Biomedical Signal Processing

Visit the link below to download the full version of this book:

https://medipdf.com/product/biomedical-signal-processing/

Click Download Now

xiv Preface

T h e second part of the b o o k c o v e r s the classical spectral estimation

techniques including discrete F o u r i e r transform (DFT) and fast F o u r i e r
transform ( F F T ) in C h a p t e r 4, and the periodogram and B l a c k m a n - T u k e y
m e t h o d in C h a p t e r 5. In addition to t h e s e , the c e p s t r u m m e t h o d is in­
cluded in C h a p t e r 6. T h e relevant applications and programs for t h e s e
techniques are included.
T h e third part of the b o o k c o v e r s the adaptive filters such as the a d a p ­
tive noise cancelling ( A N C ) m e t h o d in C h a p t e r 7, adaptive line e n h a n c e r
m e t h o d in C h a p t e r 8, and adaptive z e r o tracking m e t h o d b a s e d on the
linear prediction m e t h o d s in C h a p t e r 9, with the biomedical applications.
Finally, the last part discusses m o d e r n signal processing m e t h o d s such
as autoregressive (AR) in C h a p t e r 10, autoregressive moving average
( A R M A ) in C h a p t e r 11, and P r o n y m e t h o d s in C h a p t e r 12. N u m e r o u s
applications of t h e s e m e t h o d s are explored h e r e .
Please be advised that the c o m p u t e r p r o g r a m s given in the b o o k w e r e
written for the U n i x s y s t e m and may require s o m e modifications for in­
stallation to o t h e r c o m p u t e r s y s t e m s .
I thank Dr. S. Yucel and g r a d u a t e students Louis M e n n a and I. Marsic
for transferring s o m e of the c o m p u t e r p r o g r a m s from F O R T R A N 77 to
A N S I C and debugging t h e m .
I thank the g r a d u a t e students w h o took my c o u r s e and stimulated m a n y
valuable discussions during t h e s e s e m e s t e r s . I a m indebted to g r a d u a t e
students B . A r e n s and A . Smith for typing the manuscript. I am grateful to
graduate student J. Redling for typing and editing the manuscript and her
instrumental support t h r o u g h o u t the preparation of the b o o k . I a m grate­
ful to graduate s t u d e n t s E . J. Ciaccio, G. Ciresi, D . Cursi, A. D o w i d o -
wicz, R. Fisher, F . P h a n , L . - C . P h a n , and W. T h o m s o n for drawing s o m e
of the figures.
I wish to e x p r e s s m y gratitude to M . Yelles, D . Irey, and L . A s b u r y at
A c a d e m i c Press for their valuable help throughout the preparation of the
book.
I thank the Biomedical Engineering faculty and staff, especially D r s .
W. Welkowitz and J. N e u b a u e r , for their support. I am indebted to D r s .
O. Y u r t s e v e n , B . S a n k u r , and S. J. Orfanidis for introducing m e to the
signal analysis field.
Finally, I t h a n k Υ . M . A k a y , my wife, and A. R. A k a y , my son, for
their infinite support and p a t i e n c e .

Metin Akay
P A R T 1

Fundamentals of
Digital Signals
and Systems

A review of the fundamental concepts of discrete time signals

and systems is presented in this section. The motivation behind
discrete analysis lies in the need to extract relevant information
from discrete data sequences. In particular, biomedical signal
processing encompasses discrete analysis; estimation tech­
niques applicable to these signals are discussed. Chapter 1 in­
cludes a review of sampling and reconstruction of analog sig­
nals. The properties of the z-transform and the inverse
z-transform are considered in Chapter 2. Finally, Chapter 3
introduces fundamental concepts in filter design. Realization
forms of digital filters are presented, and both finite and infinite
impulse response filter designs along with several comprehen­
sive examples are discussed.
 CHAPTER 1

Digital Signals and Systems

In this c h a p t e r we discuss important c o n c e p t s of digital signals and

s y s t e m s . T h e topics p r e s e n t e d include stability, causality, linear time-
invariant s y s t e m s , convolution, and auto- and crosscorrelation functions.
T h e representation of signals in the frequency and time domains is also
introduced. Finally, analog-to-digital conversion (sampling) and digital-
to-analog conversion (reconstruction) are also c o v e r e d .

