| Traditional wavelength division multiplexing techniques and their extension methods cannot meet the growing communication needs in the context of big data.Looking for other new technologies to further expand the channel has become a hot direction in the field of optical communication.The vortex beam has a helical phase that changes linearly along the angular direction,and the different eigenmodes are orthogonal to each other.The possibility that we can use the orbital angular momentum to increase the channel width of optical communication has attracted great attention of researchers.Deep learning is widely used in image processing and other fields.Optical imaging can be used as an input image for processing,and effective information can be obtained.We introduce deep learning algorithm as the modal analysis method of composite vortex beams.The main innovations of this thesis are as follows:(1)A dual-output convolutional neural network(Y-Net)based modal analysis method for multimode vortex beams is proposed,which can not only output the weight of each mode based on the input light intensity distribution image,but also restore the unknown phase information.When multiple modes of vortex beams are superimposed on each other,we obtain the weights and relative phases of the different modes simultaneously by proposed modal analysis method.We first compare different model structures and find that the dual-output structure gives better predictions than the single-output structure,indicating that the dual-output structure retains the link between amplitude and phase for the light intensity distribution,while distinguishing between their different physical meanings.Then,we reconstruct the vortex beams according to the predicted weights and phases,and the correlation coefficients before and after the reconstruction are over 0.9999,demonstrating the powerful modal analysis capability of the neural network.We also study the performance of neural network under different mode numbers and different propagation distances.The results show that when the mode number is 9,the weight mean absolute error(MAE)is about 1.4×10-3 and the relative phase MAE is about2.9×10-3,and the weight error is also only 5.6×10-3 when the beam is propagated120 meters.For unknown data sets and noise interference,the neural network still maintains more accurate predictions,which has certain application value for the future optical system.(2)A convolutional neural network(CNN)-based mode occupation analysis method for Laguerre-Gaussian beam multiplexing systems containing radial indices is proposed,which both generalizes well to channel variations and can be used with U-shaped networks to detect the mode occupation of blocked beams.In optical communication,Laguerre-Gaussian beams with non-zero radial and angular quantum numbers not equal to 0 can further broaden the channel number of information transmission.Our proposed method has good generalization in the dimension of data composition,so we choose the number of channels at communication to investigate the generalization performance of neural networks over channel variations.We start with the case where the channel pattern composition is fixed,and gradually extend the applicability of the neural network to the case where both the channel pattern composition and the number of patterns are not fixed.The results show that when the mode composition is fixed,the average absolute error of the neural network is about 0.002,and even when the maximum absolute error allowed is 0.005(i.e.the deviation of the mode share prediction is up to 0.5%),the accuracy is 94.86%.When both the mode composition and the number of patterns are not fixed,the prediction accuracy is 93.26%(when the maximum allowed absolute error is 0.01),demonstrating good generalization is existed.In optical communication,the carrier wave may be obstructed.In this case,we introduce U-shaped network(U-Net)to restore the occluded light intensity distribution and analyze the result of restoration field.Within a given obstruction range,the restored light intensity distribution image is very close to the original image and its PSNR value is above 33d B,indicating that U-shaped network has a good restoration effect.Secondly,we analyze the mode proportion of the restored intensity distribution,and the results show that the mean absolute error of the restored intensity distribution decreases by one order of magnitude compared to the unrecovered image,and is in the same order of magnitude as the original intensity distribution,which proves the feasibility of mode proportion analysis by the combination of U-Net and CNN.We also investigate the influence of noise on the stability of neural network performance.The results show that when using a maximum absolute error value of 0.009 as the threshold for prediction accuracy,we have a prediction accuracy of around 83%even when the instantaneous noise intensity reached 0.03,demonstrating that the neural network has some resistance to noise. |
PDF Full Text Request