Abstract
Deep learning has revolutionized signal classification, yet its reliance on vast labeled datasets remains a bottleneck in dynamic environments like 5G and electronic warfare. This paper introduces Spectral-Guided Self-Supervised Signal Classification (SGSSC), a framework that leverages unlabeled data by predicting a signal's frequency spectrum from masked time-domain inputs. By bridging the gap between temporal and spectral domains, SGSSC achieves superior performance with minimal labels and demonstrates remarkable robustness across varying channel conditions.
Introduction
In modern communication systems, accurately identifying modulation schemes—such as 8PSK or CPFSK—is vital for spectrum management and security. While supervised deep learning models excel at this task, acquiring expert-labeled radio frequency (RF) data is prohibitively expensive. Existing self-supervised learning (SSL) methods, often adapted from computer vision, frequently fail to capture the intrinsic physical properties of waveforms, as they lack a mechanism to understand the underlying frequency-domain semantics.
The SGSSC Framework
The proposed SGSSC framework addresses these limitations through a novel cross-domain pretext task and an adaptive fine-tuning mechanism.
Cross-Domain Pretext Task
The core innovation lies in treating the frequency spectrum as the "semantic label" for unlabeled time-domain signals. During pre-training, the model is presented with a signal where 70% of the data points are masked. The objective is not merely to reconstruct the missing time-domain samples but to predict the Discrete Fourier Transform (DFT) of the original, unmasked signal.
This high masking ratio (0.7) prevents the model from relying on local interpolation. Instead, it forces the encoder to learn global periodic structures and modulation-specific spectral characteristics. As shown in the figure above, different modulation schemes like 8PSK and CPFSK possess distinct spectral signatures that the model must internalize to succeed at the pretext task.
Token Attention Block for Fine-Tuning
Standard SSL pipelines often discard the final layers of a pre-trained encoder during downstream task adaptation. However, the authors found that for signal processing, these deeper layers contain critical modulation-specific knowledge. SGSSC employs a Token Attention Block that adaptively aggregates features from all encoder layers, ensuring that the semantic richness developed during spectral prediction is fully utilized for the final classification.
Experimental Results
The framework was rigorously tested on the RML2016.10a dataset under various channel conditions (AWGN, Rayleigh, and Rician fading) and on general time-series benchmarks from the UCR Repository.
Label Efficiency and Robustness
SGSSC demonstrates exceptional label efficiency, outperforming state-of-the-art SSL methods like TimesURL and InfoTS by an average of 2.3% when using only 0.5% of labeled data. Furthermore, the model exhibits strong cross-channel generalization; a model pre-trained on simple AWGN noise maintains high accuracy when fine-tuned on complex Rayleigh fading channels, significantly surpassing purely supervised alternatives.
Conclusion
SGSSC represents a significant step forward in applying self-supervised learning to the RF domain. By using the frequency spectrum as a guiding semantic signal, the model learns deep physical representations of communication signals. Future research will focus on evaluating the framework against real-world hardware impairments and non-stationary interference in "over-the-air" environments.
Summary updated