DTIC ADA350698: An Investigation into the Passive Detection of Point Targets Using FM Broadcasts or White Noise

Survival, Water, Medical Field Manuals

Military Manuals

Defense Technical Information Center

Document text

REPORT  DOCUMENTATION  PAGE 


form  Approved 
0MB  No.  07040m 


Public  rsporling  burden  for  tN's  collection  of  information  it  ettirrMted  to  average  1  hour  per  reiponte,  Including  the  time  for  reviewing  Intiructiont,  aearcNng  eiisting  data  sourcet,  gathering  and  maintaining  the  data  needed,  and  completing  and 
reviewing  the  collection  of  information.  Send  commonti  regarding  this  burden  eetimata  or  any  other  aipect  of  tWi  collection  of  information,  including  euggettion*  for  reducing  tNs  burden,  to  Washington  Headquarters  Services,  DIrectorata  for 
Information  Operations  and  Reports,  1215  Jefferson  Davis  Highway,  Suite  1204,  Arlington,  VA  222024302,  and  to  the  Office  of  Management  and  Budget,  Paperwork  Reduction  Project  (07044)188),  Washington,  DC  20503. 


1.  kGEHOX  USE  OHli  (Leave b/ank/ 


3.  HEPORT  TYPE  AND  DATES  COVERED 


2.  REPORT  DATE 

I  4  August  1998 


4.  TITLE  AND  SUBTITLE 

AN  ENVESHGATION  INTO  THE  PASSIVE  DETECTION  OF  POINT  TARGETS 
USING  FM  BROADCASTS  OR  WHITE  NOISE 


6.  AUTHOR(S) 

Chadwick  D.  Lindstrom 


7.  PERFORMING  ORGANIZATION  NAME(S)  AND  ADDRESSIES) 

University  of  Washington 

8.  PERFORMING  ORGANIZATION 

REPORT  NUMBER 

98-038 

9.  SPONSORING/MONITORING  AGENCY  NAME(S)  AND  ADDRESSIES) 

THE  DEPARTMENT  OF  THE  AIR  FORCE 

AFTT/CIA,  BLDG  125 

2950  P  STREET 

WPAFB  OH  45433 

10.  SPONSORING/MONITORING 

AGENCY  REPORT  NUMBER 

11.  SUPPLEMENTARY  NOTES 

12a.  DISTRIBUTION  AVAILABILITY  STATEMENT 

Unlimited  distribution 

In  Accordance  With  AFI 35-205/ AFIT  Sup  1 

12b.  DISTRIBUTION  CODE 

17.  SECURITY  CLASSIFICATION 
OF  REPORT 


18.  SECURITY  CLASSIFICATION 
OF  THIS  PAGE 


19.  SECURITY  CLASSIFICATION 
OF  ABSTRACT 


20.  LIMITATION  OF  ABSTRACT 


Standard  Form  298  (Rev.  2-89)  EG) 

Prescribed  by  ANSI  Std.  239.18 

Designed  using  Perform  Pro,  WHSfDKHt,  Oct  94 


AN  INVESTIGATION  INTO  THE  PASSIVE 
DETECTION  OF  POINT  TARGETS  USING 
FM  BROADCASTS  OR  WHITE  NOISE 


by 

Chadwick  D.  Lindstrom 


A  thesis  submitted  in  partial  fulfillment  of  the 
requirements  for  the  degree  of 


Master  of  Science  in  Electrical  Engineering 


University  of  Washington 


1997 


Approved  by 


Chairperson  of  Supervisory  Committee 


Program  Authorized 
to  Offer  Degree  _ 

Date 


19980810  089 


Master's  Thesis 

In  presenting  this  thesis  in  partial  fulfillment  of  the  requirements  for  a  Master's 
degree  at  the  University  of  Washington,  I  agree  that  the  Library  shall  make  its 
copies  freely  available  for  inspection.  I  further  agree  that  extensive  copying  of  this 
thesis  is  allowable  only  for  scholarly  purposes,  consistent  with  "fair  use"  as 
prescribed  in  the  U.S.  Copyright  Law.  Any  other  reproduction  for  any  purposes 
or  by  any  means  shall  not  be  allowed  without  my  written  permission. 


Signature  0. 

Date  0. 1 


TABLE  OF  CONTENTS 


LIST  OF  FIGURES . iii 

LIST  OF  TABLES . vi 

INTRODUCTION . 1 

CHAPTER  1:  System  Overview  and  Requirements . 3 

1 . 1 ;  Receiver  Topology . 3 

1.2;  Basic  Signal  Processing  Algorithm . 4 

1.3:  Receiver  Geometry . 5 

1 .4;  Sensitivity  Requirements . 7 

CHAPTER  2:  The  Scattering  Model  and  the  Ambiguity  Function . 13 

2.1:  The  FM  Radio  Broadcast . 13 

2.2:  Scattering  Models . 14 

2.3:  The  Ambiguity  Function . 18 

2.4:  Discrete  Computation  of  the  Ambiguity  Function . 20 

CHAPTER  3;  The  Bias  of  the  Self  Ambiguity  Function . 21 

3.1;  Notation . 21 

3.2:  The  Bias  of  the  Self  Ambiguity  Function . 22 

CHAPTER  4:  The  Probability  Distribution  of  the  Self  Ambiguity  Function . 27 

4  .1:  Derivation  of  the  Probability  Density  of  the  Clutter  Floor . 27 

4.1.1  Comparison  to  Simulated  Data . 34 

4. 1 .2  Model  Comparison  to  Experimental  Data  . 38 

4.2:Determination  of  Target  Distribution . 51 

4.3:  Ideal  Operating  Charactersitics  of  the  White  Gaussian  Signal . 57 

4.4:  Adjustments  for  the  FM  Waveform . 63 

CHAPTER  5:  Adaptive  Beamfoming/Directional  Antennas . 70 


5. 1 ;  System  Overview . 70 

5.2:  Analysis  of  Adaptiv  Beamforming  Effects  on  Sensitivity . 73 

5.3:  Experiments  results . 76 

CHAPTER  6:  Results  and  Conclusions . 79 

6. 1 :  Experimental  Results . 79 

6.2:  Conclusion . 81 

BroUOGRAPHY . 82 


11 


LIST  OF  FIGURES 


Figure  Description  Page 

1 . 1  The  Direct  Conversion  Receiver . .  •  ■  4 

1.2  The  Multistatic  Receiver  Geometry . 5 

1.3  The  Bistatic  Radar  Geometry . 6 

1.4  The  Bistatic  Triangle  . 8 

1.5  Plot  of  ovals  of  Cassini  about  the  receiver  as  a  function  of  scattered  to 

direct  signal  ratio  (SDR) . 9 

1.6  Plot  of  ovals  of  Cassini  about  the  receiver  as  a  fimction  of  baseline  length . 10 

2. 1  Geometry  for  a  single  moving  target  showing  the  directions  of  the  incident 

wave,  scattered  wave,  and  motion  of  the  particle  from  a  reference  point  . 15 

2.2  An  illustration  of  the  discrete  computation  of  the  ambiguity  function  at 

range  . . 20 

4. 1  Mesh  plot  of  the  ambiguity  function,  |X(r,f)p,  for  a  White  Gaussian  Signal ...  34 

4.2  Mesh  plot  of  the  clutter  floor  of  the  ambiguity  fimction  plotted  in  Figure 

4.1 . 35 

4.3  Mesh  plot  of  the  clutter  floor  for  |X(r,f)p  for  a  White  Gaussian  Signal 

Computed  with  128  Fourier  frequencies  instead  of  64  as  in  4.2 . 35 

4.4  Histogram  of  the  clutter  floor  of  ambiguity  fimction  displayed  in  Figure  4.2 

vs.  a  xi . 

4.5  Histogram  of  the  clutter  floor  of  the  ambiguity  fimction  in  Figure  4.3  vs.  a 

. 37 

4.6  Plot  of  the  theoretical  vs.  empirical  cumulative  distribution  fimction  for  the 

data  of  figure  4.5.  The  two  curves  are  essentially  indistinguishabel . 38 

iii 


L 


4.7  Mesh  plots  of  signals  A1  =  106.9  Mhz  &  A2  =  93.3  Mhz  sampled  at 

different  times . 41 

4.8  Mesh  plots  of  |X(r,f)|  of  signals  A3  =  94.9  Mhz  &  A4  =  94.9  Mhz  sampled 

at  different  times . 41 

4.9  Mesh  plots  of  |X(r,f)|  of  signals  A5  =  97.3  Mhz  &  A6  =  97.3  Mhz  sampled 

at  different  times . 42 

4.10  |X(r,f)|^  for  A3&A4  for  ranges  16.5  km  -  50  km  and  frequencies  -500  Hz 

to -15  Hz . 42 

4.11  |X(r,f)|^  for  A3&A4  for  ranges  16.5  km  -  50  km  and  frequencies  -500  Hz 

to -15  Hz . 43 

4.12  |X(r,f)p  for  A5  &  A6  for  ranges  16.5  km  -  50  km  and  frequencies  -500  Hz 

to -15  Hz . 43 

4.13  Histograms  of  |X(r,f)|^  for  ranges  16.6  -  50  km  and  frequencies  -500  Hz  to 

-  15  Hz . 45 

4.14  Empirical  Cumulative  Distributions  corresponding  to  the  histograms  of 

figure  4.13  plotted  against  the  theoretical  distributions . 46 

4.15  Plot  of  |X(r,-f)|  -  |X(r,f)|  for  the  White  Gaussian  Signal . 48 

4.16  Histogram  of  figure  4. 1 5  vs.  the  theoretical  Gaussian  curve . 48 

4.17  Empirical  Cumulative  Distribution  corresponding  to  the  histogram  of  figure 

4. 16  plotted  against  the  theoretical  distribution . 49 

4.19  Histogram  plots  of  |X(r,-f)|  -  |X(r,f)|  for  signals  A1  -  A6 . 50 

4.20  Empirical  Cumulative  Distributions  corresponding  to  the  histograms  of 

figure  4.19  plotted  against  the  theoretical  distributions . 52 

4.21  Plot  of  noncentral  chi-square  with  k  =  102.4,  2M  =  16,  versus  a  chi-square 

of  2M  degrees  of  freedom . 55 

4.22  Plot  of  Gaussian  with  mean  »  6.84,  variance  «  1.5  versus  a  Gaussian  with 

mean  «  0  and  variance  «  1 . 57 


IV 


4.23  Plot  of  Pc  vs.  Pf  for  the  White  Gaussian  Signal  with  M  =  8,  N  =  64,  S  = 

250,  and  201og(a) . 59 

4.24  Plot  of  Pd  vs.  a  for  constant  Pf  =  lO"®  and  A  =  128,000.  M  is  indicated  on 

the  plot . 60 

4.25  Pd  vs.  Pf  for  the  White  Gaussian  Signal  with  using  Gaussian  densities  and 

M  =  8,  N  =  64,  S  =  250,  and  20  log(a) . 61 

4.26  Pd  vs.  Pf  for  As  =  128,000  with  M  =  8  and  p  =  0.067 . . 66 

4.27  Plot  of  lX(r,-f)-X(r,f)|  for  T  =  0.5,  M  =  8,  Pf  =  10■^  and  Pd  =  0.5  for  a 

White  Gaussian  Signal . 68 

4.28  Plot  of  |X(r,f)P  for  T  -  0.5,  M  =  8,  Pf  =  10'^  and  Pd  =  0.5  for  a  White 

Gaussian  Signal . 68 

4.29  Plot  of  |X(r,-f)|  -  |X(r,f)|  for  T  =  0.5,  M  =  8,  Pf  =  10•^  (3=0.15,  and  Pd  = 
0.645  for  signal  A2  with  target  at  21  km  and  -125  Hz. 

4.30  Plot  of  |X(r,-f)i  -  |X(r,f)|  for  T  =  0.5,  M  =  8,  3=  0.15  for  signal  A2  with  the 

target  injected  at  the  same  position  as  figure  4.29  except  whose  magnitude 
was  decreased  by  3  dB . 69 

5 . 1  Geometry  of  adaptive  beamforming  array . 71 

5.2  Plot  of  |X(r,-f)|  -  |X(r,f)l  for  the  cross  ambiguity  plot  of  the  adaptive 

beamformed  signal  with  signal  A2.  T  =  0.512  s,  M  =  8,  and  the  amplitude 
level  is  the  same  as  4.28 . 78 

5.3  Plot  of  |X(r,-f)|  -  |X(r,f)|  with  the  same  parameters  as  figure  5.2  except  the 

scattering  to  direct  ratio  is  3.3  dB  less . .  78 

V 


L 


LIST  OF  TABLES 


Table  Description  Page 

4. 1  KolmogrofF-SmimofF  Statistics  for  A1 -A6  for  the  . 44 

4.2  KolmogrofF-SmimofF  Statistics  for  A1-A6  for  the  Gaussian . 5 1 

4.3  Minimum  detectable  scattering  amplitude  for  the  White  Gausssian  Signal 

sampled  at  250  kHz . 62 

4.4  List  of  P  for  signals  A1-A6 . 65 

4.5  Minimum  detectable  scattering  amplitude  for  FM  signal  sampld  @250  kHz  66 


vi 


L 


ACKNOWLEDGMENTS 


I  wish  to  thank  John  Sahr,  my  thesis  advisor,  for  encouragement  and  insight  during  this 
research  project.  Also,  I  would  like  to  thank  Frank  Lind  for  providing  the  data  sets  that 
were  used  throughout  this  thesis,  and  for  providing  a  practical  perspective  on  radar 
receiver  operation. 


Vll 


INTRODUCTION 


Over  the  past  three  years,  researchers  at  the  University  of  Washington  have 
been  developing  a  multistatic  radar  system  that  uses  commercial  FM  radio  broadcasts  to 
detect  plasma  waves  in  the  ionosphere.  This  research  has  shown  that  the  short 
autocorrelation  times  of  FM  broadcasts  make  it  possible  to  achieve  good  range  and 
doppler  resolution  despite  the  continuos  operation  of  the  transmitter  However,  because  of 
the  great  distances  of  ionospheric  targets  (>600  km)  most  research  has  focused  on  a 
multistatic  system  consisting  of  a  reference  receiver  and  another  receiver  isolated  from  the 
direct  broadcast  signal  by  the  Cascade  Mountains.  Although  some  research  was  done  on  a 
bistatic  system,  it  was  quickly  seen  that  the  clutter  limited  dynamic  range  would  make  the 
observation  of  plasma  waves  very  difficult  if  not  impossible.  This  analysis  though  did  not 
rule  out  the  detection  of  aircraft  at  distances  close  to  the  receiver-transmitter  baseline.  In 
fact,  the  first  detection  using  FM  broadcasts  was  that  of  a  bistatic  detection  of  an  aircraft 
displaced  1.5  km  from  the  baseline  [1]. 

The  primary  problem  with  the  detection  of  aircraft  with  the  bistatic  radar  is 
discriminating  the  scattered  signal(s)  against  the  strong  direct  broadcast.  This  is  similar  to 
the  detection  of  weak  signals  against  strong  interfering  white  noise  sources  as  encountered 
in  sonar  and  radar  environments  with  strong  ECM  (electronic  countermeasures).  The 
optimal  solution  is  known  to  be  matched  filtering  combined  with  spatial  filtering  technique 
such  as  adaptive  beamforming  [2].  However,  because  the  signal  we  are  seeking  to  identify 
is  not  known  before  reception  it  is  impossible  to  use  the  traditional  matched  filtering 
technique.  Instead,  we  compute  the  self-ambiguity  function  of  the  signal.  This  is  still 
optimal  if  the  signal  itself  is  white  noise,  but  it  requires  that  the  scattered  signal  compete 
with  the  clutter  of  the  direct  signal.  A  performance  analysis  of  this  technique  is  done  for 
the  case  of  white  noise  which  is  then  phenomenologically  extended  to  FM  radio 


2 


broadcasts.  The  key  results  of  this  analysis  are  that  detection  of  aircraft  will  probably  not 
be  possible  at  FM  broadcast  bandwidths.  However,  the  theory  developed  for  white  noise 
processes  shows  that  a  spread  spectrum  radar  which  uses  10-20  Mhz  of  bandwidth 
could  be  developed  with  a  range  radius  of  4  -  5  km  from  the  receiver.  Although  not 
interesting  by  itself,  a  network  of  the  receivers  could  be  formed  whose  presence  would  be 
impossible  to  detect  since  they  would  be  completely  passive.  Also,  because  the  spectral 
density  of  the  transmitted  signal  would  be  close  to  the  noise  floor,  it  would  even  be 
difficult  to  determine  the  radar’s  presence.  This  would  allow  an  active  transmitter  to  be 
used. 


CHAPTER  1 :  SYSTEM  OVERVIEW  AND  REQUIREMENTS 


As  with  any  system  analysis  problem,  before  we  begin  the  detailed  analysis 
of  the  capabilities  of  the  radar  it  is  necessary  to  ensure  that  the  reader  understand  at  the 
broad  system(s)  level  the  radar  operation.  A  radar  is  a  system  which  generally  consists  of 
a  transmitter,  receiver,  antenna(s),  and  either  analog  or  digital  equipment  which 
implements  a  signal  processing  routine  on  the  raw  receiver  data  [3].  The  purpose  of  the 
signal  processing  routine  is  to  use  the  principles  of  electromagnetic  theory  along  with  that 
of  statistics  to  detect  and  determine  characteristics  such  as  range,  direction,  or  velocity  of 
a  target(s). 

The  major  differences  between  a  system  which  uses  commercial  FM  radio 
broadcasts  and  that  of  a  conventional  radar  come  from  the  operator  not  controlling  the 
transmitted  waveform  and  from  the  separation  of  the  receiver  and  transmitter.  The 
inability  to  control  the  waveform  requires  that  the  signal  must  be  modeled  in  order  to 
analyze  its  performance.  We  will  be  modeling  the  transmitted  signal  as  white  gaussian 
noise  which  is  primarily  done  for  analytical  convenience,  but  we  will  also  show  how  the 
model  can  be  extended  to  colored  gaussian  noise  which  more  accurately  describes  the  FM 
broadcast.  The  separation  of  the  receiver  and  transmitter  means  that  the  geometry  and 
location  of  the  receivers  will  be  very  important.  It  also  has  the  effect  that  the  receiver 
topology  can  be  simplified  to  direct  conversion. 


