Session Architectures forCollaborative Orbit Determination

using Ground Station Networks

Srinagesh Sharma, James W. CutlerUniversity of Michigan

May 2, 2018

(Sharma, Cutler, 2018) (1/17)

Initial acquisition with cluster launches isdifficult

• There is increased access to space -cluster launches and TLE Lottery

• We can construct Autonomous GroundStation Networks (AGSNs).

• Initial acquisition through Dopplerbased OD.

• Deep space orbit determination

(Sharma, Cutler, 2018) (2/17)

An ML Algorithm to do OD• Orbital parameters Γ

• Uncertainty distribution PΓ

• Observation interval T

• Observed by ground stationnetwork

• Machine Learning Algorithm toestimate orbits 1

1Sharma et. al IPNPR 2015(Sharma, Cutler, 2018) (3/17)

Can a practial GSN satisfy AlgorithmicConditions?

The machine learning algorithm requires some conditions to be met

• Guaranteed observation of samples of any orbit from uncertaintydistribution

• Observability

• Continuity of map

(Sharma, Cutler, 2018) (4/17)

Spectrum analysis in a GSLow Gain Antenna + LNA

SDR

Interface

TIMING

SYSTEM

FILTER

BANKS

CODECSPECTRUM

MONITORING

FEATURE

VECTOR EST

(COG. RADIO)

LEARNING

CLUSTER

SCHEDULING

&

OPERATIONS

ANTENNA

CONTROL

SJ

OJ

FEATURE

SELECTION

{XJ,i ,TJ,i }

Autonomous Software Ground Station(Sharma, Cutler, 2018) (5/17)

The global network has a learning cluster

ASGS 1

TIMING

SYSTEM

LEARNING

CLUSTER

NETWORK

OPERATIONS

CENTER

ASGS 2 ASGS 3 ASGS nGSN

PΓ Ť PX|F,z zfeature

{Oj }j=1,2..nG

zts

{X1,i ,T1,i }{X2,i ,T2,i } {X3,i ,T3,i }

{XnGSN,i ,TnGSN,i }

LG1 LG2 LG3 LGnGSN

{Sj }j=1,2..nGSN Dop

Global Architecture

(Sharma, Cutler, 2018) (6/17)

We can integrate OD into GSN Sessions

• GSNs work by reserving nodes(ground stations) through sessions.

• Orbit determination will be doneinside of an “ODTrack Session”.

• The ODTrack Session is a group ofGS Sessions. ODTrack Session

(Sharma, Cutler, 2018) (7/17)

The learning cluster determines the OD session

• Learning Cluster samples the uncertainty distribution.

• Selects ground stations based on visibility.

• Coordinated antenna pointing.

(Sharma, Cutler, 2018) (8/17)

Best explained by an example scenario

sc1 deploy cone

sc2 deploy cone

sc3 deploy cone

sc4 deploy cone

A simple model of random deployment Samples of initial state deployment

(Sharma, Cutler, 2018) (9/17)

Intervals of uncertain passes• Break down OD into intervals of observation• One pointing direction per interval

Wellington Doppler distribution has passintervals Wellington azimuth elevation distribution has

bands

(Sharma, Cutler, 2018) (10/17)

Point at the expected zero Doppler point

Collect zero Doppler time points

• Antenna is approximated as acone (cone angle: 3dBbeamwidth)

• Antenna will point to theexpected value of the azimuthelevation distribution

(Azim., elev.) probability of zero Doppler

Antenna cone points along expected AzEl(Sharma, Cutler, 2018) (11/17)

A Pointing Profile• Training Data is limited to learning with the profile information• Profile + Interval information is sent to ASGS• ODTrack Session is reserved for interval periods• ASGS points antennas and examines spectrum during ODTrack

intervals

Azimuth and elevation of pruned training data(Sharma, Cutler, 2018) (12/17)

Simulated Scenario

• Total Uncertainty in initial position of satellites: 863 km• Observation Interval: 6 hours• Deployment Interval: 200s• Average Separation at initial state: 36km• Noise in Doppler shift estimation: 2.5%∆fmaxDoppler

• Synchronization error between Ground stations: ≤ 1ms

• Probability of transmission = 0.08 ( 1 in 10 seconds)

(Sharma, Cutler, 2018) (13/17)

ResultsAverage error in initial position estimation: 21.44km

Histogram of position error Histogram of velocity error

(Sharma, Cutler, 2018) (14/17)

Conclusion and Future Work

• ODTrack Session Architecture for GSNs• Algorithms for session instantiation for OD• Coordinated Pointing Profiles for GSNs

Future Work

• Multiple antenna ground station nodes• Integrating Learning with independent RADAR and Ground

station based Doppler observations

(Sharma, Cutler, 2018) (15/17)

Thanks

(Sharma, Cutler, 2018) (16/17)

ReferencesSharma, S. and Cutler, J.W., 2015. Robust orbit determination andclassification: A learning theoretic approach. Interplanetary NetworkProgress Report, JPL, 203, p.1.

(Sharma, Cutler, 2018) (17/17)