Algorithm

Residual

Statistics

Optical fiber

Telecommunications

Computer science

Pattern recognition (psychology)

Mathematics

Bit error rate

Polarization mode dispersion

Artificial intelligence

Channel (broadcasting)

Artificial neural network

Mean squared error

Optics Express

We propose a transfer learning assisted deep neural network (DNN) method for optical-signal-to-noise ratio (OSNR) monitoring and realize fast remodel to response to various system parameters changing, e.g. optical launch power, residual chromatic dispersion (CD) and bit rate. By transferring the hyper-parameters of DNN at the initial stage, we can fast response to the channel variation with fewer training set size and calculations to save consumptions. For feature extraction processing, we use amplitude histograms of received 56-Gb/s QPSK signals as the input for DNN at the initial stage, which shows the root mean squared error (RMSE) of OSNR estimation is less than 0.1 dB with the OSNRs ranging from 5 to 35 dB. Then, we change several system parameters and find superior capabilities of fast remodeling and data resource saving with the proposed method. The required training epochs have about four times reduction, and the required training set size is only one-fifth compared to retraining the network without any accuracy penalty. The DNN assisted by transfer learning can save resources and will be beneficial for real-time application on OSNR estimation.

Transfer learning assisted deep neural network for OSNR estimation