2 T
T /2 T /2
The coefficients in this case can be written as shown on the rhs rather than the lhs.
1 a0 = 2 an = bn = 1
f ( x)dx
f ( x) cos nxdx
n = 1,2, L n = 1,2, L
49
f ( x) sin nxdx
T /2
P. Gundes Bakir,
Fourier series expansion of the periodic loading

For a function x(t) defined on the interval [-/2, /2], we have the representation on that interval
The coefficients are obtained as follows: Consider the integral
When x(t) is substituted from the first equation, equation this integral breaks down into Im(1), Im(2) and Im(3). The first is:
P. Gundes Bakir,
50
but this is zero as sin(m)=0 for all m. The second integral is:
Assuming we can change the order of integration and summation we obtain
Using the identity

Now if n and m are different integers then n-m and n+m are both nonzero integers and the sine terms in the last expression vanish. If n and m are equal, we have a problem with the second term above. We could use a limit argument but it is simpler to go back to the first equation with n=m.
P. Gundes Bakir,
52

This breaks down to: An orthogonality relation has been proved.
2n 2m cos t cos d =0 t dt
2 2
f for mn
for m = n 0 for m = n = 0
A similar analysis for the third integral gives
P. Gundes Bakir,
53

Derives along the way the orthogonality relation As
We have
I m = am
for m = n 0
I m = a 0
for m = n = 0
for m = n 0
Or in terms of the original expression:
P. Gundes Bakir,
54

If m=n=0 0i in the h b below l equation: i
a m =
x(t ) cos
2
2m t dt
for m = n = 0
Performing the same operations using a multiplier of sin(2m/) gives: via the orthogonality relation:
where mn is called the Kronecker delta and has the properties:

mn =
1 if n = m 0 if n m
55
P. Gundes Bakir,

Fourier Series Expansion in Exponential Form Recall the standard Fourier series in terms of
Now suppose we apply de Moivres Theorem
P. Gundes Bakir,
56

This allows us to write the equation
in the following form: where
P. Gundes Bakir,
57

U i th Using the equations ti
2 2nt an = x ( t ) cos dt T T T/ 2
T /2
n = 1,2, L n = 1,2, L
T /2
in
+T / 2 +T / 2 1 1 2 2nt 2nt x(t ) cos cn = (an ibn ) = dt i x t dt ( ) sin T T 2 2 T T / 2 T / 2
Gives:
1 = T
+T / 2
2nt 2nt x ( t ) cos i sin dt T T T / 2
1 cn = T
P. Gundes Bakir,
+T / 2
T / 2
i 2nt / T x ( t ) e dt
58

Alt Alternatively, ti l using i th the orthogonality th lit equation ti
T /2 int imt e e dt = 0 for n + m 0
T / 2
= T for n + m = 0
and multiplying the equation
by exp(imt), and integrating directly gives:
P. Gundes Bakir,
59
Fourier transform
If we want t to t look l k at t the th spectral t l content t t of f nonperiodic i di signals i l we have to let as all the interval t [-, ] contains important information. Recall the exponential form of the Fourier series
where
C bi i th Combining the above b t two equations ti give: i
and we now have to let

P. Gundes Bakir,
0
60
Fourier Transform
If we suppose pp that the p position of the t axis is adjusted j so that the mean value of x(t) is zero. Then according to the first of the below equation the coefficient ao will be zero.
T /2
1 a0 = x(t )dx T T/ 2
2 2nt an = x ( t ) cos dt n = 1,2, L T T/ 2 T

T /2
2 2nt bn = x ( t ) sin dt n = 1,2, L T T/ 2 T

T /2
The remaining Th i i coefficients ffi i t an and d bn b will ill in i general l all ll b be diff different t and their values may be illustrated graphically as shown.
P. Gundes Bakir,
61
Fourier Transform
Recall that the spacing between the frequency lines is so that the kth spectral line is at From the first equation, we see that The equation becomes
P. Gundes Bakir,
62
Fourier Transform
As 0, A 0 the h n become b closer l and d closer l together h and d the h summation turns into an integral with =d (assuming that x(t) is appropriately well behaved. In the limit
It follows f ll th that t if we d define fi
where F denotes the Fourier transform then the first equation implies that
and this is the inverse Fourier transform transform.

Fourier Transform
Note N t the th f formal l similarity i il it t to th the L Laplace l t transform f i in f fact t we obtain bt i the Fourier transform by letting s= i in the Laplace transform. The main difference between the two is the comparative simplicity of the inverse Fourier transform transform. {x(t),X()} are a Fourier transform pair. As they are uniquely constructable from each other they must both encode the same information but in different domains. X() expresses the frequency content of x(t). It is another form of spectrum. However note that it h t has to be b a continuous ti f function ti of f in i order d t to represent t nonperiodic functions.
P. Gundes Bakir,
64
Discrete Fourier Transform

In reality I li not only l will ill the h signal i lb be of f fi finite i d duration i b but i it will ill b be sampled. It is not possible to store a continuous function on a computer as it would require infinite memory. What one usually does is take measurements from the signal at regular intervals say t seconds apart so the signal for manipulation takes the form of a finite vector of N samples
where t is a reference time. If we take t0=0 from now on we will have
How do we compute the spectrum of such a signal? We need the Discrete Fourier Transform DFT.

Recall R ll the h exponential i lf form of f the h F Fourier i series i the h spectral l coefficients are
In keeping with our notation for the Fourier transform we will relabel cn by Xn from now on. Also the equation above is not in the most convenient form for the analysis so we will modify it slightly. Recall that x(t) is assumed periodic with period . Consider the integral
P. Gundes Bakir,
66

L t t= t+ Let t , the th i integral t lb becomes
but x(t)=x(t-) by periodicity. Also,
e 2in (t ) / = e 2nit /
P. Gundes Bakir,
67

Now, as we have only x(t) sampled at tr=r t, we have to approximate the integral by a rectangular sum,
and as =N t, this becomes,
As we started off with only N independent quantities xr we can only derive N independent i d d t spectral t l lines li at t most. t This Thi means we must t have h relations l ti between the Xn . The simplest one is periodicity. Consider,
P. Gundes Bakir,
68

Therefore, it has been confirmed that we have at most N independent lines. In fact, there must be less than this as the Xn are the complex quantities. Given N real numbers, we can only specify N/2 complex numbers. Looking at the exponent in
if this is to be identified with the exponent int of the Fourier transform, we , must have,
or
P. Gundes Bakir,
69

Alt Alternatively, ti l if we specify if th the f frequency spacing i i in H Hertz t
When n=0, the spectral line is given by: which is the arithmetic mean or DC component of the signal signal. Therefore X0 corresponds to the frequency =0 as we might expect. This means that the highest frequency that we can represent is
N N f 1 f = = = s 2 2 Nt 2t 2
where fs is the sampling frequency. This frequency is very important in signal processing and is called the Nyquist Frequency.
P. Gundes Bakir,
70

This argument Thi t says th that t only l th the first fi t half h lf of f th the spectral t l li lines are independent-so what are the second half? Consider:
1 X N 1 = N
xr e
r =0
N 1
2i ( N 1) r N
1 = N
i 2r + i 2r / N x e r e r =0
N 1
but r is an integer, so e-i2r=1, this means,
1 X N 1 = N
i 2r / N e x r r =0
N 1
1 = N
i 2r / N * x e = X r 1 r =0 N 1
A similar argument shows that generally,

* X N k = X k
P. Gundes Bakir,
71

Recall that it is a property of the Fourier Transform that
X ( )* = X ( )
This means that the spectral coefficient XN-1 corresponds to the frequency - or more generally XN-k corresponds to the frequency -k. So the array Xn stores the frequency representation of the signal xr as follows:
P. Gundes Bakir,
72

What h happens h if there h is i a value l of f k for f which hi h N-k=k? k k? This hi implies i li that h N=2k 2k so the number of sample points is even. Actually this turns out to be the most usual situation, in fact it is a requirement of the Fast Fourier Transform which we shall meet in the next lecture. In this case, we have
* XN /2 = X N /2
so this spectral line is real. This finally justifies our assertion that the maximum frequency represented is N/2.
P. Gundes Bakir,
73

Proof of the inversion theorem Let us proove that
xn = X r e i 2rn / N
r =0
N 1
where
1 Xr = N
x e
p =0 p
N 1
i 2rp / N
If the second equation is substituted in the first equation, we obtain:
xn =
r =0
N 1
1 N
x e
p =0 p
N 1
i 2rp / N
i 2rn / N e
74
P. Gundes Bakir,

If we change h the th order d of f the th summations, ti we obtain: bt i
xn =
p =0
N 1
1 N
i 2rp / N i 2rn / N x e e p r =0 N 1
which can be rewritten as:
xn =
p =0
N 1
1 N
i 2r ( n p ) / N x e p r =0 N 1
which can be reexpressed as follows provided the term in brackets is called q: N 1 N 1
1 xn = p =0 N
x p e i 2 ( n p ) / N
r =0
]
r
=q

Thus, if np p, within the equation q
xn =
p =0
N 1
1 N
i 2 ( n p ) / N x e p r =0
N 1
]
r
=q q
we have a geometric series as follows:
q
r =0
N 1
= 1 + q + q 2 + ...
If we call the left hand side of the above equation sr, we have:
sr = 1 + q + q 2 + ...
If we multiply both sides of the above equation by q, and subtract the resulting equation from the above equation:
qsr = q + .... + q r + q N 1+1 sr qsr = 1 q N


