Acoustic Source Loc.
-
- Beamforming Basics
I Steering Vector
Delay - Sum
- MUSIC
- G- CC / T Do A Spectrum- Blind Channel Estimation
IcfRx
Assume far field,
i.e.,
all µ . . /rays a parallel at Rx . -
¥. - -
- -ET
, Tz Tz . . . TN.: Path difference of nfnf -d
TN with T,
= @- Ddcoso
@ This causes a phase diff . between TN and T,
on ( N - 1) DWSO z-
a-.
^.
@ say transmitted signal = cos Zaftreuivedyfggnal=
T.eywsfaft.li#doso.2a)
@ at 0 denote Clay.za )
Also,
note than wso = Re{ ejo }
.
'
. ylt ) = My Re { eitaft.
eieD0 }
;,Re{eioft Eyeing }2e{ ejtft
.
| - e' INO
an~
( ¥0 ) }
@ As N increases , side lobes get smaller,
main lobesharper
:#I#N= 2 N=4 N=8
@ Also note than the same R× at direction 0,
experiences different gain / SNR with
increasing N .
@ Beam Rotation
¥##@ now
,how do you rotate the main lobe of
the beam towards Q so Rx gets Max SNR .
@ Bani idea : Add a phase shift to each
Ti such that the
dcoxscf .
2h phase difference
gets compensated .
: T,
transmits cos @aft )Tz transmits Cos @aft - dco¥2h)Tz " Cos ( Zaft - 2dcosyB=)
:
÷N' ' cos (2aft-i@Dd_qB.2.a)
@ Revived signalYH ) =Re{emftp.yneilmmxddo.eiswytzdejp
= Re { eithft
MyeIln¥.2ad( coso -
copy
when a =D , we get ylt ) - Re { ejtnft }
which Is the Max possible signal power .
@ Why Max at o =p ?
Recall,
YLH = Re { Fy,
cos ( zhft + 2×1 .d( coso . wsBD}
when Q =p ,then an Cos 2 aft coherently
add up; else its incoherent .
@ this in caned -DELAY-SVMB#MFoRM1NG_because each Tx is getting umvhi plied bya phase factor ( ej C it B) which one
then an adding up at the Rx .
@ Ang6ofAmivd'(A# TO
#40R,Rz- - -
-Rn. ,
Ypp-
wszhft u ft-
|
YR•,Gsthftt 0 ej0left;;¥±f" leg
lay
ejln. Dg
YRn.ie0
HE.tl#:Iendnn
for multiple sources :
*#Ro 14 Rz . - - - Rn .
YR ,e
°
ej 02
d;dt¥i%;::→.tt#dejln - DO ,
in - in
Array Steering matrix ( m ) source
output signals .
@ so how do you pull out individual signals ?
Ee " ' e. '. .
.dk#Tfggrng..)=e*e*i..em*Tien.
= Large value ifAOA is ialons 01 ,
MUSIC- m antenna
se = As theon samples
+ \A. strewing matrix
. . . . . .
In i :Lja . .
]+ !!m
I 2. . . .
d ~
T Gunman
multipathH
sen = ( As + n ) ( As + myH
= ( As ) ( As )"
t Asn" mfs)
"- inn
't
= As SHAH+ Asn 't-1ns HAH +n④
R×× = A RssAH
t Asn "t n SHAH + Run-
why Efren" ) -
- htt
Ef Asn't ) t EJNSHAHJ
mxmfmtxd
) I dxdlldxm) t
MX me
mxm diagonal matrix
but rank -- d rank = m
MUSIC :
-
⑦ Ima→ AoA spectrum
'
d'
→-
③ ifI39. - - - . .MP
Im×,
-_ [ M Nz Nz - -. am )
T
③ Model : Xm×,
= Am xd. Sdx ,
+hmm
③ Amid =
( Algo, Alya-
. -
Agha)where Aoi =
-
Lj duos Oi
l:÷÷÷:::band Sdn
=/ ÷:)si = source signal
③ nma
=
IFgm) n ; is noise at antenna i
⑨ EIXXH) -
-
EffAs + on ) C As + n )
"
]=ELLA.
Stn ) ( SHAH + NH ) ]= Ef As SHAH + A
.sn#+msHAH+mnHJ--ELAssHAHJto t O + E
fun 't )
R×× = A Rss AH + r ? I
- --
mxm r ? IAM xd . Rssdxd A "d×m m×m
- -
this matrix isQmxm Rank ( I ) .
- M
and Rank(9) = d
⑦ def ( A. Rss .A
't
) = O since Osman hasrank dem
.
.
'
.det ( R×× - r
' I ) = O = det ( AR ss AH)
This implies : TZ is a eigenvalue .
Also,
This is minimum eigenvalue .
