Modulation Classification of Satellite Communication Signals by Contrastive Learning of Dual Time-Frequency Representations

Modulation classification occupies a crucial part in spectrum management of satellite communication signals. Nowadays, deep learning brings great potential for its performance improvement. This paper suggests an automatic modulation classification approach by contrastive learning of dual time-frequency representations. First, signals are converted into two kinds of time-frequency representations by Short-Time Fourier Transform and Smoothed Pseudo Wigner-Ville Distribution. Second, a feature extraction module with coordinate attention is employed to capture dual time-frequency representation features. The features are projected to the contrastive learning space by two projection heads for contrastive learning which can improve the model. Finally, the extracted features are transmitted to a fusion classifier for classification. The model is trained on a public dataset and compared with the current advanced modulation classification models. Experimental results indicate that the proposed approach has significant performance advantages and application feasibility in the modulation classification of satellite communication signals.