Description: We propose to extend our Phase II aircraft flight behavior anomaly detector by adding a classifier to filter out operationally irrelevant statistical anomalies. This work will benefit NASA's System-Wide Safety (SWS) project by developing predictive analytics "safety intelligence" technology based on machine learning algorithms. Subject matter expert reviews in Phase II identified that alerting on some, but not all, statistical anomalies would improve situational awareness and could help reduce or avoid safety incidents. This work will reduce the nuisance alarms for air traffic controllers and pilots that would otherwise lead to distractions and desensitization. We will apply active learning to reduce the labeling time and expense associated with subject matter expert review of statistical anomalies used to train our classifier. We will also apply semi-supervised learning to improve the ability to learn from small labeled training datasets. Subject matter experts will review selected statistical anomalies as part of experiments designed to measure the extent to which active learning improves the training process and reduces labeling costs.

Benefits: Our flight anomaly detection and risk prediction software will provide a key capability for NASA’s In-time System-wide Safety Assurance (ISSA) initiative. Research into active learning and semi-supervised learning for aviation data can inform internal NASA research. Integration with NASA's Digital Information Platform (DIP) will enable new predictions that can be further processed by NASA and third party systems for situational awareness and historical analysis.

Air traffic controllers can receive operationally relevant alerts to flights exhibiting anomalous behavior in time to take corrective action. Pilots can be alerted to nearby aircraft with anomalous flight behaviors.