STA

TIST

ICA

L SI

GN

AL

PRO

CE

SSIN

G F

OR

UX

O D

ISC

RIM

INA

TIO

N F

OR

ST

ATI

STIC

AL

SIG

NA

L PR

OC

ESS

ING

FO

R U

XO

DIS

CR

IMIN

ATI

ON

FO

R

NE

XT

GE

NE

RA

TIO

N S

EN

SOR

DA

TAN

EX

T-G

EN

ER

ATI

ON

SE

NSO

R D

ATA

Lesl

ie C

ollin

s, S

tacy

Tan

tum

, Cha

ndra

Thr

ockm

orto

n, J

erem

iah

Rem

us,

Did

Wi

ht#

Eik

Gik

*d

LC

iD

avid

Wrig

ht# ,

Erik

a G

aspe

rikov

a*, a

nd L

awre

nce

Car

inD

epar

tmen

tofE

lect

rical

and

Com

pute

rEng

inee

ring

Duk

eU

nive

rsity

;#U

SGS;

*Law

renc

eB

erke

ley

Nat

iona

lLab

orat

ory

Dep

artm

ent o

f Ele

ctric

al a

nd C

ompu

ter E

ngin

eerin

g, D

uke

Uni

vers

ity; #

USG

S; L

awre

nce

Ber

kele

y N

atio

nal L

abor

ator

y (M

M-1

442)

()

INTR

OD

UC

TIO

N1

UX

O C

HA

LLE

NG

ES

2 A

PPR

OA

CH

Util

tld

tti

lith

ldt

libl

diti

ih

bt

bi

dU

XO

dl

ttl

dit

INTR

OD

UC

TIO

N1

UX

O: C

HA

LLE

NG

ES

•Usi

ngc

rren

ttec

hnol

ogie

sth

eco

stof

iden

tifin

gan

ddi

spos

ing

ofU

XO

inth

eU

nite

dSt

ates

ises

timat

edto

rang

ep

to$5

00bi

llion

2 A

PPR

OA

CH

Util

ie

stat

istic

alap

proa

ches

and

the

theo

rof

optim

ale

perim

ents

tode

sign

the

best

sens

orsa

ndse

nsin

gm

odal

ities

Unt

ilre

cent

ly,d

etec

tion

algo

rithm

sco

uld

notr

elia

bly

dist

ingu

ish

betw

een

burie

dU

XO

and

clut

ter,

lead

ing

tom

any

fals

eal

arm

s.O

ver

the

last

seve

ral

year

sm

oder

nge

ophy

sica

lte

chni

ques

have

been

deve

lope

dth

atm

erge

mor

eso

phis

ticat

edse

nsor

sun

derly

ing

phys

ical

mod

els

and

stat

istic

alsi

gnal

proc

essi

ngal

gorit

hms

Thes

ene

wap

proa

ches

•Usi

ng c

urre

nt te

chno

logi

es, t

he c

ost o

f ide

ntify

ing

and

disp

osin

g of

UX

O in

the

Uni

ted

Stat

es is

est

imat

ed to

rang

e up

to $

500

billi

on

•190

0 Fo

rmer

ly U

sed

Def

ense

Site

s (FU

DS)

and

130

Bas

e R

ealig

nmen

t and

Clo

sure

(BR

AC

) ins

talla

tions

that

nee

d to

be

clea

red

•Det

ectio

nis

nott

hebo

ttlen

eck

prim

ary

cont

ribut

orto

cost

sand

time

isth

ehi

ghfa

lse

alar

mra

te

•Util

ize

stat

istic

al a

ppro

ache

s and

the

theo

ry o

f opt

imal

exp

erim

ents

to d

esig

n th

e be

st se

nsor

s and

sens

ing

mod

aliti

es,

both

tem

pora

l/fre

quen

cy d

omai

n sa

mpl

ing

and

spat

ial s

ampl

ing.

•Use

phen

omen

olog

ical

and

empi

rical

mod

elsi

nth

ein

vers

ion

proc

esst

oge

nera

tea

rich

and

dive

rse

seto

ffea

ture

sso

phis

ticat

edse

nsor

s,un

derly

ing

phys

ical

mod

els,

and

stat

istic

alsi

gnal

proc

essi

ngal

gorit

hms.

