Exploring Self-supervised Learning for Radio Signal Recognition
Xingtong Yun, Xin Zhou · 2021
Data annotation is indispensable but difficult and expensive for supervised learning. While Radio signal is ubiq-uitous and available naturally, few approaches are designed to utilize large amounts of unlabeled data onto radio signal recognition. To handle this problem, we propose a self-supervised representation learning framework with two steps: pre-training for extracting features and fine-tuning for radio signal recognition. The radio signal from our dataset Rayl-0.5 contains 10 modulation methods for signal recognition task. Based on careful inspection of underlying regularities in the data, our model pre-trains radio signal without modulation methods labels, and provides competitive results by linear evaluation on recognizing radio signal with only 10% labels. A strict baseline method is considered using supervised learning for recognition task, which training radio signal dataset with 100% and 10% labels. Finally, our self-supervised model Self-RadioNet shows better performance compared to supervised learning methods trained for 10% labeled data.