Advanced Multidisciplinary System Engineering or How I learned to - - PDF document

SLIDE 1

1

Advanced Multidisciplinary System Engineering

• r

“How I learned to think outside of MY box!”

• Dr. Joseph R. Guerci

Director DARPA/SPO

Joe.Guerci@DARPA.MIL

All material cleared for Public Release

SLIDE 2

2

Outline

• Breakthrough systems/technologies are almost always multidisciplinary

– Arise from cross-fertilization – “Cross-fertilization” occurs in someone’s mind

• “Thinking outside the box” = “Thinking outside your box”

– Examples:

• KASSPER
• HISS
• New Trend in Multidisciplinary Systems Engineering

– Level 1: System = Interconnected set of single-purpose subsystems – Level 2: System = Interconnected set of multi-purpose subsystems – Level 3: System = Embedded multi-purpose subsystems w/o clear boundaries

• Example: ISIS
• Summary

Sample SPO Projects

(A Multidisciplinary Systems Technology Office)

PRODUCT

IAR IRSG

PRODUCT

IDA

Pathogen DNA DNA Polymerase 5’ 3’ 5’ 3’ 5’ 3’ 5’ 3’ PRODUCT Pathogen RNA Nicking Enzyme

RNA Polymerase RNA Polymerase

Toxin PRODUCT

IAR IRSG

PRODUCT

IDA

Pathogen DNA DNA Polymerase 5’ 3’ 5’ 3’ 5’ 3’ 5’ 3’ 5’ 3’ PRODUCT Pathogen RNA Nicking Enzyme

RNA Polymerase RNA Polymerase

Toxin

Next Generation Chem/Bio Sensors & Protection Advanced Intelligent Signal Processing & Embedded Systems Revolutionary Space and Near-Space Technologies

SLIDE 3

3

mD Adaptive Signal Processing

• Example: Space-Time Adaptive Processing (STAP)

Space-Time Adaptive Beamformer “Ideal” Adapted Pattern

Optimum Solution

• Weiner-Hopf

1

R− = w s

(Optimum space-time beamformer weights) (Desired signal “steering vector”) (Inverse of total interference covariance matrix)

,

NM NM NM

C R C

×

∈ ∈ w s

~10' 100' NM s s −

SLIDE 4

4

Covariance Estimation Problem

• Practical implementation example and real data

example (White Sands DARPA Mountain Top Radar)

Sample Covariance Estimation Measured Data

ˆ

i i i

R

∈Ω

′ =∑x x

Ideal (Stationary) Data

• Heterogeneous Clutter

– Rapidly varying terrain

• Mountainous (rapid elevation/reflectivity variation)
• Rapid land cover variations (e.g., littoral)
• Dense “Target” Backgrounds

– “Moving Clutter”

• Military/civilian vehicles
• Large Discretes and “Spiky” Clutter

– Urban clutter – Power lines, towers, steep mountainous terrain

• Range-Varying (Nonstationary) Clutter Loci

– Bi/Multistatics – Nonlinear array geometries (e.g., circular arrays)

Welcome to the Real-World!

Extremely suboptimal radar performance can occur if one or more of the following occurs:

(High false alarm rates and/or low Pd)

One or More of the Above is Almost Always Present in Real-World Ops!

SLIDE 5

5 Serious Performance Impacts!!

(KASSPER ’02 Data Cube & APTI Data Set)

x distance (km) −> (longitude) y distance (km) −> (latitude) Rx

(35.73°,118.5°) 20 40 60 80 100 10 20 30 40 50

Doppler (Fraction PRF) Range Bin # GMTI Range−Doppler Data (dB−thermal) −0.5 −0.4 −0.3 −0.2 −0.1 0.1 0.2 0.3 0.4 10 20 30 40 50 60

−50 50 100 150 −70 −60 −50 −40 −30 −20 −10 SINR/SNRo (dB) Doppler (m/s) rang bin 240 (38.6km)

• ptimal

PCI−40 MWF−40 post−Doppler (3 bin)

SINR Loss High False Alarm Rates

−10 −5 5 10 15 20 25 30 10

−5

10

−4

10

−3

10

−2

10

−1

10 Pixel SINR (dB) Fraction Exceeding Value AMF Exceedance STAP Only STAP w/ Pre−Whitening

Radar Environmental Knowledge Bases (DTED/DFAD/LCLU, SAR, etc.)

