Autonomous anomaly detection via physics-regularized machine learning

Manual inspection of telemetry data in the search for anomalies is a time-consuming threat detection technique. Most multi-signal systems send back extensive data that a single person cannot easily monitor in real time. Machine learning techniques that autonomously scan data and flag anomalies are attractive alternatives. The autonomous anomaly detection problem can be divided into two sub-problems: regression analysis and a classification process. In the regression analysis, a machine learning model is trained to reconstruct a given signal, and the classification process categorizes the reconstruction error as anomalous or nominal. This report examines the autonomous anomaly detection problem and proposes improvements to both the regression and classification sub-problems. With regard to the regression analysis, it was found that including the physics of the target signal in the machine learning model yielded a lower reconstruction error compared to a purely data-driven model. The classification approaches studied showed that cluster-based thresholding techniques accompanied by a pruning procedure can outperform non-parametric dynamic thresholds.Aerospace Engineerin