W can obtain We bt i the th value l of f sr from f here h as:
1 qN sr = 1 q
Now qN=1, Now, =1 which can be proved as: qN=ei2(n-p)r=cos2r(n-p)+isin2r(n-p)=1+0=1 which results in the below result for np :
1 qN =0 sr = 1 q

F n=p, the For h term i in b brackets k i in the h equation i b below l b becomes 1 1.
xn =
p =0 N 1
1 N
i 2 ( n p ) / N x e p r =0
N 1
]
r
The summation of all ones r times gives N. Ns cancel each other in the above equation and letting p=n gives:
1 N
x
p =0
N 1
N pn = xn
Thus, , the above equality q y is p proved to be satisfied which consequently q y proves the inversion theorem. This takes us to the final proven Discrete Fourier Transform formulas in the next slide.
P. Gundes Bakir,
78

Di Discrete F Fourier i T Transform f
X n = xr e i 2nr / N
r =0
N 1
Inverse Discrete Fourier Transform
1 xn = N
i 2nr / N X e r r =0
N 1
P. Gundes Bakir,
79
Gibbs Gibb s phenomenon

The first figure shows a time domain signal being synthesized from sinusoids. The signal being reconstructed is shown in the graph on the left hand side. Since this signal is 1024 points long, there will be 513 individual frequencies needed for a complete reconstruction. The figure on the right sho s a reconstr shows reconstructed cted signal using sing freq frequencies encies 0 thro through gh 100 100. This signal was created by taking the DFT of the signal on the left hand side, setting frequencies 101 through 512 to a value of zero, and then using the Inverse DFT to find the resulting time domain signal. The figure in the middle shows a reconstructed t t d signal i l using i f frequencies i 0 th through h 30 30.
P. Gundes Bakir,
80

When only some of the frequencies are used in the reconstruction, reconstruction each edge shows overshoot and ringing (decaying oscillations). This overshoot and ringing is known as the Gibbs effect, after the mathematical physicist Josiah Gibbs, who explained the phenomenon in 1899. The critical factor in resolving this puzzle is that the width of the overshoot becomes smaller as more sinusoids are included. The overshoot is still present with an infinite number of sinusoids, but it has zero width. Exactly at the discontinuity the value of the reconstructed signal converges to the midpoint of the step. As shown by Gibbs, the summation converges g to the signal g in the sense that the error between the two has zero energy.
P. Gundes Bakir,
81

Problems related to the Gibbs effect are frequently encountered in DSP. For example, a low-pass filter is a truncation of the higher frequencies, resulting in overshoot and ringing at the edges in the procedure is to truncate the ends of time domain. Another common p a time domain signal to prevent them from extending into neighboring periods. By duality, this distorts the edges in the frequency domain. These issues will resurface in future chapters on filter design design.
P. Gundes Bakir,
82
Example: An important DFT pair

Figure is an example signal we wish to synthesize synthesize, an impulse at sample zero with an amplitude of 32. Figure b shows the frequency domain representation of this signal. The real part of the frequency domain is a constant value of 32. The imaginary part (not shown) is composed of all zeros. eros As discussed in the next chapter, this is an important DFT pair: an impulse in the time domain corresponds to a constant value in the frequency domain. For now, , the important p p point is that ( (b) ) is the DFT of ( (a), ), and (a) ( ) is the Inverse DFT of (b).
P. Gundes Bakir,
83
Bandwidth
As shown in the figure figure, the bandwidth can be defined by drawing dividing lines between the samples. For instance, sample n mber 5 occurs number occ rs in the band between 4.5 and 5.5; sample number 6 occurs in the band between 5.5 and 6.5, etc. E Expressed d as a f fraction ti of f the th total bandwidth (i.e., N/2), the bandwidth of each sample is 2/N. An exception to this is the samples l on each h end, d which hi h h have one-half of this bandwidth, 1IN. This accounts for the 2/N scaling factor between the sinusoidal amplitudes and frequency domain, as well as the additional factor of two needed for the first and last samples. p
P. Gundes Bakir,
DFT can be calculated by the fast Fourier transform (FFT), which is an ingenious algorithm that decomposes a DFT with N points, i into N DFTs DFT each h with i h a single i l point.
84
Frequency responseresponse-Impulse response

A systems frequency response is the Fourier transform of its impulse response. Keeping p g with standard DSP notation, , impulse p responses p use lower-case variables, while the corresponding frequency responses are upper case. Since h[ ] is the common symbol for the impulse response, H[ ] is used for the frequency response. response That is, is convolution in the time domain corresponds to multiplication in the frequency domain.
P. Gundes Bakir,
85
How much resolution can y you obtain in the frequency response?

The answer is: infinitely high, if you are willing to pad the impulse response with an infinite number of zeros. In other words, there is nothing limiting the frequency resolution except the length of the DFT. Even though the impulse response is a discrete signal, the corresponding frequency response is continuous. An N point DFT of the impulse response provides N/2 + 1 samples of this continuous curve. If you make the DFT longer, the resolution improves, and you obtain a better idea of what the continuous curve looks like. This can be better understood by the discrete time Fourier transform (DTFT). Consider an N sample signal being run through an N point DFT, producing an N/2 + 1 sample p frequency q y domain. DFT considers the time domain signal g to be infinitely f y long g and periodic. That is, the N points are repeated over and over from negative to positive infinity. Now consider what happens when we start to pad the time domain signal with g number of zeros, to obtain a finer and finer sampling p g in the an ever increasing frequency domain.
How much resolution can y you obtain in the frequency response?

Adding zeros makes the period of the time domain longer, longer while simultaneously making the frequency domain samples closer together. Now we will take this to the extreme, by adding an infinite number of zeros to the time domain signal. This produces a different situation in two respects. First, the time domain signal now has an infinitely long period. In other words, it h turned has di into an aperiodic i di signal. i l Second, the frequency domain has achieved an infinitesimally small spacing between samples. samples That is, is it has become a continuous signal. signal This is the DTFT, DTFT the procedure that changes a discrete aperiodic signal in the time domain into a frequency domain that is a continuous curve. In mathematical terms, a system's frequency response is found by taking the DTFT of its impulse response. Since this cannot be done in a computer computer, the DFT is used to calculate a sampling of the true frequency response. This is the difference between what you do in a computer (the DFT) and what you do with mathematical equations (the DTFT).
P. Gundes Bakir,
87
Close peaks
Suppose there are peaks very close together together, such as shown in the figure. There are two factors that limit the frequency resolution that can be obtained-that is,how close the peaks can be without merging into a single entity. The first factor is the length of the DFT. The frequency spectrum prod ced b produced by an N point DFT consists of N/2 + 1 samples eq equally all spaced between zero and one half of the sampling frequency. To separate two closely spaced frequencies, the sample spacing must be smaller than the distance between the two peaks. For example, a 512-point DFT is sufficient t separate to t the th peaks k in i the th figure,while fi hil a 128 128-point i t DFT is i not. t
P. Gundes Bakir,
88
Output of a system
What are you going to do if given an input signal and impulse response, and need to find the resulting output signal? Transform the two signals into the frequency domain, multiply them, and then transform the result back into the time domain.
P. Gundes Bakir,
89
Properties of the Fourier Transform

Homogeneity means that a change in amplitude in one domain produces an identical change i amplitude in lit d in i the th other th domain. d i When Wh the th amplitude lit d of f a time ti domain d i waveform f is i changed, h d the amplitude of the sine and cosine waves making up that waveform must also change by an equal amount. In mathematical form, if x[ ] and X[ ] are a Fourier Transform pair, then kx[ ] and d kX k [ ] are also l a Fourier i Transform f pair, i for f any constant k.
P. Gundes Bakir,
90

Additivity of the Fourier transform means that addition i one d in domain i corresponds d t to addition dditi in i th the other th domain. An example of this is shown in the figure. In this illustration, (a) and (b) are signals in the time domain called ll d x1[n] and d x2[n], ] respectively. ti l Adding Addi th these signals i l produces a third time domain signal called x3[n], shown in (c). ( )
P. Gundes Bakir,
91

In spite p of being g linear, the Fourier transform is not shift invariant. In other words, a shift in the time domain does not correspond to a shift in the frequency domain. Let h(t) ( ) be the impulse p response, p , i.e., , the systems y response p to a Dirac impulse, it can be proved that the response g(t) of the system to an input f(t) is the convolution of f(t) and h(t):
+
g (t ) =

f ( )h(t )d = h(t ) f (t )
The convolution Th l ti th theorem i in F Fourier i analysis l i states t t th that t convolution l ti in i one domain corresponds to multiplication in the other domain. Hence the frequency response function H(f) is the ratio between the response and d th the i input t as a f function ti of f the th f frequency.
g (t ) = h(t ) f (t ) G( f ) = H ( f ) F ( f )

Differentiation Diff i i in i the h time i d domain i i is equivalent i l to multiplication l i li i b by a factor i in the frequency domain.
& (t )] = (i ) X ( ) F [x
Integration in the time domain is equivalent to division by i in the f frequency domain. d i

t X ( ) F x( )d = i
If the input is the harmonic probe eit, the output is eit multiplied by the FRF evaluated at .
y (t ) = H ( )eit
Fourier transform
Measured M d signals i l are time i d domain i f functions. i I It i is i important to investigate the signals in the frequency domain in order to study their frequency content. The Fourier tansform is a tool to transform signals from the time domain to the frequency domain.
The signal can be transformed back to the time domain using the inverse Fourier transform:
P. Gundes Bakir,
94
Fast Fourier Transform

Th FFT is The i a clever l algorithm l ith for f rapidly idl calculating l l ti the th DFT. DFT The N p point DFT of an N p point sequence q xn is g given by: y
nk X k = xnWN where WN = e j 2 / N r =0 N 1
Because xn may be either real or complex, evaluating Xk requires on the order of N complex multiplications and N complex additions for each value l of f k. Therefore, Th f because b there th are N values l of f Xk, computing ti an N point DFT reqires N2 complex multiplications and additions. The basic strategy that is used in the FFT algorithm is one of divide and conquer, which involves decomposing an N point DFT into succesively smaller DFTs.

