How Is Starlink Manoeuvring? An Analysis of Patterns in the

Manoeuvres of Starlink Satellites

David P. Shorten

, Wathsala Karunarathne

1,2

and Matthew Roughan

1,2

Department of Computer and Mathematical Sciences, The University of Adelaide, Australia

Teletrafﬁc Research Centre, The University of Adelaide, Australia

Keywords:

Starlink, Two-Line-Element, Satellite, Manoeuvre, Space Debris, Space Situational Awareness, Particle Filter.

Abstract:

The rapid increase in the number of active satellites orbiting earth along with the simultaneous increase in

the amount of space debris is causing earth’s exosphere to become ever more crowded. This crowding forces

satellites to perform a rising number of collision-avoidance manoeuvres. At the time of publication, of the

roughly 7700 active satellites orbiting earth, over 5000 belonged to the Starlink constellation. These satellites

not only substantially contribute to the crowding of space, but are required to perform tens of thousands of

collision-avoidance manoeuvres per year. As Starlink does not publish information on the timing of these

manoeuvres, little is known about them beyond their total number. This work uses a recently-proposed algo-

rithm for detecting satellite manoeuvres from the publicly-available 18

Space Defence Squadron TLE data

to study the patterns in the manoeuvres of this constellation. Rich structure was found in the patterns of these

manoeuvres, including regular synchronous bursts of station-keeping manouevres within launch groups (the

groups of satellites launched on a single day) and a cyclical pattern of station keeping amoung the launch

groups.

1 INTRODUCTION

Starlink is a constellation of over 5,000 satellites that

provides high-speed internet access to users with an

appropriate antenna (Michel et al., 2022). It has the

unique advantage of being able to provide this service

to users at sea or in remote regions.

The number of satellites orbiting the earth is in-

creasing at a rapid rate. As of 1 January 2023, the list

of satellites maintained by the Union of Concerned

Scientists (Union of Concerned Scientists, 2023b),

contained 6,718 operational satellites, an increase of

nearly 2,000 satellites over the previous year (Union

of Concerned Scientists, 2023a). This rate of increase

is likely to accelerate. Starlink alone, which currently

operates over 5000 satellites, has ﬁled for permis-

sion to launch an additional 30,000 satellites with the

Federal Communication Commission (FCC) (Boley

and Byers, 2021). Other companies, including Ama-

zon, OneWeb, Telesat and GW have announced sim-

ilar plans (Boley and Byers, 2021). A total of over

100,000 satellites in orbit by 2030 is considered plau-

sible (Venkatesan et al., 2020). This situation is fur-

ther exacerbated by the large amount of space de-

bris in orbit, which includes over 22,000 tracked ob-

jects (Pelton, 2015). The crowding of space increases

the need for satellite operators to be aware of space

trafﬁc and take evasive manoeuvres when collisions

are predicted. Techniques for monitoring the orbits

and activities of satellites fall under the ﬁeld of Space

Situational Awareness (SSA) (Lal et al., 2018).

SpaceX (the operators of Starlink) are already re-

quired to perform a large number of evasive manoeu-

vres. According to ﬁlings with the FCC (Goldman,

2023), between the 1

of December 2022 and the 31

of May 2023, Starlink satellites performed 25,299

propulsive manoeuvres. This equates to an average

of around 12 manoeuvres per satellite per year. Given

that the Starlink constellation currently forms the ma-

jority of satellites in orbit, an analysis of the patterns

of these manoeuvres will provide insight into the fu-

ture of space trafﬁc management.

Although Starlink is comparatively generous in

their sharing of data (Goldman, 2023), they do not

publish the timestamps of satellite manoeuvres. How-

ever, the 18

Space Defence Squadron provides daily

updates on the orbits of all starlink satellites via

Space-Track (Space-Track, 2023). This paper utilises

a recently-proposed algorithm (Shorten et al., 2023a)

for satellite orbit anomaly detection from this public

174

Shorten, D., Karunarathne, W. and Roughan, M.

How Is Starlink Manoeuvring? An Analysis of Patterns in the Manoeuvres of Starlink Satellites.

DOI: 10.5220/0012619900003705

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 9th International Conference on Internet of Things, Big Data and Security (IoTBDS 2024), pages 174-184

ISBN: 978-989-758-699-6; ISSN: 2184-4976

data in order to infer the timestamps of these manoeu-

vres. This approach has been thoroughly validated on

a benchmark dataset (Shorten et al., 2023b) contain-

ing ground-truth manoeuvre timestamps.