1.1 Introduction

A signal can be r e p r e s e n t e d in both time and frequency d o m a i n s . Time

and frequency may be either c o n t i n u o u s or discrete. A c o n t i n u o u s time
signal is r e p r e s e n t e d in analog form (usually as an analog voltage). A
discrete time signal is represented by a set of quantized analog voltages
which c o r r e s p o n d to sampled instances in time.

1.1.1 Digital Signals

A discrete signal can be represented by a data s e q u e n c e x(n) w h e r e — oc

< η < oc. E a c h element x(n) r e p r e s e n t s the A i t h sample of the s e q u e n c e ,
w h e r e η is an integer. Figure 1.1 s h o w s a typical discrete data se­
quence.

3
4 1. D i g i t a l S i g n a l s a n d S y s t e m s

-6-5-4 4 5 6 η
-3 -2 -1 1 2 3

Fig. 1.1. A typical discrete data sequence.

Let us n o w introduce s o m e well-established s e q u e n c e s . We begin with

the unit impulse function,

fl forA2 = 0
δ(η) = \ (1.1)
[0 for η Φ 0,

which is p r e s e n t e d in Fig. 1.2. In addition, we present the unit step func­

tion, u(n), which is given as

il for>z>0
u(n) = \ (1.2)
[0 for η < 0.

Figure 1.3 s h o w s the unit step function.

δ(η)

-6 -5 -4 -3 -2 -1 1 2 3 4 5 6

Fig. 1.2. Unit sample function.

1.1 Introduction

U(r ι)
1 <ρ C ρ C ρ Cρ C

η
ο ο ο ο ο ο
1 2 3 4 5 6
- 6 - 5 - 4 - 3 - 2 - 1

Fig. 1.3. Unit step function.

T h e unit impulse and step functions can be written in t e r m s of each

other:

u(n) = 2 8(p) = Σ δ(ρ) for ρ = 0, 1, 2, . . . (1.3)

δ(/ι) = ι/(/ι) - w(az - 1 ) . (1.4)

A n o t h e r very useful s e q u e n c e is the exponential s e q u e n c e ,

x(n) = a\ (1.5)

It is important to note that if a > 1 the s e q u e n c e increases with n, and if

a < 1 it d e c r e a s e s with n. Figure 1.4 s h o w s a real exponential for the case
w h e r e a < 1 (decreasing function). In some signal processing applica-

l
cρ X,(n)
ρ
ρ
c}
c c 1
? <

ΐ ? ? ? ? 9 "r
-6-5-4-3 -2 -1 1 2 3 4 5 6

F i g . 1.4. Exponential signal.

6 1 . Digital Signals and S y s t e m s

tions, it is n e c e s s a r y to combine constituent data s e q u e n c e s . Addition is

based on the sample-by-sample summation of two data sequences such as

* i + x2 = xx(n) + x2(n). (1.6)

Multiplication of digital signals is performed by multiplying the current

samples of the independent variables:

x\x2 = x\(n)x2(n). (1.7)

Equivalently, multiplication of a signal by some constant a requires the

multiplication of every sample by that constant:

ax = ax(n). (1.8)
Finally, the signal x2(n) can be defined in terms of a shifted or delayed
version of a n o t h e r signal x\(n),

xi(n) = χάη - n0), (1.9)

w h e r e n0 is an integer.
A p a r a m o u n t role in digital signal processing is a s s u m e d by the periodic
digital signal. A periodic signal is defined as any sequence x(n) with a
period TV, w h e r e x(n) = x(n + N). F o r e x a m p l e , a sinusoidal signal with
period Ν is periodic and can be written as

x(n) = A COS[2TT(« + n0)IN], (1.10)

w h e r e n 0 is a c o n s t a n t . Figure 1.5 shows the digital sinusoidal s e q u e n c e .