Knowledge-Aided Sensor Signal Processing & Expert Reasoning (KASSPER)

Clutter Knowledge Base

I C C R

i i i

KA

2 2

Cells Clutter Over Sum

σ γ + ∑ ′ =

{ }

i

γ

Clutter Cell Returns GPS/INS

N Elements [ x1 x2 x3 • • • xM ] . . . T T M P ulse s . . . . . . T T M Pulse s Array S napshots

Sensor Characteristics

{ }

i

C

Clutter Steering Vectors Clutter Knowledge Base

I C C R

i i i

KA

2 2

Cells Clutter Over Sum

σ γ + ∑ ′ =

{ }

i

γ

Clutter Cell Returns GPS/INS

N Elements [ x1 x2 x3 • • • xM ] . . . T T M P ulse s . . . . . . T T M Pulse s Array S napshots

Sensor Characteristics GPS/INS

N Elements [ x1 x2 x3 • • • xM ] . . . T T M P ulse s . . . . . . T T M Pulse s Array S napshots

Sensor Characteristics GPS/INS GPS/INS

N Elements [ x1 x2 x3 • • • xM ] . . . T T M P ulse s . . . . . . T T M Pulse s Array S napshots N Elements [ x1 x2 x3 • • • xM ] . . . T T M P ulse s . . . . . . T T M Pulse s Array S napshots

Sensor Characteristics

{ }

i

C

Clutter Steering Vectors

• 60
• 50
• 40
• 30
• 20
• 10

10 20

• 0.5

0.5

• 0.5

0.5

• 60
• 50
• 40
• 30
• 20
• 10

10 20

• 0.5

0.5

• 0.5

0.5

X

Nonstationary Clutter (plus Signal)

X R Y

KA 2 1 −

=

2 1 − KA

R

2 1

ˆ −

SMI

R

Reduced-Rank Conventional Filter KA Pre-Filter

Y R Z

SMI 2 1

ˆ − =

Detector 1st Stage Knowledge-Aided Pre-Filter Response 2nd Stage Conventional Filter

KASSPER

Clutter Knowledge Base

I C C R

i i i

KA

2 2

Cells Clutter Over Sum

σ γ + ∑ ′ =

{ }

i

γ

Clutter Cell Returns GPS/INS

N Elements [ x1 x2 x3 • • • xM ] . . . T T M P ulse s . . . . . . T T M Pulse s Array S napshots

Sensor Characteristics

{ }

i

C

Clutter Steering Vectors Clutter Knowledge Base

I C C R

i i i

KA

2 2

Cells Clutter Over Sum

σ γ + ∑ ′ =

{ }

i

γ

Clutter Cell Returns GPS/INS

N Elements [ x1 x2 x3 • • • xM ] . . . T T M P ulse s . . . . . . T T M Pulse s Array S napshots

Sensor Characteristics GPS/INS

N Elements [ x1 x2 x3 • • • xM ] . . . T T M P ulse s . . . . . . T T M Pulse s Array S napshots

Sensor Characteristics GPS/INS GPS/INS

N Elements [ x1 x2 x3 • • • xM ] . . . T T M P ulse s . . . . . . T T M Pulse s Array S napshots N Elements [ x1 x2 x3 • • • xM ] . . . T T M P ulse s . . . . . . T T M Pulse s Array S napshots

Sensor Characteristics

{ }

i

C

Clutter Steering Vectors

• 60
• 50
• 40
• 30
• 20
• 10

10 20

• 0.5

0.5

• 0.5

0.5

• 60
• 50
• 40
• 30
• 20
• 10

10 20

• 0.5

0.5

• 0.5

0.5

X

Nonstationary Clutter (plus Signal)