We apply this anomaly detection algorithm to the

historic publicly-available data of all Starlink satel-

lites active on the 27

of November 2023. We ﬁnd

that the pattern of manoeuvres within the Starlink

constellation exhibits rich structure. Section 3.2 ex-

plores some of the global properties of this struc-

ture, including the high rate of manoeuvres of satel-

lites post launch. The pattern of manoeuvres across

the constellation is strongly associated with the con-

stellation’s structure. The Starlink constellation is

formed of seven orbital shells, where all satellites

within each shell share a common orbital altitude and

inclination (see table 3 for a list). These shells can

be further divided into their launch groups, consist-

ing of the satellites launched on a single day. Sec-

tion 3.3 examines the manoeuvres of a representa-

tive launch group, ﬁnding regularly-spaced stereo-

typed bursts of station-keeping manoeuvring where

all the satellites within the launch group manoeu-

vre near-simultaneously. Moreover, the relative order

of manoeuvring within these groups is mostly main-

tained over time. Section 3.4 subsequently shows

how these patterns degrade over time in older satel-

lites. This analysis is extended to the relationship in

station-keeping manoeuvring between launch groups

in section 3.5, where it is shown that SpaceX performs

station-keeping manoeuvres on the satellites in a shell

in a cyclical fashion, iteratively moving through its

launch groups. Finally, section 3.6 presents an anal-

ysis of the rate of collision-avoidance manoeuvres

across the constellation.

2 METHODS

A list of the SATCAT numbers (Kelso, 1998) of the

5044 Starlink satellites active on the 27

of Novem-

ber 2023 was obtained from N2YO (N2YO, 2023).

The TLE data for all of these satellites was then down-

loaded from Space-Track (Space-Track, 2023). ta-

ble 1 contains a summary of the data used in this

work. TLE data consists of the mean Keplerian orbital

elements of satellites, along with metadata such as a

ballistic drag coefﬁcient, recorded roughly daily (Val-

lado and Cefola, 2012). The timestamp associated

with each record is usually referred to as the epoch.

The orbital elements are ‘mean’ in the sense that they

exclude high-frequency non-Keplerian components.

That is, they specify the elliptical orbit which most

closely approximates the true (non-elliptical) orbit of

Table 1: Summary statistics of Starlink satellite data.

Feature Value

Number of Starlink satellites anal-

ysed

4,998

Total number of TLE data points 7,363,925

Total number of detected manoeu-

vres

129,860

Median number of manoeuvres per

satellite

First launch date included 11/11/2019

Last launch date included 29/10/2023

Table 2: Parameters for optimal proposal ﬁlter.

See (Shorten et al., 2023a) for an explanation of each

parameter.

Parameter Description Value

N Number of particles 250

α Variance inﬂation

factor

Threshold on

effective

for reg-

ularisation

0.2

shift

Threshold on neg-

ative log predictive

density for ensemble

shift

anom

Threshold on neg-

ative log predictive

density for anomaly

detection

100

the satellite.

The 46 satellites from the two most recent

launches (on the 22

and 27

of November 2023

(SpaceX, 2023)) were discarded as there was insuf-

ﬁcient data for the operation of the anomaly detection

algorithm.

A recently-proposed (Shorten et al., 2023a) satel-

lite orbit anomaly detection algorithm was then ap-

plied to the TLE data of the remaining 4998 satel-

lites. This approach operates by applying an optimal-

proposal particle ﬁlter (Snyder, 2011) to the mean Ke-

plerian orbital elements present in the TLE data. The

idea of ﬁltering, in this context, is to assume that the

mean elements present in each TLE line pair are noisy

How Is Starlink Manoeuvring? An Analysis of Patterns in the Manoeuvres of Starlink Satellites

175

measurements of the true underlying values of these

elements. Filtering then seeks to infer the true val-

ues of these elements at each epoch, along with an

uncertainty. Particle ﬁlters are well-suited to the non-

linear nature of satellite orbit evolution and allow us

to compute a non-Gaussian uncertainty. The ﬁltering

setup makes use of the SGP4 (Vallado et al., 2006)

model for orbit evolution. However, it only makes

use of the initial part of this model which propagates

the mean Keplerian elements before incorporating the

high-frequency non-Keplerian components. More-

over, a non-standard implementation of SGP4 is used.

Standard implementations ﬁrst convert the mean mo-

tion from the Kozai to the Brouwer formulations be-

fore propagation (Vallado et al., 2006). Instead, we

perform this as an initial pre-processing step, and per-

form no such conversions during propagation.

Once the ﬁlter has arrived at an estimate (with un-

certainty) for the mean elements at a given epoch, the

estimated mean elements and associated uncertainty

can be propagated to the subsequent epoch. They can

then be compared with the observed mean elements

(in the TLE) at the subsequent epoch. If the observed

elements are deemed sufﬁciently unlikely given our

propagation and uncertainty, then the epoch is desig-

nated as anomalous. More speciﬁcally, we compute

the negative logarithm of the predictive density of the

observation at the subsequent epoch. This ﬁgure is

taken to be our anomaly statistic. After a threshold is

chosen, epochs with an anomaly statistic greater than

this threshold are designated as anomalous. We ex-

pect that the majority of detected anomalies will be

the result of manoeuvres, although they could also be

the result of changes in the processing of TLEs by the

Space Defence Squadron, among other reasons

(see section 4 for further discussion).