N o t e that the sum of t w o periodic signals (composite signal) is periodic if
the ratio of the periods is a rational n u m b e r .
Period signals are of special interest due to their statistical nature or
" p r e d i c t a b i l i t y . " A signal w h o s e statistical and spectral behavior can be

Xi(n)
Αφ
9

-6 -5 -4 4 5 6 η
-3 -2 -1 1 2 3

Fig. 1.5. Sinusoidal signal.

1.1 Introduction 7

predicted is a deterministic signal. T h e p o w e r of a deterministic signal can

be estimated as
X

2
E= Σ Ι*(»)Ι · (ΐ·ΐΐ)

T h e m e a n p o w e r of a discrete signal x(n) of length Ν can be obtained from

2N l
_ 1
E = lim- Σ Ι*(")Ι ·
2
(1-12)

T h e signal w h o s e behavior c a n n o t be predicted from mathematical

transforms is not deterministic but is considered r a n d o m [ 1 - 5 ] .

1.1.2 Properties of Digital Signals

Linear Shift-Invariant Systems

T h e r e p r e s e n t a t i o n of a s y s t e m which m a p s an input signal x\{n) into an

output signal x2(n) has the general form [4, 5]

x2(n) = rU,(/i)]. (1.13)

The system o p e r a t o r T[] relates the input s e q u e n c e x\{n) to an output
sequence x2(n) through some equation or system of e q u a t i o n s . A system is
referred to as a linear system if the principle of superposition is satisfied
such that

T[ax\{n) + bx2(n)] = aT[xx{n)] + bT[x2(n)] = ax3(n) + bx4(n), (1.14)

w h e r e xx and x2 are the inputs and JC3 and x4 are the o u t p u t s .
A shift-invariant system can be defined as a system which m a p s x\{n) to
x2(n) and x\(n - d) to x2(n - d) w h e r e d is the delay. Since η r e p r e s e n t s a
sample in time, t h e shift-invariant s y s t e m can also be referred to as time-
invariant. F o r e x a m p l e , if h{n) r e p r e s e n t s a r e s p o n s e to δ(η), then the
r e s p o n s e to δ(η - k) will be h(n - k). In this c a s e , the impulse r e s p o n s e
h{n) characterizes a linear time-invariant system as discussed in the next
section on convolution.

Convolution

If the values of a s e q u e n c e x2(n) are given in t e r m s of the values of the

t w o s e q u e n c e s x\{n) and h{n) as

xi(n) = Σ *\{k)h{n - *), (1.15)

8 1. D i g i t a l S i g n a l s a n d S y s t e m s

then representation is called the convolution sum and can be written as

x2(n) = X](n) * h(n), (1.16)

w h e r e h(n) r e p r e s e n t s the impulse r e s p o n s e of the system.

N o t e that rearranging the input signal and the unit sample r e s p o n s e
does not change the output signal r e s p o n s e for

x2(n) = h{n)*xx(n) (1.17)

or
X

x2(n) = Σ h(k)x,{n - k). (1.18)

Convolution is a c o m m u t a t i v e , associative, and distributive operation.

T h e equations that follow exhibit t h e s e properties:

x\{n) * [x2(n) * xi(n)] = [χλ(ή) * x2(n)] * x^n) (1.19)

x\(n) * [x2(n) + x3(n)] = x^n) * x2(n) + xx(n) * x3(n). (1.20)

Steps in Computing Convolution [5]

