Recognition of Low-Probability-of-Intercept Radar Signals Based on a Hybrid CNN-MLP Model

Abstract

This paper proposes an intelligent recognition system for low-probability-of-intercept (LPI) radar signals based on a hybrid neural network model that combines a complex-valued convolutional neural network (Complex 1D- CNN) for processing in-phase and quadrature (IQ) components with a multilayer perceptron (MLP) for extracting parametric signal features. Unlike conventional approaches that rely on time-frequency transforms such as the short- time Fourier transform (STFT) and Choi–Williams distribution (CWD), the proposed model operates directly on complex-valued data while preserving the phase structure of the signal. To improve robustness under noisy conditions, a Hankel Cross-Attention mechanism and SNR-aware loss weighting are employed. Experiments conducted on the RadChar Baseline dataset (5 classes) achieved an accuracy of 90.0% and a macro F1-score of 0.90, outperforming single-branch CNN models and demonstrating stable performance at SNR levels down to -12 dB. The obtained results confirm the effectiveness of the proposed integrated approach for LPI radar signal recognition in noisy environments.