Date Approved

2-25-2026

Embargo Period

2-25-2026

Document Type

Thesis

Degree Name

M.S. Electrical and Computer Engineering

Department

Electrical and Computer Engineering

College

Henry M. Rowan College of Engineering

Advisor

Huaxia Wang, Ph.D.

Committee Member 1

Ben Wu, Ph.D.

Committee Member 2

Qianqian Zhang, Ph.D.

Committee Member 3

Chen Shen, Ph.D.

Abstract

Automatic modulation recognition has found its way into the core modern-day wireless communication systems whereby intelligent receivers estimate signal formats without any prior coordination with the transmitters. Challenging situations arise in any working radio frequency environment for the conventional modulation recognition systems due to insufficient labeled data, rapidly fluctuating channel conditions, and appearance of plenty other unseen modulation types. Taken together, these challenges raise questions concerning the ability to institute learning frameworks that could perform well on few-shot and open-set conditions. Hence the thesis attempts to address the same by proposing two complementary learning-based frameworks for few-shot open-set modulation recognition by using signal constellation representation, meta-learning, and contrastive learning methods.The first chapter provides a few-shot open-set automatic modulation classification frame- work based on meta-learning and signal constellation diagrams. The complex IQ samples are transformed into a constellation image and are processed through an encoder based on ResNet18 in an episodic meta-learning framework. The strategy employs a prototype-guided learning approach to enable rapid adaptation from a small set of labeled samples, while an uncertainty-aware open-set loss has been integrated into the framework to alleviate overconfident predictions on unseen modulation types. The proposed technique is trained for 30,000 episodes of meta-learning over 48 modulation classes and multiple signal-to-noise ratio conditions, ensuring stable convergence, strong prototype formation, and reliable rejection of unknown signals. Experimental results prove that the framework maintains a balance between classification accuracy and open-set awareness, making it suitable for dynamic spectrum environments. The second part of the thesis is dedicated to developing a Meta Supervised Contrastive Learning framework (MSCL-X) for further improving few-shot open-set modulation recognition. Rather than relying upon the traditional supervised classification schemes, MSCL-X learns a discriminative embedding space applying supervised contrastive learning to constellation images derived from raw IQ samples. This encourages strong intra-class compactness and large inter-class separation in the feature space. Gaussian prototypes are fitted to the embeddings learned from known modulation classes enabling likelihood-based classification and principled rejection of unknown signals. Extensive experimental evaluations under realistic channel impairments show that MSCL-X consistently outperforms state-of-the-art open-set baselines such as OpenMax-Lite, ICS-Lite, and GE2E-Lite regarding macro-F1 score and normalized accuracy. To conclude, this thesis demonstrates combining constellation-diagram representations with meta-learning and supervised contrastive learning as a powerful, data-efficient method of few-shot open-set modulation recognition. The proposed frameworks attain robustness to channel perturbations, reduce dependency on large labeled datasets, and achieve reliable detection of unfamiliar modulation types. These contributions push intelligent wireless signal recognition further along its SOTA and provide practical value for cognitive radio, spectrum monitoring, and electronic warfare applications in dynamic and non-cooperative environments.

Download

Share

COinS