3.3. Modulation formats classification

In this technique, we also used the MLP3 for the modulation format identification by assigning output nodes to represent each format. Four output neurons in the output layer represent the 40 Gbps NRZ-OOK, 160 Gbps OFDM DP-16QAM, 400 Gbps DC PDM-QPSK and 1 Tbps WDM-Nyquist NRZ-DP-QPSK modulation formats. As designed previously, before training the network, the multilayer perceptron output can be considered as the highest probability, which represent one from the four modulation types. Eigenvalues after the SVD for each modulation format are represented at the input of the ANN. The multilayer perceptron used in this architecture requires four output neurons representing the number of format types.

The four used modulation formats are generated with carrier frequency equal to 193.1 THz. Using a variable optical attenuator (VOA), the injected power is tuned and passed through an optical amplifier (OA) that undergo the ASE noise effect at higher gain amplifier. As a result, the OSNR of the signals is adjusted in the range of 10–35 dB (steps of 5 dB). Then, using a CD/ PMD emulator, chromatic dispersion is varied from 85 to 510 ps/nm (steps of 85 ps/nm) to reach 30 km on SMF, and the DGD between 0 and 20 ps (steps of 5 ps). The appropriate carrier to be classified is selected by an OBPF. As described in Figure 8, in MFR block, using the CWT, each received signal is processed by scaling factor up to 128 without amplitude normalization. Moreover, with length equal to 3, the median filter was applied to extract the features set and remove the peaks. In addition, we make the SVD to the time-scale parameters of the wavelet coefficients matrix and obtain the eigenvectors as the final characteristic vectors for each received signal. Finally, we reach the ANN classifier as described in the previous section, where for each modulation format we generate 150 realizations with 65,536 symbols

3.2. Design of the proposed method

22 Advanced Applications for Artificial Neural Networks

The setup of the proposed MFR technique is shown in Figure 11.

corresponding to different combinations of CD, DGD and OSNR.

Figure 11. Proposed modulation recognition system.

The proposed classification method is demonstrated at high bit rates for the four used optical modulation formats including 40 Gbps NRZ-OOK, 160 Gbps OFDM DP-16QAM, 400 Gbps DC PDM-QPSK and 1 Tbps WDM-Nyquist NRZ-DP-QPSK. For each modulation type, to verify the robustness of our method, 150 realizations with 65,536 symbols have been used. All signals are transmitted through ideal (B-to-B) and impaired channels. The simulation results are given Figures 12–14.

The plot of the probabilities of recognition against OSNR is given in Figure 12. The OSNR is in the range of 10–35 dB. 100% correct identification was observed for 160 Gbps OFDM DP-16QAM and 1 Tbps WDM-Nyquist NRZ-DP-QPSK at higher OSNR values. In contrast, for

Figure 12. Percentage recognition versus OSNR values for NRZ-OOK, OFDM DP-16QAM, DC PDM-QPSK and WDM-Nyquist NRZ-DP-QPSK modulation types.

Figure 13. Percentage of correct identification for four used modulation formats depending upon residual dispersion for OSNR = 20 dB.

Figure 14. Percentage of correct identification for four used modulation formats depending upon DGD where OSNR = 20 dB.

lower OSNR values (lower than 15 dB), a challenging problem remains for NRZ-OOK identification, which reached values between 80 and 90%. This misclassification is because of its coding properties, where to extract its features, the wavelet transform located ambiguities.

The modulation format can also be recognized correctly when varying the CD, as described earlier in Figure 13. From the simulation, the residual dispersion varies between 85 and 510 ps/ nm with OSNR = 20 dB. The 160 Gbps OFDM DP-16QAM and 1 Tbps WDM-Nyquist NRZ-DP-QPSK still having the highest recognition probabilities (~100%). Furthermore, 40 Gbps NRZ-OOK and 400 Gbps DC PDM-QPSK reached the minimum values of identification accuracies, that is, 90% for residual dispersion more than 350 ps/nm. In the main, when increasing the residual dispersion values, all probabilities of recognition decreases. This is interpreted by the fact that signal characteristic are widely modified in terms of phase and amplitude with the chromatic dispersion. For that, the wavelet transform will lose its principal features.

In optical transmissions, in the presence of several impairments that broadly modified the signals features and makes the formats recognition issue more difficult. We can cite in this case the impairment caused by the DGD. In Figure 14, we show the probabilities of identification versus this parameter.

The DGD is between 0 and 20 ps with steps of 5 ps. The misclassification percentage for 40 Gbps NRZ-OOK signal remains with a small difference for 160 Gbps OFDM DP-16QAM and rises when the DGD is greater than 15 ps. 400 Gbps DC PDM-QPSK and 1 Tbps WDM-Nyquist NRZ-DP-QPSK modulation formats still having the highest accuracies when the DGD is less than 10 ps and begin to decrease for DGD values more than 15 ps. It is suspected that this error is due to the effect of SOP induced by the PMD. It is based on the DGD and the frequency, and it varies differently for each modulation format with the optical carrier, after fiber transmission.

Table 2 shows the recognition results of the neural network when OSNR is equal to 30 dB. For 160 Gbps OFDM DP-16QAM and 1 Tbps WDM-Nyquist NRZ-DP-QPSK formats, we obtain 100% correct classification, while 40 Gbps NRZ-OOK and 400 Gbps DC PDM-QPSK were identified at 96.6% and 98%, respectively. An ambiguity of identification for these formats is increased to 3% at higher DGD and CD. This misclassification is due to the fact that the signal features (phase, amplitude, etc.) are infected in the presence of signal impairments. Thereby, CWT proves a difficulty to sign these formats and extract their eigenvalues.


Table 2. Recognition rates at OSNR = 30 dB.

Figure 13. Percentage of correct identification for four used modulation formats depending upon residual dispersion for

Figure 14. Percentage of correct identification for four used modulation formats depending upon DGD where

OSNR = 20 dB.

24 Advanced Applications for Artificial Neural Networks

OSNR = 20 dB.

The robustness of this technique is evaluated and the simulated results show that the correct modulation identification scheme is possible even at all channel parameters ranges: OSNR between 10 and 35 dB, CD from 85 to 510 ps/nm and DGD in the range of 0–20 ps for all modulation formats mentioned previously. We reached higher identification accuracies at high bit rates which facilitates the performance monitoring process (OPM) for the network management.

From the results obtained, it is found that the method employing the amplitude histograms is simple and more robust to impairments link, than the other one using CWT, for features extraction. Moreover, the asynchronous amplitude histograms (AAH) generation followed by ANN is rapid on time response compared to the method using CWT for features extraction. Besides the modulation formats, identification ambiguities is more apparent for the second method.