1 . 1  RECEIVER  TOPOLOGY 

The  direct  conversion  receiver  can  be  used  in  this  radar  since  there  is  no 
transmitter  sharing  the  same  antenna  as  the  receiver.  Hence,  the  traditional  interference 
problem  which  requires  that  the  signal  first  be  mixed  with  an  intermediate  frequency  is  not 
required.  Thus,  the  receiver  topology  for  this  radar  is  given  by  figure  1.1  [3]: 


Antenna. 


iBPy  I  co5(wt) 


J 

-my- 

-ra — 

Digital  Signal 

Amp'Ll 

> 

sm(wt) 

-111- 

— 

Processing 

Figure  1.1:  Direct  Conversion  Receiver 


1 .2  BASIC  SIGNAL  PROCESSING  ALGORITHM 

The  optimal  linear  detection  algorithm  in  radar  and  also  for  other 
applications  which  must  detect  signals  in  noise  is  matched  filtering  [2],  In  the  time 
domain,  because  the  impulse  response  of  the  matched  filter  is  the  time  reversed  conjugate 
of  the  signal,  matched  filtering  is  equivalent  to  correlating  the  received  signal  with  the 
signal  that  was  transmitted.  This  intuitively  makes  sense  because  we  would  expect  the 
algorithm  to  compare  the  transmitted  signal  to  what  was  transmitted.  The  matched 
filtering  allows  us  to  quantitatively  determine  the  degree  of  correlation  between  the  signal 
that  was  transmitted  and  what  has  been  received  and  it  does  this  in  the  optimal  sense  of 
maximizing  the  SNR  (signal  to  noise  ratio). 

The  bistatic  FM  radar  uses  just  one  receiver  which  records  a  vector  sum  of 
both  the  direct  and  the  scattered  signal  as  shown  in  figure  1.3.  To  be  able  to  record  both 
signals  simultaneously  requires  sufficient  dynamic  range.  This  is  primarily  determined  by 
the  number  of  bits  of  the  analog  to  digital  converter.  The  matched  filtering  is  implemented 
by  correlating  the  received  signal  with  itself  There  will  be  a  large  spike  in  the  ambiguity 
function  at  zero  delay  and  doppler  shift  because  the  signal  perfectly  matches  itself 


5 


Targets  must  now  compete  with  the  noise  plus  the  clutter  of  the  direct  signal. 
Characterizing  the  clutter  of  the  direct  signal  is  the  main  problem  that  must  be  understood 
to  determine  the  sensitivity  of  the  radar. 


1.3  RECEIVER  GEOMETRY 

The  location  of  the  receivers  is  important  in  understanding  the  operation  of 
the  radar.  Several  options  are  open  to  the  designer  each  of  which  offers  its  own 
advantages  and  problems. 


Figure  1.2:  Multistatic  Receiver  Geometry.  The  direct  signal  is  received  at  RXl  and 
correlated  with  the  scattered  signal  received  at  RX2.  The  direct  signal  is  blocked  by  the 
Cascade  Mountains  which  allows  RX2  to  receive  only  the  scattered  signal. 

The  first  design  is  given  in  figure  1.2.  This  is  the  multistatic  radar  which  is  examined 
extensively  in  Paul  Hall’s  thesis  [4].  It  has  the  advantage  that  the  mountain  range  blocks 
the  direct  signal  to  the  second  receiver.  This  alleviates  the  djmamic  range  and  some  of  the 
clutter  problems  which  are  present  in  the  bistatic  geometry.  Unfortunately,  this  method  of 
implementation  requires  suitable  geographic  features  which  limits  its  use  to  a  few  select 
regions  of  the  world.  It  also  has  the  problem  that  a  relatively  high  speed  data  link  be 


6 


established  between  the  two  receivers.  Finally,  this  geometry  has  been  extensively 
investigated  by  researchers  at  the  University  of  Washington. 


Figure  1.3:  Illustration  of  Bistatic  Geometry.  In  both  cases,  the  receivers  must  receive  a 
vector  sum  of  the  direct  and  scattered  signals. 

Figure  1.3  shows  the  single  receiver  bistatic  radar.  It  requires  the  least  amount  of 
hardware  with  only  a  single  receiver.  However,  the  simplicity  in  hardware  components  is 
offset  by  the  requirement  to  distinguish  targets  against  the  strong  direct  signal  whose  self 
clutter  will  obscure  targets. 

This  problem  may  be  partially  dealt  with  by  using  two  directional  antennas 
one  pointed  at  the  receiver  and  the  other  in  the  direction  of  the  target.  This  can  also  be  be 
done  by  using  two  or  more  closely  spaced  receivers  (within  several  wavelengths  of  each 
other).  It  is  then  possible  to  use  the  phase  differences  due  to  the  difference  in  the  path 
lengths  to  cancel  the  direct  signal  which  can  be  implemented  through  the  spatial  filtering 
technique  called  adaptive  beamforming.  Of  course  this  requires  two  or  more  receivers  to 
be  built,  as  well  as  an  increase  in  the  computation  burden  which  are  the  systems  main 
drawbacks. 


7 


be  built,  as  well  as  an  increase  in  the  computation  burden  which  are  the  systems  main 
drawbacks. 


1.4  SENSITIVITY  REQUIREMENTS 

The  sensitivity  requirements  can  be  determined  from  the  bistatic  radar 
equation.  This  is  given  by  (ignoring  polarization  or  impedance  mismatch)  [5] : 


where  Pr,  Pt  are  the  power  at  the  receiver  and  transmitter  respectively.  The  two  gain 
factors,  Gt(i )  -  transmitter  antenna  gain  and  Gr(o)  -  receiver  antenna  gain,  are  dependent 
upon  direction  patterns  which  is  the  reason  for  the  vector  arguments.  The  bistatic  cross 
section,  is  the  parameter  determined  by  the  target  and  is  also  dependent  upon 

direction.  The  factor  is  the  wavelength  of  the  carrier  frequency  and  results  from  the 
receiver  cross  section.  The  final  two  factors,  Ri  and  R2  are  defined  in  the  figure  1 .4. 
Similarly,  the  received  power  of  the  direct  signal  can  be  written  as; 


(1.2) 


where  the  prime  indicates  the  directions  will  often  be  different  from  the  scattered  signal. 


The  two  quantities  which  are  pertinent  to  the  problem  of  interest  is  the 
ratio  of  1.1  to  1.2,  and  the  ratio  of  1.1  to  noise  power.  First,  we  write  the  ratio  of  the 
received  power  of  the  scattered  signal  relative  to  the  direct  signal  (1 . 1  to  1 .2): 


GXbGAo) 

GXhGAo') 


i47!:)R^R^ 


(1.3) 


8 


Target 


Transmitter  L  Receiver 

Figure  1.4;  The  bistatic  triangle. 

We  will  call  this  quantity  the  scattered  to  direct  ratio  (SDR).  It  will  often  be  the  case  that 
we  will  be  operating  with  Pr  »  N  where  N  is  the  noise  power.  Thus,  it  is  useful  to 
determine  the  performance  of  the  system  for  a  minimum  SDR  and  specific  values  of  L,  p, 
and  crj,  (6,7") .  The  requirements  on  Ri  and  R2  are  then  that  they  satisfy  the  law  of  cosines 
with  L  being  the  third  side  of  the  triangle  as  shown  in  figure  1.3,  and  that  they  satisfy 
equation  1 .3  for  the  given  values.  It  is  then  easy  to  show  R2  must  obey  the  following 
quartic  equation  with  0  being  defined  as  indicated  in  figure  1.4: 

- ILRl  cos(^)  +  I'i?'  -  =0  (1.4) 

(4^)«nim 

Notice  that  for  small  R2  that  1.4  reduces  to  the  familiar  equation  of  a  circle  in  polar 
coordinates.  This  is  one  possible  equation  that  can  be  used  to  describe  the  Ovals  of 


9 


Cassini  that  result,  and  other  solutions  that  use  a  different  origin  or  coordinate  system 
exist  [2],  This  equation  although  difficult  to  solve  analytically  can  easily  be  plotted  with 


Figure  1.5:  Plot  of  ovals  of  Cassini  about  the  receiver  with  L  =  20  Km,  <Tj,;(o,/)=  40  m^, 
p  =  1,  and  c)r^„  as  given  on  the  curves. 


the  implicit  plot  function  of  Maple.  The  result  for  various  values  of  is  shown  in 
figure  1.5.  This  indicates  that  we  must  be  capable  of  detecting  scatter  -55  dB  below  the 
transmitter  signal  to  be  able  to  detect  a  target  with  a  cross  section  the  size  of  an 


Figure  1.6:  Ovals  of  Cassini  for  indicated  baseline  lengths  and  — 65  dB,  p-1 

and (Tj,, (6, i)  =  40  m^. 


11 


airliner,  40  at  1  km  from  the  receiver  without  directional  antennas.  However,  a  similar 
mirror  image  of  the  ovals  occur  about  the  transmitter  which  means  it  is  possible  to  detect  a 
target  21  km  away.  It  is  also  important  to  see  as  shown  in  figure  1.6  that  the  ovals 
described  by  equation  1 .4  are  quite  close  for  varied  baseline  lengths.  The  reason  for  this  is 
that  both  the  scattered  signal  and  the  direct  signal  depend  upon  the  baseline  length. 
Hence,  when  the  power  ratio  is  taken  the  change  due  to  the  baseline  length  changing  is 
typically  small. 

Next,  we  use  the  model  that  the  noise  is  white  and  hence  its  power  can  be 

written  as  [5]: 

N  =  kBTfB  (1.4) 

kB  =  1.38  X  10'^^  J/K  is  Boltzman’s  constant,  T  is  the  noise  temperature  of  the  receiver 
plus  the  cosmic  background  radiation,  and  fe  is  the  bandwidth  of  the  receiver.  This  means 
that  the  ratio  of  the  scattered  signal  to  the  noise  (SNR)  is  given  by; 

P,  _  X^G,{i)GXo)<^Alo)Pt  .15) 

Normally  T  is  taken  close  to  the  effective  cosmic  temperature  which  can  obtain  a 
maximum  of  10,000  K,  however,  in  this  case  it  was  found  the  measured  noise  power  was 
closer  to  -60  dBm  because  of  the  loud  radio  environment  in  the  Seattle  area  due  to  factors 
such  as  FM  sidebands  [1].  This  means  that  T  should  be  taken  as  2.89  x  10*  K.  However, 
this  still  only  results  in  an  SNR  of  about  -25.5  dB  for  a  baseline  length  of  80  km  and  100 
kW  transmitter  power  which  is  much  higher  than  the  corresponding  SDR.  Indeed,  if  we 
can  handle  the  SDR  then  it  will  be  possible  to  easily  deal  with  the  noise  at  almost  all 
baseline  lengths  of  interest. 


12 


We  have  overlooked  the  improvement  in  gain  that  may  be  possible  with 
directional  antennas/adaptive  beamforming  and  the  effect  of  physical  clutter.  We  will 
examine  those  effects  as  degradation  or  improvement  factors  in  the  signal  processing. 
This  means  that  problem  that  we  must  pursue  is  determining  if  it  is  possible  to  achieve 
through  signal  processing  the  detection  of  a  target  whose  SDR  is  near  -60  dB  if  we  wish 
to  pursue  passive  radar  nets  and  even  smaller  if  we  wish  for  a  passive  radar  to  have  the 
range  of  a  short  range  surveillance  radar.  It  is  obvious  that  pursuit  of  passive  radar  nets 
will  be  easier  to  pursue.  Thus,  a  SDR  of  about  -60  dB  is  what  we  should  be  looking  for  in 
the  analysis  that  follows. 


CHAPTER  2:  THE  SCATTERING  MODEL  AND  THE  AMBIGUITY  FUNCTION 


This  chapter  discusses  the  scattering  models  that  will  be  used  in  this  thesis 
and  their  relation  to  the  ambiguity  function.  The  emphasis  is  on  developing  the  correlation 
properties  of  the  fields  which  are  incident  upon  the  receiver,  and  not  upon  detailed  analysis 
of  scattering  cross  sections.  These  properties  will  then  be  used  to  describe  the  output  of 
the  ambiguity  function. 


2. 1  THE  FM  RADIO  BROADCAST 

The  FM  Radio  broadcast  uses  voice  and  music  signals  to  modulate  the 
carrier  signal.  This  is  done  through  the  use  of  instantaneous  frequency.  Suppose  the  voice 
and  music  signal  is  given  by  a(t)  and  that  the  carrier  frequency  is  given  by  fo.  Then  the 
modulation  for  a  monophonic  radio  would  be  given  by  [6]: 

t 

v(t)  =  cos(2^o^  +  Kja{t')dt'  +  (2.1) 

—  00 

where  k  is  a  constant  called  the  frequency  sensitivity  which  is  used  to  spread  the  signal  out 
over  the  given  bandwidth.  Now  differentiating  the  argument  of  the  cosine  gives  us  the 
instantaneous  frequency  which  yields: 

1 

=  (2.2) 

Thus  the  instantaneous  frequency  is  indeed  modulated  by  the  voice  or  music  signal.  Now 
real  FM  radio  broadcasts  are  done  in  stereo  so  that  a(t)  is  replaced  by  a  multiplexed 
version  of  the  sum  and  difference  of  the  left  and  right  speaker  channels  [6]. 


14 


However,  for  the  purposes  of  this  thesis  the  importance  of  the  FM  radio 
modulation  is  that  the  transmitted  signal  can  be  expressed  as  a  modulated  time-harmonic 
wave  as  follows: 

v(0  =  Re{M(OexpO'2^oO} 

t 

u(t)  =  Qxp(j[K  J  aiOdt'+Oo  ])  (2.3) 

—  00 

The  modulation  function,  u(t),  in  the  following  analysis  will  be  assumed  to  be  a  Gaussian 
random  process  with  a  Gaussian  autocorrelation  function.  This  assumption  may  not  be 
valid  in  all  cases,  however,  experimental  evidence  presented  by  Hall  [4]  suggests  it  can  be 
made. 


2.2  SCATTERING  MODELS 

We  describe  two  scattering  models.  The  first  is  the  simple  model  which 
assumes  one  target.  The  second  model  will  include  multiple  targets  and  can  be  extended 
to  distributed  targets. 

First,  we  start  by  considering  the  single  target  model.  We  know  the  field  at 
the  transmitting  antenna  is  given  by  equation  2.3  which  means  that  we  can  solve  for  the 
harmonic  amplitude  terms  at  zero  range; 


1 

^  j  E(0,(i))e\p{j(a)d(i)  =  u{t)exp{j(o^t) 

£(0,m)  =  C/(a>-u>o) 


(2.4) 


Next,  since  the  radio  antenna  can  be  modeled  as  a  collection  of  oscillating  dipoles  with 
field  pattern  /,(i)  the  plane  wave  approximation  can  be  used.  This  allows  the  time- 
harmonic  wave  to  be  written  at  the  target  a  distance  Ri  fi'om  the  transmitter  as  indicated 


15 


Figure  2.1:  Geometry  for  a  single  moving  target  showing  the  directions  of  the  incident 
wave  i  ,  scattered  wave  d ,  and  the  motion  of  the  particle  from  a  reference  point  with 
velocity  V  [7], 

in  figure  1.4  as; 

X,.  .exp(-#7?,) 

E{R^,co,i)  =  U{co-o)f,) - - - /,(i)  (2.5) 

^1 


The  wave  is  then  scattered  from  the  target  as  indicated  in  figure  2.1  with  scattering 
amplitude  /,(©,  i) .  This  allows  the  field  at  the  receiver  to  be  written  as; 


E{R„R^,coXo)^U{(o-co,) 


exp(-y^(7?i  +R^))exp(-jk^r) 


/,(i)/,(o.i)/,(6)  (2.6) 


where  R2  is  the  range  from  the  target  to  the  receiver,  ks  =  A:(i  -  6) ,  r  is  the  displacement 
vector  from  the  reference  point  as  in  figure  2.1,  and  X(d)  is  the  receiving  pattern  [7]. 
Now  since  U(co-coo),  the  baseband  modulation,  is  approximately  gaussian  with  a  width  of 
30  kHz  for  FM  broadcasts,  the  narrowband  approximation  that  the  scattering  amplitudes 
and  antenna  patterns  are  the  constant  can  be  made  [7].  Next,  since  k=  o/c,  we  separate 


16 


2.6  into  three  parts; 


E{R^,R2,co,i,o)  =  C/(<»-6?o)exp 


JO>\ 


^R,  +R2  +f,-r' 


a 


(2.7) 


Where  a  is  the  scattering  amplitude  divided  by  the  range  and  =  i  -  6.  Now  we  take 
the  inverse  fourier  transform  to  get  the  field  as  a  function  of  time  at  the  receiver; 


t- 


i?,  +  7?2  +  fs  ■ 


exp(j6>oOexp(-A(^i  +'^))exp(-Ar. ' 0  (2.8) 


Now  mixing  this  signal  and  making  the  approximation  that  •  r  «  i?,  ,/?2  allows  it  to  be 
written  as; 


oa4\ 


t- 


7?1  +i?2 


exp(-A(7?i  +/?2))exp(-Ar.  -r) 


(2.9) 


Since  r  is  the  time- varying  displacement  vector  from  the  reference  point  in  figure  2.1,  it 
follows  that  if  we  let  t=  0  be  the  time  when  r  =  0  then  if  the  objects  velocity  is  V  it  is 
possible  to  write  the  third  term  as  exp(-7Aof^  V/) which  is  the  bistatic  doppler  shift. 
With  this  change  and  the  addition  of  the  direct  signal  and  the  receiver  noise  the  single 
scatter  model  is  given  by; 


L  R 

x{t)  =  Pu{t  -  — )  exp(-yA:o A )  +  «*/(/  -  — )  ex^irjk^R^ )  exp(-yA:of ,  •  \t)  +  n{t)  (2. 1 0) 

c  c 


f  (i')  f  (-i') 

where  |3  =  — - - -  with  i'  being  the  unit  vector  of  the  direct  line  of  site  between 


transmitter  and  receiver,  L  being  the  baseline,  R*  =  Ri  +  R?  called  the  range  sum,  and  n(t) 
being  the  receiver  noise. 