Thes

ene

wap

proa

ches

have

dram

atic

ally

redu

ced

fals

eal

arm

rate

s,al

thou

ghfo

rthe

mos

tpar

tthe

yha

vebe

enap

plie

dto

data

colle

cted

atsi

tes

with

rela

tivel

ybe

nign

topo

logy

Toad

dres

sth

ese

prob

lem

sSE

RD

Pan

dES

TCP

have

been

supp

ortin

gef

forts

to

•Det

ectio

n is

not

the

bottl

enec

k -p

rimar

y co

ntrib

utor

to c

osts

and

tim

e is

the

high

fals

e-al

arm

rate

•Dig

ital g

eoph

ysic

s and

stat

istic

al p

roce

ssin

g ha

ve sh

own

som

e pr

omis

e•H

owm

uch

can

perf

orm

ance

been

hanc

edus

ing

mul

ti-ax

issy

stem

sand

optim

alpr

oces

sing

?

•Use

phe

nom

enol

ogic

al a

nd e

mpi

rical

mod

els i

n th

e in

vers

ion

proc

ess t

o ge

nera

te a

rich

and

div

erse

set o

f fea

ture

s.•U

se a

dvan

ced

feat

ure

sele

ctio

n al

gorit

hms t

o as

sess

the

best

and

mos

t rob

ust f

eatu

res.

•Use

stat

istic

alte

chni

ques

durin

gth

ein

vers

ion

proc

esst

om

itiga

tepo

sitio

nalu

ncer

tain

ties

with

rela

tivel

ybe

nign

topo

logy

.To

addr

ess

thes

epr

oble

ms,

SER

DP

and

ESTC

Pha

vebe

ensu

ppor

ting

effo

rtsto

deve

lop

ane

wge

nera

tion

ofU

XO

sens

ors

that

will

prod

uce

data

stre

ams

ofm

ulti-

axis

vect

oror

grad

iom

etric

mea

sure

men

tsTh

efo

cus

ofth

ere

sear

chth

atw

ew

illpr

esen

the

reis

onde

velo

pmen

tof

new

phys

ics-

base

dsi

gnal

How

muc

h ca

n pe

rfor

man

ce b

e en

hanc

ed u

sing

mul

ti-ax

is sy

stem

s and

opt

imal

pro

cess

ing?

•Use

stat

istic

al te

chni

ques

dur

ing

the

inve

rsio

n pr

oces

s to

miti

gate

pos

ition

al u

ncer

tain

ties.

•Des

ign

optim

al in

vers

ion

stra

tegi

es fo

r mul

ti-ax

is d

ata

(the

mul

tiple

loca

l min

imum

pro

blem

is m

ore

pron

ounc

ed in

th

isric

herd

ata

set)

LBL

BU

D A

EM

USG

S A

LLTE

M

mea

sure

men

ts.

The

focu

sof

the

rese

arch

that

we

will

pres

ent

here

ison

deve

lopm

ent

ofne

wph

ysic

sba

sed

sign

alpr

oces

sing

appr

oach

esap

plic

able

toth

epr

oble

min

whi

chve

ctor

data

isav

aila

ble

from

such

sens

ors.

this

rich

er d

ata

set).

•Wor

k w

ith se

nsor

dev

elop

ers t

o un

ders

tand

the

field

dat

a, in

clud

ing

prep

roce

ssin

g ne

cess

ary

for o

ptim

al u

se.

•Con

side

r a w

ide

varie

ty o

f cla

ssifi

ers,

both

trad

ition

al a

nd a

ctiv

e, to

est

imat

e pe

rfor

man

ce g

ain

with

mul

ti ax

is se

nsor

s M

ulti-

Axi

s Sy

stem

Mul

ti-A

xis

Syst

em

Spec

ifica

lly,

we

will

pres

ent

mod

elin

gan

dpr

oces

sing

resu

ltsob

tain

edus

ing

stat

eof

the

art

mul

ti-ax

isse

nsor

sde

velo

ped

byLB

Lan

dU

SGS.

Firs

t,w

ede

mon

stra

teth

atut

iliza

tion

ofth

eph

enom

enol

ogic

alm

odel

sdev

elop

eddu

ring

INVE

RSI

ON

MET

HO

DO

LOG

Y

y,

,p

gan

d as

sess

per

form

ance

on

real

istic

site

s.p

y,

pg

pg

this

prog

ram

for

data

inve

rsio

nre

sults

inim

prov

eddi

scrim

inat

ion

perf

orm

ance

over

inve

rsio

nst

rate

gies

that

use

sim

plifi

edm

odel

s.W

eal

soco

nsid

erth

eim

pact

ofre

laxi

ngth

eas

sum

ptio

nof

asy

mm

etric

obje

ctin

the

inve

rsio

n

INVE

RSI

ON

MET

HO

DO

LOG

Y

proc

ess,

and

dem

onst

rate

impr

oved

clas

sific

atio

nre

sults

.W

eca

refu

llyco

nsid

erop

tions

fort

hein

vers

ion

proc

ess,

and

dem

onst

rate

that

care

fuld

ata

sele

ctio

nca

nim

pact

perf

orm

ance

quite

sign

ifica

ntly

.In

addi

tion,

we

also

repo

rton

new

Raw

Sen

sor

Dat

aPh

enom

enol

ogic

al

Mod

elO

bjec

t Fe

atur

escl

assi

fierw

ork.

