Robust Emitter Classification Using A Scanning Receiver

In this research effort, Vadum will evaluate multiple classification techniques to mitigate the effects of corrupted measurements from the new scanning receivers in the SLQ-32 electronic warfare suite. Vadum will create an Automated Optimization Environment (AOE) to train three advanced classification techniques using existing electronic intelligence (ELINT) databases; these techniques will be tested using corrupted radar emitter intercept measurements that are noisy, biased, and missing pulses. Vadum"s innovative AOE approach, which optimizes parameters of the classifiers, reduces risk by allowing a wide range of classification techniques to be quickly optimized and evaluated. Many classification candidate techniques exist, each with advantages and disadvantages; during Phase I Vadum will evaluate three renowned techniques and determine which of these best solves the problem presented in this SBIR topic. Classification techniques to be evaluated include: Neural Network (NN) (tried and true), Support Vector Machine (SVM) (currently best in class), and Random Forest (RaFo) (state of the art). NNs are known to be robust and generalize well when not over-trained. SVMs are optimal, in the sense of maximized decision boundary margin, when classes are linearly separable. RaFos have been shown to have similar classification performance to SVMs but with less computational complexity. The study performed in Phase I will answer the question of which classification technique minimizes the emitter candidate list in the presence of biased, noisy, and incomplete scanning receiver measurements.