Suppose pp that the length g of xn is even, , (i.e., ( , N is divisible by y 2). ) If xn is decimated into two sequences of length N/2, computing the N/2 point DFT of each of these sequences requires approximately (N/2)2 multiplications and the same number of additions.Thus, the two DFTs require 2*( 2 (N/2)2=N2/2 multiplications and the same number of additions. Therefore, if it is possible to find the N point DFT of xn from these two N/2 point DFTs, great computational time will be saved because N2/2 operations ti will ill b be required i di instead t d of f N2. Let xn be a sequence of length N=2v, and suppose that xn is split (decimated) into two subsequences subsequences, each of length N/2. /2 The first sequence gn is formed from the even-index terms,
g n = x2 n
N n = 0,1,..., 1 2
96
P. Gundes Bakir,

Th second The d sequence hn is i f formed df from th the odd-index dd i d t terms,
hn = x2 n +1
N n = 0,1, ,..., , 1 2
In terms of these sequences, the N-point DFT of xn is

nk X k = xnWN = n =0 N 1 2 l =0 N 1 nk x W n N + N 1 2 l =0 nk x W n N
n even
n odd
2 lk ( 2 l +1) k = g lWN + hlWN
P. Gundes Bakir,
97

Si Since
WN = e j 2 / N
2 lk lk WN = WN /2
Substituting the above result into the equation:

nk X k = xnWN = n =0 N 1 2 l =0 N 1 nk x W n N + N 1 2 l =0 nk x W n N
n even
n odd
2 lk ( 2 l +1) k = g lWN + hlWN
We obtain:
lk k lk X k = g lWN + W h W /2 N l N /2 l =0 l =0
N 1 2
N 1 2
P. Gundes Bakir,
98

N t th Note that t in i th the equation ti
lk k lk X k = g lWN / 2 + WN hlWN / 2 l =0 l =0 N 1 2 N 1 2
the first term is the N/2 point DFT of gn and the second is the N/2 point DFT of hn:
k X k = Gk + WN Hk
k = 0,1, ,...., , N 1
Although the N/2 point DFTs of gn and hn are sequences of length N/2, the periodicity of the complex exponentials allows us to write:
Gk = Gk + N / 2 Hk = Hk+N /2

Th f Therefore, Xk may be b computed t d from f the th N/2 point i t DFTs DFT Gk and d Hk. If N/2 is even, gn and hn may again be decimated. For example, Gk may be evaluated l t d as follows: f ll
nk Gk = g nWN /2 = n =0 N 1 2 nk g W n N /2 + N 1 2 nk g W n N /2 N 1 2
n even
n odd
As before, this leads to

N 1 4 n =0 N 1 4
nk nk k Gk = g 2 nWN + W g W /4 N / 2 2 n +1 N / 4 n =0
where the first term is the N/4 point DFT of the even samples of gn, and the second is the N/4 point DFT of the odd samples.
P. Gundes Bakir,
100

If N is a power of 2, the decimation may be continued until there are only two-point DFTs. A block diagram showing the computations that are necessary for the first stage of an eight point decimation in time is shown in the figure. g This diagram g is called the FFT butterfly y.
P. Gundes Bakir,
101

Computing C ti an N point i t DFT using i FFT i is much h more effective ff ti th than calculating the DFT directly. For example, if N=2v, there are log2N=v stages of computation. At each of these stages, we are required to carry out N multiplications. The total cost of the algorithm is then Nlog2N. The saving in moving from the DFT to FFT is:
N2 N N = = N log 2 N log 2 N v
Suppose N=1024, we get a saving of computational effort of the order 100:1, and this saving increases with N.
P. Gundes Bakir,
102
Problems associated with analog analogg-to to-digital g conversion

Sampling Aliasing g
Problems associated with ADC

We will W ill now explore l th the i issues of f aliasing li i and d apparent t signal i l distortion associated with choosing a sample rate for digital conversion of analog signals. This process is known as analog to digital conversion (A/D) (A/D). The process of sampling reduces an infinite set to a finite set of data resulting in a loss of information which is responsible for the i issues d detailed t il d i in thi this section. ti Since the original analog signal contains infinite information (knowledge of the signal at any point in time), the frequencies within th signal the i l are also l k known. Information lost in the sampling process is also lost in the frequency representation. We will explain how this leads to the problem of leakage where high frequency components can not be distinguished from low frequency components.
P. Gundes Bakir,
104
Sampling
Nearly N l all ll d data acquisition i i i systems sample l d data with i h uniform if time i intervals. For evenly sampled data, time can be expressed as:
T = ( N 1)t
where N is the sampling index which is the number of equally spaced d samples. l F For most tF Fourier i analyzers l Ni is restricted ti t dt to a power of 2. The sample rate or the sampling frequency is:
1 fs = = ( N 1)f t
P. Gundes Bakir,
105
Sampling
S Sampling li f frequency i is th the reciprocal i l of f th the ti time elapsed l d t from one sample to the next. The unit of the sampling frequency is cycles per second or Hertz ( (Hz), ) if the sampling p gp period is in seconds. The sampling theorem asserts that the uniformly spaced di discrete samples l are a complete l representation i of f the h signal if the bandwidth fmax is less than half the sampling rate. The sufficient condition for exact reconstructability from samples at a uniform sampling rate fs (in samples per unit time) (fs2fmax).
Aliasing
One problem encountered in A/D conversion is that a high frequency signal can be falsely confused as a low frequency signal when sufficient precautions have been avoided. This happens when the sample rate is not fast enough for the signal and one speaks of aliasing. Unfortunately, this problem can not always be resolved by just sampling faster, the signals frequency content must also be limited. Furthermore, the costs involved with postprocessing and data analysis increase with the quantity of data obtained. Data acquisition y have finite memory, y, speed p and data storage g capabilities. p systems Highly oversampling a signal can necessitate shorter sample lengths, longer time on test, more storage medium and increased database management and archiving requirements.
Aliasing
The central concept to avoid aliasing is that the sample rate must be at least twice the highest frequency component of the signal (fs2fmax) ( ). We define the Nyquist yq or cut-off frequency q y
fs 1 = fN = 2 2t
The concept behind the cut-off frequency is often referred to as Shannons sampling criterion. Signal components with frequency content above the cut-off frequency are aliased and can not be distinguished from the frequency components below the cut-off cut off frequency.
P. Gundes Bakir,
108
Aliasing
Conversion of analog frequency into digital frequency during sampling is shown in the figure figure. Continuous signals with a frequency less than one one-half half of the sampling rate are directly converted into the corresponding digital frequency. Above one-half of the sampling rate, aliasing takes place, resulting in the frequency being misrepresented in the digital data. Aliasing always changes a higher frequency into a lower frequency between 0 and 0.5. In addition, aliasing may also change the phase of the signal by 180 degrees.
P. Gundes Bakir,
109
Aliasing
What happens if the original signal actually has a component above the Nyquist frequency?
Now if the spectrum of the continuous signal extends beyond the Nyquist frequency we see overlap
P. Gundes Bakir,
110
Aliasing
If any energy in i the h original i i l signal i l extends d b beyond d the h N Nyquist i frequency, it is folded back into the Nyquist interval in the spectrum of the sampled signal. This folding is called aliasing
fs2fmax
P. Gundes Bakir,
111
Aliasing
Just as aliasing can change the frequency during sampling, it can also change the phase. phase For example, example the aliased digital signal in the figure is inverted from the original analog signal; one is a sine wave while the other is a negative sine wave. In other words, aliasing has changed the frequency and introduced a 180" 180 phase shift. shift Only two phase shifts are possible: 0" (no phase shift) and 180" (inversion).
fs2fmax
P. Gundes Bakir,
112
Leakage
When converting Wh i a signal i lf from the h time i d domain i to the h f frequency domain, the Fast Fourier Transform (FFT) is used. The Fourier Transform is defined by y the equation: q
which requires a signal sample from to . The Fast Fourier Transform however only requires a finite number of samples (which must be a value of 2n where n is an integer. i.e. 2, 4, 8, 16 512, 1024). The FFT is defined as:
P. Gundes Bakir,
113
Leakage
The Fast Fourier Transform is commonly used because it requires much less processing power than the Fourier Transform. Like all shortcuts, there are some compromises involved in the FFT. The signal must be periodic in the sample window or leakage will occur. The signal must start and end at the same point in its cycle. Leakage is the smearing of energy from the true frequency of the signal into adjacent frequencies. Leakage also causes the amplitude representation of the signal to be less than the true amplitude of the signal.
P. Gundes Bakir,
114
Leakage
An example p of a nonperiodic p signal g can be seen in the Figure. g
P. Gundes Bakir,
115
Leakage
By comparing B i the h Fi Figures, i it can be seen that the frequency content of the signal is smeared into adjacent frequencies when the signal is not periodic. p In addition to smearing, the amplitude lit d representation t ti of f the th signal is less than the true value.
P. Gundes Bakir,
116
Leakage
The only solution to the leakage problem, is to make sure that the signal is periodic or completely observed within the observation window. Generally, this is very difficult to achieve. For systems with a perfect linear behaviour, it can be accomplished by exciting the structure periodic signal g . Excitation signals g as burst random also with a p minimize this problem. Decreasing g the frequency q y step p f increases the observation time T and hence will improve the periodicity of the signal. The use of a time window other than a rectangular g one offers an approximate solution to the leakage problem.
P. Gundes Bakir,
117
Windowing
In signal processing processing, a window function is a function that is zero zerovalued outside of some chosen interval. Applications of window functions include spectral analysis and filter design. The first type of window is called the rectangular window; it does not weight the signal in any way and is equivalent to saying that no window was used. Rectangular window is used whenever frequency resolution is of high importance. This window can have up to 36% amplitude error if g is not p periodic in the sample p interval. It is g good for signals g the signal that inherently satisfy the periodicity requirement of the FFT process.
P. Gundes Bakir,
118
Windowing
A rectangular t l window i d i af is function ti th that t is i constant t t inside i id th the interval and zero elsewhere, which describes the shape of its graphical representation. When another function or a signal (data) is multiplied by a window function function, the product is also zero zero-valued valued outside the interval: all that is left is the "view" through the window. It can be shown that there is no window with a narrower main lobe than the rectangular window window.
P. Gundes Bakir,
119
Windowing
Windows work by weighting the start and end of a sample to zero while at the same time g the amplitude p of the increasing signal at the center as to maintain the average amplitude of the signal. The effect of a Hanning window on a non-periodic signal in the Frequency Domain can be seen in the Figure. Figure shows that the window reduces smearing and better preserves the amplitude of the g signal.
Windowing
Th The effect ff t of f the th same Hanning Window on the time domain signal g can be seen in the Figure. Figure shows how the Hanning window weights the beginning and end of the sample to zero so that it is more periodic i di during d i th the FFT process.
Windowing
The Flat Top window is used whenever signal amplitude is of very high importance. The flat top window preserves the amplitude of a signal very well; however it has poor frequency resolution so that the exact frequency q y content may y be hard to determine, , this is particularly an issue if several different frequency signals exist in close proximity to each other. The flat top window will have at most 0 1% amplitude error 0.1% error.
P. Gundes Bakir,
122
Windowing
The Hanning Th H i window i d i a compromise is i b between the h Fl Flat T Top and d Rectangular windows. It helps to maintain the amplitude of a signal while at the same time maintaining frequency resolution. This window can have up to a 16% amplitude error if the signal is not periodic.
P. Gundes Bakir,
123
Windowing
The most common window Th i d used d for f random d excitations i i exerted d by shakers is the Hanning window.
P. Gundes Bakir,
124
Windowing
P. Gundes Bakir,
125
Windowing
The exponential Th ti l window i d i is used to make a measurement from a vibrating structure more accurate. accurate It is used when the ringing of a structure does not attenuate adequately during the sample interval. An example of the p window can be exponential seen in the figure.
P. Gundes Bakir,
126
Windowing
The exponential window can cause some problems if not used properly. l As an example example, a very simple lightly damped structure was subjected to an impact test test. The signal processing parameters were selected for a 400 Hz bandwidth which resulted in a 1 sec time window.
Exponential Window
O On the th right i ht more damping is applied and the peaks are much wider now! If an excessive amount of damping is needed to minimize the effects of leakage then you run leakage, the risk of missing closely y spaced p modes.
Exponential Window
Before any window is applied applied, it is advisable to try alternative approaches to minimize the leakage in the measurement such as: Increasing the number of spectral lines Halving the bandwidth which both result in increased total time for measurement.
1 fs = = ( N 1)f t
Windowing
Impact testing always causes some type of transient response that is the summation of damped exponential sine waves. If the entire transient event can be captured such that the FFT requirements can be met, leakage will not be a problem. But for lightly damped structures, structures in many impact testing situations situations, the use of an exponential window is necessary. However, the use of exponential window can cause some difficulties However when evaluating structures with light damping and closely spaced modes. The use of windows may also hide or distort the modes in the measurement.
P. Gundes Bakir,
130
Windowing
The effects of leakage can only be minimized through the use of a window. It can never be eliminated! All windows distort data! Almost all the time when performing a modal test, the input excitation can be selected such that the use of windows can be eliminated. e.g.,signals such as pseudo random, burst random, sine chirp, and digital stepped sine.
P. Gundes Bakir,
131
Spectra and FRFs

