Research On Co-phasing Error Sensing Of Segmented Telescope Based On Deep Learning

To improve the telescope’s light gathering ability and imaging resolution,increasing the telescope’s aperture is the most direct means.However,with the increase of aperture,there are great difficulties in the mirror processing,detection,lightweight,transportation and launch process of the traditional monolithic telescope.Segmented primary mirror structure has become an effective way to solve these problems,and has become the development trend of large-aperture astronomical telescopes.However,it is difficult for the segmented primary mirror to achieve the same imaging effect of a monolithic mirror.It is necessary to correct the co-phasing error between each submirror in real time to make sure its RMS value is less than 1/40 wavelength.At this time,only the traditional mechanical adjustment cannot achieve the detection accuracy at all,so the high-precision wavefront detection technology is an indispensable step to ensure the imaging quality of the telescope.In recent years,the deep learning method has been introduced into the wavefront detection field.This method uses the training data generated by Fourier Optical theory to train the network,and uses its strong nonlinear fitting ability to establish the mapping relationship between the aberration and the PSF image.Compared with iterative algorithms such as traditional phase recovery,deep learning does not need iterative calculation,and it is easy and simple to implement,which is convenient to obtain phase information quickly.However,at present,deep learning method is rarely applied in the research field of the co-phasing error detection of the segmented telescope,and there are still many problems need to be solved,such as the low co-phasing detection accuracy,the low universality of the co-phasing error detection for the extended scene,and the detection is vulnerable to noise.Therefore,it is necessary to conduct in-depth exploration and research on deep learning method to improve accuracy and practicability of this method in the field of co-phasing error detection of the segmented telescope.The main research work is as follows:Aiming at the problem of not high accuracy of co-phasing error detection,this thesis proposes a Bi-GRU network model,which decomposes two defocused PSF images into sequence data as network input,calculates the system co-phasing error,so that it can achieve the fine phasing of each sub-mirror of the segmented telescope.The method in this thesis does not need iteration and optimization process,and can correct the system phase error in real time,and detection accuracy is better than the CNN deep learning model.Besides,this thesis analyzes that the increase of the number of submirrors leads to the increase of the detection difficulty.By further improving the BiGRU network model,it can still effectively detect the co-phasing errors between each sub-mirror.Aiming at the problem of the low universality of extended scene co-phasing detection,this thesis proposes a useful feature image extraction method.By innovatively designing the form of pupil mask,and further updating the OTF in the frequency domain,we obtain a new decoupled independent feature image that can simultaneously detect the piston error and tip/tilt error of all sub-mirrors,which is effectively decoupled,and get rid of the dependence of the data set on the imaging object.Then the Bi-GRU network is used to recover phase error information with high accuracy from the feature image proposed in this thesis.We also quantitatively discuss the influence of the wavefront sensing accuracy of our method under broad spectrum bandwidth,which is more in line with the actual observation process of the telescope.The method proposed in this thesis is still effective when the bandwidth is less than 200 nm.Aiming at the problem that co-phasing error detection is easily affected by noise,this thesis designs and constructs a multi-scale residual network(MSRNN),which is introduced into the preprocessing process of co-phasing detection to denoise the defocused image,so as to improve the noise robustness of the co-phasing error detection.The multi-scale feature fusion layer and dense connection blocks in MSRNN enhance the network’s ability to sense noise and the network convergence speed.The successful trained MSRNN has the ability of blind noise reduction.It is no need to evaluate the noise before image processing,and the operation speed is fast,so that the operation time of the co-phasing error detection process will not be increased additionally.Finally,according to the actual situation of the laboratory project,a concrete and implementable experimental scheme is designed.In order to solve problem of getting a large number of real experimental images for network training,a simulation-driven experimental-verification mechanism is proposed,which verifies the feasibility and calculation effectiveness of the segmented primary mirror co-phasing error detection based on deep learning.