Optical network performance evaluation using a hybrid neural network
Aspects of the present disclosure describe systems, methods. and structures in which a hybrid neural network combining a CNN and several ANNs are shown useful for predicting G-ONSR for Ps-256QAM raw data in deployed SSMF metro networks with 0.27 dB RMSE. As demonstrated, the CNN classifier is trained with 80.96% testing accuracy to identify channel shaping factor. Several ANN regression models are trained to estimate G-OSNR with 0.2 dB for channels with various constellation shaping. Further aspects include the tuning of existing optical networks and the characterization of retrofit/upgraded optical networks to estimate capacity—both aspects employing our inventive hybrid neural network methodology.
1. A method for tuning a wavelength division multiplexed (WDM) optical fiber network comprising:
recording raw data from real-time probabilistic-shaped quadrature amplitude modulation (PS-QAM) on selected WDM channels;
determine, for each of the selected channels, a general optical signal to noise ratio (G-OSNR) of optical signals exhibiting the PS-QAM, through the effect of a two-stage, hybrid neural network including a convolutional neural network (CNN) classifier in a first stage, followed by a second stage including a plurality of artificial neural networks (ANN) that estimates the G-OSNR after the CNN Classification;
determining, for each of the selected channels, if the estimated G-OSNR meets or exceeds a predetermined threshold/margin.
2. The method of claim 1 further comprising:
adjusting a PS-QAM transponder data rate for each of the selected channels when the G-OSNR for the transponder does not meet or exceed the predetermined margin.
3. The method of claim 2 further comprising:
repeating the recording, determining, determining, and adjusting steps until the estimated G-OSNR for all of the selected channels meet or exceed the predetermined margin.
4. The method of claim 3 further comprising:
determining a capacity estimate of the WDM optical network from the adjusted data rate(s); and
outputting an indicia of the network capacity.
5. The method of claim 3 wherein the two-stage, hybrid neural network uses only the raw data as input for the estimation.
6. The method of claim 5 the PS-QAM raw data is associated with a shaping factor that is used to generate different format spectral efficiency.
7. The method of claim 1 wherein the CNN comprises several layers including 2 convolutional, 1 pooling, 1 flatten, and 2 fully-connected layers and 6 outputs for different shaping factor categorizations.
8. The method of claim 7 wherein the ANN includes a plurality of ANN regressions to estimate the G-OSNR for various probabilistic-shaped channels.
9. The method of claim 8 wherein the ANN includes two hidden layers each having 50 and 20 neurons, respectively.
10. The method of claim 9 wherein a plurality of ANN training models, at least one for each differently classified group is employed for training.