Definition Pads d and df filters l
Averaging
Suppose that now we want to estimate the spectrum of a random signal. In the limit as , we would get an accurate spectrum but for finite , we have a problem. Any finite realisation of a random process will not represent p p exactly y the long g term frequency q y content precisely because it is random. Assuming no problems with aliasing we will find
where X is the true spectrum and is an error term associated with the finite sample size. Now for each spectral line is a random variable and is just as likely to cause an underestimate as an overestimate. This means we can remove it by averaging
P. Gundes Bakir,
133
Averaging
The averaging Th i can be b implemented i l t d by b taking t ki N segments t of f ti time data xi(t) and transforming to
then
For a signal
The frequency of the sine wave is chosen such that it is periodic over the window, so we dont have to worry about leakage from the sine wave.
P. Gundes Bakir,
134
Averaging
The first Th fi i is one average- a one-shot h measurement. Al Although h h the h sine wave (at 10.24 Hz) is visible, there is a lot of background noise from the single average.
P. Gundes Bakir,
135
Averaging
Th next figure The fi shows h the h result l of f taking ki 10 averages.
P. Gundes Bakir,
136
Averaging
Fi ll we see the Finally, h effect ff of f taking ki 100 averages.
P. Gundes Bakir,
137
Power spectral density

Often we plot Oft l t E[|X|] 2 as this thi is i proportional ti l to t power. It is i called ll d the th power spectral density and is denoted
The autocorrelation of a random process is defined by:
It is a measure of how much a signal looks like itself when shifted by an amount . It is used to find regularities in data. Suppose that x(t)=sin(2t/ ), then there will be regular peaks in xx () when =n. So the autocorrelation function can also be used to detect periodicities.
Power spectral density

If x(t) is zero mean, then
and if x is not zero mean, xx (0) is the mean square of the process. As x(t) is stationary, we can change the origin of t to t- without changing the autocorrelation, i.e.
So xx () is an even function of
Filters
Definition Pads d and df filters l
Filters
Assume that we are trying to build a Fourier transforming device which can give us the spectrum of a given time signal. Suppose that we have a maximum sampling frequency of 1000 Hz i.e. a Nyquist q y of 500 Hz. frequency If the time signal has a broadband spectrum which is flat up to 750 , what will the estimated spectrum p look like? So energy gy is aliased Hz, into the range 250-500 Hz from the range 500-750 Hz and we obtain a completely fictitious spectrum. How can we help this?Suppose we had a device which removed the part of the signal at frequencies between 500 and 750 Hz. Then we would have changed the signal admittedly but the FFT would at least give us an accurate spectrum all the way up to 500 Hz Hz.
P. Gundes Bakir,
141
Filters
S Such had device i which hi h passes parts t of f a signals i l f frequency content and suppresses others is called a filter. The particular filter described above is called an antialiasing p g filter for obvious reasons.
P. Gundes Bakir,
142
Filters
A filter fil i af is function i that h i in the h f frequency d domain i h has a value l close l to 1 in the range of frequencies that the analyst wishes to retain and close to zero in the range of frequencies that the analyst wishes to eliminate. The filter can be applied in the time domain domain, by convolution of its transform with the time history, or in the frequency domain by multiplying the filter frequency response function with the F Fourier i amplitude lit d spectrum t (FAS) of f the th time ti history hi t , and d th then obtaining the filtered time history through the inverse Fourier transform.
y (t ) = IDFT [H ( f ) X ( f )]
Filters
Equally unimportant is the choice of the actual generic filter: users are faced with a wide range of filters to choose from, including Ormsby, elliptical, Butterworth, Chebychev and Bessel. The correct application of the chosen filter is much more important than the choice of a particular filter. The terminology used to describe filters can be confusing, especially for engineers more accustomed to thinking in terms of periods than q frequencies. A filter that removes high frequencies (short periods) is usually p filter because motion at lower frequencies q referred to as a low-pass gets through and higher frequencies are, in effect, blocked by the filter. For such a filter civil engineers prefer the term high-cut, which refers directly to the frequencies being removed.
Classification of filters
If it is judged that there is significant high-frequency noise in the record, or if for some other reason it is desirable to reduce or remove high frequencies introduced by interaction effects at the recording station, this can be easily achieved by the application of filters. Filters can be applied in the frequency domain or the time domain but their function is best understood in the frequency domain. If the filter is a mechanical or electrical device which operates on the continuous time physical signal it is called an analogue filter. If the filter is a numerical algorithm or mechanical device which operates on sampled data it is called a digital filter.
High passpass-low pass filters

The purpose of Th f a low l pass filter filt is i to t remove that th t part t of f the th signal i l that is judged to be heavily contaminated by high-frequency noise that is often observed in strong-motion records. A low-pass filter passes low-frequency components in the signal with minimal distortion and attenuates the high-frequency components. The so-called so called cutoff or corner frequency divides the pass band and the stop band. A low pass causal Butterworth filter is an all pole filter with a squared magnitude response given by:
P. Gundes Bakir,
146