17 


The  scattering  by  multiple  targets  is  simply  the  sum  of  the  scatter  by  the 
targets  given  by  equation  2.9,  the  direct  signal,  and  the  receiver  noise  since  the  scattered 
fields  are  added  as  vectors.  This  means  that  the  measured  field  can  be  written  as: 

T 

x{t)  =  Pu{t  — )  exp(-yAo ^)  +  Z  ~  ~) exp(-A^« )  exp(-Ar«  •  VO  +  «(0  (2.11) 

C  j  c 

An  important  assumption  that  will  be  made  in  this  thesis  that  has  been  shown 
experimentally  to  be  true  is  that  the  correlation  between  scattered  fields  at  different  ranges 
is  zero. 

The  previous  model  does  not  account  for  targets  which  are  distributed  in 
range  and  it  does  not  account  for  the  changing  environment.  The  first  of  these  problems 
can  be  handled  by  changing  the  summation  to  an  integration.  The  second  requires  that  the 
coefficient  of  the  modulation  signal,  u(t-R«/c),  in  equation  2.11  be  replaced  by  a  quantity 
called  the  scattering  amplitude,  <I>(Rs,t),  which  is  a  random  variable.  Its  value  at  any  given 
time  at  particular  range  is  given  by  the  targets  that  are  there.  Now  this  means  it  is  possible 
to  write  the  measured  field  as: 


x{t)  =  +n{t)  (2.12) 

The  variation  in  phase  due  to  doppler  shift  means  that  over  long  time  periods  the 
scattering  amplitude  goes  to  zero.  Also,  it  will  go  to  zero  because  of  range  migration. 
These  properties  along  with  those  given  above  can  be  summarized  as  follows: 


{<^R.,.iyt>\R,„t-  T)}  =  ^  exp(-Af.  •  Vr) 


(2.13) 


18 


Notice  that  the  time-correlated  scattering  amplitude  is  the  radar  equation  1.1  normalized 
by  the  transmitter  power  times  the  bistatic  phase  shift  due  to  the  target’s  motion.  The 
bistatic  phase  shift  is  the  result  of  the  bistatic  doppler  shift  of  the  target. 


2.3  THE  AMBIGUITY  FUNCTION 

The  ambiguity  function  is  the  output  of  a  filter  matched  to  a  delayed, 
doppler  shifted  version  of  the  transmitter  wave.  It  is  a  function  of  range,  r,  and  doppler 
frequency,  v,  with  the  ideal  shape  of  the  self-ambiguity  function  being  a  delta  function  at 
the  origin.  The  following  is  a  slightly  modified  version  taken  from  Levanon  [8]: 


\z(.r,v)\  = 


I y(t)x*{t  -  r)  Qxp(-j2^vt)dt 


(2.14) 


Notice  that  it  looks  nearly  like  the  fourier  transform  in  doppler,  except  it  is  a  fimction  of 
range.  Indeed  if  we  define  the  yx  sequence  as  being: 

yx^(t)=y(t)x\t-r)  (2.15) 

then  for  a  constant  range  the  ambiguity  fimction  is  the  fourier  transform  of  the  yx 
sequence.  Now  if  we  square  the  ambiguity  function  then  at  a  constant  range  we  obtain  the 
power  spectrum  of  the  yx  sequence.  Then  it  can  be  shown  that  the  inverse  fourier 
transform  of  the  ambiguity  fianction  at  a  constant  range  results  in  the  autocorrelation 
sequence  of  the  yx  -  sequence  [4].  This  can  also  be  shown  through  the  Wiener-Khintchine 
theorem  in  the  case  of  stochastic  transmitter  waveforms.  The  result  is  that  the  ambiguity 
function  in  the  correlation  domain  is  given  by: 


Q(r,  t)  =  E[yx^  {t)yx*{t  -  r)}  =  E[yit)y\t  -  T)x{t  -  r  -  r)x* (t  -  r)}  (2.16) 


19 


The  yx  sequence  provides  a  natural  way  of  interpreting  the  ambiguity 
function.  The  question  then  is  what  is  the  physical  meaning  of  the  yx  sequence.  The 
following  will  provide  the  answer  in  the  case  of  a  traditional  pulse  radar.  Consider  the 
scattered  wave, 

CX) 

y(t)=  j^(R,t)u(t--)dR  +  nit)  (2.17) 

0  ^ 

and  the  transmitter  waveform  x(t)  =  u(t).  Take  the  time  average  of  the  yx  sequence; 


Ryxif)  =  -r))  =  (  j -  ^)dR  +  «(/) 


=  +  (2.18) 
»  J  {<t>(R,l))L(t  -  -)»•(/  -  r)W  =  I  {^R,t))R..(r  -  ^)dR 


where  we  have  used  the  independence  of  the  noise  and  transmitter  waveform  along  with 
their  zero  mean  value.  Also,  the  independence  of  the  transmitter  waveform  and  the 
scattering  amplitude  was  used.  The  average  of  the  scattering  amplitude  is  zero  mean  over 
long  time  intervals  for  moving  targets  because  of  the  bistatic  doppler  shift.  However,  if 
the  averaging  is  done  only  for  relatively  short  intervals  it  is  constant.  The  averaging  over 
short  intervals  is  coherent  averaging.  Since  the  noise  and  transmitter  waveforms  average 
to  zero  much  quicker  than  the  scattering  amplitude,  coherent  averaging  is  useful.  The 
sequence  that  results  from  coherent  averaging  at  a  particular  range  is  samples  of  the 
convolution  of  the  autocorrelation  of  the  transmitter  waveform  with  the  scattering 
amplitude(s)  of  the  target(s)  at  that  range.  Ideally,  it  is  desired  that  this  autocorrelation 
would  be  a  delta  function  centered  at  the  origin.  This  would  mean  the  resulting  sequence 
consists  of  samples  of  the  scattering  amplitude  at  the  desired  range.  The  power  spectrum 
of  these  samples  could  then  be  computed  to  find  the  spectrum  of  the  scattering  amplitudes 


20 


of  the  target(s)  at  that  range.  This  is  what  the  ambiguity  function  does.  In  an  ideal  sense, 
the  ambiguity  function  at  a  given  range  should  give  us  estimates  of  the  power  spectrum  of 
the  target(s)  at  that  range. 

2.4  DISCRETE  COMPUTATION  OF  THE  AMBIGUITY  FUNCTION 

The  computation  of  the  square  of  the  ambiguity  function  is  easy  to  infer 
from  the  discussion  above.  First,  we  form  the  yx  sequences  from  the  data  at  the  ranges  of 
interest.  Next,  we  perform  coherent  averaging  which  can  also  be  thought  of  as  low  pass 
filtering  of  the  yx  sequences.  This  is  then  followed  by  computing  the  periodogram  using 
block  averaging.  The  block  averaging  step  is  called  incoherent  averaging  in  the  radar 
community  because  it  involves  averaging  power  quantities  which  are  phase  independent. 
Figure  2.2  illustrates  the  discrete  computation. 


x(t) 


=  ^S 


. I  I  I  I  I  I  I  i-i  1. 1  I  I  M  M  mca: 


x*(t-r) 


. Ill  I~rm 


izxiixnnnnQxiixnnnniim^ 

N  \  \  \  /  / 


N 


I 


I 


^  'A  ^ 


I  I  :  izn 


I  I  I  1  M=2 


Output  of  Ambiguity 
Function  at  Ranee  r  - 


N 


N 


>  I  I  I  1 . 1 


Figure  2.2:  An  illustration  of  the  discrete  computation  of  the  ambiguity  fimction  at  range  r. 
As,  is  the  length  of  the  sequence  (also  will  be  defined  as  time-bandwidth  product  later),  S 
is  the  number  of  coherent  averages,  M  is  the  number  of  incoherent  averages,  and  N  is  the 
number  of  Fourier  frequencies. 


CHAPTER  3:  THE  BIAS  OF  THE  SELF  AMBIGUITY  FUNCTION 


In  this  chapter  we  evaluate  the  bias  of  the  self  ambiguity  function  of  the 
FM  radio  broadcast  in  the  correlation  domain.  This  can  then  be  Fourier  transformed  to 
the  frequency  domain  to  give  the  bias  of  the  self  ambiguity  function.  By  doing  this,  we 
learn  about  the  general  shape  of  the  self  ambiguity  function  of  signal  composed  of  a  direct 
plus  scattered  signal.  It  is  found  that  because  the  bias  is  concentrated  origin  of  the  self 
ambiguity  function  that  it  does  not  limit  the  sensitivity  of  the  radar. 


3.1  NOTATION 

It  is  quite  inconvenient  to  write  out  all  the  terms  in  equation  2.16  since  81 
terms  result  when  the  received  signal  is  described  by  equation  2.12.  Instead,  the  following 
convention  from  section  2.3  will  be  adapted; 

x(t)  =  fiuit  -  —)  =  direct  signal 

c 

CO 

y{t)  =  f  ,  t)u{t - -)dR^  =  scattered  signal  (3.1) 

n{t)  =  noise 

In  the  analysis  of  multiple  receivers  a  subscript  (1,2,. .n)  will  denote  the  receiver.  A  capital 
X(t)  means  the  sum  of  all  three  signals:  X(t)  =  x(t)  +  y(t)  +  n(t).  It  too  will  have  a 
numbered  subscript  in  the  case  of  multiple  receivers. 

For  four-product  correlations,  it  will  be  necessary  to  expand  them  out  and 
use  a  simple  convention  for  the  four  products  which  result.  Consider  the  fourth-order 


22 


moment; 


X(t-r-T) 


"»ini  =  {x(t)x*(t  -  r)x*(t  -  T)x{t-r- 1)) 
/w,2,3  =  ix{t)y\t  - r)x\t  -  T)n(t  -r-r)) 


(3.2) 


The  two  terms  mnii  and  mms  are  the  examples  of  the  notation  used  for  the  four  product 
terms.  Notice  that  an  index  being  one  indicates  it  is  the  direct  signal,  two  indicates  the 
scattered  signal,  and  three  means  the  noise  term.  The  position  of  the  index  indicates  the 
time  shift  of  the  term  with  the  first  index  having  no  shift,  the  second  being  shifted  by  the 
range  r,  the  third  by  autocorrelation  lag  t,  and  the  fourth  by  both  range  and 
autocorrelation  lag. 


The  final  simplification  is  to  represent  an  autocorrelation  as  Ruu(r)  where 
the  subscript  is  the  random  variable  which  is  being  correlated. 


3.2  THE  BIAS  OF  THE  SELF-AMBIGUITY  FUNCTION 

To  evaluate  the  bias  of  the  self-ambiguity  function  in  the  correlation 
domain  we  must  evaluate; 

^r, r)  =< {x(t)+y{t)+n{t)){x{t-r)  +y{t-r)+iit -r))\x(t -  r)  +y(t -  T)+n{t -  r))’ 

(x{t  -T-r)  +y{t  -  r-r) +«(/  -T-r))> 

Instead  of  writing  down  all  81  four  products  it  is  easier  to  first  identify  terms  which  go  to 
zero.  To  do  this  let  us  review  the  assumptions  being  made; 

1)  The  noise  and  the  transmitter  waveform  are  zero  mean  random 
variables.  This  means  the  gaussian  moment  theorem  may  be  applied 


23 


[4],  namely  the  second  moment  of  the  same  variable  without 
conjugation  on  one  goes  to  zero,  third  moments  are  zero,  and  the 
fourth  moment  with  two  conjugated  variables  may  be  expanded  using 
equation  2.29,  otherwise  they  are  zero  as  well. 

2)  All  three  random  variables:  the  noise,  the  transmitter  waveform,  and  the 
scattering  amplitude  are  independent  of  each  other. 

3)  The  averaging  time  is  long  enough  so  that  any  product  of  the  scattering 
amplitude  involving  an  odd  number  of  terms  or  an  odd  number  of 
conjugations  will  go  to  zero.  This  is  a  result  of  the  doppler  shift  of  the 
moving  target.  This  does  not  apply  to  stationary  targets.  However,  we 
ignore  these  since  the  power  spectrum  should  allow  us  to  distinguish 
moving  targets  from  them. 

We  use  these  assumptions  to  eliminate  terms  in  equation  3.3.  Notice  the  third  term  of  the 
received  signal  is  noise.  Using  this  combined  with  the  assumptions  above  means  the 
following  48  terms  average  to  zero: 

miii3,  mii23,  mii3i,  mii32,  mi2i3,  mi223,  mi23i,  mi232,  mi3ii,  mi3i2,  m^i, 

mi322,  mi33i,  mi332,  mi333,  m2U3,  m2123,  ni2131,  ni2132,  ni2213,  m2223,  Itl2231, 

m2232,  m23n,  m2312,  m2321,  m2322,  m2331,  m2332,  m2333,  m3iii,  m3ii2,  m3ii3, 

m3l21,  m3i22,  m3i23,  m3i33,  m3211,  m3212,  m3213,  m3221,  m3222,  m3223,  m3233, 

m33i3,  m3323,  m333i,  m3332 

The  scattering  amplitude  appears  only  in  the  scattered  signal  so  we  can  eliminate  18  more: 

niiu2,  mu2i,  mi2n,mi22i,  mi222,  mi233,  mi323,  m2m,  m2n2,  m2i22,  m2i33,  m22i2 

m2221,  m2313,  m3i32,  m3231,  m33i2,  m3321 

This  reduces  the  problem  to  the  following  1 5  four  products;  notice  that  every  subscript 
appears  two  or  four  times: 

mini,  mii22,  mms,  mi2i2,  m^n,  m2i2i,  m22ii,  m2222,  m2233,  m2323,  m3i3i, 

m3232,  m33ii,  m3322,  m3333 


24 


Unfortunately,  all  these  terms  must  be  evaluated.  First,  let’s  look  at  the  terms  not 
requiring  the  use  of  the  gaussian  moment  theorem.  These  all  involve  the  autocorrelation 
of  the  noise  times  either  a  correlation  of  the  scattered  or  direct  signal.  The  terms  are: 


"^1133  ="^*311  ={x(Ox\t-r)){n\t-T)nit-T-r))=\a\\^(r)R^(-r) 
^313  =^*31  ={x(Ox\t-T)){ri\t-r)n(:t-r-T))  =  \a\^R^^(T)R^(-T) 


^2233  =  =  {y(0y\t  -r))(n{t  -  T)n(t  - 

=  "432  =  (yit)yit  -  r)){nit -r)n(t  - 


T-r))=  \R^{R„r)R^{r)dR^ 


ILi-r) 


r-'r))=  \R^{R,,t)KuiT)dR^ 


^Rd 


JLi-r) 


(3.4) 


Since  the  autocorrelation  at  a  positive  lag  is  equivalent  to  the  complex  conjugate  of  the 
autocorrelation  at  the  negative  lag  we  see  that  the  terms  above  appear  in  conjugate  pairs. 
This  means  that  when  summed  the  bias  due  to  these  terms  only  occurs  in  the  real  part  of 
the  estimate  of  the  autocorrelation  function  of  the  scattering  amplitude.  When  the 
autocorrelation  lags  are  Fourier  transformed  to  the  frequency  domain  this  clutter  should 
be  symmetric  about  the  range  axis. 

The  last  seven  terms:  mmi,  mii22,  nti2i2,  ^2121,  1112211,  1112222,  1113333  all 
require  the  use  of  the  gaussian  moment  theorem.  The  first  six  terms  require  that  it  be  used 
on  the  four  product  of  the  transmitter  waveform  and  the  last  term  requires  its  use  on  the 
noise.  The  use  of  the  scattering  amplitude  being  uncorrelated  at  different  ranges  will  also 
be  used.  Finally,  there  is  no  need  to  calculate  m2222  since  it  is  the  four  product  of  the 
scattered  wave  so  the  small  value  of  the  scattering  amplitude  to  the  fourth  power 
eliminates  it.  We  are  left  with: 


25 


"*1111 =iarK«(^)r +i«rK(^)r 


^1122  ~  1^1 

Rd 

00 

^1212  “1^! 

Rd 

CO 

^^2121  ~  1^1  J^OO  5  ^) 

Rd 

00 

'”2211 =i«r 

^3333  “|^«n(^)|  "^|^Mn(^)| 


Kwf 


«..(r-^)| 


\dR 


i«..('--^^-^)i 


^K(4  dR, 


K(r)f 


Kui.'^- 


R-L. 


m. 


(3.5) 


The  terms  which  correspond  to  the  target  are  mi2i2  &  ni2i2i.  If  the  transmitter  waveform 
is  nearly  a  delta  function,  then  m2i2i  should  yield  an  estimate  of  the  autocorrelation  lags  of 
the  range  of  interest  and  mi2i2  yields  the  complex  conjugate  of  the  autocorrelation  lags  at 
the  negative  range  of  the  target.  This  is  the  hermite  symmetry  of  the  self  ambiguity 
function  [8].  The  terms  m22ii  &  mii22  result  from  the  interchangability  of  the  range  and 
the  lag  terms.  Notice  there  is  a  hermite  symmetry  present  here  as  well.  This  will  be  easy 
to  notice  in  the  following  plots  of  ambiguity  functions.  The  final  terms  are  the  self¬ 
correlation  of  the  direct  signal  and  the  noise. 


