Identifying orbital angular momentum modes in turbulence with high accuracy via machine learning

Abstract

The orbital angular momentum (OAM) of a Laguerre–Gaussian beam (LGB) is classified using the support vector machine (SVM) model. The scintillation index, beam width, and beam wander of the Gaussian beam and the LGB at the receiving site are taken as feature vectors, and the OAM number of a LGB is taken as the label in the training and test samples. The influences on the detection accuracy of the number of training samples, the transmission distance, and the different ways of grouping OAM values are analyzed, where the detection accuracy is defined as the percentage of correctly detected OAM. The results show that only a few training samples are needed up to a range of 4000 m to achieve saturation accuracy. When the transmission distance is within 2500 m, the detection accuracy of a single SVM model in four turbulent environments ( 10−14, 10−15, ) is close to 100%. When the distance increases, detection accuracy decreases gradually, and in the range of 4000 m the detection accuracy is more than 91%. Compared with the traditional spiral spectrum expansion method, the SVM model has obvious advantages. In addition, when the range of OAM to be detected is large, the OAM values may be grouped thereby improving the accuracy when using the multi-SVM models for joint detection.