Table 2 contains the speciﬁcation of all parameters

used for ﬁltering. τ

anom

(the threshold for anomaly de-

tection) was hand-tuned to produce roughly the same

rate of anomalies as manoeuvres reported by SpaceX.

All other parameters are the same as those used in

(Shorten et al., 2023a), apart from the number of par-

ticles N. This was halved for reasons of computa-

tional feasibility, given the large number of satellites

in the constellation.

This ﬁltering approach was evaluated (Shorten

et al., 2023a) on a benchmark dataset con-

taining the TLEs of 15 satellites along with

independently-obtained ground-truth manoeuvre

timestamps (Shorten et al., 2023b) as well as sim-

ulated data. It was demonstrated to be superior to

a baseline approach, similar to many previously-

proposed approaches for manoeuvre detection from

TLE data (Li et al., 2018; Li et al., 2019; Decoto

and Loerch, 2015; Mukundan and Wang, 2021; Zhao

et al., 2014).

3 RESULTS

3.1 Stucture of the Starlink

Constellation

Much of the following presentation of the results will

concern how the timing of manoeuvres is related to

the structure of the constellation. As such, we begin

with a brief description of this structure.

The Starlink constellation consists of seven orbital

shells (see table 3 for a list). The orbits of all satel-

lites in each shell share the same altitude and inclina-

tion. The satellites in each shell can be sub-divided

into launch groups. These are the 15 to 60 satellites

contained in the payload of a single SpaceX rocket

launch (Boley et al., 2022; Wikipedia, 2023; Mc-

Dowell, 2020). All satellites in a given launch group

are deployed to the same shell. However, in larger

launches, they are split into different sub-groups con-

sisting of around 20 satellites which are deployed to

different planes, distinguished by their longitude of

the ascending node (Cakaj, 2021; McDowell, 2020).

The orbital shells can themselves be grouped into

two constellation generations. The ﬁrst generation

consists of the satellites in the orbits licenced by the

FCC on the 28

of March 2018. These are the satel-

lites in shells 1, 2, 3, 4, and 6. The second generation

consists of the satellites in orbits licenced by the FCC

on the 1

of December 2022, namely, the satellites in

shells 5 and 7 (Federal Communications Commission,

2022; Wikipedia, 2023; McDowell, 2023).

Preliminary analysis indicated that the manoeu-

vres of the satellites in a given launch group tended

to occur close together in time. Much of our analysis,

therefore, is performed by grouping the satellites into

their launch groups. These groups were determined

by using the launch dates included in the list of Star-

link satellites obtained from N2YO (N2YO, 2023).

3.2 Global Patterns in Manoeuvre

Frequency

We ﬁrst analyse global patterns across the entire con-

stellation of satellites. In order to make this analy-

sis feasible, we investigate the rate of manoeuvres de-

tected within each launch group of satellites. ﬁgure 1

plots a heatmap showing the rate of detected manoeu-

vres (in manoeuvres per satellite per day) within each

launch group.

IoTBDS 2024 - 9th International Conference on Internet of Things, Big Data and Security

176

Figure 1: The rate of manoeuvres for the different Starlink satellite launch groups. The rate is estimated using a Gaussian

kernel with σ = 0.5 days. Note that although the y axis is ordered by the launch dates, these are not equally spaced, and so it

does not have a consistent scale. The rate of manoeuvres is particularly high shortly after the launch of each satellite group.

There are also pronounced spikes in manoeuvre activity around December 2022 and August 2023.

This plot shows that there is a high incidence of

detected manoeuvres near the beginning of each satel-

lite’s time in orbit. After launch, Starlink satellites

perform extensive manoeuvring in order to reach their

ﬁnal orbit (Ashurov, 2022). However, in the authors’

experience, TLEs often contain artifacts in the earli-

est published epochs. It is, therefore, possible that

the large number of detected anomalies shortly after

launch is partially due to such artifacts.

The top right of the plot shows a marked increase

in the number of detected manoeuvres for older satel-

lites. This could be being driven by deorbiting ma-

noeuvres or be due to increased incidence of malfunc-

tion in these older satellties. There are also two sharp

increases in the number of detected manoeuvres that

affect most satellites in the constellation. The ﬁrst,

more pronounced, spike occurs on around the 10

December 2022 and the second on around the 2

September 2023. These spikes could be due to the

simultaneous manoeuvring of all or most satellites in

the constellation. However, the authors have encoun-

tered instances where the TLE data of multiple unre-

lated satellites undergo a simultaneous change, likely

due to changes in how they are processed by the 18

Space Defence Squadron. These spikes could be the

result of such an artifact. If they are not the result

of an artifact, then it is plausible that the spike on

around the 10

of December 2022 is the result of

an adjustment to the constellation in preparation for

the ﬁrst launch of the 2

generation shell of Starlink

satellites, which occurred on the 28

of December

2022 (Wikipedia, 2023). See section 3.1 for a descrip-

tion of the two generations of the constellation.

