Robust Classification with Belief Functions and Deep Learning Applied to STM

This work proposes an approach to robust con-junction risk assessment given a sequence of Conjunction Data Messages (CDM). Dempster-Shafer theory of evidence (DSt) is used to account for epistemic uncertainty in the sequence of CDMs and derive a robust classification of conjunction events. We then use Artificial Intelligence (AI) to bypass the computationally expensive parts of DSt and directly produce a robust classification from a given sequence of CDMs. Five AI techniques are proposed: Random Forest (RF) with DSt structures, RF with CDMs, Light Gradient Boosting machine (LGBm) with CDMs, autoregressive LGBm (aLGBm), and Transformers for time series. These meth-ods were trained and tested both on synthetic and in real datasets to study their applicability to real scenarios. The results show the potential of AI techniques, especially LGBm, for robustly classifying encounters from the sequence of CDMs, provided balanced datasets are available.