Recognition and evaluation of constellation diagram using deep learning based on underwater wireless optical communication

Abstract. In this paper, we proposed a method of constellation diagram recognition and evaluation using deep learning based on underwater wireless optical communication (UWOC). More specifically, an constellation diagram analyzer for UWOC system based on convolutional neural network (CNN) is designed for modulation format recognition (MFR), optical signal noise ratio (OSNR) and phase error estimation. Besides, unsupervised learning is used to excavate a new optimization metric from various factors that affect the quality of underwater channel.The proposed new metric synthesizes several original indexes, which we termed it as multi noise spatial metric (MNSM). The proposed MNSM divides the quality of constellation from high to low into several levels and reflects the quality of UWOC channel. Through the simulation, the constellation diagrams of four widely used M-QAM modulation formats for 16 OSNR values (15dB~30dB) are obtained, with the phase error standard deviations ranging from 0{\deg} to 45{\deg}. The results show that the accuracy of MFR , the estimation of OSNR and phase noise are 100%, 95% and 98.6% accuracies are achieved respectively. The ablation studies are also carried out in order to analyze the performance of deep learning in the recognition of constellation diagrams.