Res

ults

are

pres

ente

dfo

rtes

tsta

ndda

tafr

omth

eA

LLTE

Msy

stem

and

Cam

pSi

bert

data

fort

heB

UD

syst

em.

Iter

ativ

e se

arch

3 M

ULT

IA

XIS

EM

I SY

STE

M M

OD

ELI

NG

5 LB

L B

UD

FIE

LD D

ATA

: CA

MP

SIB

ER

T D

ISC

RIM

INA

TIO

N S

TUD

Y3

MU

LTI-

AX

IS E

MI

SYST

EM

MO

DE

LIN

G5

LBL

BU

D F

IELD

DA

TA: C

AM

P SI

BE

RT

DIS

CR

IMIN

ATI

ON

STU

DY

GEN

ERA

LIZE

D T

IME-

DO

MA

IN M

AG

NET

IZAT

ION

TEN

SOR

MO

DEL

MO

DEL

CO

MPA

RIS

ON

:FE

ATU

RE

SELE

CTI

ON

MO

DEL

CO

MPA

RIS

ON

:C

LASS

IFIC

ATIO

NPE

RFO

RM

AN

CE

FEAT

UR

E G

ENER

ATIO

NFE

ATU

RE

SELE

CTI

ON

=

0)

(0

00

)(

)(

2

1

tf

tf

tM

CLA

SSIF

ICAT

ION

PER

FOR

MA

NC

EU

GO

Gen

eral

Mag

netiz

atio

nTe

nsor

Mod

el

)

(0

00

)(

0)

(

3

2

tf

tf

tM

Gen

eral

Mag

netiz

atio

n Te

nsor

Mod

el•T

he e

stim

ated

func

tions

f n(t)

in th

e m

agne

tizat

ion

tens

orD

ipol

eM

agne

tizat

ion

Tens

orM

odel

LBL

BU

D A

EM M

ulti-

Axi

s Sy

stem

USG

S A

LLTE

M M

ulti-

Axi

s Sy

stem

Gen

eral

Mod

el:

Dip

ole

Mod

elw

ithB

OR

Ass

umpt

ion:

Dip

ole

Mag

netiz

atio

n Te

nsor

Mod

el•

The

para

met

ers f

or th

e de

cayi

ng e

xpon

entia

ls in

the

mag

netiz

atio

n te

nsor

(M1,ω

1,M2,ω

2,M3,ω

3)G

ener

al M

odel

:f n(t)

are

3 ar

bitra

ry (n

on-p

aram

etric

) tim

e-do

mai

n fu

nctio

nsD

ipol

eM

odel

:

Dip

ole

Mod

el w

ith B

OR

Ass

umpt

ion:

f n(t)

are

of th

e fo

rm

an

d f 1(t)

=f 2(t)

tn

nn

eM

Cω−

+

ge

oe

so(

1,ω1,

2,ω

2,3,ω

3)•

The

ratio

s of t

he a

mpl

itude

s and

the

ratio

s of t

he p

oles

are

als

o ut

ilize

d as

feat

ures

Dip

ole

Mod

el:

f n(t)

are

of th

e fo

rmD

ipol

eM

odel

with

BO

RA

ssum

ptio

n:

tn

ne

Mω−

Exa

mpl

e In

vers

ions

for T

est S

tand

81m

m (0

°inc

linat

ion)

Dip

ole

Mag

netiz

atio

n Te

nsor

Mod

el w

ith B

OR

Ass

umpt

ion

•Th

e pa

ram

eter

s for

the

deca

ying

exp

onen

tials

in th

e D

ipol

e M

odel

with

BO

R A

ssum

ptio

n:f n(t)

are

of th

e fo

rm

and f 1(t)