Step 1. Reflect either the impulse r e s p o n s e h(k) or x\(k) about the y-
axis.
Step 2. Delay the reflected impulse r e s p o n s e h(-k) (or x(-k)) by
time delay n.
Step 3. Perform sample by sample multiplication of x\(k) by h(n - k)
[Eq. (1.15)] or h{k) by xx{n - k) [Eq. (1.18)] for all values of k.
Step 4. T h e o u t p u t x2 can be obtained by summing the multiplica­
tions for each delay n.
E X A M P L E . C o m p u t e the convolution of the data s e q u e n c e s x\(n) =
n
a u(n) and h(n) = rectyv(rt), w h e r e Ν = 4 and a = 0.5.
Answer. F o r η > Ν, E q . (1.18) can be written as

N - \ ι _ - N

x2(n) = Σ a'
k=o
nk
= "
n
_ -1,a · 0.21)

F o r 0 < η < Ν, E q . (1.18) b e c o m e s

η 1 _ n + \ n

xi(n) =Σ β" * =
-
, _ · a
d- )
1
22

k=0

Finally, for η < 0, E q . (1.18) a s s u m e s the value z e r o since all the multipli­
cations of h(k) · x\(n - k) are z e r o . Figure 1.6 shows the impulse re-
1.1 Introduction 9

kh(R)
2

1 <> cj> c c

- 6 - 5 - 4 - 3 - 2 - 1 1 2! 3I 4 5 6 k

X, (n-k)

k X 2(n)

2 c
C

1 « c

-β -5 -4 -3 -2 -1 1 2 3I <j
! t
5 6 η

Fig. 1.6

s p o n s e , h(n), the input s e q u e n c e , x\(n), and the resultant output se­

q u e n c e , x2(n).

Stability
A stable system is defined as a system w h o s e output values are
b o u n d e d for all b o u n d e d input values, otherwise referred to as b o u n d e d
/nput b o u n d e d o u t p u t (BIBO). L i n e a r shift-invariant s y s t e m s are consid-
10 1. D i g i t a l S i g n a l s a n d S y s t e m s

ered stable if the following requirement is satisfied,

E= Σ |Λ(Λ)|<°°, (1.23)

w h e r e Ε r e p r e s e n t s the b o u n d e d threshold. If this equation is satisfied and

the input is b o u n d e d , then we have

= Σ h(k)xx(n - k) < Σ \Kk)\ < oo. (1.24)

Therefore, the output s e q u e n c e x2(ri) is b o u n d e d . If the system does not

p r o d u c e a b o u n d e d output for b o u n d e d input values, then Ε = and the
impulse r e s p o n s e is not b o u n d e d [ 1 - 5 ] .

Causality

Any linear shift-invariant system exists as a causal system if the unit-

step r e s p o n s e is z e r o for η < 0. This definition can be extended to any
s e q u e n c e : if x\{ri) = 0 for η < 0, then the output of the system will be
x2(n) = 0 for η < 0.

1.1.3 Representation of Digital Systems and Signals in the

Frequency Domain

T h e F o u r i e r transform of t w o convoluted s e q u e n c e s , x\(n) and h(n), is

equal to the multiplication of the Fourier transform of the sequence x\{n)
and the impulse r e s p o n s e h(ri). This representation in the frequency do­
main is described by the equation

xi(f) = Σ XiWe-M" = Xx(f)H(f) = * h{n). (1.25)

The Fourier transform of the signal e x p r e s s e s the spectral distribution of

the amplitude and p h a s e of the signal. / / ( / ) is k n o w n as the frequency
r e s p o n s e of the s y s t e m . T h e F o u r i e r transform can also be used to repre­
sent a discrete time signal in the frequency domain. T h e Fourier trans­
form of a discrete s e q u e n c e x\(n) is given by

(1.26)
1.1 Introduction 11
T h e inverse F o u r i e r transform of t h e discrete s e q u e n c e is written as

- jw jw
JC,(/I) = τ i* X\{e )e " dw. (1.27)

If t h e frequency r e s p o n s e of t h e s y s t e m to t h e input s e q u e n c e x\(ri) is

jw
H(e ), t h e o u t p u t x2(n) c a n b e e x p r e s s e d as

w w n
xi(n) = τ ~ Γ H(ei )Xx(eJ )eJ" dw. (1.28)

T h e F o u r i e r transform of x2(n) is given as

w w w
X2{eJ ) = H{eJ )Xx{eJ ). (1.29)

This equation c a n also b e interpreted in t h e time domain,

xi(n) = Σ x\(k)h{n - k) (1.30)

* = - χ

or

^(/i) = Σ h(k)xx(n - k). (1.31)

In o r d e r to p r o m o t e further insight into t h e representation of signals in the

frequency d o m a i n , w e discuss t h e following properties [ 1 - 3 ] .