We  can  now  determine  the  bias  of  the  self  ambiguity  function.  Because  of 
the  bandwidth  of  the  FM  radio  broadcast  (~  30  kHz),  we  know  that  the  autocorrelation 
function  of  the  baseband  modulation  is  a  sharp  peak.  This  means  that  all  terms  whose 
arguments  are  either  x  or  r  must  be  concentrated  either  on  the  lag  or  range  axis 
respectively.  From  equations  3.4  and  3.5  it  is  apparent  this  is  every  term  except  those 
involving  the  target(s)  or  its  hermite  counterpart.  When  Fourier  transformed,  the  bias  due 
to  the  clutter  on  the  lag  axis  results  in  a  constant  being  added  to  the  clutter  floor,  and  the 
bias  on  the  range  axis  results  in  the  spike  of  the  self-ambiguity  function  at  zero  range  and 
Doppler.  This  means  that  the  bias  of  the  self-ambiguity  function  will  not  affect  the 


26 


detection  of  a  target  which  is  not  on  the  receiver  transmitter  baseline.  We  verified  this 
result  by  carrying  out  simulations  which  injected  a  target  into  sampled  FM  broadcasts  and 
trying  techniques  which  can  be  derived  from  the  analysis  above  to  reduce  the  bias.  In  all 
cases,  the  results  showed  that  there  was  no  difference  between  biased  self-ambiguity 
function  estimate  and  the  estimate  whose  bias  we  had  reduced.  This  means  that  the 
sensitivity  of  the  radar  is  determined  not  by  the  bias,  but  instead  is  determined  by  the 
variance  of  the  self-ambiguity  function.  We  discuss  this  problem  in  the  next  chapter. 


CHAPTER  4 

THE  PROBABILITY  DISTRIBUTION  OF  THE  SELF  AMBIGUITY  FUNCTION 


The  last  chapter  indicated  that  the  sensitivity  of  the  bistatic  FM  radar  is 
independent  of  the  bias  of  the  self  ambiguity  function.  Instead,  it  depends  upon  the 
variance  of  the  clutter  floor  of  the  self  ambiguity  function.  If  we  were  to  pursue  the 
variance  through  the  technique  of  last  chapter,  the  combinatoric  blossoming  of  terms 
would  quickly  overwhelm  us.  However,  if  we  make  the  additional  assumption  that  the 
signal’s  spectrum  is  white  it  is  possible  to  arrive  at  a  useful  description  of  the  self 
ambiguity  function’s  probability  distribution.  We  can  then  make  phenomenological 
adjustments  for  the  case  of  the  FM  radio  broadcast.  In  addition,  knowing  the  self 
ambiguity  function’s  probability  distribution  also  allows  a  performance  analysis  of  the 
radar  to  be  made,  and  it  provides  a  way  to  determine  the  threshold  for  automatic  detection 
systems. 

4. 1  DERIVATION  OF  THE  PROBABILITY  DENSITY  OF  THE  CLUTTER  FLOOR 

Calculating  the  probability  density  of  the  self  ambiguity  function  seems  a 
daunting  task  considering  the  steps  needed  to  determine  the  bias  of  the  ambiguity 
function  for  a  gaussian  signal.  However,  if  we  make  the  following  two  assumptions 
about  the  received  signal  then  it  will  be  possible  arrive  at  a  theoretical  model  that  can  be 
compared  to  experimental  results. 

1)  The  process  is  white  which  means  E{x(t)x*(t-r)}  =  0  for  r  0. 

2)  The  real  and  imaginary  terms  of  x(t)  are  uncorrelated  and  Gaussian 
distributed  with  zero  mean. 


28 


These  two  assumptions  mean  the  in-phase  (real)  component  x.(t)  and  the  quadrature 
(imaginary)  component  x^(t)  are  independent  of  each  other  and  are  independent  of 
themselves  except  for  r=0. 

To  calculate  the  probability  density  of  the  clutter  floor  of  the  we  first 
consider  just  the  direct  signal.  Form  the  new  sequence  a(t)  =  x(t)x‘(t-r),  where  r  is  the 
range  variable  which  is  greater  than  zero.  Consider  the  mean  of  the  new  sequence: 

E {x.  (t)Xi  (t-r)  +  x^  it)x^  (t-r)  +  j^x^  (t)x,.  (t  -  r)  -  x^  {t)x^  (t  -  r)]}  =  0  (4.1) 

which  follows  from  the  independence  of  the  zero  mean  gaussian  random  variables.  Next 
consider  the  variance  of  the  in-phase  component  of  a(t): 

va:{aXt)}=E{(x,(t)Xi{t-r)+x^{t)x^{t-r)f}  = 

E{xXtfxi{t-rf+2xXt)xXt-r)xXt)x^(t-r)+x^(tfxXt-rf}=  (4.2) 

E[xXtf}E{xXt-rf}+2E{xXt)}E{xXt-r)}E{x^(t)}E{x^(t-r^^^^ 

Where  the  last  two  steps  follow  from  independence  and  the  assumption  the  random 
variables  are  N(0,ct).  It  can  be  similarly  shown  that: 

var{a^(0}  =  2(7"  (4.3) 

Next,  we  show  that  the  real  and  imaginary  components  of  a(t)  are  uncorrelated: 

corr{aXt),a^(t')}  =  E{aXt)a^(t')}  =  E{x,.(0JC,(?-r)x^(f)x,.(f-r)- 
X,  (r)x,.  (f  -  r)Xi  (/’  )x^  (t'-r)  +  x^  {t)x^  (t  -  r)x^  (f  )x,.  (t'-r)  -  (4.4) 

(^'  )^,  (^'“'■) }  =  0 

Where  the  last  step  follows  from  the  independence  and  zero  mean  of  the  random 
variables  and  holds  for  all  sets  of  (t,t')  including  t  =  t'.  Finally,  we  show  that  real  and 


29 


imaginary  components  are  uncorrelated  with  respect  to  themselves  except  for  t  =  t'.  We 
show  the  proof  for  the  real  part  only  since  its  similar  for  the  imaginary  component. 

E{{Xi{t)x.{t-r)  +  x^ {t)x^{t  -  r))(x,. (t' )x,. (t'-r)  +  (f  {t'-r)) }  = 

E{Xi  {t)x.  {t  -  r)x.  it')Xi  (t'-r)  +  x^  (O-r,-  (t  -  r)x^  {t')x^  {t'-r)  +  t^^t'  (4.5) 

(0^,  it  -  r)x.  {f  )x.  (t'-r)  +  x^  (t)x^  (t  -  (f '  )x^  (t’-r) }  =  0 

The  result  follows  from  the  requirement  that  r^^O.  In  this  case  the  only  way  there  could  be 
a  nonzero  term  is  if  the  following  two  equations  are  satisfied: 

t  =  t'-r 
-t'=t-r 
t-t'=t'-t=>t  =  t' 


which  is  what  we  wanted  to  show. 

With  these  properties  of  a(t)  in  mind  we  proceed  in  a  manner  quite  similar 
to  the  computation  of  the  probability  density  of  the  spectral  density  function  given  in 
Percival  pages  220-223  [9].  First,  we  note  from  chapter  2  that  the  ambiguity  function  is 
the  power  spectrum  of  a(t).  This  is  estimated  through  the  periodogram  which  we  define 
as  [10]: 

S(f)  =  Z«(0expf^^^^)  (4.6) 

n=0  V  iV  / 

We  expand  the  inner  summation  into  its  real  and  imaginary  components: 


30 


This  means  S(f)  =  A(f)’+B(f)l  Now  we  wish  to  show  that  A(f)  and  B(f)  are  independent 
Gaussian  random  variables  over  the  set  of  fourier  frequencies  (f  an  integer  between  0  and 
N-1).  To  do  this  we  first  show  that  A(f)  and  B(f)  are  uncorrelated  over  the  Fourier 
frequencies. 


corr{A{flB{r)}  =  £{A(/)B(/')}  =  £:|z«/(0cos[^] 


^N-\ 


V  A  J; 


N-\  N-1 


•=^zz 


1  r’=0  r=0  L 


3,(0a,(Oco{^]sin(^)+a,(0a,(Osin( 

N-1 

=z 


a,(Oa^(Ocos(— Jco^  ^ 


(24' f 
N  J 


(4.8) 


(24']  (24' f 

«,(0«,(Osir(-^jcos(— 


"■"2.‘4¥1sA^ 


\  N  J 


+ 


f=0  _ 


N 


=  0 


The  third  step  follows  from  4.3,  4.4,  4.5,  and  the  linearity  of  the  expectation  operator. 
The  final  step  involves  the  use  of  the  relation  derived  in  exercise  1.4  of  Percival  [9]: 


N-1 


N-1 


X[cos(2;?f0sin(2;zf' r)]  =  — X[exp(;2;S(/'+/))  -exp(-;2;zr(/'+/)) 

f=0 


r=0 


exp{-j27aif  -  /'))  - exp(;2;7tr(/  -  /’)]  =  5^  (/  -  /’)  +  Sfj{f  +  /'). 

N  sin( NttF)  N 

S,  (/)  =  Y^^4^^sin((A  -  l);zf )  =  jD,  (/)sin((  A  -  \)4) 


(4.9) 


where  D^(f)  is  one  form  of  Dirichlet’s  kernel.  Since  we  are  interested  in  the  frequencies 
between  0  and  N-1  it  follows  from  4.9  that  4.8  is  zero.  This  is  for  all  except  f  =  0,  but  we 
are  not  interested  in  it  since  we  expect  our  targets  to  have  some  motion.  This  shows  that 
that  A(f)  and  B(f)  are  uncorrelated  for  any  combination  of  fourier  frequencies  except  for 
f,f  =  0. 


31 


Next,  we  show  that  A(f)  and  B(f)  are  Gaussian  distributed  using  the 
Central  Limit  Theorem.  The  Central  Limit  theorem  states  that  the  sum  of  n  independent 
random  variables  tends  towards  a  N(p,a)  where  p=p,+...+|j,„  and  aW',+...+a\  as  n 
approaches  infinity  [10].  This  requires  that  we  show  the  variables  in  the  summation  are 
independent.  In  the  case  of  4.7,  this  means  it  is  necessary  to  justify  that  ai(t)  is 
independent  of  ajO')  and  similarly  for  a^(t).  Consider  aj(t): 

a.  (r)  =  X,.  (r)x.  (r  -  r)  +  x  (t)x  (t  -  r) 

’  ^  (4.10) 

a^{t’)  =  Xi(t')Xi(t'-r)  +  x^(t')x^(t'-r) 

We  offer  the  following  heuristic  justification  based  on  conditional  probability.  It  is 
necessary  to  show  that  knowledge  of  aj(t)  does  not  increase  our  probability  of  guessing 
the  value  of  a(t').  Since  Utt'  it  means  Xj(t')Xj(t'-r)  is  independent  of  Xj(t)Xj(t-r)  and 
x^(t)x^(t-r)  is  independent  of  x^(t')x^(t'-r).  This  follows  from  the  fact  that  multiplying  two 
independent  random  variables  by  a  constant  does  not  change  their  independence.  Since 
this  is  true  it  quickly  follows  that  aj(t)  must  be  independent  of  aj(t’).  This  reasoning  can 
be  extended  to  showing  that  all  the  random  variables  in  the  set  {ai(0)...ai(N-l)}  are 
independent.  We  use  the  following  Bayesian  theorem  of  probability  which  states  that 
[10]: 

P(An...Aj=P(An|An.i  ...  AJ...P(A2  |  A,)P(Ai)  (4.11) 

where  An  ...  Ai  are  events  which  in  the  case  of  a  random  variable  Ai  =  Xi<Xi  which  means 
P{Ai}  is  simply  the  cumulative  distribution  of  x;.  Consider  the  case  of  N  =  3.  We  have 
already  shown  that  the  random  variables  ai(0),  ai(l),  ai(2)  are  pairwise  independent,  but 
we  must  show: 

P({ai(0)<xo}n{ai(l)<x,}n{ai(2)<X2})=P({ai(0)<xo})P({ai(l)<x,})P({a.(2)<X2})  (4.12) 

This  follows  easily  through  adding  an  additional  term  ai(t")  to  the  two  already  present  in 
4.10.  Notice  that  this  third  term  contains  a  new  random  variable  ai(t')  which  will  be 


32 


independent  of  ai(t)  and  ai(t’).  Using  equation  4.1 1  and  this  independence  between  terms 
it  follows  that  4.12  holds.  This  process  can  be  continued  inductively  to  show  that  the  set 
{ai(0)...ai(N-l)}  is  composed  of  independent  random  variables.  Similarily,  this  can  be 
shown  to  be  true  for  aq(t). 


Now  we  may  apply  the  Central  Limit  Theorem  to  the  four  sums  in  4.7. 


This  means  we  have  A(f)  ~  N 


+  N 

V  A 

V 

N-l 


0,JX2cr^  sin' 


r=0 


N 


and  B(f)  ~  N 


cos^ 


(=0 


2# 

N 


+  N\ 


where  we  have  used 


the  fact  that  the  taking  the  negative  of  a  zero  mean  gaussian  random  variable  does  not 
change  its  distribution.  Combining  the  distributions  of  A(f)  requires  that  the  correlation 
between  the  two  normal  distributions  be  computed.  However,  this  is  easily  seen  to  be 
zero  since  equation  4.4  shows  that  ai(t)  and  aq(t')  are  uncorrelated  for  any  combination 


(t,t').  This  obtains  A(f)  ~  A(0,cr^V^)and  similarly  it  can  be  shown  that  B(f) 
NiO,C7^^/2N). 


Finally,  we  can  determine  the  probability  density  of  the  self  ambiguity 
function  of  a  white  Gaussian  signal  away  from  r  =  0.  First,  we  note  that  because  A(f)  and 
B(f)  are  uncorrelated  and  Gaussian  they  must  be  independent.  Next,  we  scale  A(f)  and 

B(f)  as  —A==^A(f)  ~  N(0,1)  and  — t-7=B(/)~  N(0,1).  Now  squaring  the 
cr  V2A  cr  V2A/' 

normalized  variables  results  in  two  independent  chi-square  random  variables  with  one 
degree  of  freedom  [10].  Next,  summing  two  independent  chi-square  random  variables  of 
one  degree  of  freedom  results  in  a  chi-square  random  variable  of  two  degrees  of  freedom. 
This  gives: 


_ 

INa^ 


|X(r,/)f 


1N(t 


2Na 


■A(/f + 


2iVo- 


-Bi/y  ~  X2 


(4.13) 


33 


where  we  have  used  capital  chi  squared  to  represent  the  ambiguity  function  squared  and 
the  lowercase  chi  squared  represents  the  chi-square  distribution.  Next,  if  incoherent 
averaging  (block  averaging  with  no  overlap)  is  used  it  follows  that  the  chi-square  random 
variables  from  each  block  should  be  independent  of  each  other,  since  the  data  from  block 
to  block  is  independent.  This  means  the  ambiguity  function  squared  will  have  the 
following  distribution: 


|X(r,/)f= 


INo- 


2N(7 


M  1 


(4.14) 


Finally,  we  consider  the  effect  of  coherent  averaging.  In  this  case,  ai(t)  and  aq(t)  are  being 
pre-averaged  before  being  summed  in  4.7.  As  long  as  the  blocks  of  a(t)  that  are  being 
pre-averaged  do  not  overlap  the  independence  relations  hold,  but  the  variance  of  ai(t)  and 
aq(t)  changes.  The  averaging  is  done  as  follows: 

I  S-l  I  5-1 

(0  = +  s)  = -^,^1  (tS  +  s)  +  ja^  {tS  +  s)  (4.15) 

where  S  is  the  length  of  the  average.  It  follows  that  the  semivariances  of  as(t)  are  both 
equal  to  2a^lS.  This  means  that  to  account  for  coherent  integration  4.14  should  be 
modified  as  follows: 


(4.16) 


34 


4. 1 .2  COMPARISON  TO  SIMULATED  DATA 

Theory  must  agree  with  experimental  evidence  in  order  for  a  model  to  be 
useful.  We  start  by  examining  whether  4.16  actually  predicts  the  true  behavior  of  the 
square  of  the  ambiguity  function.  To  begin,  a  complex  sequence  whose  real  and 
imaginary  components  are  uncorrelated  and  gaussian  distributed  with  zero  mean  and  unit 
variance  is  generated  in  Matlab.  Although  it  is  not  necessary  to  attach  a  sampling 
frequency,  we  use  a  frequency  of  250  kHz  since  this  is  the  sampling  frequency  of  the 
experimental  data  that  is  presented  later.  The  length,  L,  of  the  sequence  that  is  analyzed 
is  131,072  points  which  means  it  is  approximately  a  half  second  of  data.  The  following  is 

the  plot  of  |X(r,/)|^using  a  coherent  averaging  factor  S  =  250,  block  size  N=64,  and  the 
incoherent  averaging  factor  M  =  8.  The  incoherent  averaging  factor  can  be  determined  as 
M  =  Ll/(SN)J  (L  J  is  the  greatest  integer  less  than  or  equal  to  the  quantity  in  the  brackets). 


White  Gaussian  Signal 


X  lO"" 


Figure  4.1:  Plot  of  |x(r,/)|^of  a  White  Gaussian  Signal.  Created  from  0.52  seconds  of 
data  sampled  at  250  kHz.  Block  size  (N=64)  and  coherent  integration  factor  (S=250). 


Doppler  Shift  -  Hz 


U 


Range  -  Km 


Figure  4.2:  Same  as  Figure  4. 1  except  range  0  is  not  displayed.  Shows  clutter  floor. 


Doppler  Shift  -  Hz  Range  -  Km 


Figure  4.3:  Plot  of  the  clutter  floor  for  same  white  gaussian  signal  except  this  time 
|X(r,/)l  was  computed  with  N=128  and  S=  250. 


36 


Figure  4.1  is  the  desired  thumbtack  appearance  of  the  white  gaussian  signal.  The  next 
figure  on  the  following  page,  figure  4.2,  is  the  same  plot  except  for  all  ranges  greater  than 
zero.  This  is  the  clutter  floor  that  equation  4.16  should  describe.  Figure. 4.3  is  the  plot  of 
the  clutter  floor  except  this  time  the  ambiguity  function  was  computed  with  a  different 
block  size  N=128.  The  next  step  is  to  compute  the  normalization  factor  and  multiply  it 
times  the  square  of  the  ambiguity  function.  Then  the  histogram  of  the  clutter  of  figure 
4.2  is  plotted  versus  the  predicted  chi-square  random  variable  of  16  degrees  of  freedom  in 
figure  4.4.  Similarly,  the  clutter  floor  of  figure  4.3  is  plotted  against  a  chi-square  of  8 
degrees  of  freedom  in  figure  4.5.  The  results  appear  to  fit  the  distributions  very  closely 
which  gives  us  confidence  in  equation  4.16. 

