Huang et al. Figure 11. Designed LPI signal recognition system. the time-frequency features including WVD, CWD, and wavelet features are calculated with the generated LPI signals with the length of 4096 samples, 1 kHz of carrier frequency and 7 kHz of sampling frequency according to the equations in the " Feature Extraction and Evaluation " section. For the wavelet feature calculation, the Morse wavelet is usedwith the symmetry parameter equal to three and the linear scales are used for image feature generation. For the polyphase-coded signals, the number of code phases and the number of cycles per phase code are range of(8,16) and (1,5). For the generated dataset, 96,000 signal samples are generated as the training dataset and 7000 signal samples are for testing. The signal samples are evenly distributed on each SNR level and each type ofwaveform for both the training and testing dataset. RESULTS OF LPI WAVEFORM RECOGNITION The generated datasets are used for the DL model training and testing for the proposed approach and the methods in existing works. The grid search method is used for the hyperparameters tuning when training the CNN models Table 3. Experimental Results of PCC With Different SNR Levels SNR (dB) WVD feature [8] 51 -5 -7 -9 0.98 0.92 0.91 0.90 Figure 12. LPI waveforms recognition performance comparison under different SNR levels. APRIL 2023 -11 0.83 -13 0.71 -15 0.63 -16 0.54 -17 0.50 -18 0.49 -19 0.44 -20 0.27 IEEE A&E SYSTEMS MAGAZINE CWD feature [10] 1 0.98 0.85 0.77 0.67 0.62 0.57 0.51 0.45 0.44 0.43 0.33 0.26 Wavelet feature [11] 1 0.97 0.87 0.83 0.74 0.67 0.62 0.56 0.49 0.46 0.45 0.35 0.27 Fusion feature 1 1 0.99 0.99 0.98 0.95 0.93 0.87 0.86 0.85 0.83 0.81 0.71 23