ﬁgure 1 also contains multiple lines of higher ma-

noeuvre intensity, at an angle from top left to bottom

right. These are likely the result of SpaceX perform-

ing sequential station-keeping or maintenance ma-

noeuvres across multiple launch groups. This is fur-

ther investigate in section 3.5.

How Is Starlink Manoeuvring? An Analysis of Patterns in the Manoeuvres of Starlink Satellites

177

(a) Individual manoeuvres and manoeuvre intensity.

(b) Fourier power spectrum of manoeuvre intensity.

Figure 2: A detailed analysis of the manoeuvres of satellite launch group ﬁfteen of shell one. The upper panel of (a) shows

the individual inferred manoeuvres for each satellite in the group. Each blue line is the timestamp of a detected manoeuvre

and each row contains the manoeuvres of a single satellite in the group. The satellites are ordered by their SATCAT num-

bers (Kelso, 1998), from top to bottom along the y axis. The lower panel shows an estimate of the intensity of manoeuvre

activity, estimated using a Gaussian kernel with a bandwidth of 2 days. (b) shows the Fourier power spectrum of the manoeu-

vre intensity. The fundamental frequency is at around 4.5 manoeuvres per year, representing an interval of around 80 days

between station-keeping manoeuvres.

3.3 Regular Station-Keeping

Manoeuvring

We now zoom in on a particular launch in order to get

a more precise idea of the exact timing of manoeu-

vres. By selecting a single, representative, launch

group we can inspect precise manoeuvre times, as op-

posed to only examining rates of manoeuvre detec-

tions.

The upper panel of ﬁgure 2a plots the precise

times of detected manoeuvres for each individual

satellite in the 15th launch group of shell one. These

satellites were launched on the 25

of November

2020. After an initial ﬂurry of manoeuvres post

launch, regularly-spaced near-synchronous manoeu-

vring begins — at regular intervals nearly every satel-

lite manoeuvres within a brief one to two day pe-

riod. Each such burst of manoeuvring is separated by

a gap of over 2 months. This regular pattern within

the launch group is highly indicative of the detected

anomalies being caused by manoeuvres as opposed

to artifacts in the TLE data. Moreover, their regu-

larity and synchronicity is indicative of them being

station-keeping manoeuvres. The orbits of satellites

in low earth orbit, such as the Starlink constellation,

degrade over time due to atmostpheric drag, solar ra-

diation pressure and the non-spherical nature of the

earth (Vallado, 2001). Regular station-keeping ma-

noeuvres are required to maintain the satellites in a

roughly constant orbit.

To highlight the periodic nature of these station-

keeping manoeuvres, the lower panel of ﬁgure 2a

IoTBDS 2024 - 9th International Conference on Internet of Things, Big Data and Security

178

Figure 3: The same plot as in ﬁgure 2a, but showing the inferred manoeuvres of launch group eight of shell one. In this older

launch group, the stereotyped nature of the station-keeping manoeuvres degrades over time. During the course of 2023, the

relative timing of the manoeuvres of the satellites in this group becomes less consistent and some satellites cease manoeuvring

during the station-keeping period.

Figure 4: The same plot as ﬁgure 1, however, the satellites are reordered along the y axis so that they are grouped into their

orbital shells. These shells are separated by the grey dashed lines and their number is given in grey text. The launch groups

within each shell have also been reordered so as to emphasize the sequential nature of the regular station-keeping manoeuvres.

How Is Starlink Manoeuvring? An Analysis of Patterns in the Manoeuvres of Starlink Satellites

179

Table 3: Various summary statistics and features of the orbital shells. The last column lists the order in which the launch

groups of each shell are plotted in ﬁgure 4. The numbers in this ordering correspond to the order that the satellites were

launched in. The assignment of launch groups to shells is made according to (McDowell, 2023).

