Experimental Study On Eliminating Wavefront Phase Defects By Deep Learning

In free-space optical communication,in order to obtain stable system performance,Adaptive Optical(AO)technology is used to suppress the effects of turbulence effects.When the AO system works,the signal light disturbed by turbulence will produce intermittent phase undulation and spot drift,the observation target is weak,and some subapertures do not receive effective optical signals and form detection bad spots,resulting in wavefront information is missing,which seriously affects the correction effect of distorted wavefront.In this paper,we introduce deep learning technology to effectively compensate for the bad spots of traditional detection mechanism and restore the missing wavefront phase information.The main research contents of this paper are as follows.1.This paper firstly starts from the wavefront detector devices,and analyzes the limitations and shortcomings of the Shack-Hartmann wavefront sensor(SH-WFS),which is most commonly used in AO systems.And the sources of the bad point error of the conventional adaptive optics system detection mechanism are sorted out.The Zernike polynomial is used to characterize the wavefront phase,and the mapping relationship between the subaperture light intensity distribution of the Hartmann wavefront sensor and the corresponding first 30th order Zernike coefficient is established,which provides a theoretical basis for the wavefront bad point compensation method based on deep learning.2.A convolutional neural network-based wavefront recovery model is constructed to compensate for the limitations of wavefront phase recovery based on slope information.A large number of measured aberrated light intensity images and their corresponding Zernike wavefront coefficients are obtained as the data set for training,preserving the implicit feature information in the images and learning the relationship between data and labels.The loss function and optimization function of the network are theoretically analyzed,and the specific network structure is designed.The Zernike coefficients of the output wavefront aberration can be directly predicted by inputting the subaperture bad spot light intensity image into the trained neural network model,which is further converted into a wavefront phase map.The feasibility of the algorithm is verified using a test set,and the first 30 order Zernike coefficients of the wavefront can be identified from the bad spot light intensity image,and the original distorted wavefront is recovered with high accuracy.3.The indoor experimental system and the outgoing link experimental platform are built,and the wavefront correction experiments based on the convolutional neural network prediction are completed by using the Peak to Valley(PV),Root Mean Square(RMS),optical power and signal-to-noise ratio of the corrected wavefront at the receiver as evaluation indexes.The theoretical correctness and feasibility of the algorithm for adaptive optical systems are demonstrated by verifying whether the CNN trained in this paper has the ability to compensate for the bad phase.The experimental results show that the spot gray value after indoor static compensation is improved from 220 to about 249.The corrected wavefront PV and RMS values were reduced from 5.215 μm and 0.281 μm to 0.425 μm and 0.166μm,respectively,in the 600m experiment in the external field,and from 7.952μm and 1.494μm to 3.06μm and 0.852μm,respectively,in the 1.3km experiment.The corrected PV and RMS values are reduced from 21.979μm and 10.061μm to 11.156μm and 6.963μm,respectively,in the 10km experiment,indicating that the CNN-based wavefront recovery model developed in this paper can compensate the bad point error caused by the SH-WFS subaperture missing light and improve the correction performance of the adaptive optics system.