SBIR/STTR Award attributes
Lethal threats to Navy vessels from torpedoes and surface craft can materialize quickly. Rapid detection and classification to enable effective evasive maneuvers or countermeasures is a ship-safety priority.Recent improvements in Machine Learning (ML) have enabled advances in commercial classification systems. One subset of Machine Learning is supervised Deep Learning (DL), which uses large amounts of training data to select nonlinear features that best discriminate the members of different classes. Not only do these features represent the underlying data, but, more importantly, they also isolate differences among the classes. The Metron/L3-AdaptiveMethods (L3-AM) team proposes an innovative approach that addresses the shortcomings of previous Torpedo Detection approaches by employing physics-based modeling to generate massive numbers of high-fidelity training data sets, combined with recently-available DL technologies to produce a favorable detection/false alarm trade-off. The proposed approach combines spectral and kinematic features in a holistic way for increased robustness and improved performance by making decisions based on all available input information instead of on an ad-hoc basis with individual features.
