Space sensemaking and space domain understanding: enabling data-centric AI for space flight safetyby Brien Flewelling
|
| In LEO, we are seeing unprecedented growth in the numbers of active, maneuverable, and even autonomous spacecraft, and this traffic is not yet sufficiently supported from a data perspective. |
ExoAnalytic has pioneered the Space Sensemaking market supporting geosynchronous orbit. Space Sensemaking is the derivation of actionable information from observations of space platforms characterized by the performance of Data Work such that artificial intelligence may be reliably applied in high-stakes applications of space data exploitation. Data Work is the detailed curation, cleaning, standardization, management, and verification of data from various sources such that modern data exploitation techniques, including machine learning, deep learning, and artificial intelligence, may be reliably applied to support high-stakes decision-making
Space Sensemaking includes our commercial infrastructure, which now delivers the desired level of automation described by Gen. Shelton and data collection strategies capable of supporting the space domain understanding required by Gen. Dickinson. By demonstrating persistent monitoring via a proliferated array of optical telescopes, we have shown that there are significant benefits to dedicated commercial observation of space systems.
As we support the space operational community with commercial advanced services, we have learned several lessons:
Valuable services minimize latency to preserve decision time for operators. Often this means having dedicated sensors observing persistently to ensure updates outpace the rate of change.
This could mean supporting dynamic space operations like the successful docking of the MEV-1 and MEV-2 spacecraft or providing on-demand analysis and alerting when new behaviors are observed, including launch and early operations, geosynchronous transfer orbits, stationkeeping strategies, and deorbit or disposal events.
It is valuable to be watching anyway so operators can be proactive in their decision making to optimize the use of their space system and preserve flight safety.
Our commercial services for GEO aim to make sure that data can be collected and exploited to optimize the effective and sustainable use of space systems.
In LEO, we are seeing unprecedented growth in the numbers of active, maneuverable, and even autonomous spacecraft, and this traffic is not yet sufficiently supported from a data perspective. In all orbit regimes we are witnessing a boom in dynamic space operations including on-orbit servicing, refueling, inspection, and other proximity operations. All these activities require more data to inform their decision-making. Economic activity in space is pivoting from siloed space systems, support infrastructure, and business models to networked, collaborating systems of systems, all connected by the cloud. As a result, there has never been a larger opportunity to field large numbers of remote-sensing capabilities which can make sense of the space domain to support a sufficient understanding to trust the decision making that will continue to scale within this orbital regime.
To this end we must pivot from the minimum-data-necessary approaches historically taken as part of efforts in Space Situational Awareness (SSA), beyond their natural extensions to Space Domain Awareness (SDA) and Space Traffic Management (STM), to the maximum-information strategies required by Space Sensemaking to support Space Domain Understanding.
| The algorithms will mean little, no matter how advanced, if there is insufficient data to ensure the required level of fidelity in reconstruction of space system behavior is achieved. |
As Gen. Dickinson expressed, “SDA gives us insight into activity throughout the space domain, including potential adversary activities, but perhaps more importantly, insight into the intent of those potential adversaries, too… All of this helps inform our understanding of adversary behavior.” It is no longer sufficient to simply have an awareness of space systems activity: we must understand many facets of the system behavior to most accurately infer intent and predict future actions. What is different today is that the luxury of time to make operator-in-the-loop decisions is rapidly disappearing. Gen. Shelton foresaw this more than a decade ago.
Today, the scale of distributed space systems, the rate at which they are maneuvering, and the scale of the hazards and threats that they must contend with are all unprecedented and growing rapidly. To do space sensemaking at the required scale and speed to support every active space platform in the future, we will require the coordination and effective processing of measurement data from equally large-scale networks of observation systems on the ground and in space. The algorithms will mean little, no matter how advanced, if there is insufficient data to ensure the required level of fidelity in reconstruction of space system behavior is achieved.