X R Y

KA 2 1 −

=

2 1 − KA

R

2 1

ˆ −

SMI

R

Reduced-Rank Conventional Filter KA Pre-Filter

Y R Z

SMI 2 1

ˆ − =

Detector 1st Stage Knowledge-Aided Pre-Filter Response 2nd Stage Conventional Filter

KASSPER

Measured

(DARPA Mtn Top)

Predicted

(DTED Level-1) Range Doppler

Bald Earth 1980 Physical 2000 1980

HPEC Real-Time Database EM Modeling Tools Physical Databases

SLIDE 6

6

CACFAR AGC, etc. IF Sidelobe Canceler Fully Adaptive Array Space-Time Adaptive (STAP) Radar Advanced and Real-Time STAP

50’s 60’s 70’s 80’s 90’s

Reinventing Adaptive Radar

First Gen Statistical Signal Processing KASSPER

Intelligent Adaptive Radars

“Real-world nonstationarity does NOT support conventional adaptivity”

00’s 10’s Real-time knowledge-aided KASSPER Classic

Savant

FLOPS/Throughput

Knowledge

Data type/MBytes

High-speed, single function Multi-function, slow access speeds

True “Intelligent” Processing

+

SAR Roads VMAP Discrete

Radar returns

Old New

∑ ′ =

Ω k k i

R x x ˆ

Space-Time Snapshot Vector

Range Cells Test Cell

“Guard” Cells

2 1 1 2 − − + + i i i i i

x x x x x

. . . . . .

Ω Ω

Conventional Space-Time Filtering

s w

1 −

= R

QR Factorization w/ Back substitution

(from Antenna-Based Signal Processing Techniques for Radar, A. Farina, Artech House)

Highly Parallel Systolic Array Implementation (Achieves 100’s to 1000’s of GFLOPS)

KASSPER HPEC Challenge: Optimizing adaptation by injecting environmental knowledge “intelligently” into the front-end signal flow

First Gen Real-Time KASSPER HPEC Clutter Knowledge Base Intelligent Signal Processing

• KASSPER requires memory

access interrupts

• Optimal interrupt

scheduling

• Optimized ISP
• “Look-Ahead” scheduling
• KASSPER requires memory

access interrupts

• Optimal interrupt

scheduling

• Optimized ISP
• “Look-Ahead” scheduling

Conventional vs. KASSPER HPEC Processing

SLIDE 7

7 “Look-Ahead” Scheduling Addresses Memory Latency Issues

CPU Registers

Registers

Faster Speed Higher Cost Staging Transfer Unit

100s Bytes <1s ns Cache 10s-100s K Bytes 1-10 ns

Source: Dave Patterson, Graduate Computer Architecture Course, University of California, Berkeley, Spring, 2001

Disk 10s G Bytes 10 ms Tape Infinite sec-min Main Memory M Bytes 100-300 ns

Cache Memory Disk Tape

• Instr. Operands

Blocks Pages Files

Prog./Compiler 1-8 Bytes Cache Controller 8-128 Bytes OS 512-4K Bytes User/Operator MBytes

Larger Size Lower Cost Capacity Access Time

CPU Registers

Registers

Faster Speed Higher Cost Staging Transfer Unit

100s Bytes <1s ns Cache 10s-100s K Bytes 1-10 ns

Source: Dave Patterson, Graduate Computer Architecture Course, University of California, Berkeley, Spring, 2001

Disk 10s G Bytes 10 ms Tape Infinite sec-min Main Memory M Bytes 100-300 ns

Cache Memory Disk Tape

• Instr. Operands

Blocks Pages Files

Prog./Compiler 1-8 Bytes Cache Controller 8-128 Bytes OS 512-4K Bytes User/Operator MBytes

Larger Size Lower Cost Capacity Access Time

Problem:

