Abstract: To address the issue of difficulty in identifying LPI (Low Probability of Intercept) radar signals under low signal-to-noise ratios, this paper proposes a gated fusion-based PSA convolution-Transformer hybrid neural network. Firstly, the CWD time-frequency distribution is used to extract features of 13 typical LPI signals including Barker, Costas, LFM, Frank, Rect, P1-P4, T1-T4, and perform preprocessing. The model structure adopts parallel branches of PSA convolution and Transformer to fuse local and global features, with a gating mechanism for dynamic adaptive control. Experimental results show that the proposed method achieves significant identification performance under low signal-to-noise ratios, with an average recognition rate of 93.82% at -8dB. This method provides a new approach for the intelligent identification of complex radar signals.