Research on co-phasing detection of segmented mirror based on convolutioned neural networks

To build a greater than 10 m aperture telescope, it is necessary to use the technique of segmented mirror. Although segmented mirror enables the construction of very large aperture telescopes, they also require fine cophasing of the sub-mirrors in order to achieve the imaging capability of the whole mirror with the same aperture. As a result, there is urgently necessary to develop a large range, high precision, and fast co-phasing errors measurement technique. In this paper, a new method of two-wavelength co-phasing detection based on convolutional neural network is proposed to address the problem of slow detection speeds and low accuracy in the detection of large-range co-phasing errors (piston) by the current two-wavelength co-phasing detection method based on the cross-correlation algorithm. Firstly, the co-phasing detection dataset with piston detection range [-10.5 & mu;m, 10.5 & mu;m] at wavelengths of 737 nm and 750 nm was analyzed by theory and simulation, and the rapidity and accuracy of the method of two-wavelength co-phasing detection based on convolutional neural network were verified by this dataset, and the accuracy of the training set was greater than 99.9% and the RMS value of the test sample error was 9.8 nm. This paper further discusses the tolerances for eccentricity and gap errors required for using the neural network detection method. When the gap error or eccentricity error is less than 0.2, the piston error can be directly detected using the convolutional neural network.