KASSPER “Look-Ahead” Interrupt Scheduling

t

t t ∆ +

Clutter Knowledge Base Predictor

Solution:

Next-Gen KASSPER HPEC Testbed

• Architecture:

– Base computer and I/O cards purchase order completed – Lab computer configuration complete – Various processing concepts in review – PDR planned for late June 03 – Demonstration at DARPATech 04

• Parallel Vector Library (PVL) chosen for
• pen standards programming language

– LL reviewing initial KASSPER algorithms for library impacts – Coding started on basic radar signal processing components (pulse compression, data retrieval, etc.) – Algorithm developers will program the hardware

Vendor Hardware

Portable Library

Maps Application Code Vendor Software

Open standards for real- time processing

MP-510 mercury processing Multiple high- speed RAID drives ASIC high-speed cache memory devices

Vendor Hardware Application Code Vendor Software

• Upgrades restricted to hardware

remapping & new features

SLIDE 8

8 Pre-filtering Followed by Conventional STAP

5 10 15 20 25 30 Doppler (Fraction PRF) Range Bin # GMTI AMF Output (dB−thermal) −0.5 −0.4 −0.3 −0.2 −0.1 0.1 0.2 0.3 0.4 10 20 30 40 50 60 5 10 15 20 25 30 Doppler (Fraction PRF) Range Bin # GMTI AMF Output after Whitening (dB−thermal) −0.5 −0.4 −0.3 −0.2 −0.1 0.1 0.2 0.3 0.4 10 20 30 40 50 60

−10 −5 5 10 15 20 25 30 10

−5

10

−4

10

−3

10

−2

10

−1

10 Pixel SINR (dB) Fraction Exceeding Value AMF Exceedance STAP Only STAP w/ Pre−Whitening

Adaptive Matched Filter

With Prefiltering Without Prefiltering

Better Behaved “Tail”

Pre-Filtering Reduces The “Tail” of the Exceedance Function 13 dB!

KASSPER:

“It’s an Architecture, NOT an Algorithm” KASSPER is an architecture for real-time adaptation of multidimensional sensor systems in real-world environments KASSPER is an architecture for real-time adaptation of multidimensional sensor systems in real-world environments

• KASSPER Architecture

– Environmental context is key to efficient adaptation

• Sensors, like humans, benefit from context!

– Key enablers: “look-ahead” scheduling and resource allocation – Multiresolution philosophy: blurring the boundaries between SAR and GMTI – KASSPER as a modern manifestation of the “Bayesian” method!

• KA-STAP Bayesian STAP
• The DARPA KASSPER Challenge: Creatively explore the

possibilities

– Re-examine entire adaptive signal processing paradigm with an eye towards maximizing knowledge-aided “robust” methods – Robust STAP algorithms AND KASSPER architecture

• Environmental knowledge base as “read/write” scratch memory

– What is “implementable”? 2010? 2020? – Environmentally aware sensors have a future!

• KASSPER Architecture

– Environmental context is key to efficient adaptation

• Sensors, like humans, benefit from context!

– Key enablers: “look-ahead” scheduling and resource allocation – Multiresolution philosophy: blurring the boundaries between SAR and GMTI – KASSPER as a modern manifestation of the “Bayesian” method!

• KA-STAP Bayesian STAP
• The DARPA KASSPER Challenge: Creatively explore the

possibilities

– Re-examine entire adaptive signal processing paradigm with an eye towards maximizing knowledge-aided “robust” methods – Robust STAP algorithms AND KASSPER architecture

• Environmental knowledge base as “read/write” scratch memory

– What is “implementable”? 2010? 2020? – Environmentally aware sensors have a future!

SLIDE 9

9

Emerging Field

• Special Issue of IEEE Signal Processing Magazine

Handheld Isothermal Silver Standard Sensor (HISSS)

The goal of the HISSS program is to develop a handheld sensor that is capable of identifying biological threats including bacteria, viruses and toxins.