However,  an  even  better  test  for  determining  the  closeness  of  a  probability 
distribution  to  a  theoretical  model  is  to  use  the  Kolmogroff-Smimov  (K-S)  test  [10].  The 
K-S  test  is  a  hypothesis  test  with  null  hypothesis  that  the  distributions  are  equal  and  the 
alternative  hypothesis  is  that  they  are  different.  The  test  statistic  is; 

q  =  max|F(x)-Fo(x)|  (4.17) 

where  Fo(x)  is  the  hypothesized  cumulative  distribution  and  r(x)is  the  empirical  estimate 
of  the  cumulative  distribution.  Estimating  the  cumulative  distribution  is  done  by  first 
sorting  the  data  into  ascending  order  and  then  assigning  the  value  of  i/N  to  F(x,  )  where  i 
is  the  position  of  the  data  point  in  the  sorted  sequence.  N  is  the  length  of  the  sequence. 
Obviously,  it  follows  from  equation  4.17  that  if  the  two  distributions  are  equal  then  q 
should  be  near  zero  if  a  sufficient  number  of  points,  N,  are  used.  This  is  expressed  as 
[10]: 


- 1  a 
2 


Accept  H„  iff  q  < 


(4.18) 


38 


Figure  4.6:  Plot  of  theoretical  vs.  empirical  cumulative  distribution  function  for  data  of 
figure  4.5,  N=2,240.  The  two  curves  are  essentially  indistinguishable. 

where  a=P{q>clH(,} «  is  the  significance  level.  A  value  of  a  near  one  indicates 
that  we  have  correctly  identified  the  distribution.  Figure  4.6  shows  that  the  empirical 
cumulative  distribution  of  the  data  in  figure  4.5  is  essentially  indistinguishable  from  the 
theoretical  cumulative  distribution  of  a  xl  ■  this  case,  q  =  0.00825.  A  value  so  small 
that  for  all  values  of  a<l  it  easily  passes  4.18.  This  means  we  should  accept  the  null 
hypothesis,  and  it  gives  strong  justification  that  equation  4.16  correctly  models  the 
simulated  data. 

4. 1 .2  MODEL  COMPARISON  TO  EXPERIMENTAL  DATA 

The  data  that  will  be  used  is  six  different  samples  each  of  length  0.5 
seconds  of  FM  radio  broadcasts  taken  at  a  sampling  frequency  of  250  kHz.  The 
broadcast  stations  that  were  sampled  were  A1 =106.9,  A2=93.3,  A3=94.9,  A4=94.9, 
A5=97.3,  and 


39 


A6  =  97.3.  A3  and  A4  were  sampled  on  different  days  as  well  as  A5  and  A6.  The  three 
sequences  Al,  A4,  and  A5  were  sampled  at  approximately  the  same  time  on  the  same 
receiver,  and  the  sequences  A3  and  A6  were  also  taken  at  nearly  the  same  time  on  the 
same  receiver.  The  purpose  of  choosing  these  samples  was  to  ensure  that  any 
characteristics  that  could  be  due  to  receiver  error  could  be  identified  by  showing  up  in 
sequences  taken  at  nearly  the  same  time. 

Before  we  begin  the  comparison  to  empirical  data,  we  give  a  brief 
justification  why  we  should  expect  similar  behavior.  The  assumptions  that  the  radio 
broadcast  is  white  Gaussian  noise  is  obviously  incorrect.  However,  it  is  well  known  that 
the  autocorrelation  time  of  the  broadcast  is  less  than  10  ps  [4].  Also,  since  the  spectrum 
of  the  FM  broadcasts  are  known  to  be  approximately  gaussian  which  implies  symmetry 
about  the  carrier  frequency,  it  can  be  shown  that  the  in-phase  and  quadrature  components 
are  uncorrelated  [2].  Thus,  we  appeal  to  the  same  argument  that  Percival  uses  to  justify 
the  density  of  the  periodogram  for  non-white/non-gaussian  processes  which  requires 
certain  higher-order  moments  to  be  finite  [9].  It  applies  because  that  analysis  is  being 
used  here  to  determine  the  probability  distribution  of  the  ambiguity  function,  and  because 
the  FM  waveform  has  finite  moments  of  all  positive  orders  [11]. 

First,  we  look  at  the  ambiguity  plots  of  the  six  sequences.  The  plots  are  all 
of  IX(r,f)l  not  its  square  since  this  makes  the  clutter  problems  more  apparent.  By  range 
we  mean  the  difference  between  the  direct  and  scattered  path.  Also,  the  Doppler  shift  is 
given  in  Hz  instead  of  velocity  since  the  carrier  frequencies  change  from  station  to 
station. 

Out  to  about  the  range  of  10  km  there  appears  to  be  quite  a  bit  of 
unexpected  clutter  in  all  the  signals.  Signals  Al,  A4,  and  A5  all  have  a  strip  of  apparent 
clutter  out  to  50  km  at  zero  doppler.  Since  the  data  was  taken  at  different  frequencies,  it 
is  probable  that  this  is  due  to  a  constant  DC  offset  in  the  receiver.  However,  if  we  look 
on  either  side  of  zero  doppler  the  plot  appears  relatively  flat.  The  flatness  of  the  clutter 
plane  also  appears  in  A2  and  A6.  However,  signal  A3  appears  to  have  considerable 


40 


clutter.  A  possible  explanation  for  this  is  that  it  is  known  that  the  station  at  94.9  MHz  has 
its  transmitter  located  on  Capitol  Hill  which  is  only  1-2  km  from  the  University  where  the 
data  was  collected.  In  contrast,  the  rest  of  the  stations  are  known  to  have  transmitters 
located  further  than  20  km  away.  However,  we  also  know  A4  is  a  sample  of  the  station  at 
94.9  MHz,  but  does  not  appear  to  have  the  clutter  at  ranges  greater  than  10  km.  One 
important  difference  is  that  A3  was  generated  at  approximately  4  pm  opposed  to  A4 
which  was  sampled  at  1  am.  It  is  possible  that  the  change  in  daily  activity  along  with  a 
possible  change  in  the  transmitting  mode  at  night  could  make  the  difference. 

With  the  clutter  problems  of  the  signal,  we  would  not  expect  our  model  to 
hold  for  ranges  less  than  10  km.  Additional  inspection  showed  that  there  was  still  some 
clutter  in  the  ranges  out  to  16.5  km.  That  the  waveforms  of  the  IX(r,f)l‘  resembled  the 
clutter  floor  of  the  white  Gaussian  signal  for  this  set  of  ranges  can  be  seen  in  figures  4.10 
-  4.12.  Notice  that  there  is  still  a  bit  of  clutter  in  Al,  A2,  A4,  A5,  and  A6  with  a 
considerable  amount  present  in  A3.  However,  we  continue  by  plotting  the  histograms  of 
the  data  versus  xfe  •  O*'®  problem  is  proper  normalization  of  the  data.  In  the  simulation, 
we  controlled  the  variance  of  real  and  imaginary  components  of  the  signal,  but  here  there 
is  no  control.  This  means  it  is  quite  likely  that  the  real  and  imaginary  components  will 
have  a  different  variance.  Also,  we  know  it  is  not  a  white  noise  process.  This  means  the 
normalization  constant  calculated  in  4.16  is  incorrect  because  when  summing  the 
variances  the  cross  correlation  between  terms  must  be  included.  Instead,  we  compute  the 
normalization  constant  by  forcing  the  mean  of  the  data  to  take  the  same  value  as  the  mean 
of  ajfg.  Since,  the  mean  of  a  chi-square  is  equal  to  its  degrees  of  freedom  this  is  a 
simple  procedure.  The  histograms  can  be  seen  in  figure  4.13  on  the  following  page.  The 
number  of  data  points  used  in  each  histogram  is  806.  Notice  that  there  is  definite 
mismatch  for  signal  A3.  However,  Al,  A4,  and  A5  also  appear  to  have  a  problem  with 
their  means  being  located  in  the  wrong  place.  The  general  shape  of  the  histograms 
though  appear  close  to  the  chi-square  density.  We  gain  confidence  by  looking  at  A2  and 
A6.  Notice  that 


Doppler  Shift  -  Hz 


Range  -  Km 


Figure  4.7:  Mesh  plots  of  IX(r,f)l  of  signals  A1  =  106.9  Mhz  &  A2  =  93.3  Mhz  sampled  at 
different  times. 


X  10 


5 


Signal  A3 


3-n 


Doppler  Shift  -  Hz  Range  -  Km 


Figure  4.8:  Mesh  plots  of  IX(r,f)l  of  signals  A3  =  94.9  Mhz  &  A4  =  94.9  Mhz  sampled  at 
different  times. 


44 


these  two  data  sets  nearly  match  the  chi-square  distribution.  Indeed,  when  we  compute 
the  cumulative  distribution  and  plot  it  versus  the  theoretical  cumulative  distribution  of  a 
Xis  A2  and  A6  appear  quite  close  to  the  theoretical  distribution.  This  is  shown  in  figure 
4.14. 

However,  when  the  significance  level  of  the  Kolmogroff-Smimoff  test 
statistic  is  computed  we  find  that  it  indicates  that  we  should  reject  the  hypothesis  that  the 
distributions  are  the  same  for  all  the  signals. 


Table  4.1:  Kolmogroff-Smimoff  Statistics  for  A1  -  A6 


Signal; 

Ai 

A2 

A3 

M 

A5 

A6 

q 

0.1603 

0.0429 

0.3783 

0.1603 

0.1438 

0.0795 

a 

00 

1 

o 

(N 

0.103 

1.3- 10-'“ 

2.05- 10-'* 

o 

1 

7522- 10'’ 

The  extremely  small  values  of  the  K-S  significance  level  definitely 
indicate  that  despite  looking  quite  similar  the  distributions  are  indeed  different.  The 
reason  for  the  difference  is  that  the  clutter  introduces  points  which  lie  far  outside  the 
region  of  probability  for  the  chi-square  distribution.  For  example,  the  probability  of  a 
point  being  greater  than  45  for  the  chi-square  distribution  is  1/7205,  but  we  see  in  figure 
4.13  that  all  the  signal’s  histograms  contain  points  beyond  this  value.  This  shows  that  this 
theoretical  model  is  unable  to  adequately  describe  the  characteristics  of  the  self  ambiguity 
function  of  the  FM  waveform.  This  stands  even  if  A3  is  considered  an  anomaly. 

However,  a  technique  does  exist  that  can  help  eliminate  some  clutter 
variation  at  the  expense  of  increasing  the  noise  variance.  The  idea  can  be  seen  by 
recalling  from  section  3.2  that  the  clutter  due  to  noise  is  symmetrical  about  the  range  axis 
in  the  frequency  domain.  There  is  strong  evidence  of  this  in  the  Figures  4.7, 4.8,  and  4.9. 
It  is  also  known  that  the  target  signal  is  a  complex  sinusoid  and  should  therefore  appear 
asymmetrically.  Normally,  subtracting  two  random  variables  is  not  advised  because  it 
increases  the  variance  due  to  noise.  However,  in  this  case  it  is  apparent  that  the  clutter 


45 


Signal  A1 


Signal  A2 
lOOi - ^ - 


300 
200 
100 
0 

0  50  100  150  200 

Signal  A5 


150 
100 
50 
0 

0  50  100 

100 

50 

0 

0  50  100 


Signal  A3 


Signal  A4 


Signal  A6 


Figure  4.13:  Histograms  of  the  IX(r,f)P  for  ranges  16.6-50  km  and  frequencies  -500  Hz  to 
-15  Hz. 

contribution  is  dominant  so  the  overall  effect  is  to  decrease  the  variance.  This  can  be 
seen  in  the  following  calculation: 

EiiSif)  -  S{-f)f  -  E{S(f)  -  S(-f)f}  =  E{S{fy }  -  2E{Sif)S(-f)]  + 
E{S{-ff }  -  E{S{f)}^  +  2E{S(f)}E[S{-f)}-  E{S{-f)}^  =  (4.19) 

(jf  +0-2  -cov{5(/),5(-/)} 


46 


A1 


A2 


A6 

1 

0.5 

0 

0  20  40  60 


Figure  4.14:  Cumulative  Distribution  corresponding  to  data  and  theoretical  curves  of 
figure  4.13. 


where  a,  is  standard  deviation  of  S(f)  and  a^is  the  standard  deviation  of  S(-f).  Notice  that 
if  the  covariance  is  quite  large,  then  the  overall  variance  of  the  resulting  random  variable 
is  reduced. 


Since  the  distributions  are  correlated,  it  may  seem  quite  difficult  to 
calculate  the  resulting  probability  density  function.  However,  it  can  be  shown  that  if  we 
take  the  square  root  of  the  IX(r,f)l^  then  the  result  is  a  chi  density  of  2M  degrees  of 
freedom.  Its  mean,  variance,  and  distribution  are  given  by  [10]: 


2M  +  1 

r(M) 


Xi 


2M-\ 


M 


exp(-y)t/(x)  (4.20) 


47 


where  r(x)  is  the  gamma  function  and  U(x)  is  the  unit  step  function.  The  chi  density  can 
be  approximated  by  a  Gaussian  with  mean  and  variance  given  by  equation  4.20  [12].  It  is 
well  known  that  if  two  gaussian  random  variables  are  summed  then  even  if  they  are 
correlated  the  result  is  a  gaussian  [10].  This  means  IX(r,-f)l  -  IX(r,f)l  will  be  a  Gaussian 
with  a  mean  close  to  zero  and  variance  given  by  equation  4.19. 

First,  we  verify  that  this  works  with  the  simulated  data.  Figure  4.15  shows 
the  result  of  subtracting  the  positive  frequencies  from  the  negative  frequencies  for  the 
white  Gaussian  signal.  Notice  that  it  appears  to  have  increased  the  variance  while 
making  the  mean  near  zero.  Figure  4. 16  shows  the  histogram  from  the  data  in  figure  4. 1 5 
and  it  shows  the  curve  representing  the  theoretical  gaussian  model  with  mean  zero  and 
variance  given  by  2a  =  2(2M-p^)  where  p  is  given  by  equation  4.20  and  M  is  the  number 
of  incoherent  averages  (M=8  in  this  case).  Notice  that  the  data  in  figure  4.15  has  to  be 
properly  scaled  by  the  square  root  of  the  factor  given  in  equation  4.16  for  the  model  to  be 
applied  correctly  which  is  the  reason  for  the  different  scale  in  figure  4.16.  Next,  figure 
4.17  shows  the  plot  comparing  the  theoretical  vs.  the  empirical  cumulative  distribution. 
The  distributions  appear  to  nearly  identical  from  the  plot.  The  q  =  0.0157  which  results  in 
a  significance  level  for  the  806  points  of  a  «1.  Let’s  see  if  this  procedure  achieves  the 
desired  result  with  experimental  data.  Figure  4.18  shows  the  negative  frequencies  minus 
the  positive  frequencies  for  ranges  16.5  -  50  km  for  signals  A1  &  A2.  The  key  fact  is  that 
the  large  hump  is  gone  in  A1  and  that  both  data  sets  look  more  like  identically  distributed 
noise  than  in  Figure  4.7.  Figure  4.19  shows  histograms  for  the  negative  minus  positive 
frequency  data  sets  for  all  of  the  experimental  signals  that  we  have  looked  at.  Notice  the 
definite  improvement  with  all  except  A3  being  closely  approximated  by  the  theoretical 
Gaussian  curves.  The  parameters  for  the  densities  were  estimated  from  the  data  since  we 
already  know  that  the  normalization  of  equation  4.16  is  not  adequate  for  the 


8 


Figure  4.15:  Plot  of  the  negative  frequencies  minus  the  positive  frequencies  for  the  white 
Gaussian  waveform. 


Figure  4.17:  Cumulative  distribution  function  of  the  theoretical  (smooth  curve)  vs 
empirical  cumulative  distribution  of  figure  4.16. 


Frequency  -  Hz  Range  -  Km 

Signal  A2 


Figure  4.18:  Mesh  plot  of  IX(r,-f)l  -  lX(r,f)!  for  signals  A1  &  A2 


50 


A1 


jj 

A 

h 

Q|  . . . . . . I 

-5  0  5 


A3 


A2 


1 

ttw. 

Ql  . . Ill . . . I 

-5  0  5 


A4 


jI 

ikbfw _  _ 

5  0  5 

A6 

_ 

Ql  . . LI . .  - _ I 

-5  0  5 


Figure  4.19:  Histogram  plots  of  IX(r,-f)l  -  IX(r,f)l  for  signals  A1  -  A6. 

experimental  data.  Figure  4.20  shows  the  cumulative  distributions  that  result  from  the 
histograms  and  the  theoretical  densities.  Notice  that  all  except  A3  are  indistinguishable 
from  their  theoretical  distributions.  Indeed,  the  K-S  statistics  are  all  close  to  one  except 
for  signal  A3.  However,  the  values  of  the  significance  of  the  K-S  statistic  should  be 
taken  rather  heuristically  since  we  estimated  the  parameters  of  the  theoretical  distribution 
from 


Signal 

Al 

A2 

A3 

A4 

A5 

A6 

q 

0.0266 

0.0309 

0.0791 

0.0266 

0.0221 

0.0238 

a 

0.639 

0.429 

8.33E-5 

0.639 

0.910 

0.803 

the  data.  This  seems  to  indicate  that  the  distributions  are  close  to  being  Gaussian.  Also, 
the  histograms  of  figure  4.19  show  that  the  histograms  of  the  data  are  quite  similar  to  that 
of  the  simulated  white  Gaussian  signal.  For  these  reasons,  and  because  the  Gaussian 
distribution  is  simple  to  calculate  we  use  it  as  the  basis  for  establishing  the  detection 
threshold. 

