A piston detection method for segments via convolutional neural networks and its robustness analysis

To achieve a diffraction-limited imaging, the piston errors between the segments of the segmented primary mirror telescope should be reduced to λ/40 RMS. The piston detection method using convolutional neural network (CNN) is an advanced technology with high precision and simplicity. However, such methods based on the deep learning strategy usually have generalization problems, that is, the network prediction precision will inevitably decrease if there is a certain difference between the test image and training set used in the network. This will directly affect the scope of application of the method. In this letter, we propose a CNN-based high-precision piston detection method and analyze its robustness. The point spread function (PSF) images acquired under the wide-spectrum light source are used to construct the dataset to overcome 2π ambiguity. In addition, a set of neural networks system including the classification CNN and the regression CNN with good generalization ability is designed to extract the piston value directly from the PSF image. Under the ideal condition, the piston detection precision can reach about 8.4 × 10-4 λ0RMS in the capture range of the interference length of the operating light. Finally, we focus on testing the effect degree of the main disturbance factors in the actual system on the accuracy of the method, such as surface error, residual tip-tilt error, and CCD noise, so as to evaluate the robustness of the method. This method is robust and does not require complex hardware. It can be widely applied in segmented and deployable primary mirror telescopes. We believe that the study in this letter will contribute to the applications of the CNN-based technique for piston sensing.