Date of Award

2017-01-01

Degree Name

Master of Science

Department

Electrical Engineering

Advisor(s)

Benjamin C. Flores

Abstract

A Neural Network (NN) used to classify radar signals is proposed for the purpose of military survivability and lethality analysis. The goal of the NN is to correctly differentiate Frequency-Modulated (FM) signals from Additive White Gaussian Noise (AWGN) using limited signal pre-processing. The FM signals used to test the NN approach are the linear or chirp FM and the power-law FM. Preliminary simulations using the moments of the signals in the time and frequency domain yielded better results in the frequency domain, suggesting that time domain training would not be as effective frequency domain training. To test this hypoThesis, we developed a training procedure for the NN using either spectra or autocorrelation sequences as inputs as they require a similar amount of signal preprocessing. Classification performance was done in terms of the probability of false alarm (PFA), probability of detection (PD), and probability of error (PE) as a function signal-to-noise-ratio (SNR). In one case, the NN is trained with a set of spectra with either a noisy FM signal with random carrier frequency and bandwidth or strong bandlimited white noise. Simulations show that at an SNR of 5dB, the NN consistently performs signal classification with a PFA close to 0% and a PD higher than 85%. At a SNR of 10dB, the NN reaches a PE of 0%. In another case, the NN is trained with a set of autocorrelations of either a noisy signal or bandlimited noise. At an SNR of 5dB, the NN consistently performs signal classification with a PFA close to 0% and a PD higher than 99%. At a SNR of 10dB, the NN reaches a PE of 0%. In a third case, the NN is trained with a set of signals, which are either linear FM, power-law FM, or bandlimited white noise. Here, at an SNR of 5dB, the NN consistently performs signal classification with a PE close to 0% for both the spectra and the autocorrelation. The conclusion is the NN at a high SNR level performs exceedingly well for either case. However, at very low SNR, the NN radar signal classifier performs better when its input is the autocorrelation of the signal.

Language

en

Provenance

Received from ProQuest

File Size

49 pages

File Format

application/pdf

Rights Holder

Ariadna Estefania Mendoza

Share

COinS