Deep Learning, Neural Networks and Aerospace Applications

Abstract

The ability to rack an enemy missile while in-flight, whether to accurately predict the point of impact (POI), the point of origin (POO) or to destroy the missile in-flight, is greatly enhanced if the missile type and/or physical characteristics (type, size, payload, etc.) are known as a priori. Given the missile type/characteristics, a conditional missile tracking system can be developed using simulated missile fly-outs (6-DOF) based on the physics of that missile type. If the missile type or characteristics are unknown, assuming the known classes of missiles in the enemy’s arsenal, a rapid missile classification must be incorporated into the refined posterior tracking system. This tracking system is a two-stage tracking system. In the first stage, within a few milliseconds of radar detection, the missile is rapidly and accurately classified within into one of k classes. In the second stage, the specific tracking system tailored to that specific class is engaged for more accurate tracking. In this thesis, we focus on deep learning neural networks (DNN) to solve the rapid missile classification problem in this application. We demonstrate the superior performance of DNNs over single layer neural networks, as well as, classical generalized linear model, using 6-DOF-fly-outs of three similar short-range rocket classes. We show that we can achieve 100% corrected classification within milliseconds of flight on our testing data (independent fly-outs).