T Two considerations id ti are i important t t when h applying l i a hi highcut h t filt filter. The first is that the application of the filter will act in a contradictory manner to any instrument correction and at least in some frequency ranges the two will counteract each other. The second consideration is that an upper frequency limit on the usable range of high frequencies in the motion is imposed by the sampling rate: the Nyquist frequency, which is the highest frequency at which characteristics of the motion can be correctly determined, is equal to (1/2 equa ( / t) where e e t is s the t e sampling sa p g interval. te a A high-cut filter applied at frequencies greater than the Nyquist will h have no effect ff t on the th record. d
Analogy between mechanical and electrical systems y

Electrical system Input Output Voltage, e(t) () Charge q(t) Current i(t)=dq/dt Constant parameters Inductance, L Resistance, R Capacitance,C Mechanical system Force, F(t) () Displacement y(t) Velocity v(t)=dy/dt Mass,m Damping,c Compliance, 1/k
P. Gundes Bakir,
148
Analogue filters
We will W ill start t t th the di discussion i with ith an electrical l t i l analogue l filt filter. C Consider id th the circuit below with an alternating voltage input,
Vi (t ) = Vi cos(t )
P. Gundes Bakir,
149
Analogue filters
Elementary El t circuit i it th theory gives i th the output t t voltage lt Vo(t) as th the solution l ti of f th the differential equation:
RC
dVo + Vo = Vi (t ) dt
where R is the resistance and C is the capacitance. Passing to the frequency domain gives:
iRCVo ( ) + Vo ( ) = Vi ( )
Vo ( ) = H ( )Vi ( ) where 1 H ( ) = 1 + iRC
S So
P. Gundes Bakir,
150
Analogue filters
The gain of the system is:
1 + R 2C 2 2 and the phase is : H ( ) = tan 1 ( RC )

When RC=0.1, , |H( | ()| )|=0.995 When RC=10, |H()|=0.0995 As we have a filter that attenuates high frequencies and passes low frequencies, it is called a low-pass filter.
H ( ) =
P. Gundes Bakir,
151
Digital low pass filters

How can we implement a filter of the sort described above on sampled data? First let us adopt a more general notation. Let x(t) or xi be the input to the filter and y(t) or yi be the output. The differential equation of the electrical analogue filter is then,
RC
dy y + y = x(t ) dt
Now suppose that x and y are sampled with an interval t, so x(t) xi=x(ti)=x(i t) and y(t) yi=y(ti)=y(i t). t) The derivative above can be approximated by:
dy yi yi yi 1 = dt t

Thi results This lt i in a diff difference equation ti
RC ( yi yi 1 ) + yi = xi t
With a bit of arrangement:
RC 1 xi yi 1 + yi = yi = RC RC +1 +1 t t yi = a1 yi 1 + b0 xi with appropriate definitions for a1 and b0. Consider the signal
x(t ) = sin( i (2 .5t ) + sin( i (2 .50t )
sampled with t=0.003 as shown in the next slide.
P. Gundes Bakir,
153

N i sine Noisy i wave
P. Gundes Bakir,
154

After one pass through the digital filter
yi = 0.6655 yi 1 + 0.3345 xi
The resulting noisy sine wave after one pass through the filter is:
P. Gundes Bakir,
155

The resulting noisy sine wave after five passes through the digital filter is:
P. Gundes Bakir,
156

So the S h fil filter i is able bl to remove the h hi high hf frequency sine-wave i effectively. However, note that the amplitude of the carrier signal has also been attenuated. Also, importantly, the phase of the signal has been altered. The next question we should ask is is-what what is the frequency response of the new digital filter derived from the original analogue filter? We use a theorem from an earlier lecture which says that if the input to a system is x(t)=eiwt, then the response is y(t)=H(w) eiwt. Now define the backward shift operator Z-1 by:
Z 1 yi = yi 1

Thi allows This ll us to write i the h equation i
yi = a1 yi 1 + b0 xi
As: Let Th Then
(1 a1Z 1 ) yi = b0 xi
xi = eiti
yi = H ( )e iti
Z e
P. Gundes Bakir,
1 iti
=e
iti 1
=e
i ( ti t )
=e
it iti
158

Th Then becomes:
(1 a1Z 1 ) yi = b0 xi
(1 a1e
it
) yi = b0 xi
It follows that the FRF is:
In terms I t of f the th FRF derived d i d above, b we have h enough h now t to obtain bt i a general result. A general digital filter would then be:
b0 H ( ) = 1 a1e it
ny
yi = a j yi j + b j xi j
j =1 j =0
nx
A high pass filter

Recall the formula for the low pass filter:
1 H ( ) = 1 + iRC
The desired properties for a low pass filter are that:
H ( ) 1 as 0 H ( ) 0 as
The desired properties for a high pass filter would be:
H ( ) 0 as 0 H ( ) 1 as
A high pass filter

Now we can obtain this characteristic by making a simple t transformation f ti on the th below b l equation: ti
H ( ) = 1 1 + iRC
1 1 or i The FRF of the low pass filter becomes: i
Namely:
i 1 H ( ) = = RC RC i +1 1+ i RC
and the gain is H ( ) = RC 1+
2
R 2C 2
P. Gundes Bakir,
161
A high pass filter

Thi filter This filt is i high hi h pass as required. i d
When When
RC
= 0.1, = 10,
H ( ) = 0.0995 H ( ) = 0.995
RC
One of the most useful families of analog filters is that of Butterworth filters. These are controlled by two parameters for the low-pass filter. The FRF gain is specified as:
f 1+ 1+ c fc where c is the cut-off frequency and n is a steepness factor which specifies p how fast the signal g should die away y after the cut off frequency. q y
H ( ) =
2
1
2n
1
2n
where n is the order of the filter (number of poles in the system function) and c is the cut off frequency of the Butterworth filter which is the frequency where the magnitude of the causal filter |H(c )| is 1/2 regardless dl of f the th order d of f the th filter. filt The purpose of a low-cut filter is to remove that part of the signal that is j dged to be hea judged heavily il contaminated by b long-period long period noise noise. The key ke issue iss e is selecting the period beyond which the signal-to-noise ratio is unacceptably low. Applying a filter that abruptly cuts out all motion at periods above the desired cut-off can lead to severe distortion in the waveform, , and therefore a transitionsometimes referred to as a ramp or a rolloff is needed between the pass-band, where the filter function equals unity, and the period beyond which the filter function is equal to zero.

The figure shows an illustration of a low-cut Butterworth filter as a f function ti of f frequency f and d period. i d Th The filter filt f frequency i is 0 0.05 05 H Hz, which hi h means that periods above 20 s are at least partially removed. The different curves are for different orders of filter: the higher the order, the more abrupt the cut-off cut off (the more rapid the roll-off roll off (but with increased filter-response oscillations for the higher order filters)). For the lower order filters, information will be removed from periods as low as 10 s.
P. Gundes Bakir,
164
CausalCausal -acausal filters

Although the choice of filter type is less important important, the way in which the filter is applied to the accelerogram has been shown to be very important. The fundamental choice is between causal and acausal filters, the distinguishing feature of the latter being that they do not produce yp phase distortion in the signal, g , whereas causal filters do result any in phase shifts in the record. The zero p phase shift is achieved in the time domain by y passing p g the transform of the filter along the record from start to finish and then reversing the order and passing the filter from the end of the record to the beginning.
P. Gundes Bakir,
165
CausalCausal -acausal filters

The imlementation Th i l i of f a zero phase h filtering fil i scheme h i is shown h i in the h figure.
P. Gundes Bakir,
166

|H(c )| takes t k the th value l of f0 0.5 5f for acausal l filt filters. Th The mathematical th ti l reasoning behind this can be explained as follows: If we let Hc() be the frequency response of the causal Butterworth filter given by the second equation, this filter has unity gain at frequency = 0. The cutoff frequency c is the frequency at which the power of the filter ouput is half the power of the filter input,i.e. |Hc(c )|2 = 1/2. The frequency response of an acausal Butterworth filter Ha() is given by:
The previous equation shows that |H | a(c )|=1/2 )| 1/2 . The frequency response of a Butterworth filter decreases monotonically with increasing frequency, and as the filter order increases, the transition band becomes narrower.
P. Gundes Bakir,
168
P. Gundes Bakir,
169
P. Gundes Bakir,
170
P. Gundes Bakir,
171
Overview of excitation signals
Classification
Th The commonly l used d excitation it ti signals i l can be b categorized in several ways. For practical purposes, it is easy y to consider two main groups: g p broad band or single g frequency signals. The signal frequency group contains:
Swept sine Stepped sine
The broadband group consist of three subgroups:

