An unsupervised method for clustering of unknown emitter signals

The classification and identification of unknown emitter signals have garnered significant attention due to the quick development of spectrum sensing techniques in cognitive radio networks. Nevertheless, the majority of existing clustering techniques for emitter signals from unknown emitters have poor recognition accuracy. To address this challenge, we propose a novel unsupervised framework for clustering unlabeled emitter signals, combining empirical mode decomposition (EMD) and contrastive predictive coding (CPC) to extract signal features and clustering of unlabeled emitter signals. Our approach systematically decomposes non-stationary waveforms through EMD, subsequently employing CPC to learn discriminative feature representations from the intrinsic mode functions (IMFs). The extracted features are then clustered using density-based spatial clustering with noise (DBSCAN), effectively handling complex signal structures while mitigating noise interference. These quantitative improvements substantiate the framework's capability to establish new state-of-the-art performance in unsupervised emitter identification tasks, particularly under low signal-to-noise ratio (SNR) conditions.