Spatiotemporal Precursors of Safety Incidents

We propose to develop, train, and validate machine learning methods to automate the discovery of safety incident precursors so that incidents can be avoided. This supports NASArsquo;s In-Time System-Wide Safety Assurance (ISSA) focus area. We will extend recent research by NASA that discovers the times at which precursors occurred for individual flights. Our extension will discover spatiotemporal precursors, i.e., both the times and data feature values that precede the incidents. Our method will find local precursors for individual safety incidents and global precursors specifying general rules of thumb that are patterns across many incidents. Currently, such pattern finding is a manual process. This work will generate new insights into the causes of unstable approaches to inform accident investigators, pilots, ATC, policy makers, and machine learning model developers. This initial Phase I study lays the groundwork for finding precursors of more complex safety incident types such as anomalies detected by black box models developed by Metron and NASA. Additionally, our precursor models will enable in-time safety incident prediction. Our team includes experts in air traffic, machine learning, and airspace data processing.