The exponential increase in the number of objects in orbit, particularly in low earth orbits, is putting a major strain on satellite security activities. It requires continuous monitoring and increased responsiveness, which places a heavy operational burden on the orbital mechanics teams.

In recent years, solutions have been developed to enable on-board autonomous collision risk management. This on-board risk management approach has been followed by CNES and implemented in the so-called ASTERIA system. This system of high responsiveness accounts for changes in collision risk and implements appropriate mitigation solutions. To improve the on-board decision-making process, the so-called ALIGATOR algorithmic module has been developed. The latter, based on the use of machine learning, adds a step to strengthen decision-making on board. 

Before being functional and implemented in the satellite flight software, the AI model of ALIGATOR must be trained on ground, benefiting from enormous training datasets and high computing power.

High-quality training data is crucial to ensure model's performance and avoid biases. In this paper, we show how to address this challenge by combining real data with generated counterparts to minimize any bias in the training dataset. The real data arises from a dataset handcrafted by the French space surveillance team CAESAR. It consists of a huge set of collision data messages detailing close encounters between satellites monitored by CAESAR and other flying objects. Each conjunction corresponds to a time series of collision data messages leading up to the Time of Closest Approach (TCA). Using this dataset, we have generated additional data using an innovative simulator. The latter allowed us to densify each time series with additional collision data messages, hence consolidating our training dataset.

The first part of the paper will detail the breakdown of the simulator and the pipeline used to generate the training data. Then, a presentation of the the feature selection process for model training will be done, as well as the selection of the model itself, with a focus on how to minimize biases. Then, the choice of the model will be discussed starting from models that have been thoroughly studied. Finally, the initial training results will be shared, demonstrating the effectiveness of the approach.