Classifying LPI radar waveforms with time-frequency transformations using multi-stage CNN system

Abstract

As the number of radar waveforms in the cognitive electronic warfare applications increases, individual detection and classification performances of each waveform vary furthermore due to their different characteristics. To provide a supervised signal classification in an efficient framework, we propose a multi-stage waveform classification system, where multiple modular blocks are combined to classify 18 different radar waveforms. In the first stage, we transform the signals into time-frequency images (TFIs) using Fourier-based Synchrosqueezing Transform (FSST) and SqueezeNet to classify the signals into two subsets: P1-4 and others. Then, the subsets are used as inputs to two different systems. These systems use different TFI techniques such as FSST and Smoothed Pseudo Wigner Ville Distribution (SPWVD) for processing and convolutional neural network (CNN) architectures such as Squeezenet, ResNet-50, and ShuffleNet for classification. In experiments, we provide supervised classification results at different signal-to-noise ratio (SNR) levels and achieve 98.08% classification accuracy at 10-dB SNR on a diverse set of frequency and phase modulated signals.