Orbit Number of First Last Altitude Median number Order of launch

shell active launch launch date (km) of manoeuvres groups in

number satellites date included per satellite ﬁgure 4

1 1445 11/11/2019 26/5/2021 550 36 1, 3, 2, 4, 7, 6, 9,

5, 12, 8, 13, 10, 15, 14,

11, 18, 16, 19, 23, 17, 20,

21, 22, 25, 24, 28, 26, 27

2 403 14/9/2021 31/5/2023 570 10 1, 4, 5, 6, 7, 8, 9, 10

3 233 11/7/2022 27/4/2023 560 8 1, 2, 3, 5

4 1566 13/11/2021 17/12/2022 540 17 3, 5, 1, 4, 7, 6, 8, 11, 9, 10,

12, 14, 13, 16, 15, 19, 17,

27, 21, 26, 18, 22, 25, 29,

23, 20, 2, 34, 35, 36, 31, 37

5 692 28/12/2022 6/7/2023 530 7 1, 2, 3, 4, 5, 6, 7, 9,

10, 11, 12, 13, 15

6 539 27/2/2023 8/11/2023 559 3 1, 2, 3, 4, 5, 6, 7, 8,

9, 10, 11, 12, 13, 14, 15,

16, 17, 18, 19, 20, 21, 22,

23, 24, 25, 26, 27

7 127 22/8/2023 29/10/2023 525 2 1, 2, 3, 4, 5, 6

plots an estimate of the intensity of manoeuvring in

this satellite group. This estimate was performed us-

ing a Gaussian kernel with a bandwidth of 2 days.

Outside of the increase in manoeuvre activity shortly

post launch, the majority of spikes in this intensity

correspond to the regularly-spaced bursts in manoeu-

vres, likely due to station-keeping activity.

ﬁgure 2b plots the Fourier power spectrum of the

manoeuvre intensity plotted in the lower panel of ﬁg-

ure 2a. The fundamental frequency occurs at around

4.5 cycles per year. This corresponds to a period of

around 80 days between station-keeping manoeuvres.

Both shells one and four had well-deﬁned station-

keeping bursts over signiﬁcant periods of time (see

ﬁgure 4). All inspected launch groups in these two

shells had a similar fundamental frequency and cor-

responding interval between station-keeping manoeu-

vres.

Manoeuvres are also detected outside of the

station-keeping bursts, likely the result of collision-

avoidance manoeuvring (Uriot et al., 2022).

3.4 Degradation of Station-Keeping

Manoeuvres

We investigate the long-term stability of the station-

keeping manoeuvre patterns by plotting the manoeu-

vre times of the eighth launch group of shell one in

ﬁgure 3. These satellites were launched on the 13

of June 2020 and their behaviour is representative of

satellites launched both at a similar time and earlier.

In the earlier parts of these satellites’ lifespans,

they exhibited a highly-stereotyped pattern of regu-

lar manoeuvre bursts. Moreover, the relative ordering

of the satellites within each manoeuvre burst was con-

sistent across time.

However, this stereotyped pattern begins to break

down, particularly after around February 2023. Each

burst of manoeuvres occurs over a broader time in-

terval and the relative time ordering of the satellites

is less consistent. Moreover, in the last burst of ma-

noeuvres, no manoeuvres were detected for a number

of satellites.

This degradation is likely due to this launch group

approaching the end of its lifespan.

IoTBDS 2024 - 9th International Conference on Internet of Things, Big Data and Security

180

Figure 5: The same plot as ﬁgure 1, however, the station keeping and initial positioning manoeuvres have been removed. This

was achieved by removing all manoeuvres where the launch group’s manoeuvre intensity was greater than 0.04 manoeuvres

per satellite per day. The manoeuvre intensity was subsequently re-estimated after these periods of high intensity were

removed. Note that the colours are rescaled from previous ﬁgures due to the lower rate of manoeuvres. The majority of

remaining manoeuvres should reﬂect collision-avoidance actions. There are distinct periods of higher or lower incidence of

these manoeuvres. For instance, there is a noticeable increase in collision-avoidance manoeuvres in early 2022 for shell one

followed by a period of lower activity for the remainder of this year.

3.5 Ordering of Station-Keeping

Manoeuvres

After establishing that the satellite groups per-

form periodic group-wide station-keeping manoeu-

vres, our subsequent analysis concerns the coordina-

tion of these station-keeping manoeuvres between the

groups. There is a number of ways that SpaceX could

coordinate these manoeuvres. For instance, all the

satellites in a given shell could manoeuvre simultane-

ously, or certain supergroups of launch groups could

manoeuvre together.

As shown in ﬁgure 4, it appears that SpaceX cy-

cles through each launch group within each orbital

shell, performing station-keeping manoeuvres in a se-

quential fashion. This ﬁgure plots the manoeuvre in-

tensity of each launch group, similar to ﬁgure 1. How-

ever, the launch groups have been reordered so that

they are grouped into their orbital shells. Moreover,

within shells one and four, the launch groups have

been ordered so as to emphasize the sequential nature

of the station-keeping manoeuvres, although this was

also done so as to maintain the launch order as far as

possible. This reordering was not done for the other

shells as there were either too few launch groups, or

most launch groups had not been active long enough

to make such an ordering sensible. Note that, as mul-

tiple launch groups within each shell do manoeuvre

simultaneously, this ordering is not unique — other

orderings of the launch groups will yield a similar pat-

tern to that shown in ﬁgure 1.