Geosynchronous orbit is still the most heavily populated orbital plane, and it has progressed the most in terms of the establishment of norms of behavior for station keeping and prevention of radio interference. ExoAnalytic collects the lion’s share of measurements on space systems in this regime and has demonstrated a mature, vertically integrated system which autonomously performs space sensemaking in support of space domain understanding for all operators. The fidelity with which we understand the behaviors of these systems is unmatched and, for our clients, well exploited to support safe and efficient operations.
In this context, space sensemaking can be broken down into a series of steps, including tasking, detection, correlation, feature extraction, and comparison of observed behavior to previous models. This detailed process has been carefully automated following a rigorous evaluation of the reliability of the algorithms performing each function. With each mapping from image coordinates to right ascension and declination, association of new measurements with cataloged objects, state vector update, and maneuver detection, we achieve a 99% confidence without a human in the loop. Taken together, a simple mathematical model can be used to evaluate our performance in informing the community or an autonomous system on the current state of the GEO orbital environment. As a simple example, this four-step process achieves a 96% composite performance when we multiply 99% four times (one for each step in the process).
To illustrate the difficulty of achieving trust in deployed automated systems within which multiple steps and algorithms must be employed, following the same process for a series of algorithms which achieve only 90% performance results in a composite performance of only 65.6%. This precipitous drop in performance is unacceptable and a subject of serious concern in the artificial intelligence community, which is revisiting the concept of “Data-Centric AI” that seeks to elevate the emphasis placed on data work such that the required level of performance for high stakes deployed inspection systems incorporating AI can achieve the required results.
For GEO, ExoAnalytic has done the data work. When we analyzed the problem for GEO, we developed a data-driven strategy that let us field our network of more than 400 telescopes worldwide. This enables us to ensure that we design and execute data strategies that ensure we have enough measurements from geographically and geometrically diverse vantage points to obtain state vector updates faster than the space systems we observe are changing them. For this reason, we can execute our automated processes to support faster decision making for our clients than anyone else in the world.
| Commercial solutions exist that will enable the proliferation of ground and space sensors to support space sensemaking. It should not be a debate if we should field them. |
For other orbit regimes, there is insufficient data collected to achieve this level of performance. We must increase the rate and the number of unique vantage points that we observe all space objects from and maintain an ability to accurately understand the states of objects despite more frequent maneuvers. Failure to do this will mean that safety recommendations, conjunction warnings, norms of behavior, and policy will be developed based on poor representations of actual behaviors of space systems. There are commercial solutions available today that with sufficient investment could scale to address these challenges. The approach should be multi-faceted including radar, ground and in-space sensing, and tracking aids.
Ultimately, we must achieve sufficient space sensemaking performance, including achieving the required latency to support frequently maneuvering space traffic in LEO and the geometric diversity, sensitivity, and coverage challenges of cislunar space. This will mean being able to completely automate the mapping of observation data on space systems to alerts associated with potential interactions between space systems such that onboard autonomous decision making can be afforded the benefit of up-to-date (perhaps up-to-the-minute) understanding of the space domain. While there are programs intended to address these issues (e.g., DARPA SpaceWatch), the fraction of space systems being fielded that are capable of contributing sensor data to support space sensemaking at scale is still too low.
Commercial solutions exist that will enable the proliferation of ground and space sensors to support space sensemaking. It should not be a debate if we should field them. We need to move beyond that discussion to doing the data work such that confident alerting can be performed at speed and scale. Today’s sensing infrastructure is inadequate, despite this fact being called out by Space Policy Directive 3 more than five years ago. If this policy directive was a spacecraft in LEO at the end of its life, we would be recommending it be deorbited by now and replaced. The time for aspirational visions for what should be done has passed. All of us must act to ensure that we are collecting the data necessary to support confident, trusted, autonomous decision-making at scale for all orbit regimes. It is certain that decision time will continue to be a premium, new activities will dictate modern data strategies, and more challenging and exciting days in space are ahead of us. We have a lot of data work to do!
Note: we are using a new commenting system, which may require you to create a new account.