=f 2(t)

tn

ne

Mω−

mag

netiz

atio

n te

nsor

(M1,ω

1,M2,ω

2)•

The

ratio

s of t

he a

mpl

itude

s and

the

ratio

s of t

he p

oles

are

als

o E

xam

ple

Inve

rsio

ns fo

r C

amp

Sibe

rt U

XO

Tar

get S

E2-

48ut

ilize

d as

feat

ures

Ft

lt

dt

ii

iP

@P

1

Com

puta

tiona

lly E

ffic

ient

(Sim

ple

Phen

omen

olog

ical

) Mod

el•

The

deca

y ra

tes d

efin

ing

the

mod

es a

ssum

ing

1, 2

, and

3

di

til

ith

il(

)F

tf

tti

llffi

it(

il

)d

ld

•Fea

ture

s sel

ecte

d to

min

imiz

e P F

A@

PD=1

•Fea

ture

s fro

m c

ompu

tatio

nally

effi

cien

t (si

mpl

e)

mod

elan

dco

mbi

natio

nof

allm

odel

ssho

best

deca

ying

exp

onen

tials

in th

e si

gnal

(α1,α

21,α

22,α

31,α

32,α

33)

•Th

e ra

tios o

f the

dec

ay ra

tes a

re a

lso

utili

zed

as fe

atur

es•F

eatu

res f

rom

com

puta

tiona

lly e

ffici

ent (

sim

ple)

mod

el a

nd

com

bina

tion

of a

ll m

odel

s sho

w b

est p

erfo

rman

ce•P

erfo

rman

ceith

gene

ralm

agne

tiat

ion

tens

orm

odel

sho

sm

odel

and

com

bina

tion

of a

ll m

odel

s sho

w b

est

perf

orm

ance

(low

est P

FA@

PD=1

)•P

erfo

rman

ce w

ith g

ener

al m

agne

tizat

ion

tens

or m

odel

show

s pr

omis

ing

perf

orm

ance

6 U

SGS

ALL

TEM

TE

ST S

TAN

D D

ATA

PHEN

OM

ENO

LOG

ICA

LLY

INSP

IRED

DEC

AYIN

GEX

PON

ENTI

AL

SIG

NA

LM

OD

EL6

USG

S A

LLTE

M T

EST

STA

ND

DA

TAPH

ENO

MEN

OLO

GIC

ALL

Y IN

SPIR

ED D

ECAY

ING

EXP

ON

ENTI

AL

SIG

NA

L M

OD

ELSi

ngle

Rec

eive

r M

odel

for M

Mod

es:

Mea

sure

dSi

gnal

Mat

rixK

T)

(×

=t

STE

ST S

TAN

D D

ATA

FEAT

UR

E G

ENER

ATIO

N0

1(

)m

Mt

mm

st

AAe

α−

=

=+∑

gg

Am

plitu

de M

atrix

(am

plitu

desv

ary

acro

ssal

lmea

sure

dsi

gnal

s)

)(

MK×

=A

•Hig

h sp

atia

l res

olut

ion

mea

sure

men

ts fo

r 35

clut

ter

exam

ples

and

42

UX

O e

xam

ples

•Spa

tials

tabi

lity

ofde

cay

rate

estim

ates

depe

ndso

nm

odel

orde

rand

targ

etty

peD

ipol

e M

agne

tizat

ion

Tens

or M

odel

with

BO

R A

ssum

ptio

n•

The

para

met

ersf

orth

ede

cayi

ngex

pone

ntia

lsin

the

mag

netiz

atio

n

()

0(

)(

)T

Tt

At

=+

SA

M

Mul

tiple

Rec

eive

r M

odel

for K

Rec

eive

rs a

nd M

Mod

es:

(am

plitu

desv

ary

acro

ss a

ll m

easu

red

sign

als)

Mod

e (d

ecay

ing

expo

nent

ial)

Mat

rix(m

odes

are

cons

iste

ntac

ross

allm

easu

red

sign

als)

MT

)(

×=

tM

Spat

ial s

tabi

lity

of d

ecay

rate

est

imat

es d

epen

ds o

n m

odel

ord

er a

nd ta

rget

type

•Si

ngle

dec

ay ra

te m

odel

tend

s to

prov

ide

a ta

rget

“im

age”

•Es

timat

esfr

omhi

gher

orde

rmod

elst

end

tobe

stab

lefo

rmor

eco

mpl

exta

rget

s(i.e

.,ch

ain)

,

The

para

met

ers f

or th

e de

cayi

ng e

xpon

entia

ls in

the

mag

netiz

atio

n te

nsor

(M1,ω

1,M2,ω

2)•

The

ratio

soft

heam

plitu

desa

ndth

era

tioso

fthe

pole

sare

also

utili

zed

()