Within shells one and four, at most three launch

groups are undergoing station-keeping manoeuvring

at any one point in time. The launch groups are cy-

cled through in an iterative fashion so that all satellites

within the shell undergo station-keeping manoeuvres

within the roughly 80 day period discussed in sec-

tion 3.3.

How Is Starlink Manoeuvring? An Analysis of Patterns in the Manoeuvres of Starlink Satellites

181

This behaviour makes sense within the context of

the management of the constellation. Presumably,

manoeuvring might interrupt the transmission of the

satellites and SpaceX would therefore want as even

a distribution of manoeuvring activity across time as

possible.

3.6 Trends in Collision Avoidance

Manoeuvres

The analysis performed so far has focussed on the

constellation’s station-keeping activity. However, we

know from SpaceX’s own ﬁlings with the FCC (Gold-

man, 2023) that the satellites in the Starlink con-

stellation perform a substantial number of collision-

avoidance manoeuvres.

We analysed trends in the rate of these manoeu-

vres over time by removing all station-keeping ma-

noeuvres, before re-analysing the intensity of ma-

noeuvring within each launch group. This was done

by ﬁrst estimating the intensity of manoeuvres within

each launch group using a Gaussian kernel with a

bandwidth of two days. All manoeuvres that occurred

within periods with an estimated intensity above 0.04

manoeuvres per satellite per day were removed. This

threshold was chosen through hand-tuning, with the

goal being to ﬁnd the highest threshold for which the

pattern of periodic station-keeping manoeuvring dis-

appeared. Note that this process also removes the

intense manoeuvre activity that occurs shortly after

launch.

After the station-keeping manoeuvres were re-

moved, the manoeuvre rate of each satellite was re-

estimated. The resulting manoeuvre rates are plot-

ted in ﬁgure 5. This ﬁgure reveals that the rate of

incidence of collision-avoidance manoeuvres is vari-

able across time. For instance, orbital shell one has a

particularly elevated rate of these manoeuvres around

April 2022 and February 2023. However, this rate is

substantially lower between May 2022 and January

2023. Similar to shell one, shell four has a raised rate

of manoeuvres in February 2023, but this is followed

by a period of particularly sparse manoeuvring in the

middle of 2023.

4 DISCUSSION

We applied a recently-developed anomaly detection

algorithm for TLE data (Shorten et al., 2023a), to the

TLE data of all currently-active Starlink satellites.

A question of primary concern is whether these

detected anomalies correspond to the manoeuvres of

the Starlink satellites. These anomalies could be

caused by a number of other potential factors, such as

artifacts in the TLE data or satellite malfunction. The

paper proposing the deployed ﬁltering technique for

anomaly detection (Shorten et al., 2023a) performed

a thorough evaluation of its accuracy. This was per-

formed on a benchmark dataset (Shorten et al., 2023b)

which contained the TLEs of 15 satellites and associ-

ated independently-obtained ground-truth manoeuvre

timestamps. This dataset contained a mixture of satel-

lites in geosynchronous and low-earth orbits, which

is where the Starlink constellation is situated. It was

demonstrated that this approach was able to achieve

fairly high performance on most satellites, achieving

F1 scores over 0.8 in many cases. However, these

scores are not perfect. Moreover, no validation of this

approach has been performed on Starlink satellites,

which make use of electric propulsion (Holste et al.,

2020) compared with the chemical propulsion of the

satellites in the benchmark dataset. It is therefore al-

most certain that some of the anomalies detected in

the Starlink TLE data are false positives. It is also

highly likely that some manoeuvres went undetected.

That being said, there are features of the pat-

terns of detected anomalies which should increase

our conﬁdence in the results. As discussed in sec-

tion 3.3, there are regular bursts of manoeuvres where

an anomaly is detected for all satellites in one launch

group within a narrow period (around a day). More-

over, as shown in ﬁgure 4, different satellite groups

rarely exhibit such bursts of anomalies simultane-

ously. This behaviour is easy to explain if the detected

anomalies are station-keeping manoeuvres being per-

formed across all satellites on the same or similar or-

bital plane. However, it is difﬁcult to see why arti-

facts in TLEs would line up so precisely based on the

satellite’s launch date or orbital plane. The consis-

tent relative timing of the manoeuvres of each satellite

in a launch group during the station-keeping manoeu-

vres, as discussed in section 3.3, is similarly better

explained by patterns in how SpaceX is choosing to

manoeuvre its satellites than by patterns in the distri-

bution of artifacts in TLE data.

The Starlink constellation exhibits a rich pattern

of detected manoeuvres across time and the different

satellite launch groups. ﬁgure 1 provides a global

summary of the detected manoeuvres by plotting a

heatmap of the rate of detected manoeuvres within

each satellite group. The rate of manoeuvres was

generally higher shortly after launch and in the lat-

ter parts of older satellites’ lifespans. There are also

some noticeable spikes in manoeuvring activity, such

as around the 10

of December 2022.