4.2  DETERMINATION  OF  THE  DENSITY  WITH  A  TARGET  PRESENT 

The  next  step  is  to  determine  the  distribution  of  IX(r,f)l"  when  a  target  is 
present  at  a  given  range  and  doppler  frequency.  In  the  case  of  high  SNR,  the  value  of 
IX(r,f)Pis  relatively  easy  to  compute  for  a  single  target.  First,  we  rewrite  the  signal,  x(t), 
as  jc(0  =  u(t)  +  ae\p(j2;rut)u(t  -r^).  This  means  a(t)  becomes  more  complex. 
However,  as  long  as  T^r„  the  probability  density  of  the  clutter  floor  remains  quite  similar 
to  equation  4.16  with  only  the  variance  changing.  Now  when  r  =  r„  and  we  assume  the 
energy  of  the  scattered  signal  is  much  larger  than  the  variance  of  the  clutter  floor  of  the 
ambiguity  function  it  is  possible  to  approximate  value  of  the  ambiguity  function  at  r„. 

This  assumption  allows  us  to  write  the  ambiguity  function  at  the  point  of  the  target  as: 

,  w-i  2  ^ 

|X(r„,u)|"  «  rj|  e\p{-j27tt(v-f  /  N))  (4.21) 

1=0 

which  evaluates  to  4a^a''N^  if  vwf/N.  Next,  this  value  must  be  scaled  by  the  quantity 
given  in  equation  4.16.  The  result  is  2a^NMS.  However,  we  recognize  that  NMS  as  the 
length  of  the  sequence.  We  rewrite  this  as  NMS  =  Tf3  where  T  is  the  time  length  in 
seconds  and 


52 


A1  A2 


Figure  4.20:  Cumulative  distributions  corresponding  to  figure  4.19. 

fjis  the  sampling  frequency.  We  will  call  this  the  time-bandwidth  product  of  the  signal 
processing  algorithm,  A^.  Now  in  the  case  of  a  white  noise  signal  the  time-bandwidth 
product  of  the  signal  processing  algorithm  is  equal  to  that  of  the  transmitted  signal,  A^.. 
This  means  that  for  a  white  noise  signal  the  value  of  the  self-ambiguity  function  when  a 
target  is  present  is  the  ratio  of  the  energy  of  the  scattered  signal  to  the  power  per  hertz,  N^. 
=  P/fj,  of  the  direct  signal,  2E^/Nc.  This  shows  that  there  are  two  basic  ways  to  increase 
the  sensitivity  of  the  radar.  The  first  is  to  increase  the  averaging  time.  The  second  is  to 


increase  the  bandwidth  of  the  transmitted  signal.  In  fact,  it  will  soon  be  shown  that  the 
time-bandwidth  product  of  the  FM  broadcast  signal  is  insufficient  to  realize  the  gain 
required  to  detect  aircraft. 


53 


To  show  this  for  the  bistatic  FM  radar  we  compute  an  approximate  form  of 
the  probability  density  of  IX(r,f)P  when  a  target  is  present.  To  do  this  we  reconsider 

equation  4.21  and  rewrite  the  inner  term  of  the  summation 

yv-i 

=  Yj  {[“(Om* (t-ro)  +  cai{t)u  {t  -  2ro )  • 

^  2 

exp(-;2;n/(t  -  r^))  +  a expU'2^^)\u(t  -  ro)|  +  a'  exp(j2;rvt)txp(-j2;rv(t  -  Tq))  • 
u{t  -  r^)u(t  -  2ro)]exp(^- 

-  r^f]  =  A(f)  +  2aa^N  +  jB(f) 

r=0  ^  ** 


N-\ 


<=o 


/ 

u(t)u(t-ro)txp\-j2m  — 


(4.22) 


N-\ 


x(t)x*(t  ■ 


f 

■ro)exp\-j2a  — 


where  A(f)  and  B(f)  are  given  by  equation  4.7.  It  has  been  assumed  here  that  the 
scattering  amplitude  is  sufficiently  small  that  the  variance  of  A(f)  and  B(f)  is  not  affected 
by  the  extra  terms  involving  a.  Now  when  we  square  the  real  and  imaginary  parts  the 
result  is  a  noncentral  chi-square  density  [2].  The  noncentral  chi-square  density  is  defined 
as  the  distribution  that  describes  the  random  variable: 

(4.23) 

/I=I  /2=1 

where  the  x„  are  independent  zero-mean  Gaussian  variables  with  common  variance  ol 
The  C„  are  the  constants  and  we  have  also  defined  the  quantity  Q  which  is  the  parameter 
along  with  a  and  N  which  characterize  the  distribution  of  y  given  in  the  unnormalized 
form  as: 


fiy)- 


2a^  VQ 


N-2 

y  \  4 


exp[-(y  +  Q)  /  (2(7^ 

<J 


(4.24) 


where  In^.^Y)  is  the  modified  Bessel  function  of  first  kind  and  order  N/2-1.  The  case  of 
interest  is  the  noncentral  chi-square  that  results  when  the  value  of  the  square  of  the 


54 


ambiguity  function  at  the  range  and  doppler  frequency  of  the  target  is  computed  with  both 
coherent  and  incoherent  integration.  Coherent  integration  scales  the  variance  of  A(f)  and 
B(f)  as  described  earlier  by  a  factor  of  1/S.  Incoherent  integration  requires  that  we 


rewrite  the  step  as: 

1  M-\  1  M-\  1  M-\ 

«/)  =  17  Z  S,(/)  =  ( A  (/)  +  Z(fl,  (/))  = 

M  /=0  /=0 

A,(/)  lacr^N  ,  ^(B,{f)V 

h^sfM  Vm  ^  1^1  VmJ 


(4.25) 


This  means  the  variance  is  <3^  =2cr''N/MS  and  that  Q  =  4a'a'‘N\  Next  we  normalize  the 
S(f)  by  its  variance  (notice  it  is  the  same  as  the  normalization  parameter  in  4. 1 6).  This 
results  in  the  normalized  noncentral  chi-square  distribution  with  noncentrality  parameter 
X  =  =  2a^NMS  =  This  results  in  the  standard  form  of  the  noncentral  chi- 

square  density  with  noncentrality  parameter  X  [2]: 

fiy)  =  \^yl  exp(-(y  +  X)  / 

//,.  =A  +  2M  (4.26) 

0-2  =4;i  +  4M 


where  N  has  been  substituted  by  2M  because  we  summed  2M  independent  noncentral 
chi-square  densities.  Figure  4.21  is  a  plot  of  the  normalized  noncentral  chi-square  density 
with  a=0.02,  N  =  64,  M  =  8,  and  S  =  250  contrasted  with  the  chi-square  density  of  2M 


degrees  of  freedom.  Notice  that  there  is  almost  no  overlap  between  the  densities  which 
means  probability  of  detection,  Pp  »  1,  and  probability  of  false  alarm,  PpA  ®  1-  Next,  we 
consider  the  density  of  the  modulus  of  the  ambiguity  function,  IX(r,f)l,  at  the  target.  The 


55 


Figure  4.21:  Plot  of  noncentral  chi-square  with  1=102.4,  2M  =  16,  versus  a  chi-square  of 
2M  =  16  degrees  of  freedom. 

density  is  simply  the  noncentral  chi  density  given  by  [2]: 


fiy)^y 


2M-2) 


(4.27) 


where  X  is  the  same  noncentrality  parameter  as  before  and  I^  iCy)  is  still  the  modified 
Bessel  function  of  the  first  kind.  The  mean  and  variance  of  the  noncentral  chi  density  can 
be  approximated  as  [from  equation  26.4.38  found  in  reference  12]: 


^  =  +  a  =  A  +  2M  b  =  X/a 

CT-  ^{^)-^{%b  +  {\  +  b){\-lb)\  +  0{a-^) 

L  0^3 


(4.28) 


56 


However,  if  the  noncentrality  parameter  is  moderately  greater  than  the  degrees  of 
freedom,  2M,  then  we  may  take  b»l.  This  means  we  can  approximate  the  mean  as 
pa-\//i-  +  2M-l  and  a’«l.  Although,  first  order  correction  could  be  completely  made 
this  will  be  quite  close  if  the  condition  above  is  met  (this  will  be  the  case  for  any  targets 
we  wish  to  detect  because  the  mean  must  be  greater  than  the  mean  of  the  clutter).  In 
addition  this  approximation  lowers  the  estimate  of  the  mean  and  increases  the  value  of  the 
variance.  This  is  good  because  it  means  the  true  values  will  yield  a  better  probability  of 
detection  than  the  approximation  will. 

Because  the  chi  density  can  be  approximated  as  Gaussian,  it  follows  that 
the  noncentral  chi  density  can  be  approximated  as  the  Gaussian  density, 
N(VA  +  2Af  - 1 ,1).  This  means  that  when  the  positive  frequencies  are  subtracted  from 
the  negative  frequencies  that  if  a  target  is  present  its  distribution  is  given  as  a  Gaussian 
with  mean  and  variance  of: 


//  =  Vl  +  2M-l-V2- 


r(M) 


=  1  +  2M- 


V2- 


r(M) 


(4.29) 


However,  just  as  the  noncentral  chi  density’s  mean  and  variance  can  be  approximated,  we 
can  do  the  same  with  the  chi  density.  Abramowitz  and  Stegun  (equation  26.4.34)  give 
the  approximations  as  [12]: 


/i  =  {l  +  [32M(2M  - 1)]-' }^2M  - ^  +  0{{2My^'^ )  « 


1 


1 


1 


(4.30) 


cr  =  — - 


2  16M  64M^ 


+ - ^  +  G((2M)'")«- 

512M-’  2 


Where  we  have  used  the  assumption  that  M>2,  which  means  that  we  are  overestimating 
the  mean  and  variance  of  the  clutter.  This  good  because  when  subtracted  from  the 
noncentral  chi  density  it  makes  the  mean  smaller  than  it  should  be  and  the  variance  larger. 


57 


Figure  4.22:  Plot  of  gaussian  with  mean  «6.84,  variance  »1.5  versus  a  gaussian  with 
mean  »0  and  variance  «1 

Thus,  it  will  generate  a  probability  of  detection  which  is  slightly  less  than  the  true  value. 
Using  4.30  we  may  rewrite  4.29  as: 

//»V>l  +  2M-l-V2M  (4.31) 

4.3  IDEAL  OPERATING  CHARACTERISTICS  OF  THE  WHITE  GAUSSIAN 
SIGNAL 

The  previous  two  plots  give  us  an  intuitive  feel  for  how  the  operating 
characteristics  of  the  radar  can  be  determined  from  equations  4.16,  4.20,  4.26,  and  4.29. 
The  figures  show  the  detection  of  a  target  can  be  posed  as  binary  hypothesis-testing 
problem  with  hypotheses:  Hg  -  No  target  present  and  H,  -  target  present.  The  performance 
of  the  radar  is  determined  by  finding  the  P^,  the  probability  of  detection,  and  Pp  the 
probability  of  false  alarm. 


58 


Although  many  possibilities  exist,  we  choose  to  analyze  the  performance 
of  the  radar  through  a  simple  threshold  system.  The  system  is  simply: 


\X{r,ff  >  X,  or 


X(r,-/)|-|X(r,/)||>X, 


(4.32) 


where  the  second  equation  is  a  two  sided  threshold  which  allows  detection  of  a  target  at 
either  a  positive  or  negative  frequency.  The  probability  densities  that  describe  the  first 
test  are  given  by  equations  4.16  and  4.26.  The  mean  and  variance  of  the  Gaussians  that 
describe  the  second  test  are  given  by  4.20  and  4.29.  The  probability  of  false  alarm  for  the 

simple  threshold  system  is  given  by  =  P{  |x(r,/)|^  >  X,l  H,,},  and  the  probability  of 

detection  is  given  by  P^  =  P{  |x(r,/)p  >  X,  1  H,}.  These  can  be  easily  determined  after 

X,  is  chosen  by  integration  of  the  appropriate  probability  density.  However,  in  radar  it  is 
desirable  to  chose  a  maximum  value  of  P^  and  then  to  choose  the  threshold  level  to 
maximize  P^  which  is  called  the  Neyman-Pearson  strategy  [2].  In  this  case,  it  is  obvious 
that  the  threshold  value(s)  that  will  achieve  this  can  be  determined  either  from  the 
percentage  points  of  a  chi-square  or  the  standard  deviation  of  a  Gaussian. 

First,  we  consider  the  performance  characteristics  using  the  square  of  the 
ambiguity  function.  The  distribution  when  no  target  is  present  is  the  Zim  a 

target  it  is  the  noncentral  chi-square  of  equation  4.26.  Figure  4.23  is  the  plot  of  the  P^  vs. 
Pf  with  values  of  201og(a),  the  ratio  of  the  scattered  signal’s  power  to  the  direct  signal, 
given  in  the  plot.  Notice  that  for  values  of  201og(a)  <  -38  dB  that  detection  with  a 
reasonable  value  for  P^  is  unlikely  for  the  given  values  of  M,  N,  and  S.  A  good  question 
to  ask  is  how  does  changing  M,  N,  or  S  affect  the  probability  of  detection.  First,  we 
notice  that  changing  N  or  S  does  not  change  the  density  of  H^,  however,  changing  M 
increases  the  mean  and  variance  of  the  Zim  which  is  the  density  of  H^.  The  effect  of 
increasing  the  values  N  and  S  for  the  noncentral  chi-square  can  be  seen 


59 


Figure  4.23:  vs.  P,  for  the  White  Gaussian  Signal  with  M  =  8,  N  =  64,  S  =  250,  and 
201og(a). 

to  directly  increase  the  noncentrality  parameter  X  =  2a’NMS.  Notice  that  the  sampling 
frequency,  f^,  divided  by  NS  gives  the  frequency  resolution  of  the  ambiguity  function.  If 
we  keep  the  same  frequency  resolution,  then  any  changes  of  N  or  S  which  are  governed 
by  the  equation  NS  =  f^  will  not  change  the  probability  of  detection.  Now  the  length  of 
the  time  series  from  which  we  generate  the  ambiguity  function  is  NMS.  This  means  for  a 
constant  time-bandwidth  product  the  only  interesting  tradeoff  to  investigate  is  between  N 
or  S  and  M.  It  is  obvious  that  since  the  time-bandwidth  product  is  constant  that  it  is  ideal 
to  pick  M  as  small  as  possible  because  this  will  decrease  the  variance  of  the  clutter. 
Figure  4.24  shows  this  for  =  128,000,  and  P^  =  10  ^ 

Next,  we  determine  the  performance  characteristics  for  the  threshold 
described  by  the  second  equation  of  4.29.  Notice  that  it  requires  a  two  sided  threshold. 
However,  because  of  the  symmetry  of  the  gaussian  random  variable  we  only  consider  the 
one  sided  case  of  the  target  being  located  at  a  negative  frequency.  This  test  requires  that 


60 


Figure  4.24:  Plot  of  vs.  a  for  constant  P,  =  10'*'  and  =  128,000.  M  is  indicated  on 
the  plot. 

M>2  because  of  the  approximation  of  the  chi  density  by  a  Gaussian.  However,  it  is 
useful  for  two  reasons.  First,  the  Gaussian  probability  density  is  simple  to  work  with 
especially  in  determining  the  Pf.  Second,  section  4. 1 .2  showed  that  the  actual  FM  radio 
data  appeared  to  be  more  accurately  characterized  by  a  Gaussian  density  if  the 
frequencies  were  subtracted.  Figure  4.25  shows  the  same  plot  as  figure  4.22  except  the 
densities  are  now  Gaussians  with  means  and  variances  described  by  equations  4.26  and 
4.29.  It  is  useful  to  determine  a  formula  which  can  be  used  to  determine  the  minimum 
detectable  scattered  signal  as  a  function  of  P^  P^  M,  N,  and  S.  This  task  can  be 
accomplished  by  considering  the  two  Gaussian  densities.  First  we  set  the  false  alarm  rate 
by  setting  the  threshold  level  at  a  multiple  of  the  standard  deviation  of  the  density  that 
corresponds  to  no  target  present.  Next,  we  set  the  minimum  probability  of  detection  by 
subtracting  a  multiple  of  the  standard  deviation  from  the  mean  of  the  distribution 
corresponding  to  the  target.  We  then 


61 


Figure  4.25:  vs.  for  the  White  Gaussian  Signal  with  M  =  8,  N  =  64,  S  =  250,  and 
201og(a). 


equate  the  results.  This  results  in: 

li-nG'=k<y  (4.33) 

where  and  a'  are  given  by  equation  4.3 1  and  a  is  approximated  as  one  from  equation 
4.30.  We  can  then  substitute  these  values  into  4.33  which  gives  the  following  result: 

A  =  ]  +  3.464nVM  +  2.828jfe  Vm  +  +  2A49nk  +  k '  (4.34) 


However,  we  may  rewrite  X  =  2a^NMS  =  2a^A5.  This  means  we  can  now  solve  for  the 
minimum  scattering  amplitude  that  is  detectable  as: 


M>2 


= 


1  +  2>A64n^[M  +  22>l%k^  +  \5n^  +  2.449«il:  +  k‘ 

2A. 


(4.35) 


The  probability  of  detection  is  given  by  Pj  =  erf(x)+0.5  and  the  probability  of  false  alarm 
is  given  by  P,  =  0.5  -  erf(k)  where  we  are  using  the  definition  [10]: 


The  following  table  shows  how  the  minimum  detectable  scattering  amplitude  decreases 
as  the  observation  time  increases  for  a  single  target: 


Table  4,4:  Minimum  detectable  scattering  amplitude  for  WG  sampled  at  250  kHz 


T  (seconds) 

M 

Pa 

Pr 

201og(a„.„) 

0.5 

8 

0.841 

10'" 

-34.71 

0.5 

4 

0.841 

-35.50 

1 

16 

0.841 

-36.808 

1 

4 

0.841 

-38.51 

2 

32 

0.841 

10'' 

-38.79 

2 

4 

0.841 

10'" 

-41.52 

4 

64 

0.841 

10^ 

-40.66 

4 

4 

0.841 

lO*" 

-43.742 

8 

4 

0.841 

-46.75 

16 

4 

0.841 

-50.6 

Notice  doubling  the  observation  time  T  doubles  the  sensitivity  of  the  radar  because  it 
doubles  the  time-bandwidth  product  when  no  change  in  the  number  of  incoherent 
averages  occurs.  This  is  what  we  would  expect  because  this  doubles  the  energy  of  the 
scattered  signal  while  the  increase  in  NS  appropriately  scales  the  clutter  power  to  keep  its 
contribution  constant.  This  table  also  indicates  that  even  at  a  bandwidth  of  250  kHz  that 
the  time  length  of  the  series  must  extremely  long  in  order  for  the  SDRs  of  -60  dB  to  be 


detected. 