Now, Rxx .
e ; = Xi Lei ⇒ teigen Eq?
. : ( A Rss AH + r2 1) ei = Xi Zei
For minimum Xi = Amin,
⇐Rss A 'tt Amin 2) enifAmin Zemin⇒ A Rss AHemin O
since Rss is rank d,
and A has no
will space ,
ooo AH. emin= O
i. e, emin vectors I to the mixing matrix
columns.
H.
'
. For the correct Aoi , Aoi . emin = O
HFor wrong Aoi ,
Ace, emin should have
Some won - Zero
take
! AoA spectrum PLO ;) = 1-
HA"oi .
enoisellI I
I x m m x ( m - d)-
.
no .÷'ll liftedAo Az
" " mkhhn.
Oi
③ Questions around music
9 ① If sources are not independent Leg .
, multipath )where win the math break ?
Q ② If d > M, why does the math break ?
Q③ '
Why is The A matrix independent ?
A ③ Vandeven oud matrices are of the form
l÷÷÷÷÷÷÷÷÷u÷.
the proof is that The det ( Vandegrift ) = IT Gi - Lj)which is only 0 only when ai
= Lj
@ Generalized Cross Correlation ( Gcc ) :
÷"#\@ y.lt
)A #ym@ Yzltd
#
@ The Gcc algorithm :
T.FI/GTctsTDoa7- Spectrum
@ Ga pseudo code :
Input : Yilt ), yzlt )
i. Yi ( f ) = FFT (Y , ( t ) )
2 . Yzlf ) = FFT ( YZLTD3 '
For increasing values Of T
G ( T ) = g§g
Yi(f)Yz(f)d2N=| Y, (f) Yz* ( f ) )
4 . gtt ) = IFFT ( GCE ) )End
.
@ ^ gk )
•
r
t.tt#lnsLandTDOA
E
- The main peak says that The Los paths at
2 microphones differ by delay'
a '
.
- However,
Its not clear how otherechoes differ .
- The 1st.
order reflections from The walls
mayhave a very small difference ( r ) ,
or a large difference ( t ) .
- Example :
←±=:' :÷¥§.nu#_m '
Mz
- the path length can be viewed to be from an
image of The source behind the reflectingwall
.Further away the image ,
less in
The TDOA
@ Blindchannelestimation@ lowrider 2 microphones my and mz .
@
YET.
)@ - µ
@ At each mic. we record :
Y ,= Hroony .
×
Yz = Hroomz .
×
We do not know n,
and we do not know
any hroomi
@ Question Is : Can we get a better version ofwhat the user Is saying ,
i.e.,
some estimate
of in that Is better thany .
@ obune that
Y¥ = YIHz
: Y, Hz - YZH ,
= 0
Let's also assume H , ,Hz are sparse ,
i.e.,
L,
norm Of H, , Hz are small
. ( Recall,
L,
norm
counts no. of non . zero elements
. )to norm is better but NP complete .
Thus,
now we formulate an objective f ? as
www.t?ntni+EEttktYtfntIYt#fwhere HHH =
fund(• it = 1+2=0 is minimum but infeasible
sine HHHE = 1.
• The 7 regularization adds a penalty forH A
,11
,
and HHZH,
beaming larger .
Ideally ,we should have done Lo norm
,which would reduce the # of won . zero
elements in H ,and Hz .
MUSIC Loe ab cd-
ac bdd
abe ad
② .
- ④
f
\¥F⑦
man en
TRs, , Lse )
⑨ Autocorrelation ( seem, . nm ,)
1-whnhh.tt.
BTA CTB " to "
3)It .
FEI!.
,-
- -- -
- --
-,
.
A BA C Ba Dlc . ED Dec FEED .seGFEE
Autocorrelation
A.BACBn.tt/)crsBAEDcCBaTEoDcBGFeEDc-#.A BA ( BAA Do.rs/3AEDcCBaFEoDcB
-
A BA CBAA Do.rs/3AEDcCBafEoDcB
A BA ( BAA Do.rs/3AEDcCBaFEoDcB
A BA CBAA DCBBA EDI.BA ink
A BA CBAA DCBBA D¥E⇒Dc
③ Cross correlation ( am. , Mma ) wwwLij > = speaker i
,mic
. j11 11 11 21 11 21 11 U 11 21
yeABA CBAA Do.rs/3AEDcCBaFEoDCBthis
12 12 12 12 22 12 22
f-"
fjghaA BA C Ba DCBA En BA -
Eo
✓abcdrac bd
③ abc✓ad✓↳°④←④Delay ( 11,12 ) = •
deedCd
} crossDelay Cry ,
22) = ab
Delay ( 11,4 ) = ad
Delay ( 12,22 ) = be } auto