Classification of radar jammer FM signals using a neural network

We propose an approach based on artificial Neural Networks (NN) to classify wideband radar jammer signals for more efficient use of countermeasures. A robust NN is be used to correctly differentiate Frequency-Modulated (FM) signals from bandlimited Additive White Gaussian Noise (AWGN). We compare the performance of the NN using the samples of the power spectrum versus the autocorrelation. Prior experiments showed that frequency-domain moments of the FM signal itself are better descriptors than time-domain moments. Using simulated wideband FM radar signals, we compute a set of N autocorrelation and spectra and feed them to the NN which has ten hidden layers. For training purposes, the autocorrelations or spectra sets are divided into three groups, 75% for training, 15% for validating and 15% for testing. For the power spectra set, we observe that a Signal to Noise Ratio (SNR) of 5dB allows the network to approach an average of 5% percent Probability of Error (PE). Training with the autocorrelation set yields comparable results. For an SNR of 5dB, the average PE reached an average of 0.3%. In both instances, the NN reaches zero percent PE at an SNR of 10dB.