Learnable Gaussian Filter-Based Automatic RF Waveform Modulation Recognition

In this work, we propose an automatic modulation recognition (AMR) method leveraging complex-valued convolutional neural networks (CV-CNNs) with parameterized, learnable Gaussian filters. Traditional AMR techniques often require complex pre-processing and transformation stages, which are computationally expensive and hinder real-time applications. Our proposed architecture processes raw complex IQ radiofrequency(RF) data directly, incorporating parameterized filters as learnable bandpass-like functions for efficient time-frequency feature extraction. This filter structure reduces the number of learnable parameters, enhancing interpretability and robustness in dynamic RF environments. We evaluate the model on a synthetic dataset of 15 radar waveform classes across three channel types-ideal, Rayleigh, and Rician-spanning signal-tonoise ratios (SNRs) from -20 dB to 18 dB. The proposed method outperforms traditional models in all key classification metrics, achieving an average accuracy above 85 % and demonstrating consistently high performance across varying SNRs. Additionally, it exhibits greater resilience to noise compared to baseline models. Comparative analysis shows that the proposed approach effectively distinguishes between diverse radar modulation classes, even under challenging conditions.