Properties

If a s e q u e n c e X\{f) is a c o m p l e x function, it c a n be d e c o m p o s e d into its

real a n d imaginary c o m p o n e n t s as

*,(/) = Re[*,(/)] + j Im[*,(/)]. (1.32)

In t h e c a s e of a real discrete signal X\(n), t h e real and imaginary parts of

X\(f) a r e written as
X

Re[*,(/)] = Σ cos(2w/n) (1.33)

Im[Z,(/)] = Σ sin(277/h). (1.34)

T h e s e equations c a n be s u m m a r i z e d by expressing X\(f) as a function of

12 1. D i g i t a l S i g n a l s a n d S y s t e m s

amplitude and p h a s e ,
j xm
X\{f) = \X\(f)\e ^ . (1.35)

T h e t e r m s |Λ"ι(/)| and arg\X\(f)\ are called the amplitude and p h a s e s p e c ­

tra. T h e s p e c t r u m | Α Ί ( / ) | r e p r e s e n t s the frequency distribution of the
magnitude of x\(n), while axg\X\(f)\ describes the frequency distribution
2
of the energy of x\(ri). T h e magnitude s q u a r e d , \Xx(f)\ , is called the
power spectrum.
F o r any s e q u e n c e x\(n), the following equation m a y be applied,

Re[*i(-/)] = Re[*,(/)], (1.36)

by proving that the p o w e r s p e c t r u m of the s e q u e n c e x\{n) is an e v e n

function. W e can also write

Im[*,(-/)] = -Im[*,(/)], (1.37)

which s h o w s that the p h a s e s p e c t r u m is an odd function.

Given the s e q u e n c e s x\(ri) and x2(n), w h e r e x2(n) is a delayed version of
x\(n), the F o u r i e r transform of the delayed signal is defined as
X

Xiif) = Σ xi{n)e-M" (1.38)

, I = X-

Χ
χ jl n
= Σ \(" ~ d)e- ^ ; (1.39)
taking η - d = k,
X
, _ + Λ1
= Σ Χι(Α)« ^* (1.40)

Σ X\{k)e-»"A. (1.41)
i - = -X J

T h e n , X2(f) is described as
d
X2(f) = e** Xx{f), (1.42)
where
X

W) = Σ xi(k)e-** (1.43)

is the F o u r i e r transform of X\(ri).

1.1 Introduction 13
This last operation is b a s e d on the shift t h e o r e m . According to this
t h e o r e m , the F o u r i e r transform of a signal shifted by a delay d may be
otherwise written as the multiplication of the F o u r i e r transform of χχ{ή) by
j27Tfd
the exponential function e~ .

1.1.4 Autocorrelation and Crosscorrelation Functions

F o r m a n y signal processing applications, it is n e c e s s a r y to quantita­

tively c o m p a r e t w o different signals x\{ri) and x2(n) in o r d e r to a s s e s s the
similarity of the functions. This is accomplished by applying the crosscor­
relation function CXiX2(n), where
χ
CXlX2(n) = Σ *\(k)x2{n + k). (1.44)
</=-*

T h e crosscorrelation function can be realized in the following steps:

Step 1. Shift x2(n) by time delay n.