63 


4.3  ADJUSTMENTS  FOR  THE  FM  WAVEFORM 

The  previous  section  discussed  the  ideal  operating  characteristics  for  a 
radar  with  a  white  Gaussian  signal  with  no  clutter  besides  itself  and  a  single  scattered 
signal.  Unfortunately,  this  is  not  an  ideal  world  and  neither  the  signal  of  interest,  FM 
broadcasts,  is  not  a  white  Gaussian  process  and  of  course  there  are  buildings,  cars,  and 
the  earth  all  of  which  introduce  clutter  into  the  detection  problem.  Finally,  there  are 
receiver  noise  sources  which  will  also  increase  the  variance  of  the  clutter  floor.  In  this 
section  we  discuss  how  the  non-ideal  world  effects  our  calculations  and  we  propose  a 
crude  phenomenological  correction  to  our  equations  which  should  allow  order  of 
magnitude  estimates  to  be  made. 

The  non-ideal  factors  of  the  FM  broadcast  can  be  broken  into  three 

categories: 

1)  The  FM  broadcast  is  not  White  Gaussian  noise.  However,  the  Central 
Limit  Theorem  can  still  be  applied  because  over  time  intervals  longer 
than  approximately  20  microseconds  the  data  is  uncorrelated.  The 
difference  is  that  the  calculation  of  the  variance  must  be  modified  to 
account  for  correlation.  This  can  also  be  viewed  as  bandwidth 
inefficiency.  If  the  broadcast  were  sampled  at  a  slower  rate  of  30  kHz, 
then  it  would  be  much  closer  to  being  a  white  process. 

2)  The  FM  broadcast  is  not  just  the  direct  broadcast  but  also  includes 
scatter  from  the  earth,  buildings,  and  cars  as  well  as  the  airplane  we  are 
interested  in.  This  scatter  will  be  mostly  concentrated  near  zero  range 
so  it  should  not  introduce  much  bias.  However,  it  will  increase  the 
overall  variance  of  the  process. 

3)  The  target  of  interest  is  an  aircraft  and  in  general  is  moving  (possibly 
accelerating).  This  means  the  scattering  amplitude  is  best  modeled  as 
a  Rayleigh  random  variable  with  uniformly  distributed  phase  [8]. 

First,  we  examine  the  increase  in  variance  due  to  using  a  signal  with  nonzero  correlation 
for  lags  other  than  zero.  It  can  easily  be  shown  that  the  variance  of  the  sum  of  two 
random  variables  which  are  correlated  can  be  given  as: 


64 


E{{x  +  yY}  =  E{x-}+2E{xy}  +  E{y-]  = 
<7]  +  Ircr^cr^.  +  =  2cr^  (1  +  r) 


where  r,  the  correlation  coefficient,  can  be  inferred  from  4.37  and  is  between  0  and  1.  The 
correlation  this  introduces  can  be  represented  as  follows: 


Ict^N 

MS 


■+r  = 


2N 

MS 


+ 


MS 

2N 


r)  = 


(4.38) 


where  cr^  is  the  rescaled  value  of  the  semivariance  of  the  signal.  Notice  that  this  can  also 
be  expanded  to  include  the  increase  in  variance  due  to  physical  clutter  such  as  buildings, 
cars,  ect.  This  does  not  change  the  value  of  O  that  is  generated  by  equation  4.25.  This 
means  that  these  two  effects  can  be  accounted  for  by  adjusting  the  noncentrality 
parameter  as  follows: 

X=Q.Ig‘1  =2a^As{a*  I  (j‘1)  =  2a^ (4.39) 

where  P  is  the  ratio  of  the  true  transmitted  signal’s  variance,  to  the  rescaled  variance 
given  by  4.38.  We  suggest  here  that  it  is  best  to  empirically  determine  P  from  repeated 
observations  of  radio  broadcasts.  We  can  do  this  in  two  parts.  First,  we  approximate  the 
increase  in  the  semivariance  caused  by  the  clutter.  This  can  be  done  by  using  a 
directional  antenna  or  adaptive  beamforming  to  isolate  the  scattered  signals.  Then  the 
semivariance  due  to  physical  clutter  only  can  be  estimated.  This  can  then  be  added  to  the 
direct  signal’s  variance  to  give  a  in  equation  4.38.  Another  way  is  to  use  a  worst  case 
scenario  estimate  for  the  physical  clutter. 

^  cr"*  2 

Now  we  can  estimate  the  ratio  B  -  —r  =  P{\  +  ^  where  ^  represents  the 

ratio  of  the  physical  clutter  to  the  broadcast  signal’s  variance.  Next  we  use  the  linearity 
of  the  expectation  operator  to  determine  p.  We  do  this  by  first  estimating  a"  from  either 


65 


the  real  or  imaginary  component  of  the  complex  data.  Then  we  use  this  estimate  to  scale 
the  ambiguity  function  by  equation  4.16.  Here  it  is  now  possible  to  estimate  p  either 
from  the  mean  of  the  chi-square  distribution  or  from  the  variance  of  the  Gaussian  density 
which  is  empirically  generated  from  the  clutter  floor.  This  means  we  have: 


XlM 


2M  2(2A/-/y;) 


P 


(1  +  ^^ 


(4.40) 


The  following  table  lists  the  estimates  for  P  of  signals  A1  -  A6  for  K=  128,000  and  M  = 

8. 


Table  4.5:  List  of  p.  the  bandwidth  inefficiency  factor,  for  signals  A1-A6 


Signal 

B  _  __ 

A1 

0.12 

A2 

0.16 

A3 

0.07 

A4 

0.12 

A5 

0.23 

A6 

0.22 

Average 

0.15 

This  suggests  P  is  probably  close  to  0.15  and  that  if  we  take  a  conservative  value  of  ^  = 

0.5  this  shows  that  p  is  close  to  0.067.  It  is  interesting  to  note  that  the  ratio  of  the  true 
bandwidth  occupied  by  the  signal,  30  kHz,  to  the  sampling  frequency,  250  kHz,  is  0.12. 

This  seems  to  indicate  that  yff  is  a  parameter  which  determines  bandwidth  inefficiency. 
This  value  can  now  be  used  to  estimate  the  corrected  value  of  the  noncentrality 
parameter.  No  change  occurs  to  the  chi-square  and  chi  densities  of  the  clutter.  Figure 
4.26  shows  the  vs.  P,  for  the  same  case  as  4.24  except  the  noncentrality  parameter  is 
corrected  for  the  value  of  P=0.067. 


66 


Figure  4.26:  Pj  vs.  for  T  =  0.5  seconds  and  M  =  8  with  p=0.067. 
The  correction  to  X  means  that  equation  4.35  is: 


M>2 


oc„ 


1  +  3A64n^fW  +  2.828i^  Vm  +\5n^  +  2A49nk  +  k " 


(4.41) 


The  results  for  p=0.067  appear  in  table  4.6  on  the  following  page.  From  this  table,  it 
appears  that  it  would  be  possible  to  detect  a  target  whose  power  relative  to  the  direct 
signal  is  -38.8  dB.  However,  we  must  recall  that  the  aircraft  will  move  over  4  km  during 
this  time  period  and  could  change  its  velocity  by  157  m/s  if  its  limited  to  a  one  g 
maneuver.  This  means  table  4.6  and  equation  4.41  should  still  only  be  taken  as  lower 
bounds  for  the  minimum  detectable  scattered  signal. 


67 


Table  4.6:  Minimum  detectable  scattering  amplitude  for  FM  signal  sampled  @  250  kHz 


T  (seconds) 

M 

P, 

P, 

0.5 

8 

0.841 

10'^ 

-23.0 

0.5 

4 

0.841 

10' 

-23.8 

1 

16 

0.841 

10' 

-25.1 

1 

4 

0.841 

-26.8 

2 

32 

0.841 

-27.1 

2 

4 

0.841 

10' 

-29.8 

4 

64 

0.841 

-28.9 

4 

4 

0.841 

10' 

-32.8 

8 

4 

0.841 

-35.8 

16 

4 

0.841 

-38.8 

However,  there  exist  multi-threshold  tracking  algorithms  that  could  possibly  be  used  to 
approach  this  lower  bound.  Also,  for  short  time  periods  table  4.6  should  be  close  to  the 
real  situation  because  most  aircraft  are  limited  in  velocity  and  acceleration. 

Figure  4.27  is  a  plot  of  the  minimum  detectable  signal  for  the  white 
gaussian  signal  where  the  plot  is  IX(r,-f)l-IX(r,f)l  with  a  target  located  at  21  km  and  -125 
Hz  (a  fourier  frequency).  The  length  of  the  time  series  was  0.512  s  with  =  lO  ’  and  = 
0.5.  Figure  4.28  is  the  plot  of  IX(r,f)l^  for  the  same  signal  which  shows  that  is  important 
to  have  the  statistics  in  mind  when  making  judgments  about  detectability.  Figure  4.29  is 
the  plot  of  signal  A3  with  an  computer  generated  scattered  signal  injected  at  the  level  of 
minimum  detectability,  P^  =  0.645  and  P^  =  10  ’  where  we  have  taken  p=0.15.  Figure 
4.30  is  signal  A3  with  a  scattered  signal  whose  SDR,  a,  is  3  dB  below  the  signal  injected 
to  make  figure  4.27. 


68 


Frequency -Hz  -500  10  Range  -  Km 


Figure  4.27:  Plot  of  IX(r,-f)l-IX(r,f)l  for  A3  =  128,000,  M=8,  P,  =  lO',  and  P,  =  0.5  for  a 
white  gaussian  signal  with  target  at  21  km  and  -125  Hz. 


Frequency -Hz  -500  10  Range  -  Km 


Figure  4.28:  Plot  of  IX(r,f)|-  for  A3  =  128,000,  M=8,  P,  =  lO',  and  P,  =  0.5  for  a  white 
gaussian  signal  with  target  at  21  km  and  -125  Hz. 


69 


Figure  4.29:  Plot  of  IX(r,-f)l-IX(r,f)l  for  =  128,000,  M=8,  =  lO’,  p=0.15  and  P,  = 

0.645  for  signal  A2  with  target  at  21  km  and  -125  Hz. 


Figure  4.30:  Same  as  figure  4.29  except  a  was  decreased  by  3  dB. 


CHAPTER  5:  ADAPTIVE  BEAMFORMING/DIRECTIONAL  ANTENNAS 


The  last  chapter  discussed  the  limitations  of  the  monostatic  FM  radar  with 
one  receiver.  It  is  obvious  that  isolating  the  scattered  signal  from  the  direct  broadcast 
signal  should  improve  the  sensitivity  of  the  radar.  Indeed,  this  is  the  impetus  for  using 
the  Cascade  Mountains  as  a  barrier  to  shield  the  receiver  at  Manastash  Ridge  from  the 
direct  FM  broadcast.  However,  it  is  possible  to  eliminate  through  spatial  isolation  the 
direct  signal  without  having  to  use  two  receivers  which  are  physically  separated  by  a 
barrier.  This  can  be  done  either  through  using  antennas  with  directional  gains  or  a 
number  of  omnidirectional  dipoles  whose  signals  are  phase  shifted  and  then  summed.  In 
the  case  of  dipoles,  the  phase  shifting  can  adapt  over  time  to  provide  a  time-varying 
optimal  response  to  the  radio  environment.  Hence,  the  name  adaptive  beamforming. 
However,  this  differs  from  the  use  of  a  physical  barrier  because  a  physical  barrier  will 
also  attenuate  random  scatter  of  the  direct  signal  by  buildings,  cars,  and  the  earth.  The 
question  of  interest  is  how  does  physical  clutter  limit  a  directional  antenna/adaptive 
beamforming  monostatic  FM  radar.  We  will  limit  our  discussion  to  adaptive 
beamforming  because  limited  experimental  data  was  available  to  explore  this  option. 
Also,  the  results  of  using  directional  antennas  should  be  quite  similar. 


5.1  SYSTEM  OVERVIEW 

The  basic  idea  of  adaptive  beamforming  can  be  seen  in  figure  5.1.  An 
array  of  more  than  one  antenna  is  arranged  in  linear  fashion  which  causes  the  phase  shift 
of  the  incident  radiation  between  antennas  to  be  given  by  (t)=kdsin(0),  the  electrical  angle 
[13].  It  is  also  possible  to  use  two  dimensional  arrays  which  would  of  course  allow 
cancellation  of  interference  in  two  dimensions.  We  assume  that  the  direction  of  the 
interference  is  unknown,  but  is  stationary  over  the  convergence  time  of  the  adaptation 


A. 

H -  d - ► 


Figure  5.1:  Geometry  of  adaptive  beamforming  array 

interest.  This  means  that  if  we  let  the  signal  vector  be  defined  as  x(t)  =  [x,(t)  Xj(t),  ...  , 
x„.|(t)]^  then  we  should  be  able  to  attenuate  the  interference  by  minimizing  the  variance  of 

e(t)  =  x„(t)  -  w"(t)x(t)  (5.1) 

where  w’^(t)  =  [w,(t),  w^Ct), ...  ,  w„  ,(t)]  is  a  set  of  weights  which  can  slowly  vary  in  time, 
Xo(t)  is  viewed  as  the  desired  response,  and  e(t)  can  be  viewed  as  the  error  signal. 
Traditional  adaptive  beamforming  also  places  a  constraint  on  the  weight  vectors  to 
preserve  a  unity  gain  in  a  predetermined  look  direction.  However,  this  also  requires  a 
minimum  of  three  receivers  to  implement.  Hence,  we  will  concern  ourselves  only  with 
minimizing  modulus  square  of  e(t). 

In  general,  there  are  two  traditional  approaches  of  approaching  the 
minimization  problem.  The  first  is  to  view  the  squared  error  le(t)l^  as  a  multivariate 


72 


function  of  the  adaptive  filter’s  weights.  The  instantaneous  gradient  can  be  formed  which 
is  then  used  to  update  the  filter  weights  so  that  they  follow  the  path  of  steepest  descent. 
This  is  the  reason  the  oldest  version  of  the  algorithm  is  called  the  method  of  steepest 
descent  [13].  We  will  use  a  more  widely  used  algorithm  which  is  recursive  and  is  easy  to 
use  called  the  least  mean  squares  algorithm  (LMS).  The  second  general  approach  to  this 
problem  is  to  recognize  that  this  problem  is  the  same  as  a  linear  least  squares  problem. 
This  can  be  implemented  using  algorithms  such  as  recursive  least  squares,  QR- 
decomposition,  ect  [13].  However,  these  algorithms  are  more  computationally  complex 
and  are  harder  to  implement  than  LMS..  The  reason  the  recursive  least  squares  and  QR- 
decomposition  algorithms  exist  is  because  they  offer  better  convergence,  but  it  was  found 
the  normalized  LMS  algorithm  (a  close  variant  to  LMS)  was  sufficiently  convergent  for 
this  problem.  For  these  reasons,  and  because  the  experimental  data  that  will  be  used  only 
comes  from  two  antennas  we  chose  only  to  examine  the  LMS  algorithm  and  its  close 
relative  the  normalized  LMS  algorithm 

The  complete  derivation  of  the  complex  LMS  algorithm  is  given  in 
Haykin  where  it  is  shown  that  it  results  in  three  basic  steps  [13]: 

1)  Filter  output:  g(t)  =  w"(/)x(r) 

2)  Estimation  error:  e(t)  =  x^Ct)  -  g(t) 

3)  Tap-weight  adaptation:  w(r  -i- 1)  =  w(/)  +  fjs.{t)e*  {t) 

This  is  simple  to  program  especially  in  the  case  of  two  antennas  because  in  that  case  the 
vectors  reduce  to  scalars.  The  quantity  p  is  the  step-size  parameter,  and  determines  the 
convergence  properties  of  the  LMS  algorithm.  Picking  a  value  of  p  which  is  to  large  will 
cause  the  algorithm  to  diverge  and  picking  a  value  which  is  to  small  will  make  the 
adaptation  time  to  long  to  be  useful.  It  was  found  through  trial  and  error  that  using  this 
form  of  the  algorithm  on  the  set  of  radio  broadcasts  that  were  sampled  did  not  yield 
useful  convergence  properties.  However,  if  we  modify  the  LMS  algorithm  by 


73 


normalizing  the  step-size  parameter  by  the  squared  magnitude  of  the  data  vector,  x(t), 
then  the  algorithm  satisfactorily  converged.  Thus,  the  normalized  least  mean  squares 
algorithm  that  was  actually  implemented  in  the  following  analysis  modifies  step  3  by 
[13]: 


3)  w(r -h  1)  =  w(r) + 


||x(r)||  +  k 


x(0e*(0 


where  ju  is  the  normalized  step  size  and  k  is  constant  which  insures  that  the  denominator 
cannot  become  to  close  to  zero.  The  values  of  ^  and  k  that  were  used  were  1  and  100, 
respectively. 


In  contrast  to  most  adaptive  filtering  problems,  the  result  that  we  want  is 
not  the  filter  output.  Instead  it  is  the  error  sequence  that  is  of  interest.  The  filter  adjusts 
its  tap  weights  so  that  its  output  closely  models  the  main  components  of  x„(t)  which 
should  be  the  direct  signal  and  any  strong  clutter  sources.  Subtracting  the  two  should 
leave  scattered  signals  which  arrived  from  different  directions  than  the  direct  signal.  This 
is  of  course  the  error  sequence  of  equation  5.1,  and  is  where  we  will  start  the  analysis  of 
how  adaptive  beamforming  will  affect  the  sensitivity  of  the  FM  radar. 


5.2  ANALYSIS  OF  ADAPTIVE  BEAMFORMING  EFFECTS  ON  SENSITIVITY 

To  analyze  the  effects  that  adaptive  beamforming  has  on  the  sensitivity  of 
the  FM  radar  system  we  limit  ourselves  to  the  simple  case  of  two  receivers.  We  do  this 
because  the  direct  signal  is  the  major  source  of  “interference”  in  this  problem  and  because 
it  should  be  easy  to  generalize  from  this  example  to  more  than  two  receivers.  Also,  the 
currently  available  experimental  data  was  for  only  two  receivers. 

