AI for Collision Avoidance — Go/No Go Decision-Making

Kurzfassung

The proliferation of manmade objects in orbit for commercial, defense, and research purposes has resulted in frequent collision alerts. While much of the conjunction assessment and collision avoidance pipeline is automated, the decision on whether or not to carry out an avoidance maneuver is still operationally done manually. This work highlights the factors influencing this decision-making process, and explores the potential of automating it using artificial intelligence (AI). Collision avoidance analysts from the German Space Operations Center (GSOC) have been consulted to provide insights into the decision making process. A set of high-risk Conjunction Data Messages (CDMs) is provided to multiple analysts who are tasked with manually classifying them into two classes -- 1. Maneuver recommended 2. Maneuver not recommended. These CDMs have a probability of collision in the order of 10^-4, which is low enough to not always warrant an avoidance maneuver and high enough to be concerning for operators. For each event they classify, the analysts fill out a document stating the rationale behind their decision, specifically highlighting the feature(s) that dictated their decisions. Since multiple analysts are asked to classify the same events, it is possible to obtain a list of events for which their decisions were not in alignment -- that is, two analysts arrive at two different conclusions when faced with the same CDM. These tricky cases are of particular interest and it is important for an AI decision-making system to account for these appropriately. The CDMs that are classified by the analysts are then fed into a machine learning model as the training data. Feature engineering and parameter selection is done to ensure the model is well equipped to learn from the data. The model's ability to identify relevant patters from the CDMs and correctly classify unseen CDMs into the two classes is assessed. The presentation covers preliminary results from this model and findings learned from carrying out this investigation. The development of the model is still a work in progress, as further training data (up to 250 CDMs classified by analysts) is yet to be obtained. The current small-scale model is intended to act as a proof of concept. The results obtained provide interesting insights into the scope for applying ML to this problem along with potential limitations. Planned work entailing additional training of this model and the development of additional models trained using other methods is also shared.