Step 2. Multiply the s e q u e n c e x\(k) by the s e q u e n c e x2(n + d) on a
sample-by-sample basis.
Step 3. S u m all the values obtained from step 2 to get the crosscor­
relation.
T h e function CXxX2(n) is m a x i m u m w h e n the signals are similar at that
specific delay. If the t w o signals are similar, the function CXlX2 is called the
autocorrelation function and is given by
χ

Q,2(H) = Σ *ι(*)*ι(« + *)· (1-45)

</=-x

T h e F o u r i e r transform of the crosscorrelation function CXiXXn) is re­

quired to find the c r o s s - s p e c t r a ; it can be obtained from
X

P*M= Σ CX]X2(n)e-*»» (1.46)

M = - X

PXxXl(f) = Xf(f)X2(f)9 (1.47)

w h e r e * r e p r e s e n t s the c o m p l e x conjugate.
N o t e that in the c a s e of x\(n) = x2(n),

Px W lX = X*(f)'X\(f)> (1-48)
which r e p r e s e n t s the p o w e r s p e c t r u m of the d a t a s e q u e n c e , x\(n).

E X A M P L E . C o m p u t e the correlation of the data s e q u e n c e s x\(n) =

n
a u(n) and x2(n) = rect^(Az), w h e r e Ν = 4 and a = 0.5.
14 1. Digital Signals a n d S y s t e m s

Answer. F o r η > 0,

n+k n k M 9
CXlJn) = Σ a = a Σ " = T I — • < >
a
k=0 k = 0 1

F o r -Ν < η < 0, t h e correlation function, C V .| r,(/i), can be estimated a s

yv-i

n+k
C,,» = Σ « - (1.50)

This equation c a n be further written by taking η + k = ρ such that

p
C„„(fl)= Σ « = . _ „ • CI-51)
F o r k < -TV, C^,^(«) will be z e r o
Figure 1.7 s h o w s t h e correlation function, C r,,,(/i), for various delays
In o r d e r t o analyze a n analog signal, o n e must convert t h e signal into
digital form with an analog-digital (A/D) converter. This conversion is
called digitization, a n d it consists of sampling a n d quantization of t h e
signal into a finite n u m b e r of bits. After digitization, t h e digital signal c a n
be p r o c e s s e d by a c o m p u t e r ; algorithms that m a y n o w be applied t o t h e
discretized signal include t h e most important signal processing modali­
ties: filtering, s p e c t r u m estimation, data c o m p r e s s i o n , and signal estima­
tion. In some applications, t h e result of digital signal processing is given
as a digital o u t p u t (data file) o r c o n v e r t e d back into a n analog signal with a
digital t o analog (D/A) c o n v e r t e r . Figure 1.8 s u m m a r i z e s all t h e stages
involved in sampling a n d reconstruction of signals.
T h e Fourier transform of an analog data sequence x\{t) has the general
form

j0it
Χλ(ω) = |_J X\(t)e~ dt, (1.52)

w h e r e ω is t h e physical frequency (rad/sec). T h e frequency / ( H z ) is given

by

/ = f . d.53)

T h e inverse F o u r i e r transform of the analog s e q u e n c e Χ\(ω) is written a s

+X
xi(t) = 4- f XMe* άω. (1.54)
L I T
 1.1 Introduction 15

ι Xi(k)

1 <ι C

— θ θ θ θ θ θ θ
- 6 - 5 - 4 - 3 -2 -1 1 2ί 3 4 5 6 d

ι X 2(n+k)

1 <
η=0
?
9
I—θ θ θ θ θ θ ο ' ΐ ° - 0 0 +
-6 -5 -4 -3 -2 -1 1 2 3 4 5 6 d

ι χ χ (η)
χ 2\ /

C( 2
c3

ρ 1 <:

ί ϊ 9
- θ — •
-6 -5 -4 3 -ί) 1 2 3 4 5 6 η

F i g . 1 . 7

SAMPLER X,(nT)
AND ANALOG Xt(D
COMPUTER
Analog QUANTIZER Digital ProcesMd RECONSTRUCTOR Analog
Signal Signal Digital Signal
Signal

F i g . 1 . 8