A Low-Complexity Deep Learning Model for Modulation Classification and Detection of Radar Signals

In the military environment, accurate automatic modulation classification and detection of radar signals are crucial for distinguishing between various radar signals and adversarial transmissions. Deep learning (DL) models are increasingly used for automatic modulation classification; however, existing DL-based models often suffer from high computational complexity, limiting their feasibility for real-time deployment. This article presents a radar classification and detection neural (RCDN) model designed to classify and detect 23 different radar modulation signals in an additive white Gaussian noise environment. The RCDN model employs a combination of R-blocks, each comprising multiple convolutional layers interconnected with skip connections, thereby preserving residual information and addressing the gradient vanishing problem. Furthermore, we use a weight-based pruning technique to design a PrunedRCDN model to reduce the complexity of the RCDN model. We evaluate the computational complexity and classification accuracy of the original RCDN model and its pruned variant against convolutional neural network (CNN), ResNet, and CNN-gated recurrent unit (GRU) architectures using the DeepRadar dataset. Our simulation results demonstrate the superiority of the PrunedRCDN model over CNN, ResNet, CNN-GRU, and the original RCDN model in terms of complexity reduction and classification accuracy.