Deep learning has achieved considerable progress in the field of modulation classification; however, it is often limited by the consistency of the distribution between training and testing data. A novel multi-scale attention transfer learning architecture (MATLA) has been designed to assist cross-domain identification in order to overcome the problem of modulation classification with inconsistent data distribution caused by variable sample rates of source and target domains. This framework employs three parallel convolutional kernels with varying scales to extract features at different granularities. Moreover, the architecture integrates a multi-scale attention mechanism, which bolsters the extraction of discriminative features by emphasizing the weights of salient features. To effectively align features from the source and target domains, multiple kernel maximum mean discrepancy (MK-MMD) is utilized to measure the divergence of feature distributions in the reproducing kernel Hilbert space (RKHS) between the two domains. Additionally, to mitigate internal variation within the source domain features and enhance their consistency and stability, multiple kernel center loss (MKCL) is proposed. According to experimental data, the suggested approach outperforms a number of different network models and domain adaption techniques, achieving a recognition accuracy of 83.42% when the signal-to-noise ratio surpasses 0 dB.

In view of the problems of low probability of acquisition (LPI) radar signal recognition in low signal-to-noise ratio (SNR) and complex network model, an LPI radar signal recognition method based on time-frequency reassignment and multi-scale residual network was proposed. The time-frequency reassignment approach is used to enhance the signal's aggregation based on the Wigner-Ville distribution (WVD). The resulting time-frequency distribution image is then fed into the multi-scale residual network to finish the signal's categorization. In addition, the complex electromagnetic environment simulation was completed by constructing a multi-path Rice-fading channel. According to the experimental results, when the SNR is −8 dB, the suggested approach can achieve 94% recognition accuracy for a total of 13 different types of typical LPI radar modulation patterns, including Costas, Frank, P1～P4, etc. Compared with other methods, it has better recognition performance at a low signal-to-noise ratio.