Polymerase Chain Reaction (PCR) Machine

Notional Sensor

DNA detection RNA detection Protein detection Fluid handling DNA readout RNA readout Protein readout System check

Notional Sample Cartridge

How to shrink into a handheld?

• Order-of-mag faster!
• At least as accurate!

SLIDE 10

10

PCR vs. Isothermal

∆t ~ 60 sec Anneal at 55ºC

Starting the process: Primers Polymerase Pathogen DNA

5’ 3’

Products: copies of Pathogen DNA

5’ 3’

Extend at 72ºC Denature at 95ºC

5’ 5’ 3’ 3’ 3’ 5’ 5’ 3’ 5’ 5’ 3’ 5’ 5’ 3’

Polymerase Chain Reaction

3’ 5’ 3’ 5’ 3’

Cleave

5’ 3’

Product falls off

5’ 3’

Polymerase re-binds

5’ 3’ 5’

Extend

3’

Products: copies of reporter Nicking enzyme Starting the process: Polymerase Trigger template Pathogen DNA

Isothermal

∆t ~ 3 sec

HISS DNA Amplification

SLIDE 11

11

HISSS Progress

• Progress:

– Demonstrated false alarm rates, using ROC curve analysis for HISSS assays that are equal to or better than current DNA, RNA, and protein assays – Successfully developed and utilized a flow-through testbed to test all assays

0.2 0.4 0.6 0.8 1 0.01 0.1 1

0.1 0.01 1

Pfa

DNA ROC Curves

0.8 0.6 0.4 0.2 0.0 1.0

Pd

PI (1:99) PII (1:99) PCR (1:99)

0.2 0.4 0.6 0.8 1 0.01 0.1 1

0.1 0.01 1

Pfa

DNA ROC Curves

0.8 0.6 0.4 0.2 0.0 1.0

Pd

PI (1:99) PII (1:99) PCR (1:99) PI (1:99) PII (1:99) PCR (1:99)

0.2 0.4 0.6 0.8 1 0.001 0.01 0.1 1

0.001 0.1 0.01 1 0.8 0.6 0.4 0.2 0.0 1.0

Pd Pfa

RNA ROC Curves

PI (1:82) PII (1:82) RT-PCR (1:82)

0.2 0.4 0.6 0.8 1 0.001 0.01 0.1 1

0.001 0.1 0.01 1 0.8 0.6 0.4 0.2 0.0 1.0

Pd Pfa

RNA ROC Curves

PI (1:82) PII (1:82) RT-PCR (1:82) PI (1:82) PII (1:82) RT-PCR (1:82) 0.2 0.4 0.6 0.8 1 0.001 0.01 0.1 1

Pd

0.01 1 0.1 0.001

Pfa

0.8 0.6 0.4 0.2 0.0 1.0 Protein Toxin ROC Curves PI (1:3000) PII (1:3000) ELISA (1:3000) 0.2 0.4 0.6 0.8 1 0.001 0.01 0.1 1

Pd

0.01 1 0.1 0.001

Pfa

0.8 0.6 0.4 0.2 0.0 1.0 Protein Toxin ROC Curves PI (1:3000) PII (1:3000) ELISA (1:3000) PI (1:3000) PII (1:3000) ELISA (1:3000)

PI Static (1:99) PII Flow (1:99) PCR (1:99) PI Static (1:82) PII Flow (1:82) RT-PCR (1:82) PI Static (1:3000) PII Flow (1:3000) ELISA (1:3000)

New Airship Design Philosophy

MDA Airship

Payload bay

Conventional Airship

Capability cannot be added to airship after development

Payload: ~2% of system mass

ISIS requires integration of sensor and airship

Payload: 30-40% of system mass Turn a disadvantage (large size) into an advantage (large antenna)!

SLIDE 12

12

The “First” ISIS?

