Author Affiliations
Information Science and Technology College, Dalian Maritime University, Dalian 116026, Liaoning , Chinashow less
Fig. 1. Block diagram of QAM-PPM hybrid modulation system under CNN-AE architecture
Fig. 2. Specific structure of dot_sigma layer
Fig. 3. Characteristics of sigmoid function with different δ values
Fig. 4. Comparison of SER performance at different orders. (a) M=4 and 8; (b) M=16 and 32; (c) M=64
Fig. 5. Comparison of SER performance when N=8 and 10
Fig. 6. Comparison of SER performance when M=4, 8, 16, 32, and 64
Fig. 7. Comparison of SER performance when N=8 and 10
Fig. 8. Loss functions of training set and validation set for two modulation schemes under AWGN channel. (a) M=4 and N=8; (b) M=64 and N=4
Fig. 9. Loss functions of training set and validation set for two modulation schemes under Rayleigh channel. (a) M=4 and N=8; (b) M=64 and N=4
Fig. 10. Comparison of SER performance of AWGN channel CNN and DNN
Fig. 11. Comparison of SER performance of Rayleigh channel CNN and DNN
Layer type | Activation function | Output dimension |
---|
Input | None | (B,L,) | s1 | None | (B,L,) | s2 | None | (B,L,) | E1_Conv1D | elu | (B,L,256) | E1_Conv1D E1_Conv1D | elu elu | (B,L,256) (B,L,256) | E1_Conv1D | linear | (B,L,2) | E2_Conv1D | relu | (B,L,64) | E2_Conv1D | relu | (B,L,64) | E2_Conv1D | linear | (B,L,N) | dot_sigma | sigmoid | (B,L,N) | Concatenate | None | (B,L,2+N) | QAM_PPM | None | (B,L,N) | Normalization | None | (B,L,N) | Channel | None | (B,N×L,2) | Conv1D | elu | (B,L,256) | Conv1D | elu | (B,L,256) | Conv1D | elu | (B,L,256) | Conv1D | elu | (B,L,256) | Conv1D | softmax | (B,L,2k) |
|
Table 1. Parameters for autoencoder network structure
System | Learned modulation | Training parameter | Training time /s |
---|
DNN-AE-QAM-PPM | 4QAM-4PPM | 345762 | 14067 | 16QAM-4PPM | 361170 | 14131 | 64QAM-4PPM | 422802 | 14580 | CNN-AE-QAM-PPM | 4QAM-4PPM | 345762 | 7201 | 16QAM-4PPM | 361170 | 7333 | 64QAM-4PPM | 422802 | 7556 |
|
Table 2. System complexity comparison between DNN and CNN