EE Seminar: Ambiguity Function Based Radar Waveform Classification and Unsupervised Adaptation Using Deep CNN Models
Speaker: Pavel Itkin
M.Sc. student under the supervision of Prof. Nadav Levanon
Wednesday, January 1st, 2020, at 15:00
Room 011, Kitot Bldg., Faculty of Engineering
Ambiguity Function Based Radar Waveform Classification and Unsupervised Adaptation Using Deep CNN Models
Abstract
We present a general and robust approach to classification and an unsupervised adaptation of phase and frequency modulated low-probability-of-intercept (LPI) Radar waveforms. Our model is inspired by deep convolutional neural networks (CNNs) that have been successfully used for multi-class image classification and domain adaptation.
Our method considers a non-cooperative and intercepted Radar signal to be a 1D sequence of complex samples assembling a single pulse. Each one is transformed into a square and complex Ambiguity Function (AF) matrix as a pre-processing step and associated with its waveform characteristics for correct classification.
A test signal is processed and mapped onto a 256-dimensional feature embedding vector, representing the waveform class space, following which, a waveform metric classification, as well as adaptation to an unlabeled reference target domain, is performed.
We use our method on a diverse simulated dataset, consisting of different encodings, pulse widths, bandwidths, a wide range of noise levels, and different signal and noise distributions as the unlabeled target domain.
Our method achieves state-of-the-art performance on classification problems on multi-encoding and multi-feature waveform datasets while incorporating only one Radar pulse and proved to be robust to diverse and extremely intense noise conditions, along with excellent generalization to other Radar related problems.
Furthermore, our novel approach to an unlabeled Radar waveform adaptation reveals an impressive classification performance improvement to a domain-shifted, differently distributed datasets.