0(

)(

)t

At

+S

AM

(mod

es a

re c

onsi

sten

t acr

oss a

ll m

easu

red

sign

als)

LBL

BU

DA

EMM

ulti-

Axi

sSy

stem

USG

SA

LLTE

MM

ulti-

Axi

sSy

stem

Estim

ates

from

hig

her o

rder

mod

els t

end

to b

e st

able

for m

ore

com

plex

targ

ets (

i.e.,

chai

n),

but c

haot

ic fo

r sim

pler

targ

ets (

i.e.,

BLU

-26)

The

ratio

s of t

he a

mpl

itude

s and

the

ratio

s of t

he p

oles

are

als

o ut

ilize

d as

feat

ures

LBL

BU

D A

EM M

ulti

Axi

s Sy

stem

USG

S A

LLTE

M M

ulti

Axi

s Sy

stem

Exa

mpl

e In

vers

ions

for

Cam

p Si

bert

UX

O T

arge

t SE

2-48

Exa

mpl

e In

vers

ions

for T

est S

tand

81m

m (0

°inc

linat

ion)

Com

puta

tiona

lly E

ffic

ient

(Sim

ple

Phen

omen

olog

ical

) Mod

el•

The

deca

y ra

tes d

efin

ing

the

mod

es a

ssum

ing

1, 2

, and

3 d

ecay

ing

60m

mA

lum

inum

Pla

tey

gg

yg

expo

nent

ials

in th

e si

gnal

(α1,α

21,α

22,α

31,α

32,α

33)

•Th

e ra

tios o

f the

dec

ay ra

tes a

re a

lso

utili

zed

as fe

atur

es

•Est

imat

ed d

ecay

rate

s may

be

affe

cted

by

sens

or p

ositi

on re

lativ

e to

targ

et•

Pres

ently

: Sel

ect d

ata

from

40c

m d

iam

eter

regi

on o

ver c

ente

r of t

arge

t•

Futu

re: I

nves

tigat

e us

ing

deca

y ra

tes e

stim

ated

from

mul

tiple

spat

ial

iii

li

bili

fi

mea

sure

men

ts to

miti

gate

spat

ial v

aria

bilit

y of

est

imat

es

4 FE

ATU

RE

SE

LEC

TIO

N S

EN

SITI

VIT

YC

LASS

IFIC

ATIO

N P

ERFO

RM

AN

CE

4 FE

ATU

RE

SE

LEC

TIO

N S

EN

SITI

VIT

YM

agne

tizat

ion

Tens

or M

odel

Fea

ture

sSi

mpl

e M

odel

Fea

ture

s

4Fe

atur

es6

Feat

ures

9Fe

atur

es

Feat

ure

Sele

ctio

n U

sing

the

Entir

e D

ata

Set

Feat

ure

Sele

ctio

n Se

nsiti

vity

to C

hang

es in

the

Dat

a Se

t•D

LRT

and

RVM

pro

vide

goo

d pe

rfor

man

ce fo

r m

agne

tizat

ion

tens

or m

odel

feat

ures

•DLR

T fo

r sim

ple

mod

el fe

atur

es w

ithou

t ene

rgy

prov

ides

per

form

ance

sim

ilar t

o RV

M fo

r MT

4 Fe

atur

es6

Feat

ures

9 Fe

atur

es

•Alte

r the

LB

L B

UD

Cam

p Si

bert

feat

ure

Sele

cted

Fea

ture

sPe

rfor

man

ce M

etri

c (P

FA @

PD=1

)B

LU-2

6C

hain

mod

el fe

atur

es•I

nclu

ding

ene

rgy

impr

oves

DLR

T pe

rfor

man

ce

D=1

set b

y re

mov

ing

feat

ures

ass

ocia

ted

with

a

sing

le ta

rget

at h

igh

P D

PFA@ PD

•Sel

ect f

eatu

res u

sing

onl

y LO

O c

ross

-va

lidat

ion

with

the

rem

aini

ng 9

5 ta

rget

s

P

•96

“exp

erim

ents

”•E

xam

ine

dist

ribut

ion

of th

e re

sulti

ng

96f

ti

UC

96 p

erfo

rman

ce m

etric

sPF

A p

erfo

rman

ce m

etri

c is

mor

e se

nsiti

veto

the

trai

ning

data

sett

han

AU

sens

itive

to th

e tr

aini

ng d

ata

set t

han

AU

C