Self-Supervised Adaptive Signal Representation Network for Low-SNR Wireless Signal Classification

A reliable wireless signal classification at low signal-to-noise ratio (SNR) is crucial for enabling cognitive wireless network and intelligent spectrum sensing, however, it is difficult due to deep noise, channel distortion and the lack of labeled data. Most supervised DL approaches require a lot of annotations and have poor generalisation in bad communication conditions. We introduce following an in-phase and quadrature (IQ) concept in contrastive self-supervised signal representation, an adaptive feature refinement model and supervised discriminative signal representation, named as Self-Supervised Adaptive Signal Representation Network (SS-ASRN). It is evaluated in simulation with RadioML2018.01A dataset with various SNR ranges from −20 dB to 20 dB with AWGN and fading channels. The results of simulated analysis indicate that classification accuracy is 96.7%, precision is 95.2%, recall is 94.8%, F1 score is 95.0%, and AUC is 0.987, and the accuracy is stable at low SNR. The proposed framework offers scalable, efficient, and effective solution for smart wireless signal classification in next generation cognitive wireless communication systems by leveraging data efficiently.