Transients Periodic Nonperiodic
Classification
Transients T i t
Burst random Burst chirp (or burst swept sine) mpact excitation
Periodic
Pseudo random Periodic random Chirp (fast swept sine)
Nonperiodic
Pure random
P. Gundes Bakir,
174
Classification
R Random d signals i l can only l b be d defined fi d b by th their i statistical properties. For stationary random signals, these properties do not vary with respect to translations in time. All random excitation signals are of the ergodic random type, which means that a time average on any particular subset of the signal is the same for any arbitrary subset of the random signal.
Pure random
Pure random is a nonperiodic stochastic signal with a Gaussian probability distribution. Averaging is essential when estimating the frequency spectrum. problem of the pure p random signal g is leakage g . Since the signal g is The main p not periodic within the observation time window, this error can not be avoided. The application of dedicated time windows (e.g. Hanning) to the input and output signals can not completely remove the effects of leakage without causing undesired side effects such as a decreased frequency resolution.
P. Gundes Bakir,
176
Pure random
Pure random easily averages out noncoherent noise. It yields the best linear approximation of nonlinear systems, since in each h averaged d time i record, d the h nonlinear li di distortions i will ill b be diff different and tend to cancel with sufficient averaging. Test time is relatively long due to the necessary number of averages. However, the total time becomes shorter when using overlap averaging g g. In the overlap p averaging g gp procedure, , each averaged g time record will contain the last part of the previous one.
P. Gundes Bakir,
177
Pseudo random
The pseudo random is an ergodic stationary signal with a spectrum consisting of integer multiples of the discrete Fourier transform frequency increment. Hence it is perfectly periodic within the sample time window window. Due to the periodicity of the signal, no leakage problem exists. However, since the same time block is repeated for averaging, pseudo random excites the nonlinearities the same way in each average. Therefore, averaging will not remove distortion caused by nonlinearities. For linear structures, only a few averages are necessary in general. Hence this excitation signal may be very fast fast.
Periodic random
Periodic random excitation is simply a different use of a pseudo random signal, so that non-linearities can be removed with spectrum averaging. For periodic random testing, a new pseudo random sequence is generated for each new spectrum average. The advantage of this is that when multiple spectrum averages of different random signals are averaged together, randomly excited non-linearities are removed. Although periodic random excitation overcomes the disadvantage of pseudo random excitation, it takes at least three times longer to make the same measurement measurement. This extra time is required between spectrum averages to allow the structure to reach a new steadystate response to the new random excitation signal.
P. Gundes Bakir,
179
Periodic random
Other advantages are:
Signals are periodic in the sampling window, so measurements are leakage free free. Removes non-linear behavior when used with spectrum averaging averaging.
Disadvantages are:
Slower than other random test methods. Special software required for implementation
Burst random
Burst random excitation is similar to periodic random testing,but faster. In burst random testing, a true random signal can be used, but it is turned off prior to the end of the sampling window time period period. This is done in order to allow the structural response to decay within i hi the h sampling li window. i d Thi This insures that both the excitation and response signals are p y contained within the completely sampling window. Hence, they are periodic in the sampling window window.
Burst random
Advantages : Signals are periodic in the sampling li window, i d so measurements are leakage free free. Removes non-linear behavior be a o when e used with spectrum averaging. Fast measurement time. Disadvantages : p software required q Special for implementation.
Burst random
The most Th t commonly l used d excitation it ti f for modal d lt testing! ti ! In order to have the entire transient be captured, the length of the excitation burst can be reduced. Generally, the use of windows for this type of excitation technique is not required!
P. Gundes Bakir,
183
Chirp and Burst Chirp

A swept sine excitation signal can also be synthesized in an FFT analyzer to coincide with the parameters pa a e e s of o the e sa sampling p g window, in a manner similar to the way a pseudo random signal is y synthesized. Since the sine waves must sweep from the lowest to the highest frequency in the spectrum, over the relatively short sampling window time period (T), this fast sine sweep often makes the test equipment sound like a bird chirping, hence p signal. g the name chirp

A burst chirp signal is the same as a chirp chirp, except that it is turned off prior to the end of the sampling window, window just like burst random random. This is done to ensure that the measured signals are periodic in the window.
P. Gundes Bakir,
185

The advantage of burst chirp over chirp is that the structure has returned to rest before the next average of data is taken. This insures that the measured response is only caused by the measured excitation, an important requirement for any multichannel measurement such as a FRF.
P. Gundes Bakir,
186

Advantages of Burst Chirp Excitation g signalto-noise g and RMS-to-peak p ratios. High Signals are periodic in the sampling window, so measurements are leakage free. Fast measurement time. Disadvantages of Burst Chirp Excitation Special software required for implementation. Doesnt t remove non-linear non linear behavior. Doesn
Swept sine
Th The sine i wave excitation it ti signal i lh has b been used d since i th the early days of structural dynamic measurement. It was the only y signal g that could be effectively y used with traditional analog instrumentation. Even broad band testing methods (like impact testing), have been developed for use with FFT analyzers, sine wave excitation is still useful in some applications applications. The primary purpose for using a sine wave excitation signal is to put energy into a structure at a specific frequency. Slowly sweeping sine wave excitation is also useful for characterizing non non-linearities linearities in structures structures.
Swept sine
Advantages Ad t of f Si Sine Testing T ti Best signal-to-noise and RMS-to-peak ratios of any signal. i l Controlled amplitude and bandwidth. Useful for characterizing non-linearities. Long history of use. Disadvantages of Sine Testing Distortion due to over-excitation. Extremely slow for broad band measurements measurements.
Stepped sine
Stepped St d sine i excitation it ti is i a modern d version i of f the th swept t sine i technique that makes maximum use of the developments in DSP during the last two decades. I t d of Instead f a continuously ti l varying i f frequency, stepped t d sine i consists i t of a stepwise changing frequency. It remains a rather slow procedure due to the frequency scan and wait periods needed for the transients to decay. This can be overcome by multi-channel acquisition. The application pp of stepped pp sine excitation requires q special p soft and hardware. The digital processing allows for varying frequency spacing, yielding data condensation and testing g time reduction, , and for a better control against aliasing and leakage problems. Useful for characterizing non-linearities. Excellent signal-to-noise ratios ratios.
Processing strong-motion records

The focus of this section is on the effects of noise in accelerograms, g , and the effects of correction procedures, on the peak ground-motion amplitudes and on the ordinates of acceleration and displacement response spectra.
Processing of earthquake records

High-pass filters are an effective way of removing the low-frequency noise i th that t is i present ti in many, if not t most, t analog l and d di digital it l strongt motion recordings. This low Thi l frequency f noise i usually ll appears as drifts d ift in i the th displacements derived from double integration of acceleration, making it difficult to determine the true peak displacement of the ground motion motion. It can never be claimed that a complete and accurate description of the ground shaking can be obtained from accelerograms accelerograms.
P. Gundes Bakir,
192

For engineering uses of strong-motion strong motion data it is important to be able to estimate the level of noise present in each accelerogram and the degree to which this may affect different parameters that are derived from the records. The main parameters of interest for engineering application are:
The ordinates of response spectra, both of acceleration and displacement. The peak ground acceleration (PGA),although of limited significance from both geophysical and engineering perspectives, is also a widely used parameter in engineering. The peaks of velocity (PGV),and displacement (PGD), measured from the time-histories obtained by integration of the acceleration, are also important parameters.

It is important for users of strong-motion g data to appreciate that digitized accelerograms are never pure and complete reproductions of the seismic signal. From the outset, however, it is important to be clear that it is not possible to identify, separate and remove the noise in order to recover the unadulterated seismic signal. The best that can be achieved in general is to identify those portions of the frequency content of the record where the signal-to-noise signal to noise ratio is unacceptably low and to thus identify that portion of the record, in the frequency domain, that can be used with some confidence confidence.
Analog accelerograms
In light of these considerations considerations, it is not appropriate to refer to most of the processing procedures described herein as corrections, since the term implies that the real motion is known and furthermore that it can be recovered by applying the procedures. In order to estimate the signal-to-noise ratio, a model of the noise in the digitized record is required. Most analog accelerographs, such as the SMA1 produce two fixed traces on the film together with the three traces of 1, motion (two horizontal, one vertical) and the time marks. If these fixed traces are digitized together with the motion, then any signal they contain can be interpreted as being composed entirely of noise since the traces are produced by infinitely stiff transducers that experience no vibration during the operation of the instrument. y the fixed traces are very y often not digitized g or else the Unfortunately, digitized fixed traces are not kept and distributed with the motion data, hence it is rare that a model of the noise can be obtained from this information.
P. Gundes Bakir,
195
Standard vs nonstandard noise

Some types of noise, noise particularly step changes in the baseline, can also be identified from the jerk, which is the first derivative of the acceleration trace trace. Figure shows horizontal component of the Bajestan recordings of the 1978 Tabas earthquake in Iran; spurious spikes are obvious in the acceleration record at 10 10.8 8 and 16 s. The derivative of the acceleration trace (to produce the quantity called jerk) will convert a spike into a double sided pulse pulse, making it easier to identify spikes. By doing this (bottom panel), spikes at 12.3, 26 and 33.2 s are also identified identified.

In this particular case, the spurious i nature t of f these th spikes ik was confirmed by comparison with a reproduction of the original analog record. The origin of the spikes has not been ascertained, although a possible cause in this instance was the misplacement of the decimal point in transcribing the digitized values. Once the spikes have been identified as erroneous, they should be removed from the digitized record; one way to achieve this is replace the acceleration ordinate of the spike with the mean of the accelerations of the data points either side.

