Automatic modulation classification based on deep learning

Modulation classification plays a key role in civilian and military applications, enabling cognitive radios to sense wireless environments, supporting spectrum surveillance, and enhancing electronic warfare through accurate signal identification. It also helps adaptive systems optimize modulation schemes under varying conditions. Traditional likelihood-based and feature-based approaches to Automatic Modulation Classification (AMC) depend on manually designed features and prior knowledge. Although these methods built the foundation of AMC, they struggle in complex or unseen environments. Deep learning enables end-to-end feature extraction but usually assumes a closed set of classes, which is impractical in real-world scenarios. Open-set AMC aims to identify known signals while rejecting unknown ones, yet collecting sufficient labeled data remains challenging. To address few-shot open-set AMC, the PEELER method employs episodic meta-learning with an open-set loss that maximizes posterior entropy, improving robustness to unseen classes. Building on this, the proposed Meta Supervised Contrastive Learning (MSCL) framework integrates meta-learning and supervised contrastive learning to refine embeddings. Episodic training enhances the clustering of similar signals and separation of dissimilar ones, producing more discriminative representations. Vision Transformers with Masked Autoencoders (ViT-MAE) are also explored to reconstruct masked signal images and learn contextual dependencies, improving representation quality under low-SNR conditions. To further enhance recognition, the CHEST framework integrates multiple signal representations—constellation diagrams, density heatmaps, eye diagrams, spectrograms, and time-domain waveforms—into a pretrained EfficientNet-B1 backbone using late-addition fusion. By enriching input diversity and leveraging fine-tuned foundation models, CHEST achieves state-of-the-art accuracy across many-shot, few-shot, and few-shot open-set AMC tasks, demonstrating the power of multi-representation fusion and large-model fine-tuning.