Optical Modulation Format Identification Under Hybrid Increasing Spacing OSNRs Based on Signal Constellation Diagrams and LIN Model

Modulation Format Identification (MFI) is of great research significance to ensure the normal operation of dynamic and flexible optical communication networks. In this paper, based on a lightweight convolutional neural network with improved Inception (LIN) model, a scheme of increasing spacing optical signal-to-noise ratio (OSNR) of training data set is proposed to train the network model. Meanwhile, data set of different OSNRs are tested to evaluate the generalization performance of the model. Firstly, 10Gbaud BPSK, QPSK, 8PSK, 16QAM, 32QAM and 64QAM coherent optical fiber communication systems are constructed, and the constellation diagrams of series of OSNRs are processed by the neighborhood grayscale enhanced algorithm. Next, the constellation diagrams of the OSNRs of increasing spacing are used as input features to create training data set to train the four deep convolutional neural network models (LIN, AlexNet, ResNet, and GoogLeNet). Finally, the constellation diagrams of the OSNRs different to the OSNRs in training data set are used as the test set of the deep learning network. The simulation experiment shows that the identification accuracy of LIN for modulation formats under three different OSNRs are 89.41%, 95.37% and 97.64% respectively, and LIN has the highest accuracy, and better generalization ability and can meet the requirements of dynamic optical networks.