The spectra in the Figure were obtained with the record shown in the figure in the last slide and after the spikes were removed. The spikes clearly constituted a serious noise contamination at short periods but it is also noted that their elimination appears to have led to slight modifications in the spectrum at long periods (spikes are broadband and have energy content at long as well as short periods). If the misplacement of decimal points i i is id identified ifi d as the h cause of the errors, then an exact correction could be made.
Instrument correction
As noted A t d earlier, li th the t transducer d f frequency i in analog l i instruments t t i is limited to about 25 Hz, and this results in distortions of amplitudes and phases of the components of ground motion at frequencies close to or greater than that of the transducer transducer. The digitization process itself can also introduce high-frequency noise as a result of the random error in the identification of the exact mid-point id i t of f the th film fil t trace as shown h i in th the fi figure.
P. Gundes Bakir,
199
Instrument correction
Fourier acceleration spectrum of an analog recording at a site underlain by thick sediments is shown in the figure. Natural processes along the propagation path have removed energy at frequencies much below those affected by the instrument response (see dashed line; the instrument response has been shifted vertically erticall so as not to be obsc obscured red b by the data) data), leading to the decreasing spectral amplitudes with increasing frequency up to about 26 Hz (coincidentally the same as the instrument frequency), at which point noise produces an increase in spectral amplitudes. Instrument correction ti only l exacerbates b t th the contamination t i ti of f th the signal i lb by hi high h frequency noise.
P. Gundes Bakir,
200
Fourier Spectra
The left left-hand hand plot in the figure shows an example of the Fourier spectra of high-frequency ground motion obtained at a very hard rock site in Canada at a distance of 4 km from the source of a small magnitude earthquake. Softer sites, even those classified as rock such as class B in the 2003 NEHRP guidelines, will tend to filter out such high frequency motion. Very high-frequency motions will also tend to attenuate rapidly with distance and hence will not be observed at stations even a few tens of kil kilometers t from f the th fault f lt rupture. t
P. Gundes Bakir,
201
Fourier Spectra
The figure shows the Fourier acceleration spectra of earthquakes recorded in eastern and western North America (left and right graphs, respectively). The eastern North America recording has much higher frequency content than that from western North America, even without instrument correction. The record from Miramichi was as recorded on an analog instr instrument, ment whereas hereas those from the Big Bear City earthquake were recorded on digital instruments (the response curves of the instruments are shown by the dashed lines and have been shifted vertically so as not to be obscured by th d the data). t )
P. Gundes Bakir,
202
Fourier Spectra
The plot Th l t in i the th figure fi also l shows h the th t typical i lt transducer d response f for th the instrument (SMA-1) on which the record was obtained, and the effect of applying a correction for the instrument characteristics, which is to increase slightl slightly the amplitudes amplit des at frequencies freq encies greater than 30 Hz H . The nature of such motions, at periods of less than 0.03 s, will only be relevant to particular engineering problems, such as the response of plant machinery and nonstructural components components.
P. Gundes Bakir,
203
Fourier Spectra
The right-hand right hand plot in the figure show the Fourier spectra of more typical ground motions obtained at soil sites during a moderate magnitude earthquake in California. These records were obtained g instruments and are lacking g in very y high g frequency q y motion on digital mainly because of the attenuating effect of the surface geology at these sites compared to the very hard site in Canada. The plot also shows the transducer response for these digital instruments, which is almost flat to beyond 40 Hz Hz.
P. Gundes Bakir,
204
Corrections for transducer characteristics

For digital recordings, instrument corrections should not be necessary. For analog recordings, if the engineering application is concerned with motions at frequencies above 20 Hz and the site y stiff for appreciable pp amplitudes p at characteristics are sufficiently such frequencies to be expected, a correction should be considered. However, , it should be borne in mind that the instrument corrections essentially amplify the high-frequency motions; if the digitization process has introduced high-frequency noise into the record, then the instrument correction will amplify this noise. Unless there are compelling reasons for applying a correction for the instrument characteristics, we recommend that no attempt should be made to do so. so The one exception to this may be the very earliest recordings obtained in the US with accelerographs that had natural frequencies of the order of 10 Hz.
P. Gundes Bakir,
205
Baseline adjustments
A major problem encountered with both analog and digital accelerograms are distortions and shifts of the reference baseline, which result in unphysical velocities and displacements. One approach to compensating for these problems is to use baseline adjustments, whereby one or more baselines, which may be g lines or low-order straight polynomials, are subtracted from the acceleration trace.
P. Gundes Bakir,
206
The figure illustrates the application of a piece-wise sequential fitting of baselines to the velocity trace from a digital recording in which there are clearly identifiable offsets in the baseline.
A similar procedure could be applied directly to the acceleration time-history t correct to t for f the th type t of f baseline shifts shown in the figure. The figure shows NS component of the 21 May 1979 Italian earthquake (12:36:41 UTC) recorded at Nocera Umbra Umbra, showing shifts in the baseline at 5.6 and 8.3 s.
The procedure applied in the fi figure is i to t identify id tif (b (by bl blowing i up the image) sections of the velocity that appear to have a straight baseline, and then fitting a straight li t line to thi this i interval. t l This line in effect is then subtracted from the velocity trace trace, but in practice it is necessary to apply the adjustment to the accelerations. The adjustment to the acceleration is a simple shift equal to the gradient (i.e. g ( the derivative) ) of the baseline on the velocity; this shift is applied at a time tv0, which is the time at which the line fit to the velocity crosses the zero axis.
The adjusted velocity trace is then inspected to identify the next straight line segment, y which is fit in the same way. In the particular case illustrated g , a total of four line in the figure, segments were required to remove the most severe distortions of the baseline visible in uppermost plot, plot although the baseline instabilities are not entirely , as evident in the removed, residual long-period trends.
P. Gundes Bakir,
210
Th The distortion di t ti of f the th baseline b li encountered t di in di digitized iti d analog accelerograms is generally interpreted as being the result of long-period gp noise combined with the signal. g Baselines can be used as a tool to remove at least part p of this noiseand probably some of the signal with it as a means of recovering more physically plausible velocities and displacements displacements. There are many procedures that can be applied to fit the baselines, including polynomials of different orders. A point that is worth th making ki clearly l l i is th that, t i in effect, ff t baseline b li adjustments are low-cut filters of unknown frequency characteristics.
The figure g on the left: Shaded line: velocity y from integration g of the eastwest component of acceleration recorded at TCU129, 1.9 km from the surface trace of the fault, from the 1999 Chi-Chi earthquake, after removal of the pre-event mean from the whole record. A least-squares line is fit to the velocity from 65 s to the end of the record. Various baseline corrections using the Iwan et al. (1985) scheme are obtained by connecting the assumed time of zero velocity t1 to the fitted velocity line at time t2. Two values of t2 are shown: 30, and 70 s.
P. Gundes Bakir,
212
The dashed line is the q quadratic fit to the velocities, with the constraint that it is 0.0 at t=20 s. The acceleration time series are obtained from a force-balance transducer with natural frequency q y exceeding g 50 Hz, , digitized g using g 16.7 counts/cm/s2 (16,384 counts/g). Right: The derivatives of the lines fit to the velocity are the baseline corrections applied to the acceleration trace .
P. Gundes Bakir,
213
The line fit approach is the more complex scheme proposed by Iwan et al. The method was motivated by y studies of a specific p instrument for which the baseline shifted during strong shaking due to hysteresis; the accumulation of these baseline shifts led to a velocity trace with a linear trend after cessation of the strong shaking. The correction procedure approximates the complex set of baseline shifts with two shifts, one between times of t1 and t2, and one after time t2. The velocity will oscillate around zero (a physical constraint), but the scheme requires selection of the times t1 and t2.
P. Gundes Bakir,
214
Figure shows the response spectra of the eastwest component of acceleration recorded at TCU129 from the 1999 Chi-Chi, Taiwan, earthq ake modified using earthquake, sing a variety of baseline corrections. Without a physical reason for choosing these times (for example, based on a knowledge of a specific instrument), the choices of t1 and t2 become subjective subjective. Figure shows that the long-period p spectrum p ordinates are response sensitive to the choice of t2 (t1 was not varied in this illustration). It is important to note that for this particular accelerogram the differences in the response spectrum are not significant until y 10 s oscillator period). p ) beyond
215
P. Gundes Bakir,
A commonly used simplification of the generalized Iwan et al. method is to assume that t1=t2, given by y the zero with the time g intercept of a line fit to the later part of the velocity trace. This corresponds to the assumption that there was only one baseline offset and that it occurred at a single time (for many records this seems to be a reasonable assumption). We p the v0 call this simplification correction.
P. Gundes Bakir,
216
Residual displacements
One of O f the th possible ibl advantages of baseline fitting techniques just discussed is that the displacement trace can obtain a constant level at the end of the motion and can have the appearance of the residual displacement expected in the vicinity of faults as shown in the figure figure. This character of the displacement record cannot be achieved using low-cut filters.
P. Gundes Bakir,
217
At the end of the ground shaking caused by an earthquake, the ground y must return to zero, , velocity and this is indeed a criterion by which to judge the efficacy of the record processing. The final displacement, however, need not be zero since the ground can undergo permanent deformation either through the plastic response of near-surface materials or through the elastic deformation of the earth due to co-seismic slip on the fault.
Close to the fault rupture of large magnitude earthquakes (~Mw = 6.5 and above) this residual displacement can be on the order of tens or hundreds h ndreds of centimeters. This can become an important design consideration for engineered structures that cross the trace of active faults, cases in point being the Trans Alaskan Pipeline System and the Bolu viaduct in Turkey, the former being traversed by the fault rupture of the November 2002 Denali earthquake, the latter by the rupture associated with the November 1999 Duzce earthquake earthquake.
The problem presented by trying to recover the residual placement through baseline fitting is that the resulting offset can be highly sensitive to the choice of parameters as shown in the figure. Furthermore there are few data with independently measured offsets exactly at the location of strong-motion instruments. The lack of independentlymeasured offsets is beginning to be overcome with the installation of continuous GPS stations sampling at sufficiently high rates colocated with accelerographs.
P. Gundes Bakir,
220
High pass filters

