Deep Convolutional Neural Network Classifier of Pulse Repetition Interval Modulations

Electronic support systems obtain valuable information about a pulsed radar through the analysis of its Pulse Repetition Intervals (PRIs). PRI signal is very agile and undergone complex distortion which makes its analysis non-trivial. In this paper, a Convolutional Neural Network (CNN) is proposed for automatic PRI classification. We show that standard min-max and zero-mean unit-variance data normalization is not suitable. We propose a novel normalization providing faster training and lower classification error. We show that a recently proposed CNN scheme is prone to miss-classify higher cardinality stagger so we enhance the scheme by deeper layers, batch normalization and dropout regularization.