More detailed analysis of the precise manoeuvre

times of individual satellites reveals further structure.

IoTBDS 2024 - 9th International Conference on Internet of Things, Big Data and Security

182

ﬁgure 2 plots the precise manoeuvre times for all

satellites in a representative launch group. The group

undergoes regularly-spaced bursts in manoeuvring ac-

tivity, where all satellites in the group manoeuvre

near-simultaneously. Moreover, the relative ordering

of the manoeuvres within the bursts remains consis-

tent over time. The frequency of these bursts in both

the representative launch group as well as in other

launch groups studied was around 4.5 bursts per year,

giving a corresponding period of around 80 days be-

tween bursts. These bursts are likely the result of reg-

ular station-keeping manoeuvres being performed by

SpaceX. Anomalies unique to individual satellites are

also detected outside of these bursts. It is likely that a

signiﬁcant proportion of such anomalies are the result

of collision-avoidance manoeuvres by SpaceX.

The regularity of the station-keeping manoeuvres

was shown to break down over time in certain satel-

lites. ﬁgure 3 plots the precise manoeuvre times of an

early launch group. In the earlier parts of their lifes-

pan, the satellites exhibit a highly stereotyped pattern

of manoeuvres, with the relative ordering of the satel-

lites within each burst remaining highly consistent.

This consistency breaks down close to the present day.

Moreover, some satellites no longer manoeuvre in the

bursts.

Section 3.5 studied the relationship in the tim-

ing of the station-keeping manoeuvre bursts between

the launch groups of each shell. The launch groups

rarely performed their station-keeping manoeuvres si-

multaneously. Instead, it appears that SpaceX moves

through the different launch groups in an iterative

fashion so as to minimize the number of satellites ma-

noeuvring simultaneously.

Finally, 3.6 studied the incidence of collision

avoidance manoeuvres in the constellation by remov-

ing the station-keeping manoeuvres from the anal-

ysis and examining the remaining manoeuvres. It

was found that the different shells of the constellation

go through distinct periods of increased or decreased

collision-avoidance manoeuvring activity.

5 CONCLUSION

Due to the anticipated rapid increase in their size,

large satellite constellations such as Starlink will in-

crease the difﬁculty of space trafﬁc management.

However, little is currenlty known about how SpaceX

manages this constellation and, in particular, how it

manoeuvres the satellites within it. This work ad-

dressed this lack of insight by applying a recently-

developed anomaly detection algorithm to the TLE

data of all satellites currently active in the Starlink

constellation. It was found that satellites undergo

an initial period of rapid manoeuvring before settling

into a state dominated by regular bursts in the ma-

noeuvring of all satellites within a particular launch

group to perform station keeping. These bursts are in-

terspersed with isolated collision-avoidance manoeu-

vres. This pattern of regular bursting degrades over

time in the earliest launches.

ACKNOWLEDGEMENT

We thank Will Heyne of BAE Systems for help-

ful insights throughout the course of this research.

This work has been supported by the SmartSat CRC,

whose activities are funded by the Australian Govern-

ment’s CRC Program.

REFERENCES

Ashurov, A. E. (2022). An effective method for detecting

satellite orbital maneuvers and its application to LEO

satellites. Advances in Aircraft and Spacecraft Sci-

ence, 9(4):279.

Boley, A. C. and Byers, M. (2021). Satellite mega-

constellations create risks in low earth orbit, the atmo-

sphere and on earth. Scientiﬁc Reports, 11(1):10642.

Boley, A. C., Wright, E., Lawler, S., Hickson, P., and

Balam, D. (2022). Plaskett 1.8 m observations of Star-

link satellites. The Astronomical Journal, 163(5):199.

Cakaj, S. (2021). The parameters comparison of the Star-

link LEO satellites constellation for different orbital

shells. Frontiers in Communications and Networks,

2:643095.

Decoto, J. and Loerch, P. (2015). Technique for GEO RSO

station keeping characterization and maneuver detec-

tion. In Advanced Maui Optical and Space Surveil-

lance Technologies Conference, page 42.

Federal Communications Commission (2022). FCC 22-91.

Goldman, D. (2023). SpaceX constellation status report,

December 1, 2022 to May 31, 2023. Technical report,

Space Exploration Technologies Corp.

Holste, K., Dietz, P., Scharmann, S., Keil, K., Henning,

T., Zsch

atzsch, D., Reitemeyer, M., Nausch

utt, B.,

Kiefer, F., Kunze, F., et al. (2020). Ion thrusters

for electric propulsion: Scientiﬁc issues developing a

niche technology into a game changer. Review of Sci-

entiﬁc Instruments, 91(6).

Kelso, T. (1998). Frequently asked questions: Two-line el-

ement set format. celestrak.org/columns/v04n03/.