The most widely y usedand also the most effective and least subjective j tool for reducing the long-period noise in accelerograms is the low-cut filter. The figure shows the accelerograms after the application of filters to the acceleration time-history, and the improvement in the appearance of velocity and displacement time histories is obvious. It should also be noted that there is little discernable difference between the filtered and unfiltered accelerations.
P. Gundes Bakir,
221
High pass filters

Although the benefits of applying filters are clear clear, it is important to be aware of the sensitivity of the results obtained to the actual parameters selected for the filter.
P. Gundes Bakir,
222
High pass filters

NS component of the FAT record of the Duzce earthquake: displacement and velocity time history obtained by using acausal high pass Butterworth filters with different corner frequencies (from the left column to the right: fc = 0.02 Hz, fc = 0.05 Hz, fc = 0.1 Hz, and fc = 0.2 Hz).
P. Gundes Bakir,
223
High pass filters

The reason that Th th t the th filters filt are described as acausal is that to achieve the zero phase shift they need to start to act prior to the beginning of the record, which can be accomplished by adding lines of data points of zero amplitude, known as pads, before the start of the record and after the end of the record. The length of the pads depends on the filter frequency and the filter order.
High pass filters

Pad length proposed by Brady:
The required length of the filter pads will often exceed the usual lengths of pre and post postevent memory on digital recordings, hence it is not sufficient to rely on the memory to act as the pads.
Zero pads
The figure shows the total length of the time-domain zero pad recommended by y to allow Converse and Brady for the filter response in 2-pass (acausal), nth-order Butterworth filters (these pads are needed regardless of whether the filtering is done in the time- or frequency) domain). Pre- or post-event data count part of the required q p pad as p length. Shown are the pad lengths for three values of the filter corner frequency, as a function of filter order order.
Zero pads
One of the causes for data incompatibility for the records disseminated by the Strongmotion processing centers is the remo al of the pads that are removal added for the application of the filter. This is an issue that creates some controversy because some argue that the pads are artificial and therefore do not constitute part of the data and hence should be removed. The consequence of their removal, however, is to undermine the effect of the filter and this can result in offsets and trends in the baselines of the velocity and displacements obtained by integration. integration
Zero pads
The removal of the pads p also has an influence on the long period response spectral ordinates as shown in the figure (with pads (dashed line), without pads (solid line)). For this reason, it is recommended that when acausal filters are used, sufficient lengths of zero pads should be added to the records and these pads should not be stripped out from the filtered data.
CausalCausal -acausal Butterworth filters

NS component of the CNA record of the Duzce earthquake processed by SPIDER using causal l (d (dashed h d line) and acausal (solid line) filters.
CausalCausal -acausal Butterworth filters

The application pp of causal and acausal filters, , even with very y similar filter parameters (the transfer functions will not be identical if time-domain filtering is used, since the causal filter will have a value of 1/2 at the filter corner frequency, frequency fc, whereas the acausal filter will have a value of 0.5, 05 regardless of the filter order), have been shown to produce very different results in terms of the integrated displacements
P. Gundes Bakir,
230
Causal vs Acausal Filters

Response p spectra p ( (damping p g ratio = 5%) ) of the FP component p of the Bolu record of the Duzce earthquake processed by acausal (solid line) and causal filters (dashed line).
P. Gundes Bakir,
231

R Response spectra t (damping ratio = 5%) for the FP component p of the Bolu record of the 1999 Duzce, Turkey earthquake processed by acausal filters with different corner frequencies: fc = 0.025 Hz (solid line), fc = 0.06 Hz (dashed line) line), fc = 0.1 0 1 Hz (dotted line), and fc = 0.2 Hz (dasheddotted line).

The influence Th i fl of f causal l and d acausal filters on both elastic and inelastic response spectra has been investigated. investigated It is found that both elastic response spectra and inelastic response spectra computed from causally-filtered accelerations l ti can b be sensitive iti to the choice of filter corner periods even for oscillator periods much m ch shorter than the filter corner periods.
P. Gundes Bakir,
233
Tapers
When adding zero pads to accelerograms prior to filtering filtering, a potential undesired consequence is to create abrupt jumps where the pads abut the record, which can introduce ringing in the filtered record. There are two different ways to avoid this, one being to use tapers such as a half-cosine function for the transition from the motion to the zero pad. A simpler p p procedure is to start the p pad from the first zero crossing g within the record, provided that this does not result in the loss of a significant portion of record, as can happen if the beginning or end of the acceleration time series is completely above or below zero.
P. Gundes Bakir,
234
Selection of the longlong-period cutcutoffs

As noted previously, the most important issue in processing strongmotion accelerograms is the choice of the long-period cut-off, or rather the longest response period for which the data are judged to b reliable be li bl i in t terms of f signal-to-noise i lt i ratio. ti A number b of fb broad d criteria can be employed by the analyst to infer the period beyond which it is desirable to apply the filter cut-off, including: Comparison of the FAS of the record with that of a model of the noise, obtained from the pre-event memory for digital records, the fixed trace from analog records or from studies of the instrument and digitizing apparatus.A point of clarification is appropriate here regarding signal-to-noise ratios: the comparison of the record FAS with the FAS of the noise indicates the ratio of signal-plus-noise to noise, i h hence if th the d desired i dt target ti is a signal-to-noise i lt i ratio ti of f2 2, th the ratio of the record FAS to that of the noise model should be 3.
P. Gundes Bakir,
235
Selection of the longlong-period cutcut-offs

Fourier amplitude p spectrum p of the FN horizontal component of the Bolu record of Duzce earthquake (thick solid line) and the noise spectrum (thick dashed line). Superimposed on this graph are the functions f2 (dotted line), f (dotted line), and f2 (d h d d d line). (dasheddotted li ) Seismological theory dictates that at low frequencies, the FAS of acceleration decays according to f2.

The figure shows the selection of filter parameters for a component of the Anderson Dam (analog) recording of the 1989 Loma Prieta earthq ake The FAS of the record earthquake. is compared with the model for the digitization noise proposed by Lee and Trifunac. Also shown is the gradient of the f2 line, superimposed as a best fit (by eye) on the section of the FAS where the decay at low frequencies commences. Also shown in the graph are the FAS of the record after applying filters with three different low-frequency cut-offs.
P. Gundes Bakir,
237

These decay more rapidly than indicated by the f2 model, which is the expected result of effectively trying to remove all of the record both signal and noise noiseat at periods greater than the cut-off. Designing a filter with a gradual roll-off that will produce an FAS that approximates to the f2 model is not advisable since the agreement with the theoretical seismological model would not mean that the real earthquake signal has been recovered, but only that an unknown mixture of signal and noise has been manipulated to produce the appearance of a genuine seismic motion motion.
StrongStrong -motion processing

Processing should be accomplished on a component by component basis. Analog recordings have limited usefullness at periods shorter than about 2 or 3 s. An issue to be considered in record processing is whether the same filter parameters should be used for all three components or whether optimal processing should be used to obtain the maximum possible from each of the three components. p If the same information p processing is applied to all three components, the filter cut-off will generally be controlled by the vertical component since this will usually have a lower signal-to-noise ratio than the horizontal components particularly in the long components, long-period period range range. Therefore Therefore, unless there is a compelling reason for the vertical and horizontal components to be processed with the same filter, this practice is not recommended.
Usable range of response periods

The amplitude of long long-period period response spectral ordinates are highly sensitive to the parameters of low-cut filters, and this is most clearly visible when looking at the spectra of relative displacement. Care g the range g of p periods for which the spectral p must be taken in deciding ordinates can be reliably used, which depends on both the filter frequency and the order of the filter. For a low-order filter applied at 20 s, the spectral ordinates should probably not be used much beyond 10 s. The studies by Abrahamson and Silva and Spudich et al. to derive predictive equations for response spectral ordinates only used each record for periods up to 0.7 times the cut-off period. Bommer and Elnashai, Elnashai in deriving predictions for displacement spectral ordinates, used each record up to 0.1 s less than its cut-off period, which will have inevitably resulted in underestimation of the p displacements p at longer g p periods. spectral

Lecture 2 v2 (Compatibility Mode)

Uploaded by

Lecture 2 v2 (Compatibility Mode)

Uploaded by

ERASMUS Teaching (2008), Technische Universitt Berlin

Digital Signal Processing

ERASMUS Teaching (2008), Technische Universitt Berlin

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

ERASMUS Teaching (2008), Technische Universitt Berlin

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

ERASMUS Teaching (2008), Technische Universitt Berlin

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

and such signals are defined as periodic.

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

Deterministic vs Random signals

ERASMUS Teaching (2008), Technische Universitt Berlin

ERASMUS Teaching (2008), Technische Universitt Berlin

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

ERASMUS Teaching (2008), Technische Universitt Berlin

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

Typical Signal Processing Operations p

ERASMUS Teaching (2008), Technische Universitt Berlin

Typical signal processing operations p

ERASMUS Teaching (2008), Technische Universitt Berlin

Simple Time Domain Operations

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

Simple Time Domain Operations

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

Simple Time Domain Operations

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

ERASMUS Teaching (2008), Technische Universitt Berlin

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

Categories of Fourier Transforms

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

Time and frequency domains

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

Time and frequency domains

ERASMUS Teaching (2008), Technische Universitt Berlin

Time and frequency domains

ERASMUS Teaching (2008), Technische Universitt Berlin

Time and frequency domains

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

Fourier series expansion

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin

Fourier series expansion

thus for any integer n=1,2,3,...

Vibration based structural health monitoring

ERASMUS Teaching (2008), Technische Universitt Berlin