AIMC-Spec: A Benchmark Dataset for Automatic Intrapulse Modulation Classification under Variable Noise Conditions

TL;DR

AIMC-Spec addresses the lack of a standardized benchmark for Automatic Intrapulse Modulation Classification by introducing a fully synthetic, spectrogram-based dataset with $33$ intrapulse modulations across $13$ SNR levels and ground-truth labels. The authors re-implement and benchmark five diverse deep learning architectures (including LDC-Unet and ViT) under a unified input format, revealing that frequency-modulated signals are considerably more separable than phase or hybrid types at low SNRs. Key contributions include public data and code, a reproducible evaluation baseline, and insights into how architectural choices and preprocessing affect performance under noise. The dataset aims to standardize AIMC research, enabling fair comparisons and guiding future developments in intrapulse modulation classification.

Abstract

A lack of standardized datasets has long hindered progress in automatic intrapulse modulation classification (AIMC) - a critical task in radar signal analysis for electronic support systems, particularly under noisy or degraded conditions. AIMC seeks to identify the modulation type embedded within a single radar pulse from its complex in-phase and quadrature (I/Q) representation, enabling automated interpretation of intrapulse structure. This paper introduces AIMC-Spec, a comprehensive synthetic dataset for spectrogram-based image classification, encompassing 33 modulation types across 13 signal-to-noise ratio (SNR) levels. To benchmark AIMC-Spec, five representative deep learning algorithms - ranging from lightweight CNNs and denoising architectures to transformer-based networks - were re-implemented and evaluated under a unified input format. The results reveal significant performance variation, with frequency-modulated (FM) signals classified more reliably than phase or hybrid types, particularly at low SNRs. A focused FM-only test further highlights how modulation type and network architecture influence classifier robustness. AIMC-Spec establishes a reproducible baseline and provides a foundation for future research and standardization in the AIMC domain.

Figures (11)

• Figure 1: Spectrograms of a 4-frequency shift-keying (4FSK) signal across all SNR levels in AIMC-Spec, ranging from +10 dB (top left) to –20 dB (bottom right). SNR initially decreases in 5 dB steps from +10 to 0 dB and then alters to 2 dB steps from 0 to –20 dB. The progression illustrates the degradation of time–frequency structure as noise increases.
• Figure 2: Spectrogram representations of the 4FSK modulation at 0 dB SNR following preprocessing by the six benchmark algorithms. Each spectrogram was generated using a 256-point Hann window with 50 % overlap.
• Figure 3: Accuracy (%) vs SNR (dB) for experiment SNR($x$):SNR($x$), using all modulation types.
• Figure 4: Accuracy (%) vs SNR (dB) for experiment SNR(ALL):SNR($x$), using all modulation types.
• Figure 5: Accuracy (%) vs SNR (dB) for experiment SNR($x$):SNR($x$), using only frequency modulations.