Computer-Aided Alignment Of Wide-Field Survey Telescope Based On Deep Learning

Wide-field survey telescopes are one of the important development directions of modern astronomical telescopes because they can obtain large-scale,high-quality,high-sensitivity and high-frequency star observation data.Based on the considerations of telescope image quality control and barrel design,modern wide-field survey telescopes generally adopt fast focal ratio primary mirror design,which leads to high sensitivity of secondary mirror for misalignment.In order to reduce the hardware complexity of the alignment system and improve the alignment efficiency,In this paper,we propose a new real-time high-precision alignment method without wavefront sensors,which establishes the mapping relationship between the focal plane star image deformation and the misalignment parameters by deep learning technology,and then calculates the misalignment parameters of the telescope directly from the neural network model.The details of the research work are as follows.1.A dynamic data exchange method based on ZEMAX and MATLAB is established to obtain the focal plane point spread function of the telescope under each misalignment state by simulation,and then obtain the Zernike coefficients by geometric feature extraction and fitting,and build a dataset for deep learning alignment with the misalignment parameters.2.A deep learning alignment model based on Mephisto is built to determine the structure of neural network and optimize the hyperparameters by analyzing the relationship between the depth of the neural network and the model accuracy and misalignment degrees of freedom.The results show that the model can accurately calculate each misalignment parameter of the telescope in the absence of noise,and obtain a good image quality of the point spread function image,but there is a problem of computational degradation caused by the nonlinear relationship of the misalignment degrees of freedom.3.According to the nonlinear pathological problem between different misalignment degrees of freedom of the telescope,in this paper,a step-by-step neural network model-based misalignment solution method is proposed,which achieves the high-precision solution of misalignment parameters in a wide range of misalignment cases through the structural adjustment and hyperparameter optimization of the twostep neural network.4.In order to reduce the effect of atmospheric turbulence on the Mephisto deep learning alignment model,the dataset is preprocessed by a robust scaling method to counteract the effect of outliers in the dataset.The results show that the proposed method is still robust under the influence of atmospheric turbulence,and can accurately calculate the misalignment parameters and obtain good image quality,with more than99% of the 1000 test datasets satisfying the RMS spot radius below 18 um.