Classifying Spacecraft Collision Risk: A Machine Learning Approach Using GRU Neural Networks

الملخص

As the density of space debris in low Earth orbit increases, the likelihood of or- bital collision risks in space rises significantly over time. This risk poses a critical and significant challenge for the global space sector. With the advancement of technology, particularly in artificial intelligence and machine learning, and their increasing application in solving real-world problems, our study focuses on utilizing these technologies, specifically by employing GRU (Gated Recurrent Units) networks. We utilized data from the” Collision Avoidance Challenge” released by the European Space Agency in 2019 to classify collision risks between spacecraft and other objects. By analyzing temporal patterns and the proximity of objects to satellites, we distinguished high- risk cases from low-risk ones. The data contains a large number of Conjunction Data Messages (CDMs), with over 162,000 messages included. Due to this large number, we needed to preprocess the data, which involved several steps, including imputing missing values using linear interpolation, standardizing features, and converting risk values into binary classifications. Several algorithms were applied to select the most influential features on the likelihood of a collision, combining statistical analysis (Pear- son and Spearman), mutual information, SHAP analysis, and permutation importance, ultimately resulting in the selection of the best 25 features using ensemble feature se- lection methods. The model was trained on sequential time-series data, and we used a Masking layer to address the issue of unequal lengths in the data sequences. Additionally, Dropout layers were applied, which significantly helped reduce the problem of overfitting. The model achieved high performance, with a verification accuracy of 97%, demonstrating its effectiveness in classifying collision risks. The results of this study highlight the potential of deep learning in solving orbital collision problems in space, enhancing collision prediction systems, and enabling proactive maneuver planning in satellite operations.