📝 SummarXiv

Abstract

Signal classification models based on deep neural networks are typically trained on datasets collected under controlled conditions, either simulated or over-the-air (OTA), which are constrained to specific channel environments with limited variability, such as fixed signal-to-noise ratio (SNR) levels. As a result, these models often fail to generalize when deployed in real-world scenarios where the feature distribution significantly differs from the training domain. This paper explores unsupervised domain adaptation techniques to bridge the generalization gap between mismatched domains. Specifically, we investigate adaptation methods based on adversarial learning, statistical distance alignment, and stochastic modeling to align representations between simulated and OTA signal domains. To emulate OTA characteristics, we deliberately generate modulated signals subjected to realistic channel impairments without demodulation. We evaluate classification performance under three scenarios, i.e., cross-SNR, SNR-matched cross-domain, and stepwise adaptation involving both SNR and domain shifts. Experimental results show that unsupervised domain adaptation methods, particularly stochastic classifier (STAR) and joint adaptive networks (JAN), enable consistent and substantial performance gains over baseline models, which highlight their promise for real-world deployment in wireless systems.