LPI Waveform Recognition Using Adaptive Feature Construction and Convolutional Neural Networks Table 4. Determined CNN Structure Used for the Fusion Features Layers Input layer Convolutional layer Maximum pooling layer Flatten layer Batch normalization layer Dropout layer Dense layer Dropout layer Dense layer Dropout layer Dense layer 32 N/A N/A N/A 0.5 64 0.5 32 0.5 6 Filters, neurons, dropout rate N/A Kernal size N/A 2 2 4 2 2 N/A N/A N/A N/A N/A N/A N/A N/A Output size 512 80 4 511 79 32 255 39 32 318240 1 318240 1 318240 1 64 1 64 1 32 1 32 1 6 1 Figure 13. Confusion matrix for the fusion feature on SNR levels of (a) -10 dB, (b) -15 dB, (c) -18 dB, (d) -20 dB with 100 signal samples in each type of simulated LPI signals. The diagonal entries represent the number of signal samples that are correctly classified, while other entries in the matrix show the signal classified incorrectly. Among all the types ofwaveforms, the P1 and P4 code waveforms are the most ofbeing misclassified. 24 IEEE A&E SYSTEMS MAGAZINE APRIL 2023