Artificial Intelligence Capabilities for Aviation Behavior Characterization & Anomaly Detection

The overarching goal of the Phase II project is to develop a prototype software module that can be eventually integrated with NAVSEA systems to provide operators with Artificial Intelligence/Machine Learning (AI/ML) derived insight into air traffic operations and their intent in a geographic region. Specifically, we aim to provide behavior characterization (BC) and anomaly detection (AD) aspects of a geographic region’s aviation traffic in a human-understandable format. This Human-Interpretable AI Learning System (HAILS) is encapsulated as a software service so that it can easily be integrated with multiple systems. By defining what is “normal” traffic and bounding those operations by corridors, as well as identifying anomalous behavior, we aim to provide the NAVSEA system operator with better situational awareness of the aviation operations in the area and thereby seek to reduce workload. The findings of our Phase I research indicate that the use of AI/ML technologies to determine the BC and AD of area aircraft operations using publically available trajectory data is feasible and very promising for enhancing NAVSEA operators’ situational awareness in a given geographic region. We successfully demonstrated three key technical aspects: (1) we synthesized and curated aircraft trajectories from ADS-B data so that they are analysis-ready and can be fed into AI/ML algorithms for BC and AD;  (2) we applied AI/ML algorithms in the form of the Multiple Kernel Anomaly Detector (MKAD), originally developed by NASA, to a subset (i.e., approximately 42,500 arrival flights to San Jose International Airport) for training and anomaly detection; and (3) we increased the geographic scope of the trajectory data for AI/ML input to beyond the San Francisco Bay Area by the fusing of multiple sensory inputs. In Phase II, we develop HAILS leveraging our work and findings from Phase I (described above) and further test it on additional traffic scenarios. To accomplish this, we address three primary objectives in Phase II: (1) Extend prototype BC & AD components to additional traffic scenarios and volume. This objective involves processing additional data in terms of historical timeframe and geographic scope for input to the HAILS learning process; (2) Increase the number of features considered by the AI approach when performing BC & AD. In the AI/ML context, these features are an attribute or property of the system being studied (in this case, the air traffic in a region) on which analysis or prediction is to be done. By increasing the number of features, the number of ways that HAILS can characterize behavior and determine anomalies is increased; and (3) Increase the explainability of the system BC & AD output.  This objective addresses the key challenge with AI/ML systems in that often users fail to accept the results of AI/ML models because there is no way to explain the reason for the model output in a human-understandable form.