In order to address the issue that traditional radar signal recognition techniques have trouble successfully identifying unknown modulated signals in practical situations, this paper suggests a multi-level radar signal open-set identification technique based on K-means clustering and Support Vector Machine (SVM) pre-training. After performing the multisynchro squeezing transform (MSST) on radar signals, the discrete wavelet transform (DWT) is employed to extract features from the preprocessed time-frequency images. An SVM classifier is trained using known radar signal data during the training phase. The classifier is used to distinguish between known and unknown modulation types during the testing phase in order to achieve open-set radar signal identification. Subsequently, K-means cluster analysis is applied to unknown radar signals, further classifying unknown modulation modes into different clusters, thereby expanding the recognition scope of radar signal modulation types. Experimental results demonstrate that the proposed method achieves a recognition accuracy of over 90% for both known and unknown signals at a signal-to-noise ratio (SNR) of −4 dB, effectively recognizing unknown modulation types.

To alleviate the excessive dependence of graph neural networks on prior knowledge in community discovery and improve recognition accuracy, a novel self-supervised learning model for community detection based on a deep graph convolutional network (GCN) is proposed. The model makes full use of the semantic features of a small number of nodes and obtains pseudo-labels of unknown nodes through a semantic alignment mechanism, and thus introduces a self-supervised module to alleviate the dependence on a large number of prior labels during the training of GCN. Furthermore, by stacking self-supervised modules, a deep graph self-supervised learning model is built to increase the accuracy of community detection by obtaining the global information of networks. Two strategies, identity mapping and initial residual, are employed to address the over-smoothing issues that the deep model introduces. According to experiments conducted on publicly available datasets, the suggested approach outperforms current models in terms of community recognition accuracy when a limited number of prior labels are used and the model depth is increased.