Accessed: 2022-10-26.

Lal, B., Balakrishnan, A., Caldwell, B. M., Buenconsejo,

R. S., and Carioscia, S. A. (2018). Global trends in

Space Situational Awareness (SSA) and Space Trafﬁc

Management (STM). Science and Technology Policy

Institute, 10.

How Is Starlink Manoeuvring? An Analysis of Patterns in the Manoeuvres of Starlink Satellites

183

Li, T., Li, K., and Chen, L. (2018). New manoeuvre detec-

tion method based on historical orbital data for low

earth orbit satellites. Advances in Space Research,

62(3):554–567.

Li, T., Li, K., and Chen, L. (2019). Maneuver detection

method based on probability distribution ﬁtting of the

prediction error. Journal of Spacecraft and Rockets,

56(4):1114–1120.

McDowell, J. (2020). The low earth orbit satellite popula-

tion and impacts of the SpaceX Starlink constellation.

The Astrophysical Journal Letters, 892(2):L36.

McDowell, J. (2023). Jonathan’s space pages - Starlink

statistics. planet4589.org/space/con/star/stats.html.

Accessed: 2023-12-10.

Michel, F., Trevisan, M., Giordano, D., and Bonaventure,

O. (2022). A ﬁrst look at starlink performance. In

Proceedings of the 22nd ACM Internet Measurement

Conference, pages 130–136.

Mukundan, A. and Wang, H.-C. (2021). Simpliﬁed ap-

proach to detect satellite maneuvers using TLE data

and simpliﬁed perturbation model utilizing orbital el-

ement variation. Applied Sciences, 11(21):10181.

N2YO (2023). Starlink satellites. www.n2yo.com/satellite

s/?c=52. Accessed: 2023-11-30.

Pelton, J. N. (2015). New solutions for the space debris

problem. Springer.

Shorten, D. P., Maclean, J., Humphries, M., Yang, Y., and

Roughan, M. (2023a). Optimal proposal particle ﬁl-

ters for detecting anomalies and manoeuvres from two

line element data. arXiv preprint arXiv:2312.02460.

Shorten, D. P., Yang, Y., Maclean, J., and Roughan, M.

(2023b). Wide-scale monitoring of satellite lifetimes:

Pitfalls and a benchmark dataset. Journal of Space-

craft and Rockets, pages 1–5.

Snyder, C. (2011). Particle ﬁlters, the “optimal” proposal

and high-dimensional systems. ECMWF Seminar on

Data assimilation for atmosphere and ocean, page 10.

Space-Track (2023). SPACE-TRACK.ORG. space-track.o

rg. Accessed: 2023-12-05.

SpaceX (2023). Launches. www.spacex.com/launches/.

Accessed: 2023-12-05.

Union of Concerned Scientists (2023a). How many satel-

lites are in space? the spike in numbers continues.

blog.ucsusa.org/syoung/how-many-satellites-are-i

n-space-the-spike-in-numbers-continues/. Accessed:

2023-12-04.

Union of Concerned Scientists (2023b). UCS satellite

database. www.ucsusa.org/resources/satellite-dat

abase. Accessed: 2023-12-04.

Uriot, T., Izzo, D., Sim

oes, L. F., Abay, R., Einecke, N.,

Rebhan, S., Martinez-Heras, J., Letizia, F., Siminski,

J., and Merz, K. (2022). Spacecraft collision avoid-

ance challenge: Design and results of a machine learn-

ing competition. Astrodynamics, 6(2):121–140.

Vallado, D., Crawford, P., Hujsak, R., and Kelso, T. (2006).

Revisiting spacetrack report #3. In AIAA/AAS As-

trodynamics Specialist Conference and Exhibit, page

6753.

Vallado, D. A. (2001). Fundamentals of astrodynamics and

applications, volume 12. Springer Science & Busi-

ness Media.

Vallado, D. A. and Cefola, P. J. (2012). Two-line element

sets - practice and use. In 63rd International Astro-

nautical Congress, Naples, Italy, pages 1–14.

Venkatesan, A., Lowenthal, J., Prem, P., and Vidaurri, M.

(2020). The impact of satellite constellations on space

as an ancestral global commons. Nature Astronomy,

4(11):1043–1048.

Wikipedia (2023). List of Starlink and Starshield Launches.

en.wikipedia.org/wiki/List

of Starlink and Starshiel

d launches. Accessed: 2023-12-10.

Zhao, Y., Zhang, K., Bennett, J., Sang, J., and Wu,

S. (2014). A method for improving two-line ele-

ment outlier detection based on a consistency check.

In Proceedings of the Advanced Maui Optical and

Space Surveilllance (AMOS) Technologies Confer-

ence, Maui, Hawaii.

IoTBDS 2024 - 9th International Conference on Internet of Things, Big Data and Security

184