Echo 1

Simultaneous AMTI/GMTI Operation via Dual Band (UHF/X-Band) Aperture

Most Powerful Airborne GMTI/AMTI Radar & Comms Ever Conceived

Long-range AMTI/GMTI/COMM FOPEN GMTI Cruise Missile Defense Steep Grazing Angles Detect/Track Dismounts Extremely High Capacity Comms Near Zero Platform Speed

No In-Theater Ground Support – 99% on station availability for 1+ years 600km radar horizon at 70kft operational altitude

SLIDE 13

13

ISIS Joint STARS Joint STARS AWACS AWACS Global Hawk Global Hawk Global Hawk 109 108 107 106 105 104 103 102 100 101 109 108 107 106 105 104 103 102 100 101

160,000 300,000,000

1.0 Relative Search Capability (PA/R2) 1.0 Relative Track Capability (PA2/λ2/R4)

240 5,100 3,300 15,000

ISAT

140 610

VHF X S S X X X X

Unprecedented Radar Performance

Platform Carries the Antenna Antenna Is the Platform

Sustained Operations Logistics

• Aircraft-based ISR Requires

– Local air base – Multiple aircraft to keep 1 flying – Air crews – Ground crews – Fuel supplies – Maintenance facilities

• ISIS

– Unmanned – Deploys worldwide from U.S. base – Regenerative Fuel Sources – One-year continuous ISR capability

SLIDE 14

14 Wind Conditions Drive Propulsion Power Needs

η ρ 2

3 3 / 2 v

V C P

d

⋅ =

Where ρ = air density at altitude V = volume of airship v = relative velocity of air η = efficiency of propellers Propulsion Power for V = 106 m3 (Cd=0.022)

500 1,000 1,500 2,000 2,500 10 20 30 40 50 Wind Speed (m/sec) Power Required (kw)

Propulsion Power for V = 106 m3 (Cd=0.022)

500 1,000 1,500 2,000 2,500 10 20 30 40 50 Wind Speed (m/sec) Power Required (kw)

44.95 m/s

Max winds drive power system requirements

Station Keeping

• ISIS Objective: 99% on-station availability for 1 year

– Function of airship speed (sustained and sprint) and available energy (regenerative and stored fuel)

• Need operational algorithms for maximizing availability

– Managing airship energy ala satellite delta-v

• 90
• 80
• 70
• 60
• 50
• 40
• 30
• 20
• 10

10 20 30 40 50 60 70 80 90 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75

Wind Speed (m/s) Latitude (degress)

Mean Wind Speed Average 99 Percentile Mean Wind Speed Average 99 Percentile

Maximum Sprint Speed

SLIDE 15

15

Requires Large Mass Reductions

Mass Volume P

• w

e r Mass Volume P

• w

e r

M V ∝

3 / 2

V P ∝ P M ∝

• ISIS designs are mass-centric

– Lifting gas has reached the maximum limit: – 0.061kg per 1m3 of He @ 21km – 0.066kg per 1m3 of H2 @ 21km

• ISIS focusing on:

– Removing mass from largest contributors – Integration, INTEGRATION, INTEGRATION!

avionics propulsion propulsion d air radar power power aperture hull h gas ISIS avionics propulsion power radar structure gas lifting air displaced

M M v V C P A V c V M M M M M M M M + + ⎟ ⎟ ⎠ ⎞ ⎜ ⎜ ⎝ ⎛ + + + + = + + + + + = η ρ η ρ ρ ρ ρ 2

3 3 / 2 3 / 2

Integration Components

Summary

• Breakthrough systems/technologies are almost always

multidisciplinary

– System engineers need to be continually learning about new technologies and methods across ALL disciplines

• “Be an annoying know-it-all!”

– Tactic: “Can the thermal engineer give the flight control engineer’s briefing?” – Often “cross fertilization” can occur even if with only a 1st or 2nd

• rder understanding of multiple disciplines
• Balance of depth and breadth

– How should engineering programs be structured in light of above?

• Undergraduate programs typically have the breadth, but don’t seem to

“close the deal”

– Example: Senior class semester devoted to dissecting a complex system

• Emergence of a “Level 3” systems integration

– Multidisciplinary from its inception!