First,  we  write  the  two  signals  without  including  the  noise  terms  because 
we  will  assume  the  clutter,  direct,  and  scattered  signal  are  all  much  greater  than  the  noise. 
From  chapter  2,  we  have: 


74 


XQ{t)  =  u(t)+  j^^(r,t)u(t-r)dr 

00 

x^(t)  =  e^^u(t)+  j(i>2(r,t)u(t  -r)dr 

R<, 


(5.2) 


where  ^  is  the  electrical  angle,  and  0,(t,r)  and  <l>2(t,r)  are  the  scattering  amplitudes  of  the 
clutter  and  targets  which  are  different  because  of  the  phase  shift  which  is  presumed 
unknown  but  different  from  the  direct  signal.  If  u(t)  is  significantly  larger  than  the 
clutter,  then  we  can  assume  the  single  spatial  weight  should  converge  to  approximately  e' 
,  but  there  will  be  some  error,  e.  This  means  the  resulting  error  term  e(t)  which  now  we 
will  call  y(t)  is  given  as: 


y(t)  =  su(t)+  (5.3) 

Next,  we  form  the  product  y(t)x*(t-rQ)  as  is  done  in  the  previous  chapter’s  analysis.  This 
results  in: 

CO 

y(t)x\t  -  To)  =  m{t)u  (t  -  To)  +  £  jd)*(r'-ro,OM(OM*(^  -  r'-rQ)dr'+ 

Rj 

00 

Jm*  (f  -  ro)M(t  -  r)[<I),  (r,0  -  (r,r)](ir  +  (5.4) 

Rd 
00  00 

J  \u{t  -  r)u  {t  -  2(^d)]dr 

^d 

This  can  now  be  compared  to  the  sequence  a(t)  as  defined  in  the  previous  chapter.  Notice 
that  if  we  assume  l£l«l,  then  the  dominant  terms  in  the  sequence  are: 


75 


(5.4) 

OO  00 

j  ju(t  -  r)u  (t  -  r'-r^  )<t>^  {r'-r^ , 0[<I> i  (r, 0  “  2  ^ 

This  obviously  shows  that  the  clutter  contribution  due  to  the  direct  signal  with  itself  is 
dramatically  reduced  if  e  is  small.  Instead,  now  the  important  contribution  to  the  variance 
of  the  probability  distributions  is  the  physical  clutter.  This  shows  that  in  the  case  of  near 
ideal  adaptive  beamforming  that  the  FM  radar  will  be  limited  by  the  physical  clutter  in 
the  environment.  Also,  since  the  target(s)  of  interest  is  implicitly  included  in  the 
scattering  amplitude,  we  see  that  another  effect  is  that  it  can  reduce  the  scattering 
amplitude  of  the  target  of  interest  by: 

=  (5.5) 

where  <1).^  is  the  electrical  angle  of  the  target  of  interest  and  (|)  is  the  electrical  angle 
corresponding  to  the  direct  broadcast.  The  value  of  k  ean  be  between  0  and  2.  This  also 
occurs  for  any  physical  clutter  in  the  problem  and  shows  that  adaptive  beamforming 
could  actually  increase  the  contribution  to  the  variance  by  some  of  the  physical  clutter. 

Now  we  use  the  factor  ^  from  before  to  represent  the  ratio  of  the  physical  clutter  to  the 
broadcast  signal’s  power  as  given  by: 

^  =  std[RQ{y{t)x*{t-rQ)]]l (5  6) 

V2crf  =  St J[Re  {x(/)x*(t-ro)}] 


then  we  may  write  the  variance  of  the  unnormalized  chi-square  as: 

MSr  2ecT:N 

MS  MS  ^  MSj3 


(5.7) 


Then  we  can  rewrite  the  noncentrality  parameter  as: 


76 


X  =  m  a]  =2K-a^Tpi  (5.8) 

where  k  is  the  loss  factor  due  to  the  distortion  of  the  beam  in  the  direction  of  the  target,  ^ 
is  the  gain  due  to  cancellation  of  the  direct  signal,  and  P  is  the  loss  factor  as  defined 
before.  This  means  we  can  rewrite  equation  4.46  for  the  case  of  adaptive  beamforming 
as: 


,  1  +  3.464nVM  +  2.828jt Vm  +I5n^  +  2AA9nk  +  k-  . 

M>2  aL= - (5-9) 

Notice  that  if  <  1  for  most  electrical  angles  then  adaptive  beamforming  should  be 
useful  barring  a  dramatic  change  in  p. 


5.3  EXPERIMENTAL  RESULTS 

The  previous  section  showed  that  the  key  to  understanding  the 
improvement  offered  by  adaptive  beamforming  or  directional  antennas  is  the 
characterization  of  the  typical  physical  clutter.  The  best  way  to  do  this  is  simply  to  turn 
on  the  receivers  and  make  measurements  of  the  clutter  environment  using  adaptive 
beamforming.  It  should  then  be  possible  to  estimate 

Estimating  ^  requires  that  we  first  find  the  standard  deviation  of  the  direct 
signal  plus  the  physical  clutter  and  then  use  adaptive  beamforming  to  estimate  the 
standard  deviation  of  the  physical  clutter  only.  This  results  in  the  following  estimate: 


direct  __  ^  scatt+direct 


direct 


k  - 1 


(5.9) 


where  is  the  standard  deviation  of  the  u(t)u*(t-r(,)  where  u(t)  is  the  direct  signal. 


77 


Only  one  set  of  data  was  available  to  determine  whether  adaptive 
beamforming  improves  the  sensitivity  of  the  radar.  The  results  were  encouraging.  The 
data  is  signal  A2  and  a  new  data  set  B2  which  was  synchronously  sampled  by  a  different 
receiver.  The  baseline  was  1.5  meters.  After  the  data  was  passed  through  the  adaptive 
beamforming  algorithm  the  standard  deviation  of  the  real  and  imaginary  parts  were 
computed.  It  was  found  that  they  were  both  approximately  10.  In  contrast,  signal  A 
which  was  used  as  the  reference  signal  had  a  standard  deviation  that  for  both  in-phase  and 
quadrature  was  30.  This  gives  a  ratio  of  1/3.  This  meant  that  ^  should  be  1/9.  However, 
a  calculation  of  ^  for  r^  =  21  km  showed  that  it  remained  about  1/3.  This  indicates  that 
the  contribution  of  the  physical  clutter  to  the  standard  deviation  is  more  than  is  predicted 
and  probably  occurs  because  of  more  correlation  due  to  physical  clutter. 

However,  ^  being  close  to  1/3  shows  that  some  improvement  is  possible. 
Notice  that  this  means  it  could  improve  the  sensitivity  by  a  factor  of  1/9  which  is  about 
9.5  dB.  To  test  this  we  injected  the  same  target  into  both  A2  and  B2  that  was  used  in 
figure  4.28.  Figure  5.1  shows  that  it  is  plainly  distinguishable.  Figure  5.2  shows  a  target 
3.3  dB  below  the  one  used  in  figure  5.1.  Notice  it  is  still  visible.  This  seems  to  indicate 
at  least  a  6  dB  gain  in  this  case.  It  was  also  known  that  the  signals  used  were  quite  noisy. 
With  better  receivers  it  may  be  possible  to  improve  the  gain  of  the  adaptive  beamforming 
system. 


78 


Frequency -Hz  lO  Range  -  Km 


Figure  5.2:  Plot  of  IX(r,-f)l  -  IX(r,f)l  for  the  cross  ambiguity  plot  of  the  adaptive 
beamformed  signal  with  A2.  Ag  =128,000,  M  =  8,  and  the  amplitude  level  the  same  as 
4.30. 


Figure  5.3:  Plot  of  IX(r,-f)l  -  IX(r,f)l  with  the  same  parameters  as  figure  5.2  except  the 
scattering  amplitude  is  3.3  dB  less. 


CHAPTER  6;  RESULTS  AND  CONCLUSION 


Throughout  this  thesis  we  have  been  exploring  the  sensitivity  of  the 
monostatic  FM  radar.  The  key  issue  in  determining  the  sensitivity  is  the  ratio  of  the 
scattered  signal  to  the  direct  signal.  The  introduction  shows  that  to  be  useful  the  bistatic 
system  must  be  able  to  detect  signals  whose  SDR  are  less  than  -54  dB.  Unfortunately, 
because  the  bandwidth  of  the  FM  broadcasts  is  insufficient  it  is  probably  not  possible  to 
us  FM  radio  broadcasts  to  detect  aircraft.  However,  if  control  of  the  transmitting 
waveform  is  possible  then  a  white  Gaussian  noise  signal  with  sufficient  bandwidth  should 
be  possible. 

6. 1  WHITE  GAUSSIAN  SIGNAL  RADAR  NETWORKS 

Equation  4.35  which  determines  the  minimum  SDR  shows  the  strong 
dependence  of  sensitivity  on  bandwidth.  Throughout  the  thesis  we  have  constrained 
ourselves  to  processes  with  bandwidths  on  the  order  of  250  kHz.  Notice  that  if  the 
bandwidth  is  increased  to  10-20  Mhz,  the  radar  now  becomes  sensitive  to  targets  with 
SDRs  in  the  range  of  -60  dB.  Because  of  the  wide  bandwidth  of  these  signals,  it  is 
probably  desirable  to  move  the  carrier  frequency  into  the  3  -  10  Ghz  range  so  that  the 
narrowband  scattering  approximation  still  applies.  However,  it  obvious  there  are  two 
problems  that  are  encountered  with  the  practical  implementation  of  the  above  system. 
The  first  is  the  increase  of  noise  power  with  bandwidth,  and  the  second  is  the  increase  in 
computational  burden  as  the  bandwidth  increases. 

First,  we  examine  the  increase  of  noise  power  with  bandwidth.  If  we 
increase  the  carrier  frequency  to  3  -  10  Ghz,  then  the  noise  floor  can  be  approximated  by 
the  cosmic  background  radiation  if  we  use  good  receiver  design  [14].  This  means  that 
equation  1 .4  can  be  use  to  approximate  the  noise  floor  with  T  being  approximately  5  K. 
This  means  that  with  a  10  Mhz  system  that  the  total  noise  power  is  only  6.9x10  '^  Watts 


80 


which  is  still  far  below  the  signal  power  at  distances  of  interest.  In  fact,  increasing 
bandwidth  will  result  in  gain  versus  noise  because  from  equation  4.46  the  increase  in  gain 
is  linear  with  a  slope  of  approximately  one.  In  contrast,  the  noise  power  increases 
linearly  with  bandwidth  but  with  a  slope  of  =  6.9x10'^^  W/Hz.  The  way  to  think  about 
what  is  occurring  is  to  realize  that  by  increasing  the  bandwidth  of  the  signal  its  power  is 
being  spread  in  frequency  because  we  are  forcing  it  to  be  white.  However,  the  scattered 
signal’s  energy  still  remains  the  same.  From  this  it  follows  that  as  the  bandwidth  is 
increased  the  system  will  eventually  reach  the  point  of  being  the  noise  limited  matched 
filter  output  with  SNR  given  by: 


Where  Eg  is  the  energy  content  of  the  scattered  signal  for  the  integration  time.  Notice  that 
N(,  is  not  dependent  on  bandwidth,  and  therefore  determines  the  ultimate  sensitivity  of  the 
radar  system  for  a  fixed  transmitter  power. 

The  second  problem  is  the  issue  of  computation  of  the  ambiguity  function 
for  bandwidths  of  10  -  20  Mhz.  The  computation  requirement  of  the  ambiguity  function 
can  basically  be  broken  down  into  two  steps.  First,  we  perform  coherent  integration  to 
reduce  the  bandwidth  of  the  signal.  The  bandwidth  of  the  targets  of  interest  depend  on  the 
carrier  frequency  and  the  maximum  radial  velocity  they  obtain.  For  man-made  objects, 
we  can  take  the  maximum  radial  velocity  as  about  1000  m/s.  This  means  that  at  a  carrier 
frequency  of  3  Ghz  we  can  expect  a  maximum  doppler  shift  of  63  kHz.  This  means  our 
complex  sampling  frequency  should  be  63  kHz.  Hence,  if  we  are  operating  with  a 
bandwidth  of  2^“* «  16.78  Mhz  then  we  can  decimate  by  a  factor  of  D  =  256.  We  must  do 
that  for  each  range  of  interest  with  225  ranges  required  if  we  wish  to  use  the  maximum 
range  resolution  for  0  to  4  km.  Following  this,  each  range  must  be  fourier  transformed  to 
give  the  frequency  spectrum  of  the  scattering  amplitudes.  This  means  it  requires  4.01  x 
10'*  complex  multiplies  per  second  which  since  each  complex  multiply  requires  four  real 


81 


multiplications  translates  into  16x10*’  real  multiplications.  The  general  expression  for 
algorithms  with  iJD  =  2"  where  n  is  an  integer  is  given  for  multiplications  as: 


/«/?(!  + 


log2(A/-P) 

D 


(6.2) 


where  R  is  the  number  of  ranges  to  examine.  Notice  that  the  decimation  step,  fgR, 
dominates  if  D  is  even  moderately  large.  Thus,  although  the  operations  count  is  at  first 
apparently  daunting  it  should  be  possible  because  of  the  simplicity  of  the  decimation  step 
to  effectively  implement  the  algorithm  in  most  cases. 


6.2  CONCLUSION 

The  investigation  of  the  possibility  of  using  FM  radio  broadcasts  in  a 
monostatic  version  of  the  radar  was  shown  to  be  ineffective  primarily  because  the 
bandwidth  of  the  FM  broadcast  was  not  wide  enough  to  give  sufficient  signal  processing 
gain  to  detect  point  targets  such  as  aircraft.  However,  the  results  of  the  investigation  have 
shown  that  monostatic  radars  which  use  wide  bandwidth  white  gaussian  signals  approach 
the  maximum  SNR  available  from  a  signal  processed  by  a  matched  filter  in  the 
background  of  the  cosmic  noise  floor.  This  shows  promise  in  developing  passive  radar 
networks  and  even  passive  radars  which  would  be  capable  of  operating  without  detection. 
It  has  also  been  shown  how  to  determine  the  probability  density  functions  of  the 
ambiguity  function  which  is  used  to  detect  targets.  This  is  important  because  it  allows 
the  performance  of  the  radar  to  be  determined  as  a  function  of  the  false  alarm  probability 
and  detection  probabilities. 

As  with  any  complex  problem,  there  is  still  much  that  remains  to  be 
explored.  The  power  of  the  physical  clutter  relative  to  the  direct  signal  has  not  been 
adequately  estimated.  The  geographic  location  of  the  transmitters  and  receivers  with 
relation  to  the  target  of  interest  can  also  have  a  profound  effect  on  the  radar’s 
performance.  An  analytical  model  of  the  effect  of  the  noise  being  colored  instead  of 


82 


white  could  also  give  much  insight.  Finally,  an  actual  experiment  using  wide-bandwidth 
white  gaussian  signals  could  be  conducted. 


List  of  References 


[1]  Personnel  Correspondence  with  Frank  Lind  a  Ph.D.  candidate  in  the  Department  of 
Geophysics  at  the  University  of  Washington  who  was  currently  building  the  radar 
receivers  at  the  time  this  report  was  prepared,  August  1996  -  September  1997. 

[2]  Whalen  A.  D.,  McDonough  R.  N.,  Detection  of  Signals  in  Noise,  2"^*  ed.. 

Academic  Press,  San  Diego,  1995. 

[3]  Skolnik,  M.I.,  Introduction  to  Radar  Systems,  2"‘*  ed.,  McGraw-Hill,  New  York 
1990. 

[4]  Hall,  P.  W.  “Correlative  Range-Doppler  Detectors  and  Estimators  in  Bistatic 
Radar  using  Commercial  FM  Broadcasts,”  Masters  Thesis,  Dept.  Electrical 
Engineering,  University  of  Washington,  Seattle,  WA,  1995. 

[5]  Ishimaru,  A.,  Electromagnetic  Wave  Propagation,  Radiation,  and  Scattering, 
Prentice  Hall,  New  Jersey,  1991 . 

[6]  Hansen,  J.  M.,  “A  New  Radar  Technique  for  Remote  Sensing  of  Atmospheric 
Irregularities  by  Passive  Observation  of  the  Scattering  of  Commercial  FM 
Broadcasts,”  Masters  Thesis,  Dept.  Electrical  Engineering,  University  of 
Washington,  Seattle,  WA,  1994. 

[7]  Ishimaru,  A.,  Wave  Propagation  and  Scattering  in  Random  Media,  IEEE  Press, 
New  York,  1997. 

[8]  Levanon,  N.,  Radar  Principles,  John  Wiley  &  Sons,  New  York,  1988. 


84 


[9]  Percival  D.  B.,  Walden  A.  T.,  Spectral  Analysis  for  Physical  Applications, 
Cambridge  University  Press,  New  York,  1993. 

[10]  Papoulis,  A.,  Probability,  Random  Variables,  and  Stochastic  Processes, 
McGraw-Hill,  New  York,  1991. 

[11]  Lind  F.  D.,  Sahr  J.  D.,  “The  Manastash  Ridge  Radar:  A  Passive  Bistatic  Radar 
for  Upper  Atmospheric  Radio  Science,”  Accepted  for  publication  in  Radio 
Science,  1997. 

[12]  Abramowitz  M.,  and  I.  A.  Stegun,  eds..  Handbook  of  Mathematical  Functions, 
Applied  Mathematics  Series-55  (AMS-55).  Washington,  DC:  National  Bureau  of 
Standards  (1964).  Paperback  edition.  New  York:  Dover  (1964). 

[13]  Haykin,  Simon  S.,  Adaptive  Filter  Theory,  Prentice-Hall,  New  Jersey,  1986. 

[14]  Hagen,  Jon  B.,  Radio  Frequency  Electronics:  Circuits  and  Applications, 
Cambridge  University  Press,  1996.