《蘭德公司：2024人工智能和機器學習在太空域感知（SDA）中的應用-基于兩項人工智能案例（英文版）（63頁）.pdf》由會員分享，可在線閱讀，更多相關《蘭德公司：2024人工智能和機器學習在太空域感知（SDA）中的應用-基于兩項人工智能案例（英文版）（63頁）.pdf（63頁珍藏版）》請在三個皮匠報告上搜索。

1、LI ANG ZHANG,KRISTA LANGELAND,JONATHAN TRAN,JORDAN LOGUE,PRATEEK PURI,GEORGE NACOUZI,ANTHONY JACQUES,GARY J.BRIGGSArtificial Intelligence and Machine Learning for Space Domain AwarenessCharacterizing the Impact on Mission EffectivenessResearch ReportFor more information on this publication,visit www

2、.rand.org/t/RRA2318-1.About RANDRAND is a research organization that develops solutions to public policy challenges to help make communities throughout the world safer and more secure,healthier and more prosperous.RAND is nonprofit,nonpartisan,and committed to the public interest.To learn more about

3、 RAND,visit www.rand.org.Research IntegrityOur mission to help improve policy and decisionmaking through research and analysis is enabled through our core values of quality and objectivity and our unwavering commitment to the highest level of integrity and ethical behavior.To help ensure our researc

4、h and analysis are rigorous,objective,and nonpartisan,we subject our research publications to a robust and exacting quality-assurance process;avoid both the appearance and reality of financial and other conflicts of interest through staff training,project screening,and a policy of mandatory disclosu

5、re;and pursue transparency in our research engagements through our commitment to the open publication of our research findings and recommendations,disclosure of the source of funding of published research,and policies to ensure intellectual independence.For more information,visit www.rand.org/about/

6、research-integrity.RANDs publications do not necessarily reflect the opinions of its research clients and sponsors.Published by the RAND Corporation,Santa Monica,Calif.2024 RAND Corporation is a registered trademark.Library of Congress Cataloging-in-Publication Data is available for this publication

7、.ISBN:978-1-9774-1447-2Cover:BlackJack3D/Getty Images.Limited Print and Electronic Distribution RightsThis publication and trademark(s)contained herein are protected by law.This representation of RAND intellectual property is provided for noncommercial use only.Unauthorized posting of this publicati

8、on online is prohibited;linking directly to its webpage on rand.org is encouraged.Permission is required from RAND to reproduce,or reuse in another form,any of its research products for commercial purposes.For information on reprint and reuse permissions,please visit www.rand.org/pubs/permissions.ii

9、i About This Report The U.S.Space Force(USSF)space domain awareness(SDA)mission is complex and rapidly growing in complexity.Space objects and activities have differing priorities for observation,and the SDA mission needs to optimize collection and analysis according to these priorities.For example,

10、high-interest objects may require a higher revisit rate for sensor collection than an object characterized as space debris.In addition to these varying priorities,the sheer data volume,the complexities of orbital mechanics and space operations,and the time constraints introduced in an operational wa

11、rfighting domain,present significant challenges to those tasked with the SDA mission and point to this mission as a prime candidate for support from artificial intelligence and machine learning(AI/ML)tools.The work should be of interest to USSF leadershipin particular,the Office of the Assistant Sec

12、retary for Space Acquisition and Integration and Space Systems Commandand to USSF personnel tasked with carrying out the SDA mission,notably those at the 18th and 19th Space Defense Squadrons.This work is accompanied by a companion report that describes the technical details of two AI/ML tools devel

13、oped in support of this research,Artificial Intelligence and Machine Learning for Space Domain Awareness:The Development of Two Artificial Intelligence Case Studies by Jonathan Tran,Prateek Puri,Jordan Logue,Anthony Jacques,Li Ang Zhang,Krista Langeland,George Nacouzi,and Gary J.Briggs(RR-A2318-2).T

14、he research reported here was commissioned by the Office of the Chief Scientist of the U.S.Air Force(AF/ST)and conducted within the Force Modernization and Employment Program of RAND Project AIR FORCE as part of a fiscal year 2023 project,“Artificial Intelligence for Space Domain Awareness.”RAND Pro

15、ject AIR FORCE RAND Project AIR FORCE(PAF),a division of RAND,is the Department of the Air Forces(DAFs)federally funded research and development center for studies and analyses,supporting both the United States Air Force and the United States Space Force.PAF provides the DAF with independent analyse

16、s of policy alternatives affecting the development,employment,combat readiness,and support of current and future air,space,and cyber forces.Research is conducted in four programs:Strategy and Doctrine;Force Modernization and Employment;Resource Management;and Workforce,Development,and Health.The res

17、earch reported here was prepared under contract FA7014-22-D-0001.Additional information about PAF is available on our website:www.rand.org/paf/iv This report documents work originally shared with the DAF on September 25,2023.The draft report,dated September 2023,was reviewed by formal peer reviewers

18、 and DAF subject-matter experts.Acknowledgments We are grateful to Victoria Coleman for her sponsorship of this important topic,as well as Daniel Eckhardt and Col David Montminy for their support and assistance throughout this project.In addition,we would like to thank Joel Mozer,Col Nathan Iven,and

19、 Gen David D.Thompson for sharing their ideas and feedback.This research could not have been completed without detailed discussions and input from stakeholders at the 18th SDS and 19th Space Defense Squadrons,and we are so grateful for their time with our project team.There are many other stakeholde

20、rs on this topic who shared their research and ideas,and,in particular,we would like to thank Justin Fletcher and Col Aaron Celaya for helping us scope and shape this project.We thank our reviewers,Kanna Rajan and Krista Grocholski,for their comments,which strengthened this report.Lastly,we are grat

21、eful to Sherrill Lingel and Alexander Hou for the program support for this research,as well as Brien Alkire for his mentorship and guidance as an in-stride reviewer during the study year.v Summary Issue To address the growing demands of operating in the space domain,space domain awareness(SDA)operat

22、ors must determine how to prioritize sensor observations more effectively,scale up to meet the sheer volume of resident space objects,and develop analytic capabilities that reflect the complexity of orbital mechanics and space operations,all while maintaining the responsiveness necessitated by opera

23、tions in a warfighting domain.These factors present significant challenges to those tasked with the SDA mission and point to this mission as a prime candidate for support from artificial intelligence(AI)and machine learning(ML)tools,because such tools have the potential to increase the analysis temp

24、o,expand the amount of usable data for this analysis,and free up operator time for more-complex tasks.AI/ML tools can help SDA operators meet these increasing challenges.For example,leveraging AI/ML to optimize sensor scheduling could improve the accuracy of the space catalog by allowing more-effici

25、ent use of existing sensing resources.AI/ML tools can also facilitate the calculation of close proximity risks and provide indications and warning by flagging abnormal events.Although AI/ML tools have the potential to help meet these SDA challenges,the impact of these tools on the overall success of

26、 the SDA mission is not well understood,and this lack of understanding is a barrier to planning and optimizing the integration of the tools.This research seeks to develop insight into the following research questions:How can existing AI/ML and automation technologies benefit the SDA mission?Which pr

27、ocesses within the SDA mission area can be improved by increased automation and AI/ML?What are the prerequisites and enabling technologies that are necessary for AI/ML implementation?What limitations should be considered when employing these technologies?Approach Our approach for this research focus

28、ed on characterizing the nature of the impact that AI/ML tools could bring to the SDA mission.To accomplish this,we examined the processes that make up this mission,the metrics for these processes,the opportunities for impact,and the impact of these tools on the process and the outcomes.The research

29、 presented here focuses in detail on the conjunction assessment(CA)mission to enable us to look at a singular SDA process in depth.CA is the probabilistic process by which the 18th and 19th Space Defense Squadrons(SDSs)quantify the risk of collision in space.To support this assessment,we interviewed

30、 stakeholders,reviewed existing academic and doctrinal literature,developed detailed process maps,and built exploratory AI/ML models.vi Key Findings Because of the growing demands and changing nature of the SDA mission,AI/ML tools have a high opportunity for impact if they can be force multipliers f

31、or SDA operators.AI/ML tools cannot address all the challenges of the SDA mission,but process changes can help these tools achieve greater impact.Realizing significant impacts from AI/ML requires moving to an architecture that enables more AI/ML development and fielding.AI/ML tool development that s

32、upports more-optimized sensor tasking could have a cascading impact on the rest of the SDA mission.Better quantification of risk and uncertainty tolerance can support the improved performance of AI/ML tools focused on prediction and classification.Recommendations Space Systems Command SDA Acquisitio

33、n Delta(SSC/SZG)and the Air Force Research Laboratory should provide clear guidance to AI/ML tool developers to focus on tools that address the needs of operators today but could also benefit a future SDA architecture.AI/ML tool development should focus on force multipliers that help meet the growin

34、g detection and characterization challenges faced by the 18th and 19th SDSs.SSC/SZG should look for opportunities to invest in these tools.SDA operatorsin particular,at the 18th SDS,19th SDS,and National Space Defense Centershould seek ways to articulate mission needs,sensor requirements,and accepta

35、ble uncertainty to optimize potential enhancements from AI/ML tools.Space Operations Command,via the SDA Mission Area Team,should examine where SDA processes could be modified to enable more-significant impact from AI/ML tools.AI/ML tool development should focus on those tools that enable more-effec

36、tive and more-efficient sensor tasking,and SSC/SZG should seek ways to acquire and develop those tools.Efforts at the 18th SDS to capture observation intent should also be supported,as these efforts,in combination,are key enablers.SSC/SZG should continue to support efforts in AI/ML tool development

37、and ensure that there are processes and infrastructure in place to test and validate these models.The availability of high-quality training data is another enabler for AI/ML impact,and USSF should support efforts at Space Systems Command in its Cross Mission Data team to ensure the availability of t

38、hese data to AI/ML tool developers.vii Contents About This Report.iiiSummary.vFigures and Table.ixCHAPTER 1.1Introduction:The Changing Space Domain Awareness Mission.1Study Purpose.1Methodology:Change Impact Analysis.2CHAPTER 2.4Background:The Space Domain Awareness Mission.4Defining the SDA Mission

39、:What Are the Objectives?.6SDA Functions:What Does SDA Do?.7SDA Operators.7The Changing SDA Mission and Infrastructure.8Chapter Findings.11CHAPTER 3.12Applying Artificial Intelligence in Space Domain Awareness.12Existing Efforts.13Sensor Tasking,Scheduling,and Prioritization.14Data Collection and In

40、tegration.16OD .17Catalog Maintenance and Uncorrelated Tracks.20CA Calculation.21Chapter Findings.22CHAPTER 4.25Improvement Models for Characterizing Impact.25Characterizing the Value of AI/ML for CA.25Improvement 1:AI/ML Supports More-Efficient Use of SDA Sensors,Resulting in Higher Revisit Rates.2

41、6Improvement 2:AI/ML for Improving Covariance Calculations in the OD Process.29Impact Characterization.30Simple Models to Realize Impact.32CHAPTER 5.34Implementation Challenges,Conclusions,and Recommendations.34Implementation Challenges.34Conclusions.36Recommendations.38 viii APPENDIX A.40Overview o

42、f Space Domain Awareness Functions.40APPENDIX B.42Summary of Artificial Intelligence and Machine Learning Tool Development and Proof of Concept.42 Abbreviations.47References.49 ix Figures and Table Figures Figure 1.1.Steps in Change-Impact-Analysis Methodology.3Figure 2.1.Tracked Space Objects,as of

43、 April 2022.5Figure 3.1.Hierarchy Illustrating Functional Relationships for SDA and Space C2.13Figure 4.1.The Orbit of a GEO Satellite with No Maneuvers.27Figure 4.2.The Orbit of a GEO Satellite with 50 m/s Maneuver.28Figure 4.3.Overview of Proposed Modified CA Process with Dynamic Sensor Tasking.30

44、Figure 4.4.Number of Collision Warnings Assuming No Sensor Updates.32Figure 4.5.Maximum Number of Collision Warnings as a Function of Update Rate.32Figure B.1.Performance Optimization.43Figure B.2.BNN Performance at Predicting T+1,000 s.45 Table Table 2.1.Key Enablers for Improved SDA Mission Effect

45、iveness.7 x 1 Chapter 1 Introduction:The Changing Space Domain Awareness Mission The U.S.Department of Defense(DoD)acknowledges the critical role of space in all-domain operations in future conflict and recognizes that understanding the role of space technologies is critical for optimizing investmen

46、t in and operationalization of these technologies in support of all-domain operations.Gen James Dickinson,while commander of U.S.Space Command(USSPACECOM),named space domain awareness(SDA)as the commands number one need.1 Artificial intelligence(AI)is a broad field with competing definitions from va

47、rious academic fields;for the purposes of this report,we consider it alongside machine learning(ML)as providers of the following technological capabilities:computer vision,language processing,reinforcement learning,prediction and classification,generative learning,and expert systems.2 We focus on un

48、derstanding how these capabilities can help the U.S.Space Force(USSF)in the SDA mission.AI/ML has the potential to establish superior SDA and enable rapid decisionmaking in the SDA domain,outpacing that of adversaries.Understanding how to optimally use these technologies and plan for their expected

49、impact on the SDA mission will help improve mission effectiveness,optimize resource management,and inform investment decisions.However,the potential benefits from the application of AI/ML to the SDA mission is not well understood.In particular,the U.S.Air Force and USSF would like to know what the i

50、mpact of such tools might be and,for those high-impact areas,how they can invest to support the implementation of these tools.Study Purpose To build this understanding,we developed a framework for identifying how existing AI/ML technologies could be applied to support decisionmaking in support of SD

51、A and to assess the material benefits of such implementations.Specifically,we had the following research questions:How can existing AI/ML and automation technologies benefit the SDA mission?Which processes within the SDA mission area can be improved by increased automation and AI/ML?1 Erwin,“Space D

52、omain Awareness.”2 We drew from U.S.Code,Title 15,Section 9401,which provides policy-oriented definitions for AI and ML.Artificial intelligence is defined as“a machine-based system that can,for a given set of human-defined objectives,make predictions,recommendations,or decisions influencing real or

53、virtual environments.”Machine learning is“an application of artificial intelligence that is characterized by providing systems the ability to automatically learn and improve on the basis of data or experience.”2 What are the prerequisites and enabling technologies that are necessary for AI/ML implem

54、entation?What limitations should be considered when employing these technologies?Methodology:Change Impact Analysis To characterize the potential impact of AI/ML tools on the SDA mission,we examined the detailed processes that make up this mission and the metrics for these processes,identified oppor

55、tunities for impact through AI/ML tools,and explored the potential change these tools would have on the process and the outcomes.The analysis focuses in detail on the conjunction assessment(CA)mission to enable in-depth examination of a singular SDA process.CA is foundational to other SDA processes,

56、as it includes sensor tasking,data ingestion,characterization of movement,and correlation to known objects.All of these processes support other SDA processes,such as threat characterization and launch detection.We can therefore look at CA as both a case study and a foundational element within SDA mo

57、re broadly.To first assess the impact of change from AI/ML tools,the RAND team conducted interviews with stakeholders and executed a review of existing academic and doctrinal literature to build a view of what the SDA mission is,how it is changing,and what that could mean for AI/ML tool implementati

58、on and impact.The team conducted in-person interviews with operators from each cell within the 18th and 19th Space Defense Squadrons(SDSs),including both uniformed and contractor personnel.The team also engaged with the Kobayashi Maru software factory to gain insight into the infrastructure moderniz

59、ation efforts to support the SDA mission and potential AI applications.3 To understand other SDA efforts,the team conducted virtual and in-person interviews with various organizations within the Space Systems Command(SSC)to learn about acquisitions and existing programs,including the AI and Autonomy

60、 Sandbox.In total,we interviewed 23 stakeholders across nine organizations to gain a broad understanding of the mission.Recognizing the critical importance of understanding the intricate details of the SDA mission and operator needs,we had a researcher to spend a week with the 18th SDS.Spending time

61、 with operators allowed for a more in-depth understanding of the unique challenges in the daily SDA operations.We then used this foundation to inform our impact assessment.To ensure that our impact assessment was systematic,we followed steps in the change-impact-analysis methodology used in the deve

62、lopment of change management processes to better understand the holistic impact of deliberate changes on a system.4 This methodology consists of four steps,described in Figure 1.1.3 See Chapter 2 for background about the Kobayashi Maru program.4 Rebentisch et al.,“Assessment of Changes in Technical

63、Systems and Their Effects on Cost and Duration Based on Structural Complexity.”3 Figure 1.1.Steps in Change-Impact-Analysis Methodology SOURCE:Features information from Dietzmann and Alt,“Assessing the Business Impact of Artificial Intelligence.”The first step,determine impacted elements,aims to det

64、ermine at which points AI/ML tools may be incorporated into a process.To identify these points,we engaged with the available literature and worked with stakeholders to develop a process map for a specific SDA mission:CA.We engaged with stakeholders at the operational level to identify where their ne

65、eds are most acute and therefore capture where there is high potential for impact from AI/ML tools.This research is presented in Chapter 2.The next step,determine change activities,looks at AI/ML tools to assess which tools would be appropriate for the different process points identified above.To ac

66、complish this step,we surveyed AI/ML tools more broadly and provided an overview of which tools may be best suited for the CA process.This research is presented in conjunction with the first step in Chapter 2.The third step,determine performance metrics,requires identifying metrics for the overall p

67、rocess and submetrics for each point at which AI/ML tools may be applied.We identified these metrics and used them to develop models to help understand the impact of two specific types of improvements that AI/ML could bring to the overall CA process.These models are discussed in Chapter 3.The last s

68、tep,determine change effort,seeks to capture the enablers and barriers to implementing AI/ML tools in this process.Although this report does not specifically focus on how to address these enablers and barriers,Chapter 4 discusses what they are and how addressing them could change the nature of the i

69、mpact of AI/ML tools on the CA process.The change-impact-analysis framework enabled us to identify key areas in which AI/ML could improve and support the CA process and facilitated the characterization of the impact from these tools.We then used this output to inform the design and building of two p

70、roof-of-concept tools to explore their application and potential for implementation in more detail.Determine Impacted ElementsDetermine Change ActivitiesDetermine Performance MetricsDetermine Change Effort1234Develop a process map to identify where in the process AI/ML tools could be integratedConsu

71、lt with stakeholders to identify key areas of need that AI/ML tools could fillIdentify key areas for AI/ML impact in each process based on these process mapsIdentify key metrics for decision quality at each decision point in the processDevelop a relationship map for these metrics to help determine w

72、hich metrics have the potential for greatest impactIdentify key enablers and requirements to execute the assessed changesAt which process elements could AI/ML tools be incorporated?Which AI/ML tools fit in these process elements?How can we measure impact from AI/ML tools on the SDA mission?What barr

73、iers and constraints need to be addressed to implement these tools?4 Chapter 2 Background:The Space Domain Awareness Mission SDA is vital to maintaining safe and secure access to space,deterring adversary actions in space,and achieving space superiority should a crisis or conflict occur.Gen B.Chance

74、 Saltzman,chief of space operations for the USSF,identified“comprehensive and actionable”awareness of the space domain as a core tenet of his theory of success.5 However,SDA is increasingly challenged by the launch of new constellations,maneuverable assets,and activities from noncooperative actors i

75、n space.The U.S.Space Surveillance Network(SSN)currently tracks about 9,000 active payloads,17,000 analyst objects,and 19,000 pieces of debris,for a total of about 45,000 objects overall,and maintains these in a catalog.6 That number is growing dramatically from launches,collisions,and testing of di

76、rect-ascent antisatellite weapons,as shown in Figure 2.1.5 Hadley,“Saltzman Unveils Competitive Endurance Theory to Guide Space Force.”6 Space-Track.org,homepage.5 Figure 2.1.Tracked Space Objects,as of April 2022 SOURCE:Reprinted from Astromaterials Research and Exploration Science,National Aeronau

77、tics and Space Administration,Legend.”To meet this challenge of sustaining and improving SDA,the USSF needs to incorporate new tools to better leverage SSN resources in a way that can scale with the growth of activity and complexity in the SDA mission.7 Using tools to address this change and growth

78、are crucial for maintaining safety in space and ensuring the ability to protect and defend U.S.assets in orbit.Quicker response times are necessary,and improvements in SDA decisionmaking timelines are essential for achieving USSF goals for tactically responsive space capabilities.8 Managing congesti

79、on by tracking and forecasting space object trajectories is also crucial for improving awareness in support of space defense.The increasing amount of data generated by new and maneuverable space objects will require more analytic capabilities.9 Effective defense relies on improved ability to charact

80、erize and forecast as strategic competitors increase their presence and level 7 The emerging challenges facing the SDA mission are twofold:growing congestion and increasing competition.See,for example,Brock,“Space Delta 2.”As the number of satellite constellations increases and strategic competition

81、 grows in intensity,SDA is moving beyond tracking existing space assets and debris to managing congestion and increasing focus on space defense.As a result,decisionmaking needs in support of SDA are changing and growing in complexity.8 Marrow,“Space Force Eyes 20252026 Timeframe for Tactically Respo

82、nsive Space Capabilities.”9 SDA stakeholders recognize this need.For example,Gen Stephen Whiting,former head of Space Operations Command and currently the new commander for USSPACECOM,acknowledged the potential for AI/ML to affect the SDA mission,noting that“we have a lot of data and AI/ML can help

83、us parse through that data”(quoted in Wolfe,“U.S.Space Force Looks to AI/ML to Aid Space Domain Awareness”).6 of activity in the space domain.10 The requirements for improving SDA provide an outline for how AI/ML technologies can play a critical role in establishing superior SDA and enabling better

84、and faster decisionmaking than adversaries.SDA is a priority mission area for the United States,and we have noted the changing SDA mission needs,but the United States also currently lacks strategic understanding of the operational advantages that AI/ML could provide to the SDA mission.Addressing thi

85、s shortfall and determining how to leverage AI/ML in the SDA mission requires an in-depth understanding of the benefits and efficiencies that these tools could provide,what data are needed to support these tools,and which technologies may be available in the near term to support SDA.The output of th

86、is work will support the U.S.Department of the Air Forces decisionmaking as it assesses opportunities for using AI/ML in SDA and will provide benefit to the broader space community,including the National Reconnaissance Office,the National Aeronautics and Space Administration,the National Geospatial-

87、Intelligence Agency,the Air Force Research Laboratory,Defense Advanced Research Projects Agency,and the Space Data Association.Defining the SDA Mission:What Are the Objectives?SDA is one of the five core competencies of the USSF and“encompasses the effective identification,characterization and under

88、standing of any factor associated with the space domain that could affect space operations and thereby impact the security,safety,economy,or environment of our Nation.”11 Although the concept of SDA is not new,12 the mission has evolved from maintaining catalogs and assessing conjunction threats of

89、benign objects with limited maneuverability to addressing a threat environment that now includes a greater number of highly maneuverable and potentially noncooperative objects.13 Data for SDA originate from various sources,including the SSN,commercial assets,and allies and partners that contribute d

90、ata from both ground-and space-based sensors.Contributing sensors 10 As noted by one former space control officer with the former U.S.Air Force Space Command,the ability to characterize and forecast objects in space is crucial for effective defense:“The United States cant stop China from moving its

91、own satellites.But what I certainly want to do is know when theyre moving around and be able to forecast a little bit to know if I need to defend myself or protect myself.And so that is really the crux of it”(Erwin,“Space Domain Awareness,”quoting Brian Young).11 USSF,Spacepower.12 Space situational

92、 awareness(SSA)has been around since the beginning of the space era,when the U.S.government initiated a space surveillance program for detecting,tracking,and cataloging space objects.See,for example,Weeden and Cefola,“Computer Systems and Algorithms for Space Situational Awareness”;Stares,The Milita

93、rization of Space.In October 2019,thenAir Force Space Command redesignated SSA as SDA to highlight the shift of space from a benign environment to a warfighting domain due to emerging space and terrestrial threats.This shift was deemed necessary because,historically,SSA focused on elements that incl

94、ude surveillance(astrodynamics environmental data),as well as performing such activities as data collection,catalog,CA,and support for space traffic management functions,with most of these outputs captured in a publicly available catalog(Space-Track.org,homepage).SSA was redesignated SDA to reflect

95、the growing acknowledgment of space as a warfighting domain,“beyond the Space Situational Awareness mindset of a benign environment to achieve a more effective and comprehensive SDA”(Erwin,“Air Force”).Recent doctrinal guidance(Joint Publication 3-14,Space Operations)identifies SSA as a subset of SD

96、A that focuses on the orbital segment in particular to learn about and characterize space objects for space safety and sustainability.13 These objects include space debris,orbital assets in need of servicing,or assets controlled by noncooperative owner-operators(O/Os).7 are dedicated SDA sensors and

97、 nondedicated sensors that support early warning and missile detection.SDA Functions:What Does SDA Do?SSA is a subset of SDA that focuses on the orbital segment.The functions of SDA and SSA are outlined in the January 2022 space doctrine note from the USSF and the 2020 Joint Publication 3-14.14 Thes

98、e functions include detect/track/identify(D/T/ID),characterization,threat warning and assessment,and data integration and exploitation.Descriptions of these functions are provided in Table 2.1.As the SDA mission becomes more complex,these functions are becoming increasingly important and resource in

99、tensive.Exploiting data quickly and with high confidence is critical.Each of these SDA functions has key enablers that can improve the performance of this function.These enablers are noted in Table 2.1.Table 2.1.Key Enablers for Improved SDA Mission Effectiveness SDA Function Description Key Enabler

100、s D/T/ID Observation-based function that searches,discovers,and tracks spacecraft and events;categorizes RSOs into mission types More objects observed,increased ability to track,quicker ability to identify Characterization Captures the features and employment of an RSO by integrating information fro

101、m sensors to support space operations and assess threats Improved data integration and exploitation,quicker analysis Threat warning and assessment Assesses potential threats to space operations and evaluates their impact on an RSOs mission and ability to deliver effects Enhanced anomaly detection,im

102、proved analytic capabilities,improved data integration and exploitation Data integration and exploitation Exploits multiple data sources and fuses them to support decisionmaking,providing higher confidence and earlier predictions Quicker integration and exploitation,improved data quality and confide

103、nce SOURCES:Discussions with SDA stakeholders;Joint Publication 3-14,Space Operations;USSF,Operations.NOTE:RSO=resident space object.SDA Operators Two squadrons function as SDA operators.The 18th SDS is a component of the USSF Space Delta 2,located at Vandenberg Space Force Base.The 18th SDS is task

104、ed with executing the USSPACECOM SDA mission and processing SSN data to monitor activity in space and maintain custody of RSOs.The 18th SDS does so through the aggregation of observations from the SSN:calculating current and predicted orbital states for RSOs,conducting launch detection,tracking,and

105、14 Joint Publication 3-14,Space Operations;USSF,Operations.8 identifying potential collisions.A key piece of this mission is the maintenance of a comprehensive catalog of satellites.15 Activated in 2022,the 19th SDS is also a component of Space Delta 2 and has assumed the mission of CA from the 18th

106、 SDS.Further,the 19th SDS serves as a primary backup to the mission functions at the 18th SDS.16 Together with the 18th SDS,the 19th SDS maintains and populates the publicly available space catalog Space-Track.org.17 The squadrons primary mission functions are noted in the box below.Primary Mission

107、Functions for 18th and 19th SDSs USSF Space Delta 2:Prepares and presents combat-ready SDA operations to assigned and attached forces 18th SDS:Squadron within Space Delta 2 tasked with tracking and monitoring space objects,collision prediction,maneuver detection 19th SDS:Squadron within Space Delta

108、2 tasked with backup to 18th SDS functions,CA,beyond geostationary earth orbit(xGEO)SSA Combined Force Space Component Command:USSPACECOM subordinate command responsible for the management of sensors used to conduct SSA Joint Task ForceSpace Defense:Protects and defends U.S.space assets,working with

109、 the intelligence community through the National Space Defense Center SDA units SOURCES:USSF,“18th Space Defense Squadron”;Peterson Schriever Space Force Base,U.S.Space Force,“19th Space Defense Squadron”;U.S.Government Accountability Office,Space Situational Awareness.The Changing SDA Mission and I

110、nfrastructure Meeting the challenges of the SDA mission requires higher speed to respond to operational needs in a crisis or conflict,scalability as the number of space operators grows,and more-complex analyses to address increasingly maneuverable RSOs and the growing numbers of both cooperative and

111、 noncooperative space vehicles.Evolving the SDA mission introduces a variety of challenges.18 15 Wilson and McKissock,“18 SDS Small Satellite Support.”There may be a shift in the burden for CA services on USSF Space Delta 2 in the coming years,but the growing SDA mission will likely outpace the reli

112、ef on analyst demands provided by this shift.In 2018,Space Policy Directive3,“National Space Traffic Management Policy,”directed the U.S.Department of Commerce to take over SSA and space traffic management services.However,the path and timeline for doing so remain unclear.16 USSF,“18th Space Defense

113、 Squadron.”17 18th and 19th SDSs,Combined Force Space Component Command,Spaceflight Safety Handbook for Satellite Operators.18 Although the SDA functions described above are still relevant,the overall SDA mission is evolving as space becomes increasingly congested,contested,and competitive(DoD and O

114、ffice of the Director of National Intelligence,National Security Space Strategy,p.1).These changes are introducing a variety of challenges to operators responsible for providing SDAfor example,keeping up with the pace of data ingestion and processing for SDA,with the ability to provide analysis in s

115、upport of increasingly complex RSO behavior and with the ability to scale in response to increasing numbers and maneuverability of space objects.Gen Saltzman has acknowledged the importance of addressing these timing and analysis challenges:“When I hear about 9 The major limiting factor is the USSFs

116、 reliance on legacy systems to perform the SDA mission.19 One challenge is the need to augment legacy systems with modern tools,methods,and processes.These legacy systems were designed for a mission that focused on cataloging objects rather than supporting a warfighting domain.20 Another challenge i

117、s turning the SSN sensor data into actionable information,which requires computers and software to integrate,manage,and analyze these data.The SSN is mostly composed of aging sensors,which were designed during the Cold War to address threats from nuclear weapons.Finally,upgrading computing and softw

118、are must be done without disrupting the ability to continue to monitor,detect,and assess nuclear threats or the SDA mission.There are several important efforts within the SDA mission to better address these challenges.The efforts are particularly important as we examine where AI/ML tools can have im

119、pact because the limitations and challenges of the current process and architecture might not be the limitations and challenges of tomorrows process and architecture.Therefore,what this future architecture looks like may shape how AI/ML tools can affect this mission,and if we are deciding where to i

120、mplement AI/ML tools,we need to decide whether we are designing for the present or for the future.Changes in SDA Infrastructure Legacy systems must be replaced in light of growing SDA needs.The existing systemsSPADOC and Command,Analysis,Verification,and Ephemeris Network(CAVENet)have been challenge

121、d for years now with keeping up with the growing satellite catalog.21 To replace the overwhelmed SPADOC system,the USSF has funded L3Harris to develop the Advanced Tracking a breakup that occurred of a rocket body,where one rocket body became five pieces of rocket body,but it took us a couple of day

122、s to put all that information togetherokay,thats probably not the kind of timeline that would allow the Space Force to take action”(quoted in Luckenbaugh,“Space Symposium News”).19 The recognition of this need is not new,and finding a way to modernize or replace these systems has plagued DoD since t

123、he 1980s,with these attempts facing challenges to stay on schedule and budget.See,for example,Albon,“US Space Force Preparing to Decommission Legacy Command and Control System.”Attempts to modernize Space Defense Operations Center(SPADOC)began in the 1980s.Notably in 2009,the Joint Space Operations

124、Center Mission System was yet another attempt to replace SPADOC,but only one of three proposed increments was fielded,and the program was well behind schedule and over budget.Further discussions of these challenges are reported in U.S.Government Accountability Office,Space Acquisitions.20 Warfighter

125、s need to pivot from the historical SSA approach that focuses on space catalog maintenance to one that can be operationally responsive in a congested and contested environment,as well as one that can assess intent and characterize behavior.This means that SDA needs to provide decisionmaking support

126、on a relevant timeline,as well as anticipate problems and intent.21 SPADOC was brought online in 1979 in the Cheyenne Mountain Complex as a cataloging system,and its last significant upgrade was in 1989(Clark,“What About JMS?”).Gen Jay Raymond,thenhead of Air Force Space Command,said in a speech in

127、2017:“I cant wait until we can take a hammer to SPADOC and just blow it to bits.Its an old clunker and its a catalog system:its not a warfighting command and control system”(quoted in Clark,“What About JMS?”).This prompted the USSF to develop CAVENet in the 2000s to augment SPADOCs capabilities and

128、provide data analysis and sharing capabilities on a local network(Hitchens,“Key Space Monitoring Sensors Still Rely on Outdated CAVENet Computer System”;see also Weeden and Cefola,“Computer Systems and Algorithms for Space Situational Awareness,”for a more in-depth accounting of the development of c

129、omputer systems and algorithms for SSA).CAVENet is an offline system of workstations that run the Astrodynamics Support Workstation(ASW)software,which provides analytic tools for classified historical analysis,such as collision and threat assessment(Jordan Logue,discussion with representatives from

130、the 18th SDS,AprilJuly 2023;Hitchens,“Key Space Monitoring Sensors Still Rely on Outdated CAVENet Computer System”).However,as SPADOC became increasingly unable to keep up with the growing SDA mission,CAVENet has taken on operational tasks it was never designed for:CAVENet now keeps a separate satel

131、lite catalog of special perturbation(SP)vectors and covariance for analysis,and CAVENet is now responsible for developing the daily SSN tasking plan.10 and Launch Analysis System(ATLAS),a software platform for monitoring launches,satellites,and debris.ATLAS is intended to allow for the decommissioni

132、ng of SPADOC and provide SDA and command and control(C2)capability to the 18th SDS and the National Space Defense Center.22 ATLAS development is managed by the USSF Space C2 program,nicknamed Kobayashi Maru.23 Warp Core and the Unified Data Library(UDL)are two recent efforts to modernize data manage

133、ment for the USSF.Warp Core,co-managed by SSC and Kobayashi Maru,is developed by Palantir to provide a common data interface,facilitate data sharing,and support decommissioning legacy systems.24 The UDL,managed by SSCs Cross Mission Data team,is a cloud-based data repository that provides data manag

134、ement support and facilitates the integration of multiple datasets,which Warp Core can then curate and fuse.25 Significance of SDA Process Changes The implementation of AI/ML tools to enhance SDA mission effectiveness depends on whether new systems,such as ATLAS,come online to replace SPADOC,because

135、 these new systems could expand processing and analytic capabilities and provide an interface for AI/ML tools.Continued reliance on legacy data processing systems(SPADOC and CAVENet)limits the benefit of improving on-orbit sensing technologies.For example,Space Fence and the Geosynchronous Space Sit

136、uational Awareness Program(GSSAP)are two systems that can track satellites or debris.However,both systems must route data through CAVENet before entering Warp Core.26 More-direct feeds to Warp Core would eliminate the need for CAVENet and open the door,potentially,which could improve SDA mission eff

137、ectiveness through enhanced sensing technologies.An important element of these process changes will be the ability to test and validate new AI/ML tools.Software sandboxes are necessary for building and implementing new tools and capabilities,including AI/ML capabilities,for the SDA process.Integrati

138、ng AI/ML tools into systems that support SDA will first require a software platform on which to develop,test,and integrate new applications.27 ATLAS intends to provide this platform,but there are significant challenges that need to be addressed to realize this needed replacement of the aging softwar

139、e and 22 See,for example,Erwin,“Space Force Extends L3Harris Contract to Upgrade Space Tracking System.”23 Kobayashi Maru was named after a training exercise in the television show Star Trek designed to represent a no-win scenario.Only James T.Kirk was able beat the scenario by changing the code,and

140、 this is the inspiration for using this name to represent the objective of doing things differently and changing the code for the space C2 mission.See Krolikowski et al.,“Space Command and Control ProgramKobayashi Maru.”24 SSC,“Warp Core.”25 Han and Sodders,“SSC Data-Management Software Plays Critic

141、al Role in SDA,Afghanistan Airlift”;SSC,“Warp Core.”26 Space Fence links to CAVENet via the Non-Traditional Data Pre-Processer(NDPP).GSSAP links to CAVENet via Red LAN,a distribution node for SDA sensors that can deliver at multiple classification levels.Warp Core pulls the data from CAVENet and sen

142、ds them to ATLAS(Hitchens,“Key Space Monitoring Sensors Still Rely on Outdated CAVENet Computer System”).27 Another such software development program is Defense Readiness Agile Gaming Online Network(DRAGON),an Air Force Research Laboratory program with a 14-day sprint cycle for operational and devel

143、opmental test and evaluation.This program provides software development in support of Joint Task ForceSpace Defense Commercial Operations and Joint Task ForceSpace Defense operations.See Bonnette,“JCO Executes DRAGON Army Ops Days to Ensure Commercial Space Integration.”11 hardware.28 However,our in

144、terview with the Kobayashi Maru team revealed a willingness and a budding capability to test new capabilities within sandboxes.Investing in capabilities that will be most effective in the current architecture might not produce the highest impact in the future architecture.Because the SDA mission is

145、changing and growing in complexity,SDA capabilities need to be able to scale and morph to meet these changes.Areas for impact therefore need to at the very least be able to scaleand ideally would facilitate this scaling while also enabling more-complex analysis and processing.The above efforts are s

146、etting the stage for more AI/ML tool integration while allowing for the decommissioning of legacy systems.Because of these anticipated changes,the development of AI/ML tools for the SDA mission needs to be guided by the infrastructure of tomorrow.Meeting todays challenges with AI/ML tools and meetin

147、g the likely challenges of tomorrow require different technology.Without the decommissioning and replacement of legacy systems,SDA mission effectiveness would likely be limited by processing and analytic capabilities and capacity.With the replacement of SPADOC(and CAVENet)with ATLAS or a similar pro

148、gram,there may be more processing capabilities available,allowing a shift in focus to refining the types of analysis that can be performed.Infrastructure changes may enable higher impact from AI/ML tool implementation.The proposed change in infrastructure supports the validation of additional tools,

149、potentially increasing the pace at which new AI/ML tools could be implemented.Reducing the data transport time between the 18th and 19th SDSs would eliminate a bottleneck that would in turn allow for any other process speed improvements to be realized.Chapter Findings It is important to note that th

150、e SDA mission is undergoing significant changes due to the increase in launches and maneuverable RSOs.The USSF is faced with the challenge of adapting to the changing operational environment with limited resources that were not originally designed to address these threats,although modernization effo

151、rts are currently underway.In the next chapter,we provide a detailed account of how existing SDA processes are being used to address these challenges by documenting the flow of information from sensors to decisionmaking.We identify potential bottlenecks and explore the potential for AI/ML tools to i

152、mprove the efficiency and effectiveness of these processes.28 ATLAS is the logical platform to support and host AI/ML tools.However,its first priority is to come online and decommission the legacy systems it is designed to replace.Therefore,AI/ML is potentially on the ATLAS road map but not a curren

153、t priority.See Hitchens,“Key Space Monitoring Sensors Still Rely on Outdated CAVENet Computer System.”12 Chapter 3 Applying Artificial Intelligence in Space Domain Awareness AI is a broad term that lacks a complete and commonly accepted definition.We also note that AI and ML are often perceived as i

154、nterchangeable but,though related to another,are distinct.A comprehensive discussion of AI taxonomy is not within scope of this report,and,for the purposes of this report,we use the term AI/ML inclusively when referring to concepts within this taxonomy.We avoid definitional debate and consider AI/ML

155、 as a set of technologies that broadly provides six capabilities to the USSF:computer vision,language processing,reinforcement learning,prediction and classification,generative learning,and expert systems.29 These capabilities are typically used to improve four broad areas:efficiency,efficacy,resili

156、ency,and human capital metrics.30 We consider how incorporating AI/ML into SDA can yield improvements in these areas in service of the SDA mission.AI/ML applications can vary,but there is typically a complexity and automation trade-off.The most-straightforward applications of AI/ML are to automate w

157、ell-defined and less complex tasks that are tedious for a human or to provide suggestions to support decisionmaking.We turned to developing a detailed look at the space catalog and CA processes for their foundational role in other SDA processes.The SDA mission is ultimately tasked with building an u

158、nderstanding of the space environment,activities,and behavior of RSOs to support decisionmaking.This understanding comes with producing knowledge,the fundamental output of SDA,and is accomplished by characterizing what has happened as reported via sensor observations to anticipate what might happen

159、and to support better course-of-action development and timely,and wise,decisionmaking.Figure 3.1 provides a view of these relationships and highlights that these bottom tiers of this hierarchy enable higher tiers.To build this knowledge requires the ability to pull in data and then synthesize,integr

160、ate,and analyze these data to produce information.Improving SDA requires better D/T/ID and characterization;that is,to improve knowledge of what is going on in the space domain,we need to focus on creating information out of data and using information from multiple sources.31 Therefore,this ability

161、to use SSN and other data sources to characterize space 29 Different academic fields have competing definitions of AI and ML.Refer to Chapter 1,which provides AI and ML definitions from U.S.Code,Title 15,Section 9401.We focus on understanding what automation,AI,and ML capabilities can help the USSF

162、in the SDA mission.Therefore,we consider AI/ML as one broad category that provides capabilities.These capabilities follow the standard text of Russell and Norvig,Artificial Intelligence,with additional insight from Wehbe and Ramdas,“Introduction to Machine Learning.”30 Menthe et al.,Technology Innov

163、ation and the Future of Air Force Intelligence Analysis:Volume 1,Findings and Recommendations,and Menthe et al.,Technology Innovation and the Future of Air Force Intelligence Analysis:Volume 2,Technical Analysis and Supporting Material,include a series of“measures of effectiveness”to evaluate how no

164、vel technologies can improve U.S.Air Force intelligence processing,exploitation,and dissemination.31 Rowley,“The Wisdom Hierarchy.”13 activities is foundational to SDA,and we focus in particular on CA because of its embodiment of these elements.Figure 3.1.Hierarchy Illustrating Functional Relationsh

165、ips for SDA and Space C2 SOURCE:Adapted from Ackoff,“From Data to Wisdom.”NOTE:COA=course of action.SDA is a uniquely challenging and technical mission set that grows in complexity while personnel,computer,and sensor resources stay relatively static.This challenge is felt acutely in the CA process a

166、s more RSOs need to be identified,tracked,and assessed.SDA occurs in eight-hour cycles:Approximately the first 40 minutes are spent on updating the High Accuracy Catalog(HAC)(via the special perturbations model)and the rest of the time are spent on the CA process.32 Any improvement within the CA pro

167、cess frees up computational and analytical bandwidth for other SDA tasks.We assess each aspect of the CA process to identify several key areas for process impact.Proposed tools to address these needs range from tools that facilitate an increase in frequency or speed for existing processes,incorporat

168、e new processes,or replace human tasking with AI/ML tools to free up analyst time.AI/ML tools are useful as force multipliers to address resource limitations faced by operators due to aging systems,increasing mission demands(and operations tempo),and personnel shortages.These tools can help operator

169、s maintain SDA even as the mission grows in size and complexity.Existing Efforts There are significant efforts already underway to develop AI/ML tools to support the SDA mission.We highlight some below,but this is not meant as an exhaustive list:Increasing number of sensor observations:Work to suppo

170、rt increased detection and improved accuracy of sensor observations is ongoing,including efforts within the Space 32 Within 40 minutes,the HAC is updated.The next two hours are spent transferring the data updates from the 18th to the 19th SDSs.Using the HAC,Automatic Conjunction Assessment(AutoCA),a

171、nd Super Computation of Miss Between Orbits(SuperCOMBO)(described in later sections)take up all of the computational capacity on CAVENet.Jordan Logue,discussion with representatives from the 18th SDS,April 2023.14 Systems Command SDA Acquisition Delta(SSC/SZG)AI and Autonomy Sandbox colocated with t

172、he Space Delta 2 15th Space Surveillance Squadron.Ongoing work here has focused on developing computer vision algorithms to improve the detection rate and precision for satellite observations in the Machine Learning for Space Superiority(MISS)program.33 The SensorCapsule program at SSC/SZG,for examp

173、le,is enabling the increased use of commercial sensing data by helping connect nontraditional sensors to the UDL.34 Improved sensor use would also support increasing the number of observations,and the SSC/SZG mission-driven,autonomous,collaborative,heterogeneous,intelligent,network architecture(MACH

174、INA)represents ongoing work on sensor orchestration in support of sensor management optimization.35 To deconflict from this ongoing work,we did not focus on planning capabilities in our research.Ensuring that tools can be tested and validated:SSC/SZG is developing sandboxes for AI/ML tools as part o

175、f the Pivot SDA portfolio.This sandbox is configured to interface with UDL data and enables individual providers to develop analytic tools that suit their specific needs on a government-owned platform,removing barriers to third-party development of tools and allowing more-rapid deployment of models.

176、This capability was demoed with commercial sensing data in a recent Sprint Advanced Concepts Training(SACT)exercise in 2022.36 Sensor Tasking,Scheduling,and Prioritization The 18th SDS coordinates the observation activities of the different sensor sites composing the SSN.There more RSOs spread acros

177、s too much sky than the 18th SDS can observe every day,so the 18th SDS must stagger observations to capture every object with sufficient frequency to correlate with eachs track history.Additionally,different sensor sites feature different technological configurations,reliabilities,percentage downtim

178、es for maintenance,and precisions;these factors all must be accounted for as part of the sensor tasking process.Sensor allocation refers to the deliberate distribution of taskings to individual sensor sites to achieve the twin goals of catalog maintenance and event capture.To maximize the number of

179、individual observations and minimize uncertainty in orbit determination(OD),it is necessary to optimize the allocation of these limited sensor resources.The 18th SDS orbital analysts are responsible for building the daily composite(COM)file,which is a composite of tasking orders for all sensor sites

180、,also known as the consolidated tasking list.Issues The 18th SDS has a Sensor Optimization Cell that monitors SSN sensor site performance.Performance metrics,which are based on laser calibrations,are constantly produced at each site.33 Fletcher,“AI and Autonomy for SDA.”34 Fletcher,“AI and Autonomy

181、for SDA.”35 Fletcher,“AI and Autonomy for SDA.”36 Fletcher,“AI and Autonomy for SDA.”15 However,the 18th SDS is resourced to check these metrics only periodically.Sensors that are out of calibration can produce erroneous data about a particular target,limiting the squadrons ability to characterize h

182、igh-interest behavior and polluting the space catalog.Tracking RSOs is ultimately an uncertainty(referred to as covariance)reduction problem:Covariance grows over time(from maneuvers or changes to atmospheric density solar radiation)and can be reduced only by sufficiently accurate sensor observation

183、s.Sensor allocation is a balance between uncertainty reduction and revisit rate using limited sensor resources.Unfortunately,the current paradigm emphasizes revisit rates.The daily taskings are often overridden in favor of higher-level guidance to track priority RSOs.As a notional example,guidance m

184、ay require that a specific RSO needs to be observed X times within a day.To meet this requirement,the 18th SDS may task a sensor that is unable to get the needed collection(e.g.,an optical telescope currently facing the sun or a radar with higher covariance than the current objects covariance).The o

185、pportunity cost of a tasking override is missed detections,which may lead to uncorrelated RSOs(uncorrelated tracks will be discussed in the“Catalog Maintenance and Uncorrelated Tracks”section later in this chapter).SP Tasker was not designed to handle the large quantity of objects currently in orbit

186、 and is at the limits of its abilities.Nor was it ever designed to support agile responses to changing real-time observation needs.The increases in the population of RSOs and frequency of high-interest events pose a challenge to SP Taskers processing capabilities and operations cadence.The upcoming

187、ATLAS program,as described in Chapter 2,is intended to replace SPADOC.In addition to having more processing capacity,ATLAS could potentially augment the SSN using information from non-SSN U.S.government sensors;it remains an open question as to how to optimize load balancing across these capabilitie

188、s and how to weigh and trust information from non-SSN sensors.Potential AI/ML Application Sensor calibration is a major challenge for the 18th SDS.Because of our stakeholder interviews,we believe that the 18th SDS wields the responsibility for tasking sensor sites to perform laser calibrations,proce

189、ssing the resulting calibration data,and directing the adjustments for improving subsequent observations from that site.Given the acute manpower crunch at the 18th SDS,solutions include either(1)empower sensor sites to perform their own calibrations as needed or(2)automate calibration tasking.The fi

190、rst option requires additional training and systems for sensor site staff but reduces the workload of the 18th SDS.The second option requires the development of automated calibration tools(e.g.,calibrate at regular intervals or set thresholds to identify anomalous observations).37 Either option will

191、 provide the 18th SDS with an improved ability to track sensor covariance over time,which can lead to improved sensor use.A separate tool can track these attributes for each sensor site and determine feasibility of potential taskings and optimize sensor allocations based on timeline requirements(to

192、capture a potential high-interest event)or prevent an RSO from becoming lost 37 To the best of our knowledge,calibrating observations is a manual process today.16(thereby creating an uncorrelated track and analyst object).These features are enabling steps toward dynamic sensor retasking.Automating s

193、ensor calibration is a first step toward creating a sensor validation process.As the USSF considers leveraging non-SSN capabilities(U.S.government or commercial)to augment the aging SSN sensors,creating a framework or support to validate sensors smooths the transition toward ingesting new sensor dat

194、a and new data formats.Data Collection and Integration Once SSN sensor observations are taken and transmitted to the 18th and 19th SDSs,the second stage of data integration begins.Sensors observe range,range rate,azimuth angle,and elevation angle and send the data to the SPADOC Space Data Server hou

195、sed at Cheyenne Mountain via data cables,where they are translated into a format that is ingested by CAVENet.Observation data across each sensor site must be fused because sites are not identical.The current fusion procedures are complex and put CAVENet at its maximum processing capacity.Note that t

196、he 18th SDS must access these data through CAVENet,but the 19th SDS has direct,albeit slow,data lines to the SSN because of the legacy Navy Space Surveillance System.Issues The number of RSOs is steadily growing,but processing capabilities and the number of SSN sensors are relatively static.To meet

197、this expanding need,it is vital to leverage additional sensing resources,including commercial sensors.Success will entail effectively ingesting,fusing,and validating data sourced from disparate resources.Processing power and data communication are the two primary challenges in this stage,and the new

198、 ATLAS system is intended to address these bottlenecks with both hardware and software upgrades.38 The Innovation Cell within the 19th SDS is also actively working on these issues.39 The current SSNs capabilities do not match the new reality of near-daily launches and mega-constellations of hundreds

199、 of satellites.This problem is exacerbated with the increase of noncooperative space entities,those domestic and international O/Os who do not voluntarily provide the 18th SDS with screening ephemerides.Mega-constellations,in particular,have caused a drastic shift in operational tempo.40 38 The SSN

200、is old,and the centralized system,SPADOC,has coordinated the SSN since the 1970s and relies on hard data lines for all communications with the sensor sites.Its replacement,ATLAS,is under development,but its development and uptake have been negatively affected by the failure of the last effort to rep

201、lace SPADOC:Joint Space Operations Center Mission System.39 Li Ang Zhang,George Nacouzi,and Melissa Heron,in-person visit to the 19th SDS,December 2022.40 In response to mega-constellations,the 18th SDS conducts these activities:standing up the new Mega-Constellation Analysis Cell dedicated to catal

202、oging constellations with more than 100 satellites disambiguating objects during large-scale deployments(this activity is difficult because of proximity and is possible only with O/O-provided or commercially procured ephemerides)17 Potential AI/ML Applications New sensor modalitiessuch as on-orbit o

203、bservations,passive radio frequency,or commercial sensorscan enhance SSN observational data and help address accelerating space activity.The 18th SDS needs support for ingesting new sources of sensor data,in addition to automating sensor validation.Although computational limitations may limit this c

204、apability today as a result of the reliance on legacy infrastructure for data transmission and analysis,the inclusion of data from additional sensors that are not part of the SSN is something that DoD is actively pursuingfor example,through the Joint Task ForceSpace Defense Commercial Operations cel

205、l and could be supported through the use of AI/ML tools.41 Academic research indicates that computer vision may help disambiguate individual RSOs within a mega-constellation and mimic C2 behavior based on existing constellation management.42 Data fusion is an evolving practice,and AI/ML can enable n

206、ew methods of observation and potentially even new prediction models.OD Following data collection and integration,the 18th SDS uses OD to determine the most likely state of a given RSO from sensor observations.Each sensor site takes multiple observations of a target RSO,and there are multiple source

207、s of uncertainty in these observations,including sensor noise and bias and RSO motion uncertainty.To mitigate the impact of uncertainty,the 18th SDS uses the weighted least squares method to determine the most likely state for the RSO.43 Each observation,weighted by its confidence(inverse of sensor

208、covariance),44 influences the most likely RSO state.These observations,and the associated uncertainties,feed a process known as differential corrections that estimates the most likely averaged state of the object at a given time:a so-called ephemeris.Time series of ephemerides are recorded in the HA

209、C maintained by the 18th SDS as one of the squadrons core missions.The HAC serves as the starting point for forecasting future RSO trajectories and performing CA.Improvements in OD,such as faster processing,more-frequent observations or more-accurate observations,will all improve downstream processe

210、s.employing ion thrusters in mega-constellations to maintain their orbits;the frequent station-keeping burns result in non-Keplerian orbits that violate the elliptical orbit assumptions used for OD calculations tracking the growing orbital population with the SSN,which was designed for ballistic mis

211、sile warning;senor sites are not necessarily located facing modern satellite population centers.41 The Joint Task ForceSpace Defense Commercial Operations allows for the use of commercial SSA data to support DoD.See,for example,U.S.Government Accountability Office,Space Situational Awareness.42 Mass

212、imi,Ferrara,and Benedetto,“Deep Learning Methods for Space Situational Awareness in Mega-Constellations Satellite-Based Internet of Things Networks”;Kopacz,Roney,and Herschitz,“Deep Replacement.”43 Because of computational constraints and historical reasons,CAVENet uses weighted least squares estima

213、tion to update the catalog with new observations.44 The uncertainty associated with a sensor observation is quantified as covariance,the matrix of values describing the spread of each state variable(e.g.,x,y,and z directions in a Cartesian system)with respect to itself and every other state variable

214、.The spread of values for a given observation depends on such sources of uncertainty as sensor noise and bias.18 CAVENet ASWs Automatic Special Perturbations Ephemeris Generation(ASP)script is run regularly to produce SP propagations for the updated ephemerides,45 after which the determined trajecto

215、ries are transferred over to the 19th SDS via dedicated data lines through Cheyenne Mountain.Issues ODs later stages involve propagating an objects most recent ephemeris and covariance into a different time(either into the future for CA or into the past for object association).This step is resource

216、intensive and is currently performed only by the 19th SDS on a subset of RSOs because of practical processing limitations.46 The RSO covariance propagation has the potential to provide two benefits.First,propagating an RSOs future covariance would enable the 18th SDS to proactively identify objects

217、that are about to be lost during the Automatic Differential Correction(ADC)process.47 Currently,this is a reactive process in which a percentage of daily observations fail ADC and must be recorrelated manually.These objects can be very difficult to recorrelate and may take anywhere from minutes to d

218、ays,if at all.48 Second,RSO covariance is a critical aspect of improving sensor tasking.When tracking an RSO,there are two sources of covariance:the object and the sensor.To make a useful observation,the selected sensor must have lower covariance than the object;otherwise,the observation may actuall

219、y contribute to object covariance.Tracking both RSO and sensor covariance can improve the sensor allocation process and enable a feasibility calculation(what can we improve with detections today?).A feasibility assessment might also mitigate conflicts between SP Tasker and revisit rate mandates for

220、high-interest RSOs.49 Potential AI/ML Application Most downstream components of SDA depend on accurate OD.This is a step that can be appropriate for an AI/ML application.Neural networks are universal function approximators,and academia is demonstrating that they can feasibly approximate complex phys

221、ics models.50 Although neural networks will typically yield less accurate solutions(relative to a known physics model),they 45 ASP currently takes about 40 minutes and runs three times per day at the beginning of each eight-hour cycle.46 The 19th SDS propagates refined RSO ephemerides and covariance

222、s for objects that were flagged for risk of collision(as calculated by a CA prescreening algorithm).These covariances are used to calculate a more authoritative collision probability.47 Li Ang,Krista Langeland,George Nacouzi,and Jon Tran,discussion with various analysts at the 18th SDS,April 2023;Li

223、 Ang Zhang,George Nacouzi,and Melissa Heron,discussion with various analysts at the 19th SDS,December 2022.48 Li Ang,Krista Langeland,George Nacouzi,and Jon Tran,discussion with orbital analysts at the 18th SDS,April 2023.49 As discussed earlier,there revisit-rate mandates often override SP Tasker.T

224、he result may be improper sensor tasking for the sake of meeting the mandate.Tracking both RSO and sensor covariance enables a feasibility calculation:Although we can achieve a revisit rate,will it improve our situational awareness(in terms of decreasing uncertainty on an RSO)?50 Hornik,Stinchcombe,

225、and White,“Multilayer Feedforward Networks Are Universal Approximators.”19 can achieve high-accuracy estimates at a fraction of the time.This speedup can be a highly desirable aspect for SDA.We explored whether any element of OD can be approximated by a neural network and whether a neural network of

226、fers any benefits compared with status quo processes.To accomplish this,we developed in-house models to rapidly prototype AI experiments and determine application viability.The primary AI capability we leveraged was prediction(regression)to train a neural network to closely align inputs and outputs

227、with a physics-based numerical orbit model.We developed two neural network models:one to propagate an RSO state into the future and one to propagate time-varying covariance into the future.Using standard orbit propagation parameters and an RSOs initial state,we leveraged Poliastro,51 an astrodynamic

228、s Python package,to generate ground truth trajectories and a standard architecture neural network to approximate inputs and outputs.On the basis of our testing,we determined that state-only propagation with AI/ML is feasible.Within 90 milliseconds,the neural network is able to project an RSO state t

229、en days out with a root mean squared error(RMSE)of 1 km.52 For context,the neural network is able to propagate a 40,000-sized space catalog in one hour.By our best estimates,OD takes several hours for a similar sized-catalog(due to processing constraints at the 18th SDS)and the general perturbation

230、catalog is accurate to several meters(1 km).The accuracy of the neural network can likely improve,but this invokes a speed-versus-accuracy trade-off.Additional details and discussions are provided in the companion report,which presents the results from two AI/ML case studies.53 State-only OD offers

231、several advantages for SDA:Analysis is faster,and orbital analysts within the 18th SDS spend significant time associating uncorrelated objects to known objects(e.g.,breakup events or lost objects).Fast,near-instantaneous ten-day state propagation can help analysts quickly assess potential matches or

232、 filter out unlikely candidates.Quicker and more-frequent calculations could in turn promote a more responsive(i.e.,more-frequent)OD process,which would support the maintenance of a more current HAC for screening efforts at the 19th SDS.Conjunction prescreening can be sped up,minimizing the number o

233、f false-positive RSO pair candidates that require high-fidelity calculations.State-only propagation may speed up elliptical volume screening(see the“CA Calculation”section later in this chapter for more information).We trained a second neural network to propagate an RSOs covariance.For this experime

234、nt,we used Bayesian neural networks(BNNs),an architecture designed to replicate probability distributions,to propagate an RSOs covariance into the future.However,our results showed that this approach is significantly more difficult than the state-only propagation network.It is difficult to contextua

235、lize the accuracy of the BNN because we lack an understanding of the operational requirements of current covariance propagation models,but the BNN did not fully replicate time-51 PoliastroAstrodynamics in Python,homepage.52 The 90-millisecond figure is based on our Nvidia Tesla P6 graphics processin

236、g unit(GPU)workstation.This likely can be accelerated even further with a high-performance computer or with more-modern GPUs.53 Tran et al.,Artificial Intelligence and Machine Learning for Space Domain Awareness.20 varying covariance in our test data.54 Although we could not demonstrate BNN feasibil

237、ity to propagate covariance,55 with operational requirements and sufficient training data,such a tool may one day offer a faster alternative to high-fidelity CA screens.Catalog Maintenance and Uncorrelated Tracks Catalog maintenance involves updating the HAC via OD and the association of uncorrelate

238、d tracks(UCTs).56 UCTs are those strings of observations that are unable to be automatically associated with an object in the catalog.This may be because the observed object has been lost and has been assigned to an analyst object with no track history,or it could be because the objects orbit has ch

239、anged so significantly that the system is unable to recognize its relationship with a past track.There are numerous potential causes,including maneuvering or fragmentation events.During OD,the ADC process successfully associates the majority of observations to a known object,but failure can occur wh

240、en new observations cannot be associated with an existing track within set tolerance thresholds.57 These failures trigger a manual review for analysts within the 18th SDS.Issues The orbital environment is rapidly evolving,with increased launch frequency,increased RSO maneuvering,and the proliferatio

241、n of noncooperatives throughout the orbital environment.These events lead to ADC failures at a pace that is outstripping the 18th SDSs current analyst ability.UCTs are assigned a SPADOC code for future bookkeeping.These so-called analyst objects currently require manual intervention by 18th SDS anal

242、ysts for reassociation with their past known custody.The longer analyst objects go without reassociation,the more time this process takes and the less likely it is to be successful.A UCT may become lost and reappear as a separate UCT;a competitor may be able to take advantage of this limitation to o

243、bscure deployment of spacecraft.54 Accuracy is not precisely used here.To evaluate the outputs of the BNN,we used different evaluation metrics,such as an outputs mean average percentage error(MAPE)from the mean of the ground truth covariance and Kullback-Leibler divergence of the outputs distributio

244、n relative to ground truth distribution.However,this is difficult to contextualize because we were unable to assess the operational performance of ASP during this study and therefore could not compare these metrics against operational reality.55 We attempted to troubleshoot the issue by adjusting va

245、rious parameters of the BNN,but none of these adjustments resulted in improved performance.We also considered the possibility that our dataset was insufficient or that there were issues with the quality of the data.56 We also refer to these objects as lost objects.57 A set of observations is assigne

246、d to the closest existing track.ASW verifies this assignment for a solution precision threshold,quantified by weighted root mean squared(WRMS)error.If the WRMS error of a Gaussian batched-weighted least squares solution is too high,the ADC system will reject the solution and flag the observations fo

247、r manual inspection.21 Potential for AI/ML Tools to Support Catalog Maintenance Object correlation appears to be a key area for AI/ML impact.Although the vast majority of observations are automatically assigned,the increasing numbers of noncooperative space entities and total RSOs increase the amoun

248、t of UCTs.Some aspects of correlating UCTs can be automated,enabling more-frequent object association.Orbital analysts within the 18th SDS follow a highly complex step-by-step task list to reassociate UCTs.List generation processes are typically undocumented and depend on the experience of the indiv

249、idual being tasked.We recommend codifying these tasks to preserve the knowledge and potentially automating these rules into an expert system of if-then-else rules.This tool would be able to handle the most-common scenarios for ADC failure without requiring human intervention.Not every UCT can be cor

250、related this way,but such a system can help shrink the number of UCTs and free up analyst bandwidth for more-difficult scenarios.This system can be augmented by compiling a database of previously unknown analyst objects and their eventual reassociation to known tracks.Using the expert system as a st

251、arting point,an advanced AI/ML application can be trained to recognize the relationships in UCTs and generate automatic track association recommendations.Splitting the development of this AI/ML application into layers of increasing complexity can mitigate development risks.CA Calculation CA is the p

252、robabilistic process by which the 18th and 19th SDSs quantify risk of collision in space.The process is probabilistic rather than deterministic because of the numerous and chaotic orbital forces that progressively perturb an objects trajectory away from its theoretical path over time.Because the clo

253、sing velocities involved in space collisions necessitate avoidance maneuvering far in advance of the predicted time of closest approach(TCA),it is(currently)impossible to know exactly where the two colliding objects will be at the TCA.CA thus requires a dynamic assessment of risk that supports conti

254、nuous ingestion of new information to improve confidence in the prediction and avoid unnecessary maneuvering.Probability of collision(Pc)is a measure of confidence of a predicted collision occurring.Calculations of Pc seek to minimize the number of false positives by accounting for the dynamic uncer

255、tainties built into the RSOs respective state space variables.58 Increasing Pc accuracy is thus a key objective for CA;to achieve it,the 18th and 19th SDSs leverage their OD and sensor tasking processes to aggregate raw sensor observations for bookkeeping in the space catalog and subsequent calculat

256、ions of Pc.To characterize these processes,we consulted with analysts to identify the decision elements embedded within that support the timely and accurate determination of Pc.The CA process flags pairs of RSOs that are at high risk of collision.In the early days of space operations,operators held

257、that the space domain was so voluminous that the odds of collision were remotethe Big Sky assumption.Subsequent collisions demonstrated the invalidity of this 58 Note that the need for uncertainty quantification in the form of a covariance matrix,propagated along with the state vector itself,necessi

258、tated the development of SP methods(whose numerical nature would allow incremental bookkeeping of uncertainty in a way that general perturbation could not.The current HAC of RSOs has been maintained according to SP ever since.22 assumption,as well as the outsized impact of even a single collision on

259、 the celestial environment,leading to a recognition of the need to assess the risk of collision in space.The 19th SDS is primarily responsible for CA.Using the latest RSO states from the HAC,the 19th SDS calculates CA using two ASW tools:AutoCA and SuperCOMBO.ASWs ASP uses the SP model for OD and pr

260、opagates all HAC object states into the future.For deep space objects,the 19th SDS uses an ellipsoid screening method in which covariance is estimated as a fixed-size ellipsoid.59 Potential collisions are flagged if any RSO enters the volume of this ellipsoid.For near-earth objects,the propagated co

261、variances of RSOs are used to calculate a probability of collision at the TCA.60 The probability of collision for RSO pairs is calculated,and probabilities that exceed a certain threshold are added to the attention list for closer observation by the SSN,or a conjunction message is sent to the O/O of

262、 the respective RSOs.Issues OD and uncertainty quantification are the key elements for computing a reliable Pc.The 19th SDS relies on the HAC to screen the orbital trajectories of all RSOs against all other RSOs.61 Timely and accurate updating of this catalog relies on frequent performance of the OD

263、 process and timely identification of RSOs,and there is a need here for a more responsive OD process to support the maintenance of a more current HAC for screening efforts at the 19th SDS.The current data lag between the 18th and 19th SDSs and the current eight-hour HAC update cycle can result in th

264、e 19th SDS using hours-old RSO states when starting CA computations.Furthermore,the HAC all-versus-all conjunction screen is a large and high-fidelity computational task that exponentially grows with the number of RSOs.An accurate prescreen that can minimize the number of candidates for subsequent h

265、igh-fidelity conjunction calculations can help mitigate the exponential complexity of this task.Potential AI/ML Application An AI/ML-based conjunction screening application would build on previous work by using OD and neural networks to predict distance and time of closest approach.Given state vecto

266、rs of two initial RSOs,a neural network can be trained to predict the distance of closest approach.Although a neural network estimate will never replace the accuracy of the CA process,it can serve as a rapid prescreening step to minimize the number of false-positive RSO pairs that make it to the hig

267、h-fidelity CA step.Chapter Findings The SDA mission is becoming more complex,and tools that support this mission can act as force multipliers to help the 18th and 19th SDSs meet increasing demands.Processes such as checking and 59 18th and 19th SDSs,Combined Force Space Component Command,Spaceflight

268、 Safety Handbook for Satellite Operators.60 18th and 19th SDSs,Combined Force Space Component Command,Spaceflight Safety Handbook for Satellite Operators.61 18th and 19th SDSs,Combined Force Space Component Command,Spaceflight Safety Handbook for Satellite Operators.23 verifying sensor performance,h

269、andling UCTs,and resolving broken differential corrections are currently performed manually and require significant time and effort.The growing number of mega-constellation deployments,uncorrelated tracks,lost objects,and manual differential corrections are threatening to outpace the ability of the

270、18th and 19th SDSs to address them.Resourcing the development of tools that run constantly in the background would be helpful in mitigating the untenable growth in object correlation cases that require manual intervention as the celestial neighborhood grows.Noncooperative mega-constellations present

271、 unique challenges to USSPACECOM,including the deployment of secret payloads and non-Keplerian orbits that require more-frequent observations to avoid breaking ADCs thresholds.The addition of the characterization element to SDAs mission demands requires increasing person-hours to address,whereas enl

272、isted technical expertise is constantly being lost because of tour-of-duty limitations.Launch CA processes are struggling because of the exploding tempo of launches,both cooperative and noncooperative.The changing demands in the space domain require SDA to process and respond at an operational times

273、cale,which currently presents barriers to quicker SDA responsiveness.To address these challenges,AI/ML tools should help shift human manpower to tasks that do not lend themselves to checklists and algorithms.Increasing sensor observation capacity could promote the ability to detect and respond to no

274、nnominal and noncooperative behavior more rapidly.Building in slack to sensor observation capacity would help increase responsiveness to nonnominal behavior,such as maneuvering,fragmentation events,and non-Keplerian orbits.An overview of these tools is provided in the box below.The companion report

275、offers a detailed description of the proof of concepts we developed in many of these areas,62 and Appendix B in this report provides a brief summary of these tools for the readers convenience.Summary of Potential AI/ML Tools Sensor tasking,scheduling,and prioritization These tools automate calibrati

276、on tasking.These tools track sensor attributes and determine tasking feasibility.Data collection and integration Computer vision may help disambiguate RSOs.These tools enable new methods of observation.OD Neural networks can approximate complex orbit models in a fraction of the time.Catalog maintena

277、nce Some aspects of correlating UCTs can be automated.Generate automatic track association recommendations.CA Neural networks can be a rapid prescreener for CA.AI/ML tools have several potential impacts.These tools can improve resource allocation,optimize processes,perform tasks more quickly and fre

278、quently,and reduce the need for human analysis.These 62 Tran et al.,Artificial Intelligence and Machine Learning for Space Domain Awareness.24 solutions act as force multipliers.According to our CA process analysis,the following impact characterizations can guide the development of AI/ML tools to me

279、et SDA mission needs in the following areas:optimization of sensor tasking and prioritization to increase observation capacity and support rapid and accurate OD improved detection of nonnominal behavior in sensors to more rapidly identify erroneous sensor behavior that would otherwise pollute the tr

280、ack history of any observed satellites,thereby supporting HAC health increased number and frequency of sensor observations to improve the accuracy of OD,which supports arriving at more-accurate Pc predictions quickly;this can be achieved by optimizing existing sensor resources or ingesting additiona

281、l data from commercial sensors optimization of sensor allocation,including data fusion,to take advantage of the growth in commercial sensor networks while mitigating declining sensor performance with age and the strengths and weaknesses of different sensors for different observation use cases.By lev

282、eraging the capabilities of AI/ML,higher sensor revisit rates and faster OD can be achieved,thus acting as a force multiplier on static sensor and computational resources.In the next chapter,we will assess the quantitative impact of improved sensor allocation OD speed on the SDA mission set.25 Chapt

283、er 4 Improvement Models for Characterizing Impact Through consultation with stakeholders,we identified key areas of potential impact based on the most-acute needs for meeting the growing challenges of the SDA mission and the CA process(Chapter 2).We then assessed the role that AI/ML could play in th

284、ose key areas of impact(Chapter 3).This chapter examines the nature of that impactthat is,how could improvement in one process element affect overall mission success?Characterizing the Value of AI/ML for CA A key area for potential improvement for AI/ML tools is in improved sensor optimization.Senso

285、r resources are finite,and using AI/ML tools to better leverage existing resources could facilitate higher numbers of useful sensor observations and higher revisit rates.More-effective prioritization of sensor observationsi.e.,to better track a high-interest objectwill better direct existing sensor

286、resources to high-priority intelligence needs.We examine the potential cascading impacts that AI/ML tools can bring to the SDA mission by examining two key performance parameters that represent broader CA process objectives:Improve the accuracy of the calculated Pc,which avoids collisions through ti

287、mely warnings.Decrease probability of false alarms,Pfa,minimizing the false prediction that a conjunction would occur,which can cause unnecessary satellite maneuvers that degrade lifespan and mission performance.The first improvement described below is about the impact of AI/ML tools if they can be

288、applied to optimize sensor tasking to increase revisit rates,and we examine the impact of revisit time on downstream process elements.The second improvement described below is about the impact of more-frequent sensor updates on the OD process,which in turn has downstream effects on the uncertainty a

289、ssociated with the Pc calculation.63 63 These two improvements were chosen based on current obstacles as identified by stakeholders within Space Delta 2that is,maintaining awareness as more mega-constellations are launched,maintaining tracks when more RSOs are maneuverable,and maintaining responsive

290、ness and accuracy in Pc calculations.26 Improvement 1:AI/ML Supports More-Efficient Use of SDA Sensors,Resulting in Higher Revisit Rates The evolving space environment is seeing increasing activity from U.S.strategic competitors,including recent behavior exhibited by Chinas Shijian-21 satellite.In D

291、ecember 2021,Shijian-21 approached a defunct BeiDou navigation satellite to conduct a sophisticated rendezvous and proximity operation(RPO),docking with the satellite,then conducting an engine burn that culminated with the Shijian-21 towing the defunct satellite to a graveyard orbit above geostation

292、ary earth orbit(GEO).64 Although this capability has the potential to be used for responsible debris mitigation in space,there are also growing concerns about such sophisticated RPO activities enabling offensive counterspace capabilities.For this reason,the United States places a high priority on tr

293、acking and characterizing these events,and the SDA mission will be increasingly focused on these events as competition in space continues and grows.Another example of orbital maneuvering is from analysis conducted by COMSPOC of Chinas Shiyan-12 satellite executing complex maneuvers to avoid being ob

294、served by a U.S.satellite,the USA 270.65 Maintaining awareness of activities,such as those conducted by Shijian-21,requires a high revisit rate from observing sensors and full orbit coverage capability to support U.S.space operations that require SDA data in an operationally relevant time frame.Effo

295、rts to increase sensor revisit rates can make SDA operations more responsive to maneuvering satellites,and the below proposes a model for illustrating the potential impacts of increasing revisit rates when tracking maneuverable RSOs.We present a simple analysis illustrating the potential impact of a

296、 delta-v(velocity)maneuver on the satellite position.The intent is to show the importance of the revisit rate to help maintain track on a maneuvering satellite.The example analysis considers a space object orbiting in GEO and estimates its position over a 24-hour period with and without a maneuver;t

297、he difference between the two cases illustrates the difference in the satellite position over the 24-hour period.The four panels in Figure 4.1 illustrate a nominal geostationary orbit.66 Panel(a)illustrates the trace of the circular orbit of the satellite,and panel(b)provides the radius as a functio

298、n of true anomaly(or angular position).Panel(c)shows the time history of the orbit radius,which is constant,given the analysis approach,and panel(d)provides the time history of the true anomaly of the object,which has a constant slope representing a nonaccelerating body.64 Jones,“Chinas Shijian-21 T

299、owed Dead Satellite to a High Graveyard Orbit.”65 Werner,“An In-Orbit Game of Cat and Mouse.”66 The analysis uses Keplers orbit equation and does not account for any perturbations,but these approximations are sufficient for the purpose of this discussion.27 Figure 4.1.The Orbit of a GEO Satellite wi

300、th No Maneuvers NOTE:The panels depict a satellite in geostationary orbit.Panel(a)is trace of the circular orbit of a satellite.Panel(b)is the radius as a function of true anomaly.Panel(c)is the time history of the orbit radius.Panel(d)is the time history of true anomaly.The second set of panels in

301、Figure 4.2 illustrates the impact of a 50 m/s delta-v,which increases the eccentricity of the orbit to be slightly elliptical.The instantaneous thrust results in an angular drift rate of approximately 0.74 degrees/hour,or 544 km/hour in GEOi.e.,the satellite is 544 km away from its expected position

302、 after one hour.Consequently,a ground-based sensor with a 1-degree field of viewor 650 km field of view at GEOwould likely fail to acquire the satellite if it was unaware of the maneuver,since the sensor would likely attempt to place the satellite at the center of its field of view.(b)(a)(d)(c)28 Fi

303、gure 4.2.The Orbit of a GEO Satellite with 50 m/s Maneuver NOTE:The panels depict a satellite following a 50 m/s delta-v maneuver.Panel(a)is trace of the elliptical orbit of a satellite.Panel(b)is the radius as a function of true anomaly.Panel(c)is the time history of the orbit radius.Panel(d)is the

304、 time history of true anomaly.A search pattern would likely enable a sensor to acquire the object relatively quickly,since it is within half the sensor field of view.However,recall that sensor tasking currently occurs every 24 hours and that orbit propagation occurs once every eight hours.Supposing

305、that we gain situational awareness of the maneuver immediately,there are no processes to dynamically retask a sensor to revisit this object.Assuming that the daily sensor tasking had an a priori revisit rate set at eight hours,the object would be about 4,000 km away from its expected position.67 Thi

306、s difference of expected versus actual satellite position of several thousand kilometers will result in a much longer search time and possibly an uncorrelated track that requires additional sensor and operator resources to resolve.68 The probability of the timely acquisition of a maneuvering satelli

307、te decreases as the revisit period and magnitude of the maneuver increases.Improvement 1 illustrates that improving sensor utilization has cascading effects,such as higher revisit rate,fewer lost objects(uncorrelated tracks),and faster sensor acquisition times.All these 67 The calculated value is th

308、e position error in three dimensions;however,the value would be less when projected on a two-dimensional sensor detector because of the relative angle between the trajectory trace and the sensor detector.68 The extent of the additional required resources to find the satellite after its maneuver depe

309、nds on several variables,including the characterization level of the object,the actual maneuver performed,the number and type of sensors used to perform the search,and the illumination conditions,as well as the algorithms and analytics used to support the search,and cannot be easily approximated.(b)

310、(a)(d)(c)29 effects reduce analyst bandwidth at the 18th and 19th SDSs and support an increase in Pc accuracy.There are likely numerous ways to improve sensor revisit rates,and AI/ML tools may be one such approach.Improvement 2:AI/ML for Improving Covariance Calculations in the OD Process The OD pro

311、cess is foundational for maintaining HAC accuracy and identifying collision risk.In addition,OD helps identify where sensors need to take another look;in this way,improving the OD process can help prioritize which and how different SSN sensors are used.We propose a CA process that leverages an AI/ML

312、 tool to accelerate the orbit projection estimation and prioritize the sensor scheduling based on Pc.Once the initial Pc calculation is completed,it is currently used to inform the prioritization of sensor observations.The change that we are proposing here is the use of an AI/ML tool to accelerate t

313、he subsequent OD from these sensor observations so that false alarms are more efficiently removed from the concern list.The process steps are described below and illustrated in Figure 4.3:1.The first step(labeled 1)provides the reference orbit for all the objects being tracked.2.Using the informatio

314、n from step 1,the Pc between any two objects is calculated and sorted.69 3.Sensor observations are prioritized based on the N-highest Pc and add the high-interest objects regardless of their associated Pc.70 4.Sensor observations are scheduled according to prioritized list.71 5.Sensor observations a

315、re executed.6.The OD of all observed objects is updated.This is where the AI/ML tool is introduced to accelerate this step.We propose that an AI tool would enable this step to be performed in less than one hour to support an hourly sensor scheduling tempo.72 a.Step 6a involves updating the covarianc

316、e for all the observed objects.Our assumption is that information derived from the latest observations of objects cannot be used to refine the covariance of objects that were not observedi.e.,differential corrections apply only to the observed objects and cannot be leveraged for other nearby objects

317、.7.Position error projection is reduced for objects that have been observed.8.Pc is updated using the latest data and calculations.69 For detail on the process of this calculation,see the companion report,Tran et al.,Artificial Intelligence and Machine Learning for Space Domain Awareness.70 N needs

318、to be determined separately based on sensor capacity and other relevant criteria.71 Sensor scheduling could also benefit from AI/ML and,if appropriately implemented,could improve the efficiency with which the existing sensor network is being usedthat is,a possible objective function is to minimize t

319、he covariance across the cataloged RSOs.See Stottler,“Automatic,Intelligent Commercial SSA Sensor Scheduling.”72 See Chapter 5 for a sample AI/ML-based tool to demonstrate this capability.30 Figure 4.3.Overview of Proposed Modified CA Process with Dynamic Sensor Tasking NOTE:HIO=high-interest object

320、;MOP=measure of performance;NRT=near real time;Pfa=probability of false alarms;SHIO=super high-interest object.Increasing the frequency of the sensor tasking cyclee.g.,from eight hours to one hourwould on its own improve the overall accuracy of Pc,assuming that the processing can also be executed wi

321、thin the shorter available time.However,other points in the process could be modified in parallel to keep up with this increased frequency.For example,as the tasking frequency increases,the number of objects that can be accessed and observed decreases because less time is available for the taski.e.,

322、the sensor network is provided a shorter period to execute the observation task(SSN and possibly other supporting assets,such as commercial and allied).For this reason,an update to the sensor scheduling strategy may also be needed to support the higher tempo of sensor tasking orders.Implementation o

323、f the AI/ML tool described here also requires Pc calculations to keep up with the pace of the new sensor measurements.Impact Characterization We proposed a CA process in which sensor tasking is informed by both Pc and revisit rates.We use a low-fidelity model to illustratively quantify the impact of

324、 this suggestion.For this purpose,we developed a one-dimensional model consisting of 100 randomly spaced moving objects.State(speed 8.Improve Projected Pc i.e.,Reduce Pfa,Increase True Positives6a.Reduce Covariance7.Reduce Error Projection over Time5.Sensor Measurement Update:Priority Objects3.ID/Pr

325、ioritize Object Revisit Based on Pc(along with HIO/SHIO tasking)1.Initial Orbit Determination,All Objects2.Update Pc6.Update Orbit Determination 4.Schedule Observations with Prioritized Objects(as highest priority)Additional Objective:Minimize Unnecessary Mission ImpactAI/ML reduces processing time,

326、allows running updates in NRT for reduced set of objectsAI/ML OD ToolMOPs:-Processing Time-Accuracy 31 and position),covariance,and sensors are accounted for.73 We can calculate covariance and Pc as a function of time and sensor measurement frequency.We set a Pc threshold of 105,which triggers a con

327、junction warning.In Figure 4.4,panel(a)illustrates the number of conjunction warnings that would be issued in a scenario in which a single sensor update is performed at T=0(no subsequent sensor updates are performed).When no measurement updates are performed,the number of conjunction warnings grows

328、continuously with projected time.Panel(b)illustrates the number of conjunction warnings issued in a scenario in which a sensor is tasked hourly and prioritizes objects with the highest Pc.In this scenario,the number of conjunction warnings initially grows and then decreases as the sensor resolves it

329、s queue of high-Pc objects.74 We note that all other parametersi.e.,initial object position and sensor measurement uncertaintyare the same for all the cases.Panels(c)and(d)illustrate the same scenarios for every two-hour sensor tasking and every four-hour sensor tasking,respectively.Figure 4.5 expli

330、citly shows the relationship between the update rate and maximum number of warnings for the case described here.We see a significant reduction in the number of collision warnings when the interval between updates is reduced from four hours to two hours.73 We used a simple sensor error representation

331、measurement errors for position and speedto form a covariance for each object.We then propagated speed uncertainty in time using a linear model to update the covariance of that object.If a sensor obtains a measurement of the object position,the position uncertainty is then reduced based on the senso

332、r measurement.The Pc estimate is calculated next by considering the overlap between the position uncertainty of the pair of objects being considered using a one-dimensional normal probability density function to represent the position uncertainty of an object.See National Institute of Standards and

333、Technology,“Normal Distribution,”for an explanation of normal probability density functions.The initial 100 objects are randomly distributed relative to each other and have a similar covariance based on the assumed sensor error model.The actual values of the sensor models are irrelevant for our purposes,since the effects we assessed are based on the difference between collision warnings with and w