CPAA: Self-Supervised Cross-View Prediction With Automatic Augmentation for OFDM Modulation Classification
Bing Ren, Zhengyang Su, Kah Chan Teh, Hongyang An, Erry Gunawan, Alex Chichung Kot · IEEE Transactions on Vehicular Technology · 2025
Orthogonal frequency division multiplexing (OFDM) technology, robust against frequency-selective fading channels, is widely applied in the fifth-generation (5G) communication systems, which makes automatic modulation classification (AMC) tasks critical for OFDM communication systems in military and civilian applications. This paper proposes a self-supervised learning framework, named cross-view prediction with automatic augmentation (CPAA). Unlike existing supervised methods requiring substantial labeled data, CPAA leverages massive unlabeled samples and minimal labeled data. Initially, an automatic data augmentation module is incorporated with traditional augmentation methods. Subsequently, the unlabeled pairwise signals are utilized for self-supervised pertaining through the cross-view prediction (CP) module. Finally, supervised fine-tuning using well-trained encoder is conducted on a small amount of labeled datasets. The proposed framework is capable of identifying five different OFDM modulation formats and enhancing performance through automatic data augmentation, which effectively retains data identity and generates diverse views. Simulation results indicate that CPAA can achieve robust classification performance with massive unlabeled samples and extremely limited labeled samples. Moreover, the proposed framework outperforms state-of-the-art methods across a range of signal-to-noise ratio (SNR) levels and demonstrates better generalization capabilities under low SNR conditions. Based on the simulation results, the CPAA framework achieves a classification accuracy of over$\bm {85\%}$at SNR$\bm {>}$8 dB with only one labeled sample per SNR for each modulation scheme. Moreover, the automatic data augmentation module enhances the performance of self-supervised learning when limited labeled data are available, yielding an approximate average performance gain of$\bm {48.78\%}$and demonstrating good